Labor Markets, Migration, and Mobility: Essays in Honor of Jacques Poot (New Frontiers in Regional Science: Asian Perspectives, 45) 9811592748, 9789811592744

Citation preview

New Frontiers in Regional Science: Asian Perspectives 45

William Cochrane Michael P. Cameron Omoniyi Alimi  Editors

Labor Markets, Migration, and Mobility Essays in Honor of Jacques Poot

New Frontiers in Regional Science: Asian Perspectives Volume 45

Editor-in-Chief Yoshiro Higano, University of Tsukuba, Tsukuba, Ibaraki, Japan

More information about this series at http://www.springer.com/series/13039

William Cochrane • Michael P. Cameron • Omoniyi Alimi Editors

Labor Markets, Migration, and Mobility Essays in Honor of Jacques Poot

Editors William Cochrane School of Social Science University of Waikato Hamilton, Waikato, New Zealand

Michael P. Cameron School of Accounting, Finance and Economics University of Waikato Hamilton, Waikato, New Zealand

Omoniyi Alimi School of Accounting, Finance and Economics University of Waikato Hamilton, Waikato, New Zealand

ISSN 2199-5974 ISSN 2199-5982 (electronic) New Frontiers in Regional Science: Asian Perspectives ISBN 978-981-15-9274-4 ISBN 978-981-15-9275-1 (eBook) https://doi.org/10.1007/978-981-15-9275-1 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Introduction

This volume of essays celebrates the life and work of Jacques Poot, currently Emeritus Professor of Population Economics at the University of Waikato in Hamilton, New Zealand, and Visiting Professor in the Department of Spatial Economics at the Vrije Universiteit Amsterdam. Previously, Professor Poot held academic positions at the University of Tsukuba, Japan as Foreign Professor and at Victoria University of Wellington, from which he obtained his PhD with a thesis entitled ‘Models of New Zealand Internal Migration and Residential Mobility’ in 1984. Prior to this, he had studied econometrics at the Vrije Universiteit Amsterdam. Professor Poot is an elected corresponding member of the Royal Netherlands Academy of Arts and Sciences and an elected member of the Academia Europaea. He is also an External Research Fellow at the Centre for Research and Analysis of Migration (CReAM) at University College London and at the Institute for the Study of Labor (IZA) in Bonn, Germany as well as an affiliate of Motu Economic and Public Policy Research in Wellington, New Zealand. Over his career, Professor Poot’s research interests have covered a broad array of topics including all aspects of the economics of population (such as migration, fertility, labour force, and ageing), with a particular focus on the spatial dimensions of these topics. He had a leading role in several large multi-institution research programmes, including Capturing the Diversity Dividend of Aotearoa New Zealand (CaDDANZ), Nga Tangata Oho Mairangi (NTOM), the Integration of Immigrants Programme (IIP), regional demographic change in New Zealand, and migrant diversity and regional disparity in Europe. He is a member of the editorial, scientific, or advisory boards of several respected international journals including the Asia-Pacific Journal of Regional Science, Region—Journal of the European Regional Science Association, Papers in Regional Science and Australian Journal of Labour Economics, among others. Aside from his academic roles, Professor Poot has also engaged in a career-long commitment to service to regional science, particularly the Regional Science Association International (RSAI), with whom he is a Fellow, and the Pacific Regional

v

vi

Introduction

Science Conference Organisation (PRSCO), where he is a past president (1998– 1999) and ex officio council member. The essays in this volume are written by former students, colleagues, associates, and co-authors of Professor Poot and cover some of the large number of topics addressed by him in his career to date. Together, they demonstrate the breadth and depth of the research topics that Professor Poot has engaged in. The chapters are briefly described below. Professor Poot has previously researched the impact of cultural diversity on firms (see Ozgen et al. 2013). In the first contributed chapter of this volume, Maré, Sanderson, and Fabling examine remuneration patterns among workers in foreignowned firms operating in New Zealand. They track workers as they move across jobs in different types of firms, seeking to document the extent of the ‘foreign wage premium’ distinguishing between compositional factors (e.g., differences in industry and employment composition between foreign and domestic firms) and remaining differences in wage levels and growth rates. They find that the earnings gap between workers in foreign-owned firms and domestically owned firms was largely due to compositional factors, though a gap of two to four per cent remained among apparently similar workers and firms. This premium seemed to be specific to employment in foreign firms. Human capital has also been a subject of interest to Professor Poot (e.g., see Poot et al. 2008). Poot, in a wide-ranging chapter, starts with the objective of locating human capital alongside other forms of capital by identifying two different dimensions, ‘widening’ and ‘deepening’. He finds this insufficient, arguing that the issue of the widening of human capital opens up the peopling and development dimensions of the leading crises facing humanity. Chief among these crises is climate change. However, Poot also acknowledges other crises, such as social inequality, including its downstream effects such as inter-country and intercontinental migration, which are amenable to policy interventions, and age-structural shifts which, by contrast, are inexorable and need management but cannot be altered in direction. He concludes that the causes and effects of structural ageing are misunderstood and almost no attention paid to numerical ageing (increasing numbers at any age). The determinants of migration have been a long-standing area of research for Professor Poot (e.g., see Brosnan and Poot 1987). The next chapter notes that Germany faces the prospect of a declining labour force over the next two decades, increasing its dependence on the immigration of workers from abroad. Using a threefold Oaxaca–Blinder decomposition and a large amount of information from a linked employer–employee data set, Brunow and Jost consider whether Germany is competitive for immigration—i.e., do German employers pay enough to make it attractive as a destination country? They explore the wage gap between foreign-born and German-born employees and focus on different countries of origin to better understand issues related to wage setting among these groups. Their results indicate that most of the wage gap is attributable to observed characteristics, and they offer evidence of country-specific differences that should be considered when attracting people to Germany and facilitating their integration into the German labour market.

Introduction

vii

Professor Poot has made a number of contributions in the area of empirical labour economics (e.g., Longhi et al. 2005). The next chapter begins with the observation that, for the U.S. labour market, previous research has found that among the nativeborn population, bilingual people have on average lower earnings. Using data from the ‘Understanding Society’ survey, Clifton-Sprigg and Papps examine whether a similar pattern holds true for the UK. Their approach examines between and withingender differences to identify groups in which the negative effects of bilingualism are most concentrated. They show that the negative earnings effects are restricted to bilingual, particularly Bengali, women in areas with few English speakers and speakers of relatively uncommon languages. Professor Poot has been an enthusiastic supporter of applications of the gravity model of migration (e.g., see Poot et al. 2016). In the next chapter, Grimes et. al consider whether bilateral regional migration flows are primarily driven by a city’s quality of life or quality of business. Using five-yearly census data covering 1986– 2013, they construct measures of quality of life and quality of business for 31 urban areas in New Zealand. Adopting a gravity model of regional migration—augmented by destination and origin quality of life and quality of business—they model the bilateral flows of working-age migrants (post-tertiary education and pre-retirement age). In addition, flows between urban and rural areas and flows for the urban areas to and from overseas locations are also modelled. Different attractors for international versus domestic migrants are found, with domestic migrants being attracted by consumption amenities, while international migrants favour cities with productive amenities. Understanding how population processes apply at neighbourhood level has also been a feature of Professor Poot’s research (e.g., see Maré et al. 2012). Buck and Batey’s chapter focuses on the relationship between migration and neighbourhood change. Starting with a review of the relationship between urban deprivation, residential mobility patterns, and urban regeneration policy—commenting on both theoretical concepts and previous empirical findings—they then make extensive use of small area statistics from the U.K. Population Census to link migration data with a geodemographic classification system, providing interesting insights into the structure of residential mobility in urban areas. The chapter concludes with an examination of the impact of area-based initiatives, part of the EU Objective One programme, on migration flows (both in and out) in the Merseyside region. Their primary finding is that the area-based initiatives had a significant impact on residential stability. The interplay between migration and institutions has also been a research interest of Professor Poot (e.g., see Tran et al. 2018). While research has covered in detail both the specific and the wider impacts of a migration influx into a destination or host country, the impacts of outmigration and remittances on the source country (or country of origin) have received less coverage. Incaltarau et. al address the issue of whether outmigration from (national or regional) labour markets, and the associated remittances, has impacted on the Romanian political-economic transition in the post-communist period, particularly after Romanian accession to the EU. System-wide migration impacts transmitted through both remittances and

viii

Introduction

labour market mechanisms are analysed for Romania, using a long-term time series (1991–2015) while accounting for endogeneity of migration and remittances. The period following the fall of communism is divided into three stages and the magnitude of these effects for each stage is tested. A net positive impact of migration on economic growth is found for the first and third of these stages, while remittances were also shown to enhance growth, particularly in the second stage. The main message to emerge from this study is that migration is not sufficient for economic development, as home country conditions need to continuously improve for the full development potential of outmigration to be realised. Population ageing has also featured in Professor Poot’s research (e.g., see Poot 2008). The late working-age population is often faced with the decision of when to exit the labour market. Frequently, this choice is made jointly with both a locational and a housing consumption decision. As workers priorities change, proximity to the workplace ceases to be a strong influence, while closeness to children and grandchildren, climate, and amenities takes on a pivotal role, and incomes fall with retirement, leading to downsizing in housing. Using data on recent retirees drawn from the 2005 to 2017 American Community Surveys and focusing on the interplay of retirement, migration, locational choice, and housing consumption, Kim and Waldorf aim to better understand the wide range of possible manifestations of these linked decisions. They hypothesise that income is the crucial factor driving retirees’ choices and find that spatial sorting resulting from behavioural differences sees retirees facing very different circumstances depending on their income. This would seem to be especially true of retirees choosing to move to or remain in rural areas where they experience lower housing costs but higher automobile reliance and poorer access to specialised medical care. Professor Poot has made a number of important contributions in the development of population projection methods (e.g., see Cameron and Poot 2011). Demographic projections or forecasts of future populations are a key input into decision-making processes, although the uncertainty that is associated with these projections is often underappreciated by decision-makers. In addition, a decision-maker faced with multiple population projections has no clear basis on which to decide between alternative projections. Cameron, Dunstan, and Cook examine the sources of uncertainty in population projections before discussing the development of population projection methods in New Zealand, emphasising the representation of uncertainty. Taking the example of subnational areas for the Waikato Region of New Zealand, the chapter concludes with the demonstration of a model-averaging approach that can be used by decision-makers to combine, without requiring substantial modelling capability, the information from multiple independent population projections. This approach may provide a more accurate and less risky approach to the use of population projections in decision-making. Professor Poot has been a great supporter of innovative uses of emerging data sources, such as Statistics New Zealand Integrated Data Infrastructure1 (e.g., see

1

See https://www.stats.govt.nz/integrated-data/integrated-data-infrastructure/

Introduction

ix

Alimi et al. 2018). Large or ‘big’ data sets containing information on labour markets and migration flows have become increasingly available, leading to increased interest in the application of machine learning methods to the prediction of migration flows and relevant shocks to labour markets. In the concluding chapter of this volume, Celbis presents a survey of research papers that use machine learning methods in policy-relevant regional economics and development studies and discusses the advantages that machine learning approaches may bring in the study of labour markets and migration. The chapter has a strong focus on social science perspectives and brackets out applications in other fields. The approach taken is qualitative, as the collection of available research is far from extensive, ruling out quantitative meta-analytic approaches. Overall, Celbis finds that while machine learning techniques can be useful, particularly where they are used as intermediate steps. These techniques are no substitute for econometric methods, especially if the research question requires the estimation of causal effect sizes. This collection of chapters reflects just some of the breadth of Professor Poot’s research and that of his collaborators, colleagues, students, and friends. It has been our great pleasure to work with Jacques over many years. Despite his ‘retirement’, he is showing little sign of slowing down (and may in fact be more productive now than ever, given his release from teaching responsibilities), and we look forward to continuing to collaborate with him in the future. Hamilton, New Zealand June, 2020

William Cochrane Michael P. Cameron Omoniyi Alimi

References Alimi OB, Maré DC, Poot J (2018) More pensioners, less income inequality? The impact of changing age composition on inequality in big cities and elsewhere. In: Stough R, Kourtit K, Nijkamp P, Blien U (eds) Modelling aging and migration effects on spatial labor markets. advances in spatial science (The regional science series). Springer, Cham Brosnan P, Poot J (1987) Modelling the determinants of transtasman migration after world war II. Econ Rec 63:313–329 Cameron MP, Poot J (2011) Lessons from stochastic small-area population projections: the case of Waikato subregions in New Zealand. J Popul Res 28 (2–3):245–265 Longhi S, Nijkamp P, Poot J (2005) A meta‐analytic assessment of the effect of immigration on wages. J Econ Surv 86:451–477 Maré DC, Pinkerton RM, Poot J, Coleman A (2012) Residential sorting across Auckland neighbourhoods. N Z Popul Rev 38(1):23–54

x

Introduction

Ozgen C, Nijkamp P, Poot J (2013) The impact of cultural diversity on firm: evidence from Dutch micro data. IZA J Migr 2:18 Poot J (2008) Demographic change and regional competitiveness: the effects of immigration and ageing. Int J Foresight Innov Policy 4(1,2):129–145 Poot J, Waldorf B, van Wissen L (eds) (2008) Migration and human capital. Edward Elgar, Cheltenham Poot J, Alimi O, Cameron MP, Maré DC (2016) The gravity model of migration: the successful comeback of an ageing superstar in regional science. Investig Reg 36:63–86 Tran NTM, Cameron MP, Poot J (2018) Local institutional quality and return migration: recent evidence from Vietnam. Int Migr 57(4):75–90

Contents

1

Earnings and Employment in Foreign-Owned Firms . . . . . . . . . . . David C. Maré, Lynda Sanderson, and Richard Fabling

1

2

The Widening and Deepening of Human Capital . . . . . . . . . . . . . . Ian Pool

41

3

Foreign Vs. German Wage Differentials in Germany: Does the Home Country Matter? . . . . . . . . . . . . . . . . . . . . . . . . . . . Stephan Brunow and Oskar Jost

4

Bilingualism in the Labour Market . . . . . . . . . . . . . . . . . . . . . . . . . Joanna Clifton-Sprigg and Kerry L. Papps

5

The Contrasting Importance of Quality of Life and Quality of Business for Domestic and International Migrants . . . . . . . . . . . Arthur Grimes, Kate Preston, David Maré, Shaan Badenhorst, and Stuart Donovan

55 77

97

6

Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy Initiatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 Malachy Buck and Peter Batey

7

Outmigration and Remittances as Facilitating Conditions for Economic Transition in Romania . . . . . . . . . . . . . . . . . . . . . . . 143 Cristian Incaltarau, Liviu George Maha, Karima Kourtit, and Peter Nijkamp

8

Retirement, Relocation, and Residential Choices . . . . . . . . . . . . . . . 181 Ayoung Kim and Brigitte S. Waldorf

xi

xii

Contents

9

The Development of Uncertainty in National and Subnational Population Projections: A New Zealand Perspective . . . . . . . . . . . . 197 Michael P. Cameron, Kim Dunstan, and Len Cook

10

Applications of Machine Learning Models in Regional and Demographic Economic Analysis: A Literature Survey . . . . . . . . . . 219 Mehmet Güney Celbiş

Chapter 1

Earnings and Employment in Foreign-Owned Firms David C. Maré, Lynda Sanderson, and Richard Fabling

Abstract This paper examines remuneration patterns among workers in foreignowned firms operating in New Zealand. By tracking workers as they move across jobs in different types of firms, we document the extent of the “foreign wage premium” distinguishing between compositional factors (e.g., differences in industry and employment composition between foreign and domestic firms) and remaining differences in wage levels and growth rates. We find that much of the average earnings gap between foreign- and domestically-owned firms is due to compositional factors—foreign firms tend to be larger and to employ workers who would have received relatively high wages regardless of where they worked. However, even among apparently similar workers and firms, we find a two to four percent earnings gap between workers in domestic and foreign-owned firms. This gap is primarily associated with a wage increase of around two percent on moving from a domestic to a foreign firm, augmented by higher wage growth within foreign-owned firms. However, these premia appear to be specific to foreign-firm employment, as workers who return to domestically-owned firms do not appear to retain these additional earnings in their subsequent jobs.

Access to the anonymised data used in this study was provided by Statistics New Zealand in accordance with security and confidentiality provisions of the Statistics Act 1975, and secrecy provisions of the Tax Administration Act 1994. The findings are not Official Statistics. The results in this paper are the work of the authors, not Statistics NZ, the Ministry of Business, Innovation and Employment, or Motu Economic and Public Policy Research, and have been confidentialised to protect individuals and businesses from identification. See Maré et al. (2014) for the full disclaimer. D. C. Maré (*) Motu Economic and Public Policy Research Trust, Wellington, New Zealand e-mail: [email protected] L. Sanderson Ministry of Business, Innovation and Employment, Wellington, New Zealand R. Fabling Independent Researcher, Wellington, New Zealand © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 W. Cochrane et al. (eds.), Labor Markets, Migration, and Mobility, New Frontiers in Regional Science: Asian Perspectives 45, https://doi.org/10.1007/978-981-15-9275-1_1

1

2

1.1

D. C. Maré et al.

Introduction

Foreign direct investment (FDI) has the potential to raise domestic productivity and increase incomes. FDI is often argued to be a source of direct benefits to the receiving firm, through improvements in management capability and greater access to advanced technologies, international networks, and financial capital. If the benefits of improved productivity and profitability are shared with local workers, this can in turn lead to higher incomes for domestic residents. Meanwhile, knowledge embodied in foreign-owned firms may also be available to other local firms, via observation, via transactions with local suppliers and customers, through product market competition, and through labour mobility. Such benefits (both direct and indirect) are often cited as a rationale for reducing barriers to FDI and supporting greater foreign investment. This paper explores a key potential source of economic benefits from foreign direct investment—human capital accumulation and earnings increases for employees of foreign-owned firms. International research consistently shows a significant gap between the average wages and salaries earned by workers in domestically-owned firms and those under foreign ownership or control (Lipsey 2004; Hijzen et al. 2013). We examine the drivers of this foreign wage premium for New Zealand by tracking workers as they move between foreign- and domesticallyowned firms, distinguishing between compositional factors (e.g., differences in industry and employment composition) and the foreign premium per se. Following the taxonomy developed by Malchow-Møller et al. (2013), we separately identify the role of worker heterogeneity, firm heterogeneity, and heterogeneity in the learning opportunities available in foreign-owned firms. Building on the work of Malchow-Møller et al. (2013), we provide an alternative method of estimating the relative importance of worker, firm, and learning heterogeneity, based on the observation of individual workers moving between foreign and domestic firms. We also extend the analysis to consider heterogeneous responses for different worker types and firm types. Finally, we provide an assessment of the role of selection and job-matching in explaining the lack of any apparent long-term gains to workers from foreign-firm experience. In particular, we look at whether workers who appear to be a particularly good match to a given firm, based on having a larger positive residual in their starting earnings, are more likely to stay with that firm and hence be excluded from the analysis of future job changes, and whether employees of foreign-owned firms are more likely to transition to other foreign firms rather than returning to the domestic sector. While this analysis shows evidence of selective transition out of foreign-owned firms, the bias due to this selectivity appears to be small. We find that while firm and worker composition explain most of the observed wage gap between foreign and domestic firms, a foreign premium of between 2.7 and 3.5% remains after controlling for composition. This premium is primarily the result of a firm-specific wage premium, with workers who join foreign firms gaining on average a two percent higher wage increase than those moving to comparable

1 Earnings and Employment in Foreign-Owned Firms

3

domestic firms. There is evidence of a small learning premium, with workers in foreign-owned firms experiencing slightly stronger within-job wage growth than those in domestic firms. However, workers do not appear, on average, to retain any of the extra wage gains experienced during their employment in foreign-owned firms, suggesting that the experience gained in these firms is not especially highly valued by domestic employers. Section 1.2 provides a brief review of the recent literature. Section 1.3 sets out the conceptual framework on which our analysis is based. Section 1.4 describes the data, while Sect. 1.5 describes the analysis of wage impacts. Section 1.6 concludes.

1.2

Literature Review

Significant gaps in average wages between foreign and domestically-owned firms have been documented across a range of countries, including New Zealand (Lipsey 2004; Hijzen et al. 2013; Fabling and Sanderson 2014). Although recent empirical research in New Zealand provides little evidence of productivity improvements associated with FDI (Doan et al. 2015; Fabling and Sanderson 2014), the significant wage gap between foreign and domestic firms, alongside post-acquisition increases in average wages and employment documented in Fabling and Sanderson (2014) provide an a priori indication that the presence of foreign-owned firms improves opportunities for domestic workers to gain high-income employment. However, the economic implications of this wage gap differ depending on the source of the gap. For example, if average wages rise in foreign-owned firms solely because these firms selectively hire highly skilled workers who could have earned a similar income elsewhere in the economy or because foreign owners bring in highly paid executives from offshore, the net gain to New Zealand of foreign ownership may be minimal. In contrast, if foreign-owned firms offer higher wages for a given level of skill and experience or allow workers to gain skills and knowledge which are of value to them and to their future employers, foreign investment can have a positive effect on aggregate labour market outcomes. From a policy perspective it is important to understand not only whether foreign firms are having a positive impact on earnings and human capital accumulation on average, but also whether there are differences in these impacts across different types of firms or workers. For example, Huttunen (2007) finds that positive wage impacts from foreign acquisition are concentrated among university-educated workers, and Pesola (2011) finds that more educated workers are also more likely to retain the wage premium associated with foreign-firm experience when they move to a domestic firm, implying that FDI may increase the wage gap between skilled and unskilled workers.1 Andrews et al. (2009) find that acquisition impacts are stronger for firms

Taylor and Driffield (2005) find that FDI has been a significant contributor to wage inequality in the UK manufacturing sector between 1983 and 1992.

1

4

D. C. Maré et al.

and workers in the service sector, while Girma and Görg (2007) find little difference in wage impacts between firms in the skill-intensive electronics sector and those in the low-technology food manufacturing sector. To the extent that the government is able to influence the composition of FDI flows into the country (e.g., through restrictions on foreign investment or targeted support provided to potential investors), a better understanding of the effect of different types of investment can help identify where to focus government intervention. While the existence of a substantial difference between average wages in foreign and domestic firms is well documented, much of this gap can be explained by the characteristics of the workers and firms involved (see Lipsey 2004, for a review).2 For example, simple controls for firm size, industry composition, and observable measures of workers’ human capital reduce the observed FDI premium in Ghana from 65 to 8.5% (Görg et al. 2007), and from between 10 and 19% (for non-production and production workers, respectively) to between 1.2 and 7.3% in the USA (Doms and Jensen 1998).3 However, even after controlling for observable differences, a significant wage gap remains in many countries. To address remaining compositional issues (unobserved worker quality and firm characteristics), the literature on earnings impacts of FDI has tended to take one of the two approaches: tracking either average or individual wages at firms which transition between foreign and domestic ownership (e.g., Heyman et al. 2007; Huttunen 2007; Fabling and Sanderson 2014) or tracking individuals as they transition between firms under different ownerships (Pesola 2011; Martins 2005), or both (Andrews et al. 2009; Hijzen et al. 2013). The former provides a control for selective acquisition of higher performance targets based on (time-invariant) unobservable characteristics of the firm, while the latter controls for selection into foreign firms based on the unobservable characteristics of individual workers. In this paper we focus on worker transitions between firms. This decision is driven by a combination of conceptual and practical reasons. From a purely practical perspective, annual information on foreign ownership is available for only a subset of firms in the data that we use, and relatively few firms transition from domestic to foreign ownership over the observation period (Sanderson 2013; Fabling and Sanderson 2014). As such, a focus on worker transitions provides a much larger sample and reduces the scope for measurement error compared to analysis of firm transitions. Moreover, where local MNEs (domestically-owned firms with subsidiary companies located offshore) can be distinguished from other domestic firms, the “foreign” wage premium is shown to be more strongly associated with multi-national status than with foreignness per se (Doms and Jensen 1998; Heyman et al. 2007; Iammarino and McCann 2013). 3 Foreign wage premia are generally found to be stronger in developing than developed countries (Hijzen et al. 2013), which may reflect larger differences in the characteristics of foreign and domestic firms in these countries, greater concerns about retention of trained workers in an environment with weaker intellectual property protection and/or lower levels of skill and education in the wider labour force, and international rent-sharing across countries (on the latter, see Budd et al. 2005; Egger and Kreickemeier 2013). 2

1 Earnings and Employment in Foreign-Owned Firms

5

From a conceptual perspective, a focus on worker transitions provides a long-run perspective on the impacts of foreign ownership. Employment and wage patterns are likely to differ from their long-run equilibrium in the years immediately following acquisition, as the new owners may restructure the existing operations or bring in an interim management team from offshore. Similarly, consideration of acquisition effects can shed light only on the impact of contemporary foreign investment into existing businesses, not that of earlier acquisitions or greenfields investment. By considering transitions of workers between firms, we reduce the potential for our results to be affected by short-term, transitional changes in wages and employment patterns and allow for consideration of a broader range of FDI impacts. The downside of this approach is that if FDI patterns have changed substantially over time, the wage impact in the current stock of foreign-owned firms may not accurately reflect the potential effect of the marginal investor. Cartwright (2001) and Gawith (2002) argue that while historical FDI into New Zealand has been primarily either market-seeking or resource-seeking, more recent investments have targeted high-potential technology firms with the aim of incorporating the technologies and skills held by the New Zealand firm into the wider organisation. To acknowledge the possibility that this change in motivation has affected the composition, and hence the average impact, of FDI we consider differences in the estimated foreign wage premium across industries, regions, and firm size groups (Sect. 1.5.2). Looking beyond the direct impacts of foreign-firm experience on worker earnings, a further question is whether skills and knowledge acquired in foreign-owned firms can be transferred to other domestic firms, generating productivity spillovers.4 International research shows some support for the premise that experience in foreign-owned firms is valued by workers’ future domestic employers. Balsvik (2011) finds that workers with recent experience in foreign-owned firms contribute positively to domestic plant productivity, while Görg and Strobl (2005) find that, in Ghana, new firms founded by individuals with foreign-firm experience are more successful than those run by entrepreneurs who have not worked in foreign-owned firms. Focusing on worker impacts, both Martins (2005) and Pesola (2011) find that experience in foreign-owned firms is rewarded by future employers through higher wages, while Poole (2013) finds that the share of workers with foreign-firm experience also leads to increases in the average wages of continuing workers in domestic firms. She attributes this to productivity spillovers as domestic workers interact with, and gain knowledge from, those with foreign-firm experience. Although we do not directly examine the existence of productivity spillovers through labour mobility, Sect. 1.5 addresses the question of whether earnings increases gained in foreign firms are maintained when workers move to domestic firms.

4 While Doan et al. (2015) find little evidence of productivity spillovers from FDI, their work uses industry-level measures of supplier–customer relationships based on input–output tables. These measures do not take into account other forms of interaction, such as interfirm labour mobility, which may be an important source of knowledge transfer.

6

D. C. Maré et al. Wage gap Domestic firm 1

Foreign firm

Domestic firm 3 Worker 1 F2

H

R Worker 2

F1

D2

D1

S Domestic firm 1

Domestic firm 2

Domestic firm 3 Time

Fig. 1.1 Wage growth and employment in foreign-owned firms

1.3

Conceptual Framework

The analysis in Sect. 1.5 is based on the taxonomy developed by Malchow-Møller et al. (2013), distinguishing three potential explanations for the observed foreign wage premium: 1. heterogeneous workers; 2. heterogeneous firms; and/or 3. heterogeneous learning. These effects are illustrated in Fig. 1.1. The solid black line shows the wage level of a hypothetical worker (“worker 1”) who moves from a domestically-owned firm to a foreign-owned firm and then to another domestic firm. The lower solid line shows the wage level of a worker (“worker 2”) who works for three different domestically-owned firms over the same time period. For comparison, the dotted black line represents a hypothetical wage path for worker 1, had they worked for the same employers as worker 2. The “heterogeneous worker” hypothesis refers to the possibility that foreignowned firms may selectively employ workers who would have earned relatively high wages regardless of their employer, due to above-average levels of skill or experience. This “skill gap” or “selection effect” is shown in the diagram as “S”—the ex ante wage gap between worker 1 and worker 2. Such a gap might be expected to arise if there are complementarities between skill levels and the technology or production processes applied in foreign firms. The “heterogeneous firm” hypothesis refers to the situation in which foreign firms pay the same worker a higher wage than the worker could receive in a domestic firm. This may reflect rent-sharing in the presence of productivity or profitability differentials (Katz and Summers 1989), compensating differentials for real or perceived

1 Earnings and Employment in Foreign-Owned Firms

7

lower job security (Bernard and Sjöholm 2003; Görg and Strobl 2003), or efficiency wages to promote greater work effort or to discourage workers from resigning if, for example, foreign firms face greater hiring or monitoring costs or are concerned about transfer of proprietary knowledge (Fosfuri et al. 2001; Glass and Saggi 2002).5 The effect of selection into high-paying foreign firms is shown in the diagram by the gaps (F1-D1) and (F2-D2). Both workers gain a wage increase on moving to a new firm, but the wage gain by worker 1 from moving to a foreign firm (F1) is larger than that experienced by worker 2 (D1). When worker 1 leaves the foreign firm and returns to a domestic firm, their wage falls as they lose the benefit of the foreign-firm wage premium, while worker 2 again receives a small wage increase (D2).6 Finally, the “heterogeneous learning” hypothesis allows for the possibility that workers pick up additional skills or knowledge from working in a foreign firm, which may be reflected both in their earnings trajectory within the foreign-owned firm, and their earnings levels in later jobs. This is shown in the diagram by H and R. H reflects the more rapid wage growth experienced by workers during their time at a foreign firm leading to a higher ending wage level, while R shows the “retained premium”, in which some portion of the wage gain accumulated by the worker in their time at the foreign firm is retained when they subsequently return to a domestic firm. The difference between the wage level premium at the end of a foreign-firm job and the retained premium in future domestic firm employment reflects the degree to which skills acquired in the foreign firm are valued by future domestic employers.7 These hypotheses are not mutually exclusive, and the effects may interact with each other. For example, higher potential learning opportunities may also depress the starting wage premium if workers recognise that foreign-firm experience will raise their lifetime earnings capacity and are willing to accept a lower initial wage in return for the additional learning opportunities provided through their employment. The extent to which learning opportunities affect starting wages will in turn depend on the specificity of the skills provided—workers will be more willing to accept low starting wages if the skills they expect to gain are applicable across a range of alternative workplaces, rather than being specific to the foreign employer—

Foreign-owned firms may also pay higher wages and/or provide better working conditions if they are more closely held to account by either local authorities or international customers than domestic firms, particularly in countries with lower enforcement of labour standards. 6 More correctly, the diagram could allow for factors such as the age-wage profile, with slower within and between job wage growth later in the life cycle. These refinements are omitted for simplicity. 7 An additional possibility, which we do not consider here, is that knowledge spillovers and complementarities between workers may affect the earnings of workers who remain in domestic firms as well as those who move into foreign firms. When worker 1 returns to a domestic firm, knowledge transfer and skill complementary may raise the productivity and hence the earnings of worker 2. If this is the case, the estimated residual impact of working in a foreign firm may be biased downwards, as the control group of workers who remain in domestic firms will also have their earnings raised through contact with other workers. Poole (2013) finds evidence of spillovers of this type in Brazilian firms. 5

8

D. C. Maré et al.

and the extent of credit constraints which reduce workers’ ability to smooth consumption over the life cycle.8 In practice, a range of other factors may also affect the observed earnings differential between foreign and domestic firms. For example, if foreign firms are less likely to employ part-time staff, higher average earnings may reflect longer hours worked. In addition, if foreign firms are more likely to bring in employees from offshore, some of these workers may be paid more than local staff to reflect dislocation costs or to match their earnings in their home markets. To limit the impact of such compositional differences, we restrict attention to those employees for whom we observe a clean transition between full-time jobs within New Zealand. Supplementary analysis of worker hours and labour sourcing behaviour in Maré et al. (2014) suggests that neither of these factors are materially affecting the results.

1.4

Data

We make use of monthly individual-level earnings data linked to firm characteristics from Statistics New Zealand’s Integrated Data Infrastructure (IDI). The IDI is a linked longitudinal database that brings together data on individuals and households, including wage and salary information from Inland Revenue and firm-level information from a range of survey and administrative sources held in the Longitudinal Business Database (LBD). Employment information is available over the period from 1999 to 2011. The unit of observation used in this analysis is a “job” (job spell)—a continuous period of employment of an individual at a firm.9 Spell-level observations are used in preference to a panel of monthly employment observations as the former provide a convenient method to control for both spell durations and gaps between spells, while the latter would be computationally infeasible for the full population of employeemonths. To accommodate information on an individual’s previous and future labour market status, the main analysis is restricted to job spells that commence after May 2000 and conclude before April 2010. The population for the examination of wage dynamics is restricted to a “balanced panel”—those job spells for which we observe clean transitions between full-time jobs at both the start and the end of the spell.10 Employment information is sourced

Pesola (2011) finds no evidence that Finnish workers pay for foreign experience in the form of lower starting wages. 9 We exclude all periods of employment where the employee has ever received income as a working proprietor of that firm, as there are empirical and conceptual issues with determining the appropriate measure of earnings when workers have an ownership interest. To accommodate periods of leave and other short breaks in employment, we allow for one-month gaps in income receipt within a job spell. Where an income gap extends beyond one month, the periods before and after the gap are treated as two separate jobs with the same employer and are excluded from the main analysis. 10 Full-time status is identified following Maré and Hyslop (2007). 8

1 Earnings and Employment in Foreign-Owned Firms

9

from Pay-As-You-Earn (PAYE) tax returns, which are submitted monthly by all employing firms. These capture all forms of labour income, including wages and salaries, bonuses, and commissions. Starting incomes are calculated in the second month of employment, and ending incomes in the second-to-last month of employment to avoid these measures being contaminated by part-months of employment and unusually large final payments (e.g., severance pay).11 Nominal earnings are adjusted to reflect changes in the consumer price index over the period. Tenure is defined as the total number of months that an individual is employed in a given job spell. Worker quality, or “skill”, is captured through estimates of worker fixed effects, following Maré and Hyslop (2007). These estimates are based on a separate regression of log annual full-time equivalent earnings (yi jt) on observable worker characteristics xit (a flexible function of gender and age), worker fixed effects (θi), firm fixed effects (ψ j) and annual time dummies τt, yi jt = θi + ψ j + xi jt β + τt + εi jt. The worker fixed effects provide an indication of a worker’s earnings potential, capturing a range of time-invariant characteristics not observed in the data including education and innate ability, as evidenced by the relative income of each worker across all their jobs after controlling for observable worker characteristics and the time-invariant effect associated with each firm they work for. At the firm level, foreign ownership is defined as having either 50% or higher recorded foreign ownership in the Longitudinal Business Frame (LBF), and/or a positive response to the disclosure question “Is the company controlled or owned by non-residents?” from the IR4 Company Tax return.12 While the IR4 is filed annually by almost all limited liability companies, updates to foreign ownership information in the LBF are primarily based on responses to the Annual Frame Update Survey, which is full-coverage only for the largest firms.13 As such, information on foreign investment is less reliable for small, non-corporate firms (e.g., sole-proprietors and partnerships). At the same time, the specific questions that are used to identify foreign ownership across the two sources differ, implying that some firms may legitimately respond positively to one but negatively to the other. We therefore take all point-in-time ownership statuses associated with actual survey responses, tax returns, and manual adjustments by Statistics New Zealand’s Business Frame operators,14 and use these to distinguish four types of firms based on their “permanent” ownership status over the observed life of the firm: firms that are “always” foreign-owned at every observation; firms that are “never” foreign-owned; firms that 11 In cases where an individual receives no income in the relevant month, we use earnings from the third (or third-to-last) month of employment. 12 Where firms are part of a group of parent-subsidiary enterprises, we give precedence to responses of the individual firm. If no information is available at the firm level, and the information provided by other group members is consistent, firms are allocated to domestic or foreign ownership based on the group response. 13 See Sanderson (2013) for a discussion of alternative sources of FDI information in the LBD. 14 These adjustments are made in response to information about firm ownership from other sources, including other Statistics New Zealand surveys and media reports.

10

D. C. Maré et al.

Table 1.1 Foreign ownership status, as at 31 March 2011 Proportion of employing firms Proportion of employees Proportion of full-time equivalent employment

Always 2.33 8.73 9.44

Mixed 3.72 12.16 12.59

Never 82.62 75.87 75.19

Unknown 11.33 3.24 2.78

Total 135,564 1,446,100 1,237,600

are “mixed” or “sometimes” foreign owned, across time, data sources, or both; and firms for which we have no reliable information about their ownership status (“unknown” ownership).15 All four groups are used in the analysis, with a focus on the comparison of firms that are “always” and “never” foreign-owned. Table 1.1 reports the prevalence of each firm type as at 31 March 2011.16 The link between foreign ownership status and firm size is apparent, with the two percent of firms that are “always” foreign and the four percent with “mixed” ownership accounting for around nine and twelve percent of employment, respectively. Conversely, domestic firms and those with no ownership information account for a larger share of firms than employment. A similar pattern, though less pronounced, is apparent in the comparison between headcount and full-time equivalent (FTE) employment, with foreign-owned firms showing a stronger tendency towards employing full-time staff.17 The prevalence of foreign ownership differs substantially across industries, and to a lesser extent, across labour market regions (LMRs) (Tables 1.2 and 1.3).18 The proportion of “always” foreign firms ranges between 0.1% in agriculture and 13.7% in petroleum, chemical, polymer, and rubber manufacturing (0.3 and 33.0% of employment, respectively). Missing ownership information is particularly apparent in agriculture, where there are many small, owner-operator firms, but is also noticeable in many service industries. Differences across LMRs are less pronounced, but an urban bias is apparent in foreign investment. Always foreign-owned firms account for 13.3% of employment in Greater Auckland, compared to 7.1% across other urban areas and 4.1% in non-urban regions. Considered at the firm level, the average earnings gap between foreign- and domestically-owned firms is substantial. Figure 1.2a plots the distribution across firms of the mean log monthly earnings of full-time employees as at March 2011. 15 A small number of apparent single-year transitions into and out of foreign ownership from IR4s are ignored where they are inconsistent with other sources of FDI information. 16 As our definition of ownership is based on “permanent” ownership status over the life of the firm, there is little variation in reported foreign ownership rates over time aside from an initial decrease in the proportion of firms and employment allocated to the “unknown” ownership category as firms for which we have no FDI information exit the population. 17 As our method of identifying full time employment is based on wage and benefit income receipt, rather than hours information, the distinction between foreign- and domestically-owned firms may be overstated, as high-wage employees are less likely to be identified as working part time. 18 Labour market regions are defined based on labour catchments, following Papps and Newell (2002).

NZSIOC Agriculture Forestry and logging Primary industries support services Mining Manufacturing: Food, beverage and tobacco products Textile, leather, clothing and footwear Wood and paper products Printing Petrol., chem., poly. and rubber products Non-metallic mineral products Metal products Transport equip., mach. and equipment Furniture and other Electricity, gas, water & waste services Construction Wholesale trade Retail trade Accommodation and food services Transport, postal and warehousing Information media and telecommunications Financial and insurance services Rental, hiring and real estate services Professional, scientific and tech. services

Percentage of firms Always Mixed 0.14 0.74 2.05 1.37 0.40 0.81 9.20 10.34 4.01 5.84 6.37 6.37 1.03 3.08 2.29 2.80 1.69 5.49 13.66 13.98 2.50 7.50 2.23 6.68 3.67 5.31 0.89 2.07 6.59 10.18 0.94 1.26 8.71 12.52 3.04 3.17 0.57 1.61 3.31 6.55 7.94 16.14 10.80 10.46 1.09 4.41 2.73 4.74 Never 70.13 87.67 86.41 77.01 86.01 77.62 89.38 92.11 90.30 71.12 86.88 88.87 88.87 92.31 77.84 90.35 76.65 89.48 87.30 81.12 72.49 71.09 76.37 85.44

Table 1.2 Prevalence of foreign ownership by industry, March 2011 Unknown 28.99 8.90 12.38 3.45 4.14 9.63 6.51 2.80 2.53 1.24 3.13 2.23 2.15 4.73 5.39 7.45 2.13 4.30 10.52 9.02 3.44 7.65 18.14 7.09

Total 14,673 438 2229 261 11,604 2118 876 1179 711 966 480 1887 2373 1014 501 14,019 8748 14,292 10,980 4623 1134 2667 4698 11,769

Percentage of employees Always Mixed Never 0.30 1.97 79.73 1.98 0.91 94.51 0.98 2.17 91.67 20.68 31.02 47.57 13.90 15.87 69.58 14.48 17.41 67.44 12.00 6.00 80.78 14.81 11.25 73.44 13.48 24.51 61.27 33.04 17.39 49.28 8.12 15.00 75.90 6.29 18.86 73.86 9.90 15.55 74.22 2.18 3.42 93.24 7.67 20.02 71.73 11.11 5.08 81.80 18.85 19.23 61.39 9.55 17.73 71.82 5.76 8.98 81.71 10.58 15.25 72.98 45.08 26.23 28.45 27.69 45.16 26.20 5.29 13.46 74.52 9.25 12.43 76.62 Unknown 18.00 2.59 5.18 0.72 0.66 0.67 1.22 0.50 0.75 0.29 0.97 1.00 0.34 1.17 0.58 2.01 0.53 0.91 3.55 1.18 0.24 0.94 6.73 1.70

(continued)

Total 46,660 3280 14,290 4835 177,495 68,360 8170 16,885 7345 17,250 5665 19,090 28,295 6435 11,990 84,600 78,520 154,000 90,200 64,260 27,065 46,940 20,800 94,100

1 Earnings and Employment in Foreign-Owned Firms 11

NZSIOC Administrative and support services Education and training Health care and social assistance Arts and recreation services Other services Total

Table 1.2 (continued) Percentage of firms Always Mixed 2.36 5.86 0.22 1.89 1.02 2.11 0.42 1.17 0.56 0.83 2.33 3.72 Never 80.86 85.93 86.84 83.79 76.61 82.62

Unknown 10.92 11.97 10.04 14.62 22.00 11.33

Total 4452 5565 7680 2832 11,196 135,564

Percentage of employees Always Mixed Never 8.24 28.74 61.18 0.18 2.94 85.65 2.81 4.97 90.78 1.01 2.28 93.73 2.19 3.98 82.90 8.73 12.16 75.87 Unknown 1.85 11.23 1.44 2.99 10.93 3.24

Total 59,500 132,640 167,100 26,780 50,300 1,446,100

12 D. C. Maré et al.

LMR Northland West Northland East Greater Auckland Thames—Coromandel Greater Hamilton Taranaki Rural Taranaki Urban Tauranga North Central North Island Gisborne—Opotiki Napier—Hastings Hawkes Bay—Central North Island Rural Palmerston North Wanganui Horowhenua—Wairarapa Wellington Urban Nelson—North of West Coast Marlborough—North Canterbury Greater Christchurch South Westland—Rural South Canterbury Central Otago—North and East Southland Dunedin Greater Invercargill and Stewart Island Missing/Undefined

Percentage of firms Always Mixed 0.75 1.5 1.78 2.97 3.14 5.06 0.73 1.38 2.15 2.96 0.57 1.59 3.51 4.39 2.21 3.71 1.55 2.19 1.76 1.89 2.37 3.62 0.83 1.3 3.00 4.5 2.21 3.41 1.14 2.53 2.98 4.52 1.87 2.94 1.36 3.39 2.36 3.98 1.17 2.57 0.79 1.46 2.83 3.86 1.62 2.38 10.11 5.62 Never 78.76 81.82 83.6 80.65 82.86 72.92 77.83 84.09 83.31 77.71 83.48 77.28 81.48 80.92 77.69 84.45 83.32 82.58 84.02 85.51 79.44 82.68 82.17 56.18

Unknown 18.98 13.42 8.20 17.25 12.03 24.91 14.27 9.99 12.95 18.64 10.52 20.59 11.01 13.45 18.63 8.06 11.87 12.67 9.64 10.75 18.31 10.63 13.82 28.09

Table 1.3 Prevalence of foreign ownership by labour market region, March 2011 Total 1596 3531 38,331 3705 7404 2637 2733 5016 5607 2382 4305 2535 3597 1494 2367 11,688 4497 2652 13,476 6417 2670 3498 3147 267

Percentage of employees Always Mixed Never 4.43 5.18 83.8 4.11 9.25 81.85 13.28 15.38 68.96 6.11 7.42 79.91 6.92 9.06 81.26 1.63 9.66 79.66 8.38 13.33 74.64 3.99 10.64 81.6 6.25 8.27 80.85 4.92 5.53 82.92 4.89 10.89 80.67 2.21 10.46 79.95 7.41 12.1 77.04 4.48 7.11 84.64 3.99 8.68 81.55 9.67 13.01 75.46 3.67 11.74 79.95 3.45 8.63 83.56 8.77 12.39 75.73 3.29 8.02 85.39 2.40 6.43 84.75 4.50 8.5 84.25 2.89 7.98 85.47 4.89 13.53 60.15 Unknown 6.59 4.79 2.38 6.55 2.77 9.05 3.66 3.77 4.64 6.63 3.56 7.38 3.46 3.77 5.78 1.85 4.65 4.37 3.11 3.29 6.43 2.75 3.66 21.43

(continued)

Total 10,620 29,200 492,300 22,900 79,500 16,570 26,260 45,100 49,600 18,090 45,000 16,260 40,500 14,060 17,290 167,500 40,900 18,550 157,400 48,600 17,110 40,000 30,070 2660

1 Earnings and Employment in Foreign-Owned Firms 13

Percentage of firms Always Mixed 3.14 5.06 2.42 3.68 1.06 2.09 2.33 3.72 Never 83.60 82.93 80.96 82.62

Unknown 8.20 10.98 15.89 11.33

Total 38,331 66,777 30,291 135,564

Percentage of employees Always Mixed Never 13.28 15.38 68.96 7.05 11.18 78.65 4.07 8.28 82.30 8.73 12.16 75.87 Unknown 2.38 3.12 5.35 3.24

Total 492,300 733,400 218,700 1,446,100

Labour market regions are groupings of labour market catchments as used in Newell and Callister (2009), defined using the algorithm described in Papps and Newell (2002). The classification is available from the authors on request. Individuals are allocated to LMRs according to the location of their employing firm, as recorded in the LBF. Where a firm operates across multiple regions, individuals are allocated to geographic units by Statistics New Zealand based on information about the relative employment in each plant and the residential or postal address of the employees. “Urban” LMRs are those where the majority of employment is in a major urban area

LMR Greater Auckland Other Urban Non-urban Total

Table 1.3 (continued)

14 D. C. Maré et al.

15

Density

Density

1 Earnings and Employment in Foreign-Owned Firms

7.5

8

8.5 9 9.5 Log of wages Never FDI Mixed FDI Always FDI

10

–1

–.5

1

Density

(b)

Density

(a)

0 .5 Firm fixed effect Never FDI Mixed FDI Always FDI

–1

–.5 0 .5 Mean worker fixed effect Never FDI Mixed FDI Always FDI

1

1

2 3 45 10

(c)

25 50 100 250 500 1000 Employment (log scale) Never FDI Mixed FDI Always FDI

5000

(d)

Full-time employees only. One observation per firm. Tails of each distribution compressed in accordance with Statistics New Zealand confidentiality protocols.

Fig. 1.2 Components of foreign earnings premium, March 2011

While the distribution of log earnings for domestic firms is concentrated between 7.9 and 8.7 (monthly earnings of $2700–6000), that for always foreign-owned firms is wider, and centred between 8.2 and 9.3 ($3600–9900). Looking across all job spells, the average monthly starting wage in a domestic firm is $3735, compared to $5685 in always foreign-owned firms, a gap of over 50% (Table 1.4). Figures 1.2b and 1.2c distinguish two components of the overall wage gap—that associated with differences in the firm fixed effect, which captures whether a firm is a relatively high or low wage employer, and that associated with the mean worker fixed effect across the firm’s employees, which captures whether the firm hires workers who tend to be well paid regardless of where they work. Both components play an important role in explaining the gap in average wages between foreign and domestic firms. However, measured as a firm-level average, this mean wage gap overstates the worker-level average due to differences in the distribution of firm size across ownership groups (shown in Fig. 1.2d). While domestic firms tend to be very small (most have fewer than 10 employees, with average employment of 15.9), foreign-owned firms are much larger, with average employment of 111.7 (Table 1.4)

16

D. C. Maré et al.

Table 1.4 Raw wage premia, firm- and worker level Never foreign Mean Firm level Starting wage $3735 Ending wage $3833 Firm size (E) 15.85 Firm size (lnE) 1.81 N 99,654 Worker level (balanced spells) Starting wage $4261 Ending wage $4483 Firm size (E) 480.56 Firm size (lnE) 3.79 N 699,000

Std. dev. $1581 $1502 117.57 1.09

$2645 $4023 1438.88 2.09

Always foreign Mean $5685 $6681 111.69 3.10 1848 $5109 $5781 1007.52 5.97 96,100

Std. dev. $2792 $10,496 352.30 1.73

$3915 $8795 1351.76 1.59

Sample criteria: Balanced full-time job spells commencing after May 2000 and concluding prior to April 2010. Consecutive spells within the same firm are excluded. Firm size is calculated as the mean employment across the start and end of each spell. Firm level means are calculated by taking the mean across all job spells within the firm which meet the above criteria, then calculating the mean and standard deviation across all firms with at least one applicable job.

and a substantial proportion of firms in the range of 25–1000 employees (Fig. 1.2d). As larger firms commonly pay higher wages (e.g., Oi 2004; Troske 1999), a firmlevel calculation that places equal weight on all firms accentuates the foreign wage premium. An alternative is to consider the distribution of wages at the worker level, as shown in Fig. 1.3. By placing equal weight on each worker, rather than each firm, this effectively down-weights the small firms that make up the bulk of the domestic firm population, making the two populations more comparable. This substantially reduces the gap between the two distributions, and reduces the apparent foreign starting wage premium from 50 to 20% (Table 1.4). However, there remains a substantial difference between foreign and domestic wage levels. Section 1.5 explores the source of this gap, with a focus on the three explanations outlined in Sect. 1.2—heterogeneous workers, heterogeneous firms, and heterogeneous learning. Summary statistics at the job level are provided in Table 1.5, separately for workers in always domestic and always foreign firms.19 The comparison of spells across foreign and domestic firms shows up a number of differences in the characteristics of both workers and jobs. Completed tenure is longer on average in foreign firms than domestic firms, with a mean length of 20.1 months compared to

19

Panel A is restricted to the regression population used in Sect. 1.5. This population is restricted to job spells where we observe a clean, full-time transition between jobs at both the start and the end of the relevant job spell. Workers moving in or out of part-time jobs, multiple job holders, and repeated employment spells with the same firm are excluded. Panels B and C relax these constraints, covering all observed job starts and ends within the period from June 2000 to March 2010.

17

Density

1 Earnings and Employment in Foreign-Owned Firms

7.5

8

8.5 9 Log of wages Never FDI Always FDI

9.5

10

Mixed FDI

Full-time employees only. One observation per worker. Tails of each distribution compressed in accordance with Statistics New Zealand confidentiality protocols.

Fig. 1.3 Worker-level foreign wage premium, March 2011

16.7 months.20 Panels B and C also show that workers in foreign firms are more likely to be working fulltime, and less likely to hold multiple jobs. While restricting to the balanced panel substantially reduces the sample size compared to the unrestricted sample of job spells, the two populations present a consistent picture with respect to the size of the raw wage premium (Table 1.6). The inclusion of less stable workers (including those coming from or moving into parttime or multiple job employment, but also workers who are first entering or just leaving the workforce) drives down the mean wage in the unrestricted sample compared to the balanced panel, but the relative foreign premia remain reasonably steady.21 Meanwhile, the stronger wage premium at the end of each spell compared to the start may reflect either a steeper wage profile or higher average tenure in foreign-owned firms. This stronger wage growth is also reflected in the raw wage changes associated with job transitions (Table 1.7). This table shows two things: for job starts and ends it shows the average wage change associated with a transition between employers

20

These estimates are lower than the average tenure found by Papadopoulos (2008). Longer spells are excluded from our analysis as we restrict to spells where we can observe a clean transition at either end of the job. Short spells (less than 3 months) are also excluded as monthly wage changes cannot be observed. 21 The lower panel of Table 1.6 maintains the requirement that the current job start or job end is full time to maintain comparability of earnings, but places no restrictions on either the existence or characteristics of past or future job spells or job characteristics at the other end of the current spell.

18

D. C. Maré et al.

Table 1.5 Summary statistics Ownership status of current employing firm Never foreign Always foreign Mean Std. dev. Mean Std. dev. Panel A: balanced spells Femalea Age at start Firm size (E) Firm size (lnE) Firm size (lnE) at start Firm size (lnE) at end Length of job (months) Entered this firm with no gap after previous job enda Gap after previous job before this job start (months) Change in firm size from previous job ( lnE) Previous job was in always foreign-owned firma Enters next firm with no gap between jobsa Gap after job end before next job start(months) Change in firm size in next job (lnE) Next job is in always foreign-owned firma Previous firm was smaller than current firma Previous job was in same industrya Previous job was in same regiona Next firm is smaller than current firma Future benefit recipienta Next job is in same industrya Next job is in same regiona N(obs) Panel B: all job starts after May 2000 Full-timea Multiple job holdera Current job spell in same firm as previous spella Previous job was in always foreign-owned firma Previous firm was smaller than current firma Previous job was in same industrya Previous job was in same regiona N Panel C: all job ends before December 2010 Full-timea Multiple job holdera Next job spell in same firm as current spella Next job is in always foreign-owned firma Next firm is smaller than current firma Next job is in same industrya

0.326 35.143 481 3.787 3.760 3.664 16.688 0.350 5.095
0.362 0.063 0.341 4.962 3.664 0.065 0.432 0.443 0.636 0.560 0.117 0.445 0.635 699,000

0.469 11.283 1439 2.090 2.100 2.153 16.123 0.477 9.993 2.459 0.243 0.474 9.808 2.153 0.246 0.495 0.497 0.481 0.496 0.322 0.497 0.481

0.383 34.717 1008 5.973 5.949 5.894 20.124 0.314 4.484 0.796 0.284 0.327 4.897
0.716 0.289 0.600 0.393 0.612 0.405 0.078 0.386 0.609 96,100

0.486 10.510 1352 1.586 1.589 1.630 17.615 0.464 9.513 2.444 0.451 0.469 10.064 2.436 0.453 0.490 0.488 0.487 0.491 0.268 0.487 0.488

0.403 0.491 0.176 0.381 0.095 0.294 0.030 0.170 0.274 0.446 0.295 0.456 0.408 0.492 7,893,700

0.566 0.135 0.069 0.203 0.438 0.284 0.382 688,000

0.496 0.341 0.253 0.403 0.496 0.451 0.486

0.414 0.186 0.099 0.033 0.717 0.297

0.573 0.144 0.069 0.211 0.574 0.256

0.495 0.351 0.254 0.408 0.494 0.436

0.493 0.389 0.298 0.178 0.451 0.457

(continued)

1 Earnings and Employment in Foreign-Owned Firms

19

Table 1.5 (continued) Ownership status of current employing firm Never foreign Always foreign Mean Std. dev. Mean Std. dev. 0.463 0.499 0.435 0.496 7,567,700 686,900

Next job is in same regiona N(obs)

Sample definitions: Panel A restricted to job spells where we observe a clean, full-time transition between jobs at both the start and the end of the relevant job spell – the population used in the main analysis. Workers moving in or out of part-time jobs, multiple job holders, and repeated employment spells with the same firm are excluded. Panel B covers all observed job starts from June 2000 onwards, while Panel C covers all observed job ends up to March 2010. Variables marked with “a” are binary variables set to 1 if the statement is true, 0 otherwise. Recent benefit recipient: Receiving benefit within 12 months prior to starting this job. Future benefit recipient: Receiving benefit within 12 months after this job ends Table 1.6 Raw wage premia (1) Never foreign Balanced panel Start 8.266 End 8.301 All full-time job starts/ends Start 8.170 End 8.181

(2) Always foreign

(3) Difference

(4) N(obs)

8.396 8.462

0.130 0.160

795,100 795,100

8.317 8.390

0.146 0.209

3,282,800 3,209,000

Columns 1 and 2 report mean log earnings at the start and end of each job spell, according to whether firms are always or never foreign owned. Column 3 reports the raw wage premium associated with foreign employment (difference between log average earnings). Upper section (balanced panel) restricts to job spells where a clean transition between two full-time jobs can be observed at both the start and end of the spell. Lower section (all jobs) includes all observations of full-time jobs which start after May 2000 and/or end prior to April 2010, regardless of whether the previous or subsequent job is observed

according to the ownership status of the two firms; within-jobs, it shows the average wage change between the first and last month of employment. For the balanced panel, workers moving between domestic firms (D ! D) on average experience 2.2% wage growth at the start of a job spell (relative to their earnings at the end of their previous job), followed by 3.5% growth within the spell. Moving from a domestic to a foreign-owned firm is associated with an average earnings gain of 5.5%, a 3.3 percentage point premium over the domestic-domestic average, while workers that transition from foreign-owned firms to domestic firms experience on average a 5.7% decrease in earnings.22 However, these averages are affected by a wide range of factors, including differences in tenure, and worker and firm 22 Including all clean transitions between two full-time jobs shows very similar results (lower panel, Table 1.7).

20

D. C. Maré et al.

Table 1.7 Raw wage transition premia (1)

(2)

(3)

(4)

N(obs) D!D F!F D!F Balanced panel Start 665,700 0.022 0.026 0.055 Job 798,700 0.035 0.065 End 680,800 0.014 0.008 0.040 All clean transitions between two full-time jobs Start 1,673,200 0.020 0.018 0.055 Job 1,827,900 0.037 0.063 End 1,675,600 0.019 0.018 0.056

(5) F!D
0.057
0.063
0.063
0.065

(6) (7) Relative to D ! D F!F D!F 0.004 0.030
0.005
0.002 0.026
0.001

(8) F!D

0.033


0.079

0.027


0.077

0.035


0.082

0.038


0.083

Population restricted to job transitions between firms which are “always foreign” (F) and “never foreign” (D). Upper panel restricted to job spells where a clean, full-time transition between jobs is observed at both the start and the end of the relevant spell. Workers moving in or out of part-time jobs, multiple job holders, and repeated employment spells with the same firm are excluded. Lower panel applies these restrictions only to the specific job transition in question. Columns 2–5 report the average (log) wage change associated with each of the four possible transition paths. Columns 6–8 report foreign firm premia relative to transitions between two domestic firms

composition. Differences between start and end wage changes may reflect differences in timing 23 and the tendency for wage growth to slow with age,24 as well as any impact of the composition of individuals making each transition. Section 1.5 focuses on distinguishing the underlying foreign wage premium from differences due to composition.

1.5

Wage Impacts of FDI Employment

To understand the role of worker, firm, and learning heterogeneity in explaining the foreign-firm wage premium, we now turn to a series of regression analyses, in which we consider the wage dynamics of workers moving between foreign- and domestically-owned firms while controlling for the observable characteristics of the worker, the firm, and the job spell. As a first step, we examine the size of the “true” foreign premium—the wage premium remaining after observable firm and worker characteristics have been controlled for. We then provide estimates for the various sources of foreign premium depicted in Fig. 1.1—the heterogenous worker

23

Spell ends are necessarily later in the period and thus may be affected by worsening economic conditions associated with aftermath of the Global Financial Crisis. 24 The larger difference between job starts and job ends in the balanced panel (where we are by definition observing these start and the end of the same spell) compared to the lower panel in which we observe a different set of spells for the start and end transitions, is consistent with reductions in wage growth with age.

1 Earnings and Employment in Foreign-Owned Firms

21

effect (S), the heterogeneous firm effect (F1-D1, F2-D2), and the heterogeneous learning effect (H). To understand the overall size of the foreign wage premium and the role of worker heterogeneity, we estimate two wage level regressions, for wages at the start and end of a job spell, which take the form ln W i jt ¼ α þ βa δalways þ βm δmixed þ βu δunknown þ γXit þ ϕZjt þ ψ t þ εi jt j j j

ð1:1Þ

in which log of monthly earnings of individual i in firm j at time t (lnWi jt) is regressedon a set of dummy variables capturing the permanent ownership status  of firm j δkj , k 2 falways, mixed, unknowng ,25 a set of worker-level control variables (Xit), firm-level control variables (Zjt) and a full set of time (month) dummies (ψ t). At the worker level we control for two time-invariant characteristics—gender and estimated worker fixed effects (WFE)—as well as a quadratic function of age and elapsed time and/or tenure which is adapted to suit the specific dependent variable in question. For the starting wage regression, elapsed time is defined as the gap between the end of the previous job and the start of the current job, to reflect the possibility that the length of time out of employment may affect both the wage offer made by employers and the worker’s reservation wage (Devine and Kiefer 1993; Rogerson et al. 2005).26 For the end wage regression, elapsed time reflects tenure in the current job.27 At the firm level, we control for log employment and its square, and a full set of industry and LMR dummies. In each case the βs reflect the wage premium for each type of ownership, relative to firms that are always domestically owned. Tables 1.8 and 1.9 report results for start and end wages, respectively. Column 1 shows raw wage gaps controlling only for time. Worker and firm characteristics are then introduced separately (columns 2 and 3, respectively), providing some indication of the extent to which the average gaps between foreign- and domesticallyowned firms are driven by differences in worker and firm composition. Column 4 combines both sets of covariates. The 13–15% higher average starting wage observed for “always” foreign firms in Table 1.6 is apparent in column 1 of Table 1.8, where wage is regressed only on ownership type and time effects. As might be expected, the wage premium associated with firms which are sometimes foreign owned is also positive, but weaker than that for the always foreign-owned firms, as this coefficient represents the average premium across all years, including Never foreign-owned firms form the reference group. See also Addison et al. (2009), who cast doubt on the assumption of a declining reservation wage. 27 For both start and end wages, the control variables are included as second order polynomials, containing the following terms: A1 , A2 , A21 , A22 , A1 A2 , where A1 is the worker’s age (in months) at the start of the relevant period, A2 is age at the end of the period, and the period in question is either the period of unemployment prior to a job start or the period of tenure prior to a job end. This specification controls for age, tenure and experience effects, though we cannot separately identify all three and thus cannot interpret the coefficients of the polynomial. 25 26

22

D. C. Maré et al.

Table 1.8 Starting wage premium

Foreign ownership Always Sometimes Unknown

(1) Raw

(2) +Worker chars

(3) +Firm chars

(4) All

0.142*** [0.002] 0.078*** [0.001]
0.019*** [0.002]

0.085*** [0.001] 0.054*** [0.001]
0.012*** [0.001]

0.043*** [0.002] 0.032*** [0.001] 0.007*** [0.002]

0.027*** [0.001] 0.021*** [0.001] 0.011*** [0.001]

Worker characteristics Female


0.194*** [0.001] 0.965*** [0.001]

WFE Firm characteristics ln E (ln E)2 Time dummies Age and tenure Ind dummies LMR dummies Adjusted R2 N(obs)

Yes No No No 0.032 966,900

Yes Yes No No 0.588 966,900


0.188*** [0.001] 0.923*** [0.001] 0.042*** [0.001]
0.002*** [0.000] Yes No Yes Yes 0.152 966,900

0.031*** [0.001]
0.002*** [0.000] Yes Yes Yes Yes 0.616 966,900

Significant at: *** 1%; ** 5%; * 10%. Robust standard errors in brackets. Age and tenure controls for age (at start of job), gap (since end of previous job), age2, gap2, and age* gap. Firm employment measured at start of job spell

those when the firm is domestically owned.28 Adding worker characteristics (column 2) reduces the foreign premium to around three fifths of the raw figure or 8.5%, while controlling for firm characteristics alone reduces the foreign premium to 4.3% (column 3). Together, the combined impact of firm and worker controls leaves an unexplained foreign premium of 2.7%. This reflects the role of factors such as productivity differentials between foreign- and domestically-owned firms, as well as any compensating differential or efficiency wages paid by foreign-owned firms.29

28

This relationship is consistent throughout later regressions, with the exception of the within-job wage growth premium analysis. 29 Further controls for firm performance (e.g., productivity, profitability) would likely reduce this premium further. However, as firm-specific advantages are a core part of the argument for a positive impact of FDI, controlling for these factors could be viewed as dismissing the very phenomenon we are interested in.

1 Earnings and Employment in Foreign-Owned Firms

23

Table 1.9 End wage premium

Foreign ownership Always Sometimes Unknown

(1) Raw

(2) +Worker chars

(3) +Firm chars

(4) All

0.170*** [0.002] 0.105*** [0.001]
0.022*** [0.002]

0.094*** [0.001] 0.066*** [0.001]
0.009*** [0.001]

0.059*** [0.002] 0.048*** [0.002] 0.005* [0.002]

0.035*** [0.001] 0.029*** [0.001] 0.013*** [0.001]

Worker characteristics Female


0.184*** [0.001] 0.993*** [0.001]

WFE Firm characteristics ln E (ln E)2 Time dummies Age and tenure Ind & LMR dummies Adjusted R2 N(obs)

Yes No No 0.054 966,900

Yes Yes No 0.580 966,900


0.182*** [0.001] 0.953*** [0.001] 0.035*** [0.001]
0.001*** [0.000] Yes No Yes 0.170 966,900

0.028*** [0.001]
0.002*** [0.000] Yes Yes Yes 0.604 966,900

Significant at: *** 1%; ** 5%; * 10%. Robust standard errors in brackets. Age and tenure controls for age (at end of job), tenure in job, age2, tenure2, and age* tenure. Firm employment measures at end of job spell

By introducing firm and worker controls sequentially, these regressions also provide estimated bounds for the extent to which average starting wages are affected by selective hiring of highly paid workers into foreign firms—the gap labelled S in Fig. 1.1. Comparing estimated wage premia between columns 1 and 2, and between columns 3 and 4, respectively, in Table 1.8 gives us an upper and lower bound for the role of observable and unobservable but time-invariant worker characteristics in explaining the overall foreign wage premium. The gap between the upper and lower bounds reflects the fact that worker and firm characteristics are themselves interrelated—not only are foreign-owned firms more likely to hire a particular type of workers, but across both foreign and domestic firms there are systematic differences in worker composition according to industry, firm size, and location. Without controls for firm characteristics (that is, attributing the full effect of both the worker characteristics themselves and the interdependent worker-firm characteristics to the workers), the upper estimate of the role of worker characteristics is Su = 0.142
0.085 = 0.057, a 5.7% wage gap attributable to worker heterogeneity. If we instead attribute the impact of interdependent characteristics to the firm—the

24

D. C. Maré et al.

comparison between columns 3 and 4—the remaining worker-specific component of the foreign premium is Sl = 0.043
0.027 = 0.016, a 1.6% gap attributable to worker quality. This lower bound gives a conservative estimate of the skill gap S, as it represents the pure impact of worker effects beyond those which are correlated with the composition of foreign firms. Worker and firm characteristics themselves show the expected relationship with wage levels—wages are higher for males and in larger firms, while the coefficient on worker fixed effects is close to one reflecting the construction of the variable. The foreign premium is somewhat higher for ending wages than starting wages (Table 1.9), suggesting a role for heterogeneous learning in which the return to working in a foreign-owned firm increases over time as workers gain skills.30 Worker and firm characteristics play a similarly important role in explaining start and end premia, with worker characteristics alone explaining around 45% of the raw end wage premium, reducing the coefficient on “always foreign” from 0.170 to 0.094, and worker and firm characteristics combining to explain around 80% of the raw gap ((0.170
0.035)/0.170). Having established that worker characteristics, firm composition, and unobserved factors associated with foreign ownership all contribute to the wage gap between foreign and domestic firms, we now turn to the task of distinguishing the relative roles of our other two hypotheses—firm heterogeneity and learning heterogeneity. In order to do this, we focus on worker transitions between foreign- and domesticallyowned firms. Table 1.10 reports the prevalence of each of the possible transitions between foreign and domestic ownership, showing considerable movement of workers between the two. Despite substantial movement across firm types, there is also a clear tendency for workers to transition between firms of the same type. While around 10% of all transitions involve a move to a foreign-owned firm, nearly 30% of job spells in foreign-owned firms end with a move to another foreign-owned firm (Table 1.10, column 1). Similarly, 80% of transitions from domestically-owned firms are into other domestically-owned firms, which make up 73% of all transitions (column 3). To capture the foreign-firm specific element of the earnings premium—the gaps (F1-D1) and (F2-D2) in Fig. 1.1—we estimate: m mixed u unknown Δ ln W it pc ¼ α þ βap δalways previous þ βp δprevious þ β p δprevious m mixed u unknown þβac δalways current þ βc δcurrent þ β c δcurrent þγX it þ ϕZpc þ ψ t þ εit pc

ð1:2Þ

where p refers to the previous job spell and c refers to the individual’s current spell with a different employer, Δ ln Wit pc is the log difference in earnings between the end of job spell p and the beginning of job spell c. Xit includes gender and worker fixed effects, as well as a polynomial function capturing age, tenure in the previous

30

Wage growth premia are examined directly in Table 1.12.

1 Earnings and Employment in Foreign-Owned Firms

25

Table 1.10 Prevalence of transitions between firms by ownership type, balanced panel Current employer Always Sometimes Never Unknown Total

Next employer Always Sometimes 0.289 0.228 0.191 0.247 0.065 0.095 0.040 0.087 0.102 0.128

Never 0.469 0.547 0.801 0.759 0.733

Unknown 0.015 0.015 0.040 0.118 0.038

Total 96,100 127,000 699,000 44,800 966,900

job, and elapsed time between the end of the previous job and start of the current job. The inclusion of worker fixed effects in the level regression (Eq. (1.1)) controlled for unobserved heterogeneity in wage levels. In Eq. (1.2), unobserved level differences are absorbed by differencing, and the inclusion of worker fixed effects allows for unobserved wage growth heterogeneity. The matrix Zpc includes a full set of industry and labour market dummies for both the previous and current employing firms, two additional dummies capturing whether the job transition involved a change in industry or a change in LMR, and a flexible polynomial function of firm-level employment in both the previous and current employers.31 As we include dummy variables for the ownership status of both the previous and current employer, coefficients on current firm ownership reflect the relative wage gain associated with moving to a firm of each type relative to moving to a domestic firm, while controlling for the ownership of the previous employer (i.e., βac ¼ F1‐D1). Meanwhile, the coefficients on previous firm ownership reflect the relative wage change associated with leaving each type of firm, again compared with leaving a domestically-owned firm (i.e., βap ¼ F2‐D2). These provide our estimates of the heterogeneous firm effect. Columns 2 and 4 of Table 1.11 can be thought of as two different estimates of the heterogeneous firm effect. Both control for worker characteristics (including tenure in the previous job and gap prior to joining the current employer), but while column 2 allows differences associated with the industry, region, and firm size composition of the foreign-owned firm population to be included in the overall foreign premia, in column 4 these effects are directly controlled for, leaving only the unobserved foreign premia. When firm composition is not controlled for, entering a foreignowned firm is associated with a 4.2% point greater wage increase than moving between domestic firms (Table 1.11, column 2, row 4). This premium reflects both the difference in observable firm characteristics between foreign- and domesticallyowned firms, with moves to foreign firms also more likely to entail a move to a larger firm (see Table 1.5), and the premium associated with foreign ownership itself. Controlling for firm characteristics as well (column 4), a two percent entry premium remains, reflecting the part of the foreign wage premium that cannot be explained by Employment controls include: : ln E1, ln E2, (ln E1)2, (ln E2)2, ln E1 * ln E2, where E1 is employee count at the previous firm in the month in which we measure workers’ end-of-job earnings, and E2 is employee count at the current firm, in the month in which we measure starting earnings in that job.

31

26

D. C. Maré et al.

Table 1.11 Entry and exit premia (1) (2) Raw +Worker chars Ownership of previous employer (exit penalty) Always
0.060***
0.053*** [0.001] [0.001] Sometimes
0.043***
0.039*** [0.001] [0.001] Unknown 0.012*** 0.003 [0.002] [0.002] Ownership of current employer (entry premium) Always 0.041*** 0.042*** [0.001] [0.001] Sometimes 0.028*** 0.028*** [0.001] [0.001] Unknown
0.012***
0.016*** [0.002] [0.002] Worker characteristics Female
0.011*** [0.001] WFE
0.003* [0.002] Time dummies Yes Yes Age and tenure No Yes Firm size No No Ind & LMR dummies No No Adjusted R2 0.005 0.033 N(obs) 966,900 966,900

(3) +Firm chars

(4) All


0.035*** [0.001]
0.028*** [0.001]
0.005** [0.002]


0.031*** [0.001]
0.025*** [0.001]
0.006*** [0.002]

0.020*** [0.001] 0.017*** [0.001]
0.006** [0.002]

0.020*** [0.001] 0.017*** [0.001]
0.005* [0.002]

Yes No Yes Yes 0.041 966,900


0.012*** [0.001]
0.005** [0.002] Yes Yes Yes Yes 0.054 966,900

Significant at: *** 1%; ** 5%; * 10%. Robust standard errors in brackets. Dependent variable is the change in log earnings associated with a job change. Age and tenure controls for age (at end of previous job), gap prior to commencing current job, age2, gap2, and age* gap. Firm controls include industry and LMR dummies for both the previous and current jobs, and a dummy for whether the worker has moved within the same industry or region

differences in firm composition or worker characteristics or by relative wage growth within-jobs (as the measure considers only starting wages at the new firm). As such, it is indicative of firm-specific effects such as rent-sharing, compensating differentials, and efficiency wages. As discussed in Sect. 1.3, the firm heterogeneity hypothesis implies not only that workers will gain from moving into foreign firms, but also that some or all of this wage gain will be reversed when they move back to a domestic firm. This effect is shown in row 1 of Table 1.11, which reports the differential wage change associated with leaving a foreign-owned firm relative to a domestic firm. As expected there is a penalty to leaving a foreign firm, with wage growth among exiters being 3.1 percentage points lower than that for workers leaving domestic firms, controlling for both worker and firm characteristics (column 4). This exit penalty more than fully

1 Earnings and Employment in Foreign-Owned Firms

27

reverses the two percent entry premium experienced when a worker first moves into a foreign-owned firm. Finally, to identify heterogeneous learning effects (H), we estimate a model of within-job wage growth: Δ ln W i jt ¼ α þ βa δalways þ βm δmixed þ βu δunknown j j j þγXit þ ϕZjt þ ψ t þ εi jt

ð1:3Þ

where Δ ln Wi jt is the log difference between monthly starting earnings and ending earnings within a job spell. Xi includes gender, estimated worker fixed effects, and controls for age and tenure within the job, and Zj includes a full set of industry and labour market dummies for the current employer and a polynomial expression of firm size at the start and end of the job spell.32 In this case, coefficients on foreign ownership represent the wage growth premium over the period employed by a foreign-owned firm, after controlling for tenure and other observable characteristics of the firm and worker (including an estimate of unobserved skill levels). Table 1.12 follows the same pattern as Table 1.11, sequentially adding worker and firm characteristics. On average, workers in foreign-owned firms experience an extra 2.8% growth in wages, relative to workers in domestically-owned firms (Table 1.12, column 1). This reflects in part longer average tenures in foreign firms (Table 1.5), as well as differences in worker and firm characteristics. Controlling for tenure and worker characteristics reduces the estimated foreign premium by around two-thirds, to 1.1%. Additionally controlling for firm composition, workers in foreign firms exhibit on average 0.4% higher wage growth over the course of their employment than that experienced in domestic firms (Table 1.12, column 4)—the gap labelled H in Fig. 1.1. With an average completed tenure of 20 months (Table 1.5), this implies approximately an additional 0.25% per year wage growth. This provides some support to the hypothesis that foreign firms provide learning opportunities beyond those available in domestic firms, although these do not appear to be particularly large. Unlike the between-firm transitions considered above, the within-job growth premium associated with ownership appears stronger among firms which are classed as “sometimes” foreign owned, rather than those that are always foreign owned. This may reflect transitions in ownership within a job spell, with part of the within-job earnings growth in sometimes foreign-owned firms associated with the transition from domestic to foreign ownership.33

Consistent with the cross-job regressions, firm size controls include ln E1, ln E2, (ln E1)2, (ln E2)2, and ln E1 * ln E2, where E1 is now employment count at firm j at the start of the employee’s job spell and E2 is employment in the same firm at the end of that spell. 33 Looking at firms classed as “sometimes” foreign owned across the full sample period, transitions from domestic to foreign ownership are more common than from foreign to domestic ownership. Of the 1974 firms for which we can observe ownership status in 2000/2001, and also in 2010/2011, 17.5% move from domestic to foreign ownership, while only 7.9% transition in the opposite direction. 32

28

D. C. Maré et al.

Table 1.12 Within-job wage growth

Foreign ownership Always Sometimes Unknown

(1) Raw

(2) +Worker chars

(3) +Firm chars

(4) All

0.028*** [0.001] 0.027*** [0.001]
0.008*** [0.001]

0.011*** [0.001] 0.015*** [0.001] 0.000 [0.001]

0.013*** [0.001] 0.016*** [0.001]
0.004** [0.001]

0.004*** [0.001] 0.008*** [0.001] 0.003 [0.001]

Yes No No No 0.023 966,900

0.010*** [0.001] 0.037*** [0.001] Yes Yes No No 0.069 966,900

Yes No Yes Yes 0.028 966,900

0.005*** [0.001] 0.034*** [0.001] Yes Yes Yes Yes 0.072 966,900

Worker characteristics Female WFE Time dummies Age and tenure Firm size Industry & LMR dummies Adjusted R2 N(obs)

Significant at: *** 1%; ** 5%; * 10%. Robust standard errors in brackets. Age and tenure controls for age (at start of job), tenure within job, age2, tenure2, and age*tenure. Firm size controls for employment at start and end of job, as well as squared and interacted terms. Firm controls include industry and LMR dummies for the current job only

1.5.1

Analysis of Selective Exit Behaviour

Together, these estimates paint a picture in which both worker and firm characteristics are important explanators of the overall difference between foreign and domestic firm earnings, but are not sufficient to fully explain the foreign wage premium. After controlling for both firm and worker characteristics, workers who enter foreign-owned firms gain around an extra two percentage point increase in earnings relative to workers who transition between domestic firms. In addition, workers within foreign-owned firms experience slightly stronger within-job wage growth (0.4% points above similar workers in similar domestic firms). However, the additional wage growth associated with foreign-firm employment is not retained when workers leave the foreign-owned firm, with the earnings penalty to returning to a domestic firm more than balancing out the combined wage growth associated with the entry at the start of the employee’s job spell j and E2 is employment in the same firm at the end of that spell and within-job growth premia (2.0% entry premium + 0.4% within-job premium
(3.1%) exit penalty ¼
0.7% retained premium). This contrasts with findings of positive returns to foreign-firm experience from MalchowMøller et al. (2013), Pesola (2011) and Balsvik (2011).

1 Earnings and Employment in Foreign-Owned Firms

29

One potential explanation of this finding is that domestic firms are not willing to pay a premium for experience gained within foreign firms because some types of knowledge and skills rewarded by foreign firms are not as highly valued by domestic businesses. This might be the case if jobs in foreign firms are more specialised than those in New Zealand firms (perhaps because local subsidiaries have a more narrowly defined role in the larger organisation) or because the skills that are learned are specific to the firm (e.g., developing relationships with offshore owners or customers). While we control for firm size in the New Zealand operation, foreignowned firms may have access to a broader international organisation, allowing workers within the New Zealand operation to specialise in particular tasks. Alternatively, there may be unobserved selection effects involved in the exit penalty. For example, if working conditions and expectations are stricter in foreign-owned firms, returning to a domestic firm may reflect a lifestyle choice on the part of the individual worker, with lower wage growth (or even an absolute wage decline) accepted as a tradeoff for better work-life balance. More generally, job transitions reflect endogenous choices on the part of firms and workers. As most job changes are voluntary, we observe a transition only if it is beneficial in some way for the worker concerned.34 If a more generous initial wage offer (relative to both the previous job and other alternative employers) and/or stronger wage growth are factors in workers’ employment and job-search decisions, both the wage change associated with job transitions and the wage growth withinjobs will be stronger than we would observe if workers were randomly allocated across employers. Conversely, once workers find a job that suits them, their incentive to remain in that job will also be influenced by both wage and non-wage conditions. If a more generous entry wage or stronger within-job wage growth are associated with higher incentives to remain in a job, then observed entry wage growth and within-job growth will be lower across both domestic and foreign firms, as workers with a particularly good job match will tend to remain in that job and thus will not be captured in the regression population. As an indication of the extent to which selective exit patterns matter, Table 1.13 reports mean residuals from regressions of entry wage growth and starting wage levels on the standard explanatory variables. Residuals are reported separately according to both the ownership of the current employer and workers’ subsequent destination, where the latter covers whether or not the worker remains with the firm for an extended period (more than 5 years) and, for those who leave, whether their next employer is foreign or domestically owned. That is, it provides a test of whether workers who remain with the firm long-term differ from those who leave relatively quickly and, amongst those who do leave, whether those who move to foreign firms differ from those who move to domestic firms, in terms of their unexplained entry wage growth.

34 Relevant factors may include earnings potential and employment conditions, but also lifestyle decisions such as location and job content.

30

D. C. Maré et al.

Table 1.13 Censoring and job transitions

Currently domestic Mean residual from start wage regression (Eq. (1.1)) Mean residual from entry wage growth regression (Eq. (1.2)) N(obs) Currently foreign Mean residual from start wage regression (Eq. (1.1)) Mean residual from entry wage growth regression (Eq. (1.2)) N(obs)

(1) (2) Sample: first 5 years

(3)

(4) (5) (6) Sample: balanced panel Next: Next: domestic foreign Difference

Censored

Complete

Difference


0.010 [0.002]

0.000 [0.000]

0.011***

0.000 [0.000]

0.006 [0.002]

0.006***


0.005 [0.003]

0.000 [0.001]

0.005*

0.001 [0.001]


0.004 [0.002]


0.005

12,000

324,200

302,400

24,100


0.010 [0.005]

0.001 [0.001]

0.010**


0.010 [0.002]

0.018 [0.002]

0.028***


0.005 [0.007]

0.000 [0.002]

0.005


0.007 [0.002]

0.013 [0.003]

0.020***

2100

41,000

23,600

14,000

Significant at: *** 1%; ** 5%; * 10%. Robust standard errors in brackets. Columns 1–3 use all clean transitions between two full-time jobs where the new job start occurs after May 2000 and before December 2005. Job spells are defined as “censored” if they extend longer than 5 years (i.e., over half the length of the primary analysis period). Columns 4–6 use all balanced panel spells for which there is sufficient information on past and future job spells, dividing spells according to whether the worker’s next job is in a foreign or domestic firm. Residuals are estimated based on earnings regressions at the start of the job spell. Transitions into and out of firms which are sometimes foreign owned and those with unknown ownership are not shown

Columns 1 and 2 report the comparison between workers who remain in their jobs for more than 5 years and those who move on. For both wage growth and wage levels, workers who stay long-term (those with censored spells) actually tend to be those with slightly lower than expected earnings on entry, though the difference for the wage growth regression is not significant among foreign-firm employees. The gap in mean residuals between censored and completed spells does not differ for workers in domestic firms and those in foreign-owned firms, suggesting that while censoring may be affecting the average wage level associated with job transitions, there is no evidence that it affects the gap in either wage levels or entry wage growth between foreign and domestic employers. A second possibility is that workers who move between foreign and domestic firms may be a non-random selection of workers, even controlling for observed differences. Transition patterns may be correlated either with unobserved worker characteristics or with patterns of transitory wage change, leading to biased estimates of the entry premium or exit penalty. The selection bias can take two forms. First, transition patterns may be correlated with unobserved differences between workers

1 Earnings and Employment in Foreign-Owned Firms

31

in the lifetime average level or growth rate of earning capacity. Second, transition patterns may be correlated with transitory earnings fluctuations. In this case subsequent earnings changes will reflect mean reversion, as the transitory fluctuations are reversed, in addition to the true impact of moving between firms under different ownership. The impact of mean reversion on foreign entry and exit premia will depend on the prevalence of transitory fluctuations and the strength of mean reversion for different types of transition. Time invariant unobserved characteristics associated with a worker’s average wage level are controlled for in the start and end wage regressions (Tables 1.8 and 1.9) by the inclusion of estimated worker fixed effects, and in Tables 1.11 and 1.12 by first differencing. Worker-level differences in wage growth are controlled for in Tables 1.11 and 1.12 only to the extent that worker-specific growth is correlated with unobserved worker-specific wage level components (WFE). If high latent-growth workers disproportionately enter and remain within foreign firms, the estimated impact of entry into FDI firms will be overestimated, and the earnings change associated with entering a domestic firm will be underestimated. In this case, the net exit penalty will be larger than estimated. Columns 4–6 of Table 1.13 examine whether transitions are random, conditional on observed characteristics. Specifically, they show whether the foreign ownership of a worker’s next employer is correlated with residual wage levels or wage growths on starting a job, as estimated using Eqs. (1.1) and (1.2). By construction, these residuals have zero mean conditional on the ownership type of current employer. They may, however, be non-zero by next employer if transitions are non-random. For workers leaving domestic firms, there is only a small difference in residuals for those moving to foreign firms compared with those moving to domestic firms. Residual start wage levels are only 0.6% higher, and residual entry gains are insignificantly lower (column 6). There is much stronger selection among workers leaving foreign firms. Those moving to other foreign firms have 2.8% higher residual wage levels, and 2.0% higher residual entry gains than those moving to domestic firms. The estimated entry gains and exit penalties associated with jobs in foreignowned firms may be affected by these selection patterns. Table 1.14 shows, however, that these biases are relatively small. Controlling for residual start wage or prior residual entry gains has only a small impact on the estimated exit penalty, which declines from
0.031 to
0.029 or the estimated entry premium, which rises from 0.020 to 0.023. There is, however, strong evidence of mean reversion, consistent with transitory fluctuations in earnings. Across all types of transitions, a 1% higher residual starting wage is associated with a 0.31% lower entry wage change at the next job transition (column 3). Similarly, a one percent higher residual wage gain when starting a new firm is partially reversed by a 0.14% lower wage gain at the next job-to-job transition (column 2). Finally, Table 1.15 investigates whether the persistence of wage gains varies according to the ownership of the firm. Workers are divided into quartiles based on the residual wage gains made when they entered their previous job. The relationship between entry gains and within-job growth does not differ between domestic and foreign firms, as shown by the small and mostly insignificant estimates in column 1.

32

D. C. Maré et al.

Table 1.14 Mean reversion in wage growth at job transition (1) Ownership of previous employer (exit penalty) Always
0.031*** [0.001] Sometimes
0.025*** [0.001] Unknown
0.006*** [0.002] Ownership of current employer (entry premium) Always 0.020*** [0.001] Sometimes 0.017*** [0.001] Unknown
0.005* [0.002] Wage residuals at start of previous job spell Residual Δ ln wstart

(2)

(3)


0.029*** [0.002]
0.020*** [0.002]
0.001 [0.002]


0.029*** [0.001]
0.024*** [0.001]
0.007*** [0.002]

0.020*** [0.002] 0.015*** [0.002]
0.011*** [0.003]

0.023*** [0.001] 0.017*** [0.001]
0.006** [0.002]


0.143*** [0.002]

Residual ln wstart Adjusted R2 N(obs)

0.054 966,900

0.074 513,800


0.313*** [0.002] 0.107 966,900

Significant at: *** 1%; ** 5%; * 10%. Robust standard errors in brackets. All regressions control for time, age and tenure, firm size, industry and labour market region. Residuals are estimated from earnings equations (Eqs. (1.1) and (1.3)) at the start of the previous job (i.e., the start of the job which ends with a transition between two employers over which the entry and exit premia are estimated). Column 1 repeats column 4 of Table 1.11 for reference purposes. Reduced sample in column 2 reflects the need for additional information about the previous job spell in order to estimate residuals of Δ ln wstart. A regression of the full sample using Δ ln wstart rather than the residuals produces very similar results to column 2, suggesting that this sample reduction is not strongly affecting the results

Unexpected wage gains on hiring are more persistent when the worker’s next transition is to a foreign firm. Workers who make residual wage gains at the start of a job—the unexplained gains remaining after controlling for observed characteristics—retain more of their gains if they leave a foreign firm than they would if they were to leave a domestic firm. Workers making the largest (residual) entry gains (Q4 in Table 1.15) lose only 2.0% upon leaving a foreign firm, compared with 3.3% for workers with the smallest entry gains (Q1 in Table 1.15), and 3.1% overall. Those who had previously made large (residual) entry gains also benefit more from starting their next job in a foreign firm, experiencing a 3.1% entry premium compared with 2.0% overall. In contrast, workers who had previously made the smallest (residual) entry gains receive only a fraction of the foreign premium (0.9%).

1 Earnings and Employment in Foreign-Owned Firms

33

Table 1.15 Foreign ownership premia by quartiles of residuals from entry earnings growth regression (1) Within-job Quartiles of residuals from Δ ln wstart regression Q1 (low wage growth on entry)
0.001 [0.003] Q2 0.008** [0.003] Q3
0.002 [0.003] Q4 (high-wage growth on entry)
0.001 [0.004] Overall premia 0.004*** [0.001]

(2) Exit

(3) Entry

(4) N(obs)


0.033*** [0.004]
0.035*** [0.004]
0.030*** [0.004]
0.020*** [0.004]
0.031*** [0.001]

0.009* [0.004] 0.024*** [0.004] 0.019*** [0.003] 0.031*** [0.004] 0.020*** [0.001]

128,500 128,500 128,500 128,500 966,900

Significant at: *** 1%; ** 5%; * 10%. Robust standard errors in brackets. Entry wage growth residuals (the extent to which the observed wage change for a given job spell is high/low compared to predicted wage growth given worker and firm characteristics) are estimated at the start of a job following Eq. (1.2). These are used to define four groups, based on whether transition into that job involved an unexpectedly high (Q4) or unexpectedly low (Q1) wage change. Then, within-job wage growth premia are estimated for the job spell (Eq. (1.3)), and entry and exit premia are estimated using data from the end-of-job transition (exit from the current job, entry into the next job, Eq. (1.2)), with separate regressions for the four groups. Overall premia are taken from Tables 1.11 and 1.12 for comparison, and include job spells with insufficient information on the previous spell to enable estimation of residuals

The net exit “penalty” for high residual gain (Q4) workers is actually a net exit gain of 1.1%. For low residual gain (Q1) workers, the penalty is
2.4%. That is, workers with unexpectedly high-wage growth appear, subsequently, to be more highly valued by foreign firms. It may be that foreign firms put greater weight on workers’ previous earning histories as a guide to worker productivity when hiring workers or setting wage levels. At least for workers with high residual wage growth, there is a longer term advantage to being employed in a foreign firm, even if on average workers leaving foreign firms experience a slightly larger drop in earnings than the gain made by workers joining foreign firms.

1.5.2

Heterogeneity of Foreign Premia

Alongside the average effects shown above, policymakers may also be interested in variation in the foreign premia across different groups of firms and workers. For example, if the composition of contemporary FDI differs from that experienced in the past (e.g., differences in industry focus), estimates based on employees of longstanding foreign-owned firms may not reflect the expected impacts of future FDI flows. Similarly, the relative wage and learning impacts associated with different types of firms and industries may be of interest. Political support for FDI may be

34

D. C. Maré et al.

conditional on the distribution of benefits, rather than just the average size. In particular, if gains from FDI employment are felt primarily by high-wage, highskill workers, the decision whether to support further FDI may turn on the role foreign firms play in reducing the emigration of highly skilled New Zealanders. Alternatively, if gains are felt primarily by low skill workers, equity considerations play a more obvious role in the debate. To examine the heterogeneity of potential FDI premia, Tables 1.16 and 1.17 report the foreign ownership premia associated with always foreign-owned firms from separate regressions for a range of different firm and worker groups.35 Table 1.16 reports coefficients on being employed by an always foreign-owned firm for different groups of workers, distinguished according to gender, age groups, and quartiles of the worker fixed effect distribution. Foreign premia differ by age group (panel A), with younger workers on average experiencing both a greater wage boost on entry (column 3) and a lower penalty on leaving a foreign firm (column 5). While the estimated within-job wage growth premium is increasing in age, it is significant only for “prime-age” workers (25–49 years) (column 4).36 Both entry premia and exit penalties increase with worker skill, suggesting that not only do foreign firms hire “better” workers, they also pay more to get those workers (panel B). As these workers are more highly paid to begin with, higher premia for the highly skilled are likely to exacerbate income differentials between groups. Finally, while women experience a slightly stronger wage gain on entering a foreign firm, their wage growth within those firms is low, even compared with jobs in similar domestic firms, with the positive growth premium driven by males. Overall, this yields a retained premium (column 6) of almost zero for men and
1% for women. Table 1.17 repeats the analysis above for different types of firms, distinguishing by firm size, location and a series of industry characteristics. At the firm level, we consider only start and end wage levels and within-job wage growth premia, as transitions between jobs often also involve transitions between industries or firm sizes, such that wage changes will be affected by both the characteristics of the previous firm and the new firm. Wage level premia are substantially stronger for small firms (panel A), with workers in small foreign firms receiving an average starting wage 19% higher than those in similar domestic firms. For medium-sized firms (5–49 employees), the foreign premium falls to 8.7%, and falls further to 1.5% among larger firms. Wage growth premia are concentrated among medium-sized firms (column 3). Many of the benefits that foreign firms provide (such as improved management capability, access 35

In any given subgroup there are a range of other factors involved, including both endogeneity in employment paths and potential heterogeneity across other dimensions of worker and firm characteristics. However, allowing for the full range of observable heterogeneity in premia would require a fully interacted model with impacts of FDI allowed to vary by age, by tenure, by location, etc., an approach which quickly becomes unmanageable. As such the reported results should again be treated as an observed average, which is of interest when thinking about the distributional impacts of FDI even though it does not imply a deterministic relationship between any of the binary categories considered below and the strength of the FDI premium. 36 This may in part reflect sample size differences—“prime-age” workers account for around two-thirds of the total population.

1 Earnings and Employment in Foreign-Owned Firms

35

Table 1.16 Heterogeneity by worker characteristics (1) Start level

(2) End level

(3) ΔW at entry

Worker group Panel A: age groups 24 0.022*** 0.022*** 0.025*** years [0.002] [0.002] [0.003] 25–49 0.028*** 0.036*** 0.020*** years [0.001] [0.001] [0.002] 50+ 0.016*** 0.032*** 0.015*** [0.003] [0.004] [0.004] Panel B: quartiles of worker fixed effects Q1 0.006** 0.016*** 0.003 (low) [0.002] [0.002] [0.002] Q2 0.021*** 0.027*** 0.014*** [0.002] [0.002] [0.003] Q3 0.031*** 0.037*** 0.023*** [0.002] [0.002] [0.003] Q4 0.046*** 0.052*** 0.042*** (high) [0.002] [0.003] [0.003] Panel C: gender Female 0.027*** 0.027*** 0.024*** [0.002] [0.002] [0.002] Male 0.025*** 0.037*** 0.019*** [0.001] [0.002] [0.002] Panel D: overall All 0.027*** 0.035*** 0.020*** workers [0.001] [0.001] [0.001]

(4) ΔW within job

(5) ΔW at exit

(6) Retained premium (3)+(4)+(5)

(7) N(obs)


0.001 [0.003] 0.005*** [0.001] 0.006 [0.004]


0.014*** [0.003]
0.032*** [0.002]
0.044*** [0.005]

0.010

194,800


0.007

649,900

-0.023

122,300

0.005* [0.002] 0.004 [0.002] 0.004 [0.002] 0.001 [0.003]


0.014*** [0.003]
0.029*** [0.003]
0.034*** [0.003]
0.045*** [0.003]


0.006

241,700


0.011

241,700


0.007

241,700


0.002

241,700


0.006** [0.002] 0.010*** [0.002]


0.028*** [0.002]
0.032*** [0.002]


0.010

327,500


0.003

639,500

0.004*** [0.001]


0.031*** [0.001]


0.007

966,900

Significant at: *** 1%; ** 5%; * 10%. Robust standard errors in brackets. Each reported coefficient is from a separate regression, estimated for the specified sub-sample of workers. All regressions include controls for time, industry, labour market region, age, and tenure. Columns 1 and 2 follow column 4 of Tables 1.8 and 1.9, respectively, columns 3 and 5 follow Table 1.11 (column 4), and column 4 follows Table 1.12 (column 4)

to financial capital, virtual scale and opportunities for specialisation) are potentially much harder to realise for small firms, which are unlikely to have the same access to internal and external resources as larger firms. The foreign wage level premium also appears to be an urban phenomenon, strongest in Auckland and very weak in rural areas (panel B). The knowledge, technologies, or inputs that foreign-owned firms bring may be complementary with the more skilled urban workforce or greater scale of urban activity. The higher rate of interaction within cities may also magnify the advantages that foreign-owned firms bring.

36

D. C. Maré et al.

Table 1.17 Heterogeneity by firm characteristics (1) Start level Firm group Panel A: firm size (number of employees) 5% and < 10%, > 10% and < 25%, and > 25% and < 75% Ln mean duration Log of no. of years working per firm Ln firm duration Log of years working in current firm Selectivity-related variables on location, industry, and occupation Regions (LM-REGION) Categorical variable encompassing 141 labour market regions in Germany Occupation (OCC) Categorical variable encompassing 50 occupations according to the occupational classification system KldB 2010 (related to ISCO-08) TASK level (TASK) Categorical variable representing three different task levels of the job. It consists of three groups: Auxiliary activity (unskilled task), trained/professional task (clerks), and specialist/expert task Industry (IND) Categorical variable encompassing 96 distinct industries at the 2-digit level according to the German classification Scheme WZ 2008 (NACE Rev. 2.) Supervisor Dummy variable indicating whether an employee is a supervisor Executive Dummy variable indicating whether an employee is an executive a

A correction procedure was applied for both variables (Fitzenberger et al. (2005)) The system of vocational education and training (VET) is quite unique in the international context. The training takes place in private firms and is combined with education in public schools. This training usually lasts 3 years

b

62

S. Brunow and O. Jost

Table 3.2 Firm characteristics Variable Description Key FIRM variables (FIRM) Firm size Categorical variable representing the number of employees and consisting of four groups: 1–9 employees, 10–49 employees, 50–249 employees, and above 250 employees Females Proportion of females employed in the firm Youth Proportion of employees under 35 years of age Human capital Two variables capturing the human capital intensity of the firm: First, the intensity proportion of professional assistants employed, and second, the proportion of specialists/experts employed, each as a share of total employment in the firm Firm age Categorical variable representing the firm age in years, consisting of the following groups: Under 5 years old, 5–10 years old, 10–25, and over 25 years old Characteristics for robustness checks Card–Heining–Kline Firm and individual specific effects defined by Card, Heining, and Kline effects (2013) that capture all unobserved firm and individual characteristics Proportion of The proportion of foreigners employed in the firm aims to control for foreigners firms having previous experience of employing foreigners. In addition, it controls for segregated ethnic communities that exhibit lower productivity levels on average Table 3.3 Distribution of daily gross median wages for full-time employees (in €) Foreigners Germans Wage € 120.55 No. obs. 1,514,150 Separated by tasks Auxiliary 80.30 workers Clerks 109.56 Expertsa 168.33 a

All 90.64 142,305

EU-15 105.54 44,815

EU-8 71.85 20,693

EU2 + Balkan 82.97 26,475

Turkey 99.18 41,321

Remaining World 94.10 43,628

69.15

77.05

58.81

64.31

85.09

69.04

89.24 150.08

98.18 167.79

74.66 141.81

84.83 146.89

102.36 134.34

88.85 152.08

The group of experts includes both, experts and specialists

Table 3.4 shows that the previously observed wage differences are notably reduced, especially for foreigners who have attained their VET degrees in Germany. In particular, foreigners from the EU-8, EU-2, and Balkan countries benefit considerably from VET in Germany, as their wages increase substantially. This shows that VET acquired in Germany can explain a large part of the wage difference and shows the signalling character of certificates, although selectivity in various occupations is still uncontrolled. Another decisive influence on the reported median wages is the distribution of employees within the task (Table 3.5) and educational levels (Table 3.6).

3 Foreign Vs. German Wage Differentials in Germany: Does the Home Country Matter?

63

Table 3.4 Wages of workers holding a VET certificate EUGermans Migrants EU-15 EU-8 2 + Balkan Clerks 118.05 108.23 111.18 101.59 110.22 Did not participate in VET during life-course within Germany Auxiliary 74.90 67.73 75.57 58.62 59.81 workers Clerks 99.94 83.15 92.81 72.45 71.70 Experts 144.06 117.92 142.99 99.18 102.27 Participated in VET during life-course within Germany Auxiliary 87.47 89.59 88.70 78.04 80.8 workers Clerks 113.47 104.46 106.62 96.48 91.35 Experts 156.19 132.34 142.71 116.5 120.52

Turkey 113.85

Remaining World 102.69

82.59

68.54

95.25 110.63

83.50 114.74

96.53

84.60

112.13 135.26

102.69 137.72

Notes: Median gross daily wages in €, including all individuals holding a vocational training certificate, excluding university graduates Table 3.5 Distribution of migrants within the task levels in percent

Germans EU-15 EU-8 EU-2 + Balkan Turkey Remaining world

Task level Auxiliary workers 11.68 19.34 36.21 32.05 28.87 25.81

Table 3.6 Distribution of educational levels in percent Germans EU-15 EU-8 EU-2 + Balkan Turkey Remaining world

Clerks 53.51 51.19 48.36 47.18 60.62 44.78

Specialists/Experts 34.81 29.47 15.43 20.77 10.51 29.41

Educational attainment w/o degree VET 10.27 69.42 26.96 51.38 23.93 60.04 29.55 50.99 44.64 51.20 31.74 39.90

University 20.31 21.65 16.02 19.46 4.16 28.36

Note: All values in percent. w/o degree—without degree, VET— vocational education and training, university—university or university of applied science degree

According to Table 3.5, German employees not only have the smallest share of auxiliary workers among the groups considered, they also have the highest share of specialists and experts among the employees. This partly explains the highest median wage for German employees. At the same time, it is noticeable that the groups that tend to have lower median wages, such as EU-8 foreigners or EU-2 and

64

S. Brunow and O. Jost

Table 3.7 Wage differences by age Age in All EUyears Germans migrants EU-15 EU-8 2 + Balkan < 25 85.90 65.06 71.24 56.68 58.12 25–34 109.21 88.52 99.57 71.99 77.64 35–44 128.90 94.05 113.63 74.14 79.92 45–54 132.42 94.11 109.49 70.37 76.16 55+ 125.06 95.76 108.26 77.78 79.57 Change in earnings when entering the next age group in € 25–34 23.31 23.46 28.33 15.31 19.52 35–44 19.69 5.53 14.06 2.15 2.28 45–54 3.52 0.06 4.14 3.77 3.76 55+ 7.36 1.65 1.23 7.41 3.41

Turkey 72.18 92.09 102.53 106.00 102.42

Remaining world 65.17 101.92 98.12 88.64 93.93

19.91 10.44 3.47 3.58

36.75 3.8 9.48 5.29

Notes: Median gross daily wages in €

Balkan foreigners, supply about three times as many auxiliary workers as German employees. One interesting fact is that Turkish employees, who also have a high proportion of auxiliary workers and by far the smallest proportion of specialists and experts (10.5%), earn higher wages than the two groups mentioned above. However, this can be attributed to their longer labour market experience in Germany, as they have been in the German labour market longer than the EU groups. The high percentage of specialists and experts from the remaining world is also remarkable. This is an indication of the higher hurdles for foreigners from other countries, who are only allowed to migrate to Germany under certain restrictive conditions. These figures generally show a high correlation with the level of education, which is why the next table includes this information. Table 3.6 reveals pronounced differences in educational attainment. These differences correspond to the pattern already shown in Table 3.5. German employees have the lowest proportion of workers without VET, i.e., around 10%. This proportion rises to 25–30% for the EU groups and is highest among Turkish workers (around 45%). Considering the proportion of workers with VET the variance is considerably smaller: The VET proportions show that vocational training is an important part of the qualification of German workers, as around 70% have completed such training. This share is reduced to around 50 and 60% for the remaining foreigners and the EU-8 group, respectively. In terms of the proportion of workers with a university degree, the heterogeneity between the groups is even smaller. For example, the proportion of highly skilled workers is between 16 and around 22%, with the EU-15 foreigners most often holding a university degree. It is striking here that Turkish employees have a particularly low proportion of highly qualified workers, which indicates structural differences between the considered groups. In order to better address these structural differences, which are related to the immigration cohort and thus to the age of the considered workers, the age groups and their corresponding middle incomes are considered in Table 3.7. For all groups,

3 Foreign Vs. German Wage Differentials in Germany: Does the Home Country Matter?

65

Table 3.8 Distribution across age cohorts and gender All Germans foreigners Distribution across age categories < 25 5.5 5.4 25–34 24.0 25.6 35–44 21.5 31.3 45–54 31.2 25.9 55+ 17.9 11.8 Share of 30.7 26.6 females

EU15

EU8

EU2 + Balkan

Turkey

Remaining world

4.6 22.4 26.5 29.0 17.5 26.7

7.2 31.9 31.3 19.3 10.3 33.6

7.3 32.2 33.8 19.7 7.0 32.7

6.4 21.9 33.8 30.1 7.8 17.2

3.1 25.5 32.6 25.4 13.4 28.3

Note: All values in percent

somewhat similar patterns can be found; the older the individuals are, the higher is their respective wage. The second part of the table shows the change in wages when entering another age group. Here, young EU-15 foreigners and those from the remaining world enjoy even higher wage growth compared to the Germans. However, some also experience a negative change, and for other foreign groups the increase in salaries is very small, indicating flat wage profiles. Once again, however, this only shows a bivariate picture, and therefore an analytical approach has to be chosen to consider such limitation. A more descriptive overview of the distribution across age cohorts and gender is provided in Table 3.8. As can be seen, foreigners are on average younger, and, with the exception of the EU-8, EU-2, and Balkan groups, more frequently male when compared to Germans. The group of Turkish employees shows the lowest female share: a mere 17.2%.

3.6

Empirical Analysis and Estimation Strategy

The descriptions already show large wage differences between Germans and foreigners as well as among the migrant groups themselves, even after the differentiation into task levels, age, education, and vocational training within Germany. Additional factors can be responsible for these additional differences. For example, task levels do not take into account the employee’s place of work or their occupation. Thus, certain patterns can occur here due to, e.g., certain selections or segregations. Because we employ the wage decomposition based on the method of Oaxaca (1973) and Blinder (1973), we primarily rely on a Mincer earnings function for the wage setting on the labour market. In the analysis, we have reduced issues of multicollinearity to a large extent, so that the standard errors are validly identified for the individual influences. This is important when comparing the groups, especially the meaning and significance of single effects. Furthermore, we ensure that the model itself provides robust results that are not sensitive to the usual wage equation

66

S. Brunow and O. Jost

transformations, such as adding to, leaving out, or recoding variables. With the help of the preferred wage equation, we then perform the wage decomposition. For this, we rely on the threefold wage decomposition according to Jones and Kelley (1984), which converts the wage difference between the comparison groups into three parts on the basis of separate OLS estimates providing a set of estimates, i.e. coefficients, for each migrant group and Germans. The first part of the decomposition is the endowment part, which attributes a part of the pay gap to differences in the respective observable variables. The interpretation here is how the wages of migrants would change if we adjust their endowments (i.e. characteristics) to the level of Germans, evaluated at German coefficients. The second part explains the part of the pay gap that is due to differences in coefficients. We interpret this as the extent to which a given characteristic pays off. The interpretation of the resulting wage gap is: How would the wages of migrants turn out if their coefficients are adjusted to the levels of the German coefficients evaluated at the average of the German characteristics? The third part of the wage decomposition is an interaction effect consisting of the multiplication of the two previous effects. As this part offers little insight into the wage differences, we neglect the discussion on this. The choice of the reference categories for dichotomous variables is another important point which plays a decisive role in the wage decomposition, as the interpretation of effects becomes cumbersome and can change significantly. We therefore rely on normalised groups according to Yun (2005). Then the results can be interpreted as a deviation from the average of the particular variable. The problem of endogeneity, driven, for instance, by certain displacement mechanisms in the labour market—leads to biased estimates, which results in wrong endowment and coefficient effects. This can occur, for example, if foreigners selfselect or are forced into certain “worse-paid” occupational groups and/or tasks due to a lack of language proficiency, for example. As a result, foreigners would have higher coefficients than Germans in such “less desired” occupations because of their potential “downgrading”. To identify the relevance of such crowding, we estimate all models without any selectivity-related variables. If the remaining effects do not change much, then at least such selectivity and a potential influence of endogeneity will be of little impact.

3.7

Results

The overall results of our wage decomposition are presented in Table 3.9, and the detailed decomposition of the endowment effect in Table 3.10. For comparison purposes, column 1 reports the decomposition of all foreigners relative to Germans. Columns 2–6 show the results for different countries of origin. In the analysis, we always select the German employees as the reference group to better identify differences among the migrant groups (relative to Germans). That means that we can directly compare the change in wages (in Euro) whenever a change in a specific variable occurs for the respective migrant groups. In assessing the coefficient effect,

3 Foreign Vs. German Wage Differentials in Germany: Does the Home Country Matter?

67

Table 3.9 Oaxaca-Blinder decomposition for mean wages: general view All Foreigners EU-15 (1) (2) Average gross daily wages in Euro Foreigners 91.91 108.27 Germans 125.20 125.20 Difference 0.734*** 0.865*** Oaxaca–Blinder effect decomposition Endowments 0.740*** 0.856*** (0.003) (0.003) Coefficients 0.985*** 1.008*** (0.002) (0.003) Interaction 1.008*** 1.001 (0.002) (0.003) No. 142,305 40,315 foreigners No. 1,379,013 1,379,013 Germans No. firms 124,935 109,271

EU-8 (3)

EU2 + Balkan (4)

Turkey (5)

Remaining World (6)

74.59 125.20 0.596***

81.78 125.20 0.653***

90.58 125.20 0.723***

92.68 125.20 0.740***

0.586*** (0.004) 0.950*** (0.008) 1.070*** (0.009) 18,979

0.640*** (0.004) 0.961*** (0.005) 1.062*** (0.006) 18,431

0.747*** (0.005) 0.934*** (0.005) 1.037*** (0.005) 29,033

0.766*** (0.004) 0.976*** (0.004) 0.990** (0.004) 30,749

1,379,013

1,379,013

1,379,013

1,379,013

107,383

107,446

107,646

108,573

Notes: Only full-time workers with valid information on educational attainment and qualifications are considered; * 10%, ** 5%, *** 1%; cluster-robust s.e. at firm level in (). Variations of the average wages from Table 3.3 are due to the transfer of logarithmic wages to the exponential form

Table 3.10 Detailed decomposition of the endowment effects

RAM IND OCC TASK INDIVID EDUC EXP FIRM

All foreigners (1) 1.015*** (0.001) 0.967*** (0.002) 0.959*** (0.001) 0.960*** (0.001) 1.003*** (0.000) 0.969*** (0.001) 0.904*** (0.001) 0.933*** (0.002)

EU-15 (2) 1.023*** (0.001) 0.983*** (0.001) 0.975*** (0.001) 0.981*** (0.001) 1.008*** (0.001) 0.984*** (0.001) 0.947*** (0.001) 0.948*** (0.001)

EU-8 (3) 0.989*** (0.001) 0.940*** (0.003) 0.946*** (0.002) 0.940*** (0.001) 0.989*** (0.001) 0.972*** (0.001) 0.819*** (0.003) 0.901*** (0.003)

EU2 + Balkan (4) 1.024*** (0.001) 0.951*** (0.002) 0.954*** (0.002) 0.946*** (0.001) 0.988*** (0.001) 0.961*** (0.001) 0.848*** (0.003) 0.905*** (0.002)

Turkey (5) 1.019*** (0.001) 0.982*** (0.002) 0.930*** (0.002) 0.938*** (0.001) 1.020*** (0.001) 0.922*** (0.001) 0.956*** (0.002) 0.954*** (0.002)

Remaining world (6) 1.014*** (0.001) 0.962*** (0.002) 0.976*** (0.001) 0.971*** (0.001) 1.001 (0.001) 0.998 (0.001) 0.888*** (0.002) 0.934*** (0.002)

Note: Only full-time workers with valid information on educational attainment and qualifications are considered. * 10%, ** 5%, *** 1%, cluster-robust s.e. at firm level in (). Variations to the average wages from Table 3.3 are due to the transfer of logarithmic wages to the exponential form

68

S. Brunow and O. Jost

Table 3.11 Detailed decomposition of the coefficient effects

RAM IND OCC TASK INDIVID EDUC EXP FIRM Constant

All foreigners (1) 0.978*** (0.002) 0.994 (0.004) 1.009*** (0.003) 0.981*** (0.007) 0.995*** (0.001) 1.015*** (0.003) 0.930*** (0.003) 1.113*** (0.013) 0.979 (0.014)

EU-15 (2) 0.984*** (0.006) 0.986*** (0.004) 1.000 (0.004) 0.996 (0.010) 1.005*** (0.002) 1.010** (0.004) 0.923*** (0.004) 1.082*** (0.017) 1.030 (0.021)

EU-8 (3) 0.970*** (0.004) 1.014** (0.007) 1.006 (0.008) 0.968 (0.022) 0.987*** (0.002) 1.004 (0.008) 0.901*** (0.005) 1.072*** (0.019) 1.037 (0.033)

EU2 + Balkan (4) 0.961*** (0.005) 1.015* (0.008) 1.002 (0.007) 1.044** (0.022) 0.989*** (0.002) 1.022*** (0.006) 0.932*** (0.005) 1.065*** (0.019) 0.938** (0.028)

Turkey (5) 0.966*** (0.006) 0.994 (0.005) 1.011 (0.007) 0.949*** (0.018) 0.990*** (0.002) 1.010** (0.004) 0.952*** (0.006) 1.162*** (0.019) 0.916*** (0.026)

Remaining World (6) 0.984*** (0.004) 0.992 (0.006) 1.012*** (0.004) 0.981 (0.014) 1.001 (0.002) 1.026*** (0.005) 0.928*** (0.005) 1.127*** (0.019) 0.939*** (0.022)

Note: Only full-time workers with valid information on educational attainment and qualifications are considered. * 10%, ** 5%, *** 1%, cluster-robust s.e. at firm level in (). Variations to the average wages from Table 3.3 are due to the transfer of logarithmic wages to the exponential form

we again apply and adjust Germans to the respective levels of the foreigner groups (Table 3.11). Firstly, German employees earn about 14% higher wages than employees from the EU-15 countries, which is the smallest pay gap (column 2). The endowment effect for this wage decomposition of the EU-15 foreigners amounts to almost the same value as the wage gap itself (0.856). This means that if the observed German employees had the same endowment levels in the observable characteristics as the EU-15 foreigners, they would have about 14% lower wages. About 0.8% of the wage gap is due to differences in the coefficients. Although significant, this means that there are little differences in the evaluation of “pay-off” between the two groups in economic terms, i.e., 1 € gross daily income. Consequently, the wage gap between these two groups can be explained almost entirely, and this can be attributed exclusively to differences in endowments. With regard to the EU-8 countries (column 3), there are larger differences. The average wage gap between the two groups is around 40%. Again, the endowment effect does not deviate greatly from the observed difference. Thus, if the endowment levels of German employees were adjusted to the levels of EU-8 foreigners, their wages would be 41.4% lower. However, this decomposition reveals significant differences in coefficients: Thus, Germans with the same characteristics would

3 Foreign Vs. German Wage Differentials in Germany: Does the Home Country Matter?

69

earn, on average, 5% less (about 6.25 € per day) when adjusting German coefficient levels to the levels of EU-8 employees. It is worth noting that, on average, the proportion of females and young individuals among the EU-8 foreigners is higher (see Table 3.8). Therefore, coefficients for this group tend to be usually lower, as coefficients are lower in general for these groups. A similar picture results for the groups of EU-2 and Balkan foreigners, workers from Turkey, and the remaining world—with slightly different effects: Regarding selectivity and the “downgrading” of foreigners, the wage decomposition excluding occupations and tasks does not show any major deviations from the first decomposition. In principle, the endowment effect increases in all three decompositions, so that German employees, if they had the same endowment levels as those considered migrants groups, would receive slightly higher wages than those migrants. However, these advantages are in a small range, so it does not contrast with our previous conclusions.

3.7.1

Detailed Analysis: Endowments

Which characteristics are relevant and drive the wage difference between German employees and the considered groups? Table 3.10 provides a detailed decomposition of the endowment effect. Each row represents the joint impact of a set of variables as outlined in Sect. 3.4. The interpretation of the values is analogous to the previous approach but split into subgroups here: How would a German worker earn more (> 1) or less (< 1) if the endowment level were adjusted to the endowment levels of foreigners, evaluated at German coefficients? It is worth mentioning that a 1-percent change in wages of Germans means a change in gross daily wage of approximately 1.25 € ( 37.50 € per month).2 Consideration of the regional labour markets (RAM) shows positive (> 1) results for all groups except EU-8 foreigners. This indicates that most of the migrants work in regions, which are linked to higher wages than the average regional wage for Germans. Germans would thus benefit if they were distributed in the same regions as migrants. As EU-8 foreigners are more frequently employed in East Germany, where average wage levels are lower, they show negative results for this endowment. With regard to the distribution across industries, occupations, and task levels, German employees would suffer wage losses if they had exactly the same distribution across those variables as the considered migrant groups. Foreigners are therefore more frequently employed in industries and jobs/tasks that tend to pay less, providing evidence for a partial crowding in less popular jobs. However, this does not necessarily imply discrimination, as it is not clear whether such a choice was made

2

The effect of the variable clusters in Tables 3.10 and 3.11 is composed of the additive effects of the single variables mentioned in Chap. 4.

70

S. Brunow and O. Jost

on purpose due to a lack of better alternatives. Interestingly, the individual characteristics (INDIVID: gender and age) show that the differences are economically negligible for most foreign groups and relatively higher for individuals from Turkey (+ 2%). Because foreigners’ ages are highly correlated with the immigration waves (i.e., with the guest workers of the late 1960s and 1970s), we interpret the results as comparable wages with respect to age (and gender). In terms of education and, in particular, work experience on the German labour market, however, an adjustment of German levels to the one of foreigners is associated with an economically valuable reduction in wages of 10–20% for Germans. Because age and first appearance on the labour market are already controlled for, the experience measures are not biased with respect to age and immigration time. The serious disadvantage in terms of labour market experience can be primarily attributed to the fact that these groups have not had full access to the labour market in Germany until recent years. Accordingly, they have not yet been able to gain much labour market experience in Germany or still have language deficits, for example. Large differences of 5–10% in wages relate to firm structure. Foreigners tend to be employed in firms that pay lower wages. These are usually smaller firms which have lower shares of high-skilled workers. Considering the results of the various migrant groups reveals an interesting pattern: the adjustment of endowments from German levels to the levels of EU-15 foreigners and those from remaining world are quite similar. Because immigration from outside of the EU countries is rather limited, those foreigners obviously have very individual characteristics that are potentially very valuable for Germany and comparable to the characteristics of EU-15 countries. Additionally, wages in EU-15 countries are, in real terms, partly competitive to Germany, although potentially lower; therefore, incentives to migrate from those countries to Germany are not as strong. The new EU Member States and individuals from former Yugoslavia show also rather similar patterns. The group of Turkish workers is more of a mixture. On average, they have a lower level of education than the other groups, but in terms of labour market experience, they have the smallest deficit compared to German workers, together with the EU-15 migrants.

3.7.2

Coefficients

In Table 3.11 we present details of the coefficient effect. The interpretation is as before: How would a German earn more (> 1) or less (< 1) if the German coefficients were adjusted to the coefficient levels of foreigners, evaluated at mean German endowment levels? The coefficient effect shows higher pay-offs emerging from their labour market regions for German workers than for foreigners. Hence, German workers can make better use of the conditions of a region to achieve higher wage levels. Industry coefficients show on average only minor and insignificant deviations, indicating

3 Foreign Vs. German Wage Differentials in Germany: Does the Home Country Matter?

71

negligible differences in pay-off between the groups. A similar picture emerges for the coefficients considering occupations. There are also only insignificant and negligible differences indicating equal pay-offs for occupations. These results are interesting because, in view of the high number of cases, the insignificance of the results is particularly important. The implication of this finding is remarkable: When filling a vacancy, it does not matter to the employer where the individual comes from—all the workers get offered the same pay-off for their characteristics. Concerning the tasks, there are as well no differences in pay-off for the EU-15, EU-8, and remaining world foreigners compared to Germans. The coefficients for Turkish workers are evaluated significantly lower, i.e., 5.1%, leading to a wage gap of approximately 187.50 € gross monthly income. Noteworthy is also a positive gap of about 4.4% for the task levels of EU-2 and Balkan foreigners above German coefficient levels. Again, this has remarkable implications, as foreigners from this group are often seen as individuals coming to Germany for welfare reasons. Our result shows higher pay-offs for this group, which could also indicate a certain compensation for over-qualification. This is conceivable for this group, as they have only recently gained full access to the German labour market. It may indicate that German employers struggle to assess the quality of the qualifications correctly. This could be, for example, due to different education systems that do not send a clear or known signal to German employers. As a potential consequence, German employers hire these workers as auxiliary workers at the beginning. Alternatively, it can also be that these workers chose these unskilled occupations. It may be that EU-2 and Balkan foreigners need time to reach higher task levels, as employers need time to recognise their skills and qualifications. The endowment effect in Table 3.11 shows clear disadvantages at least for this group, which once again supports our integration time argument; i.e., they are employed more frequently as auxiliary workers. The differences in the coefficients for individual characteristics and education are significant. However, these are for the most part in the economically rather negligible range of less than or about 1%. Thus, there are no noteworthy deviations in the evaluation by the employer for these characteristics. By contrast, labour market experience shows marked differences in pay-off. Migrants receive lower returns for their labour market experience than comparable German employees. These flat experience profiles are frequently found in literature; see Bosseler (2014), Brunow and Jost (2019), and Zibrowius (2012). However, non-linear relationships are also conceivable here, leading to lower experience profiles, since migrants on average also have lower values in endowment levels of labour market experience. We address this issue in the robustness chapter. With regard to the firm characteristics, there are considerable differences. It is obvious throughout the analysis that foreigners have consistently higher firm coefficients. Apparently, foreigners can make better use of the firm structures and realise a higher pay-off. Similar results are shown by Brunow and Nijkamp (2016) who provide evidence of higher productivity of firms which employ foreigners.

72

3.8

S. Brunow and O. Jost

Robustness and Discussion

Our results show that there are group-specific differences, but also similarities. More specifically, they show similarities in terms of certain agglomerations and selections in some characteristics. Foreigners tend to sort into firms offering higher wages, with their individual pay-off being assessed in the same way as that of their German counterparts. Furthermore, migrants have significantly less work experience in Germany than native workers, but that is due to the gradual opening of the labour market over the last 15 years. At the same time, their labour market experience is less valued, which leads to slower development over time. With regard to the differences between the groups, the results show different endowment levels for experience or educational attainment. In addition, there are some notable differences within the coefficients for task levels. In order to be able to better include this aspect as well as other details, we carried out the following various robustness checks (results can be obtained upon request).

3.8.1

Labour Market Experience

Labour market experience is the subject of numerous discussions led by the groundbreaking work of Chiswick (1978). We took a closer look at this by looking at potential catching-up processes of foreigners on the German labour market. Since the majority of the EU-8 and EU-2 workers in our sample belong to the newer groups which have only had access to the labour market in Germany for a few years, we can track catching-up processes for these groups here only to a limited extent. For older cohorts of these countries and foreigners from other immigration countries, the differences in coefficients compared to German workers converge clearly over time (see Brunow and Jost 2019). As a result, it can be observed in some cases that after 10 years some differences are hardly measurable. Especially with regard to labour market experience, it becomes clear that foreigners can make more and more use of their experience the longer they have been part of the German labour market.

3.8.2

Gender

Most of the literature considering the migrant pay gap is limited to men only in the analysis. This often results from the discontinuous employment biographies of women, which can be volatile due to, for example, housework, family responsibilities, or children. We followed this example and performed an analysis for men only. As for the endowment and coefficient effect, there was no significant deviation. Men and women of the migrant groups thus have similar endowments and pay-offs

3 Foreign Vs. German Wage Differentials in Germany: Does the Home Country Matter?

73

on the labour market. The results suggest that there are no differences in employment patterns between the two genders.

3.8.3

Imputation of Wages and Education

There is a possibility of potential biases occurring as a result of the wage imputation used, which is why we wanted to take this into account in our subsequent wage decomposition. To this end, we excluded all employees holding a university degree, as these persons are, to a large extent, affected by the higher, potentially imputed salaries. There were no significant differences from the previous results. High salaries as well as certain biases that may be associated with them can be excluded. Neither the endowment nor the coefficient effect changed noticeably.

3.8.4

VET in Germany

As mentioned in the descriptions, we have generated an indicator which identifies foreigners who have completed their VET in Germany. Since this does not apply to all migrants, we also carried out analyses in which this variable was not included. There were no substantial changes in the decomposition, indicating robust results without this information. We restricted the sample to foreigners without VET in Germany, and the results were as already presented above. Lastly, we considered only foreigners that held such German VET certificate and found similar results: the coefficient effect became slightly smaller, indicating again the signalling character of the certificates for the employer.

3.8.5

Individual and Firm-Specific Heterogeneity

Due to the fact that we only use a cross-section of the year 2015 for our analysis, individual-specific fixed effects can only be inadequately controlled for. We therefore used what is called the CHK effects. These effects consist of a term specific to the individual and a term specific to the firm, which are incorporated into the decomposition accordingly. As expected, the inclusion of CHK effects reduced the differences in endowment and coefficient effects between groups clearly. At the same time, however, the larger differences remained, particularly in terms of labour market experience—endowment and coefficient. This confirms the results of the previous estimates and makes it

74

S. Brunow and O. Jost

clear that time spent on the labour market is the strongest indicator of convergence between groups to date.

3.9

Conclusion

Our analysis of the wage structures of migrant workers with a focus on workers belonging to newer migrant groups from EU countries has provided some insights into the labour market structures. Migrants still earn lower wages compared to German employees. At the same time, it becomes apparent that there are some significant differences among the observed migrant groups themselves. The data shows great differences in the distribution of task levels and qualifications across the groups under consideration. On average, Turkish workers have the lowest qualification structures and a relatively low proportion of highly qualified workers. Among EU migrants, however, EU-2 and EU-8 foreigners have lower education and task distributions compared to EU-15 employees. These distributions are also reflected in the median wages, with Turkish workers in particular benefiting from their long labour market experience in Germany and compensating for some of their deficits. All these differences and deficits are considered in our analytical approach, and we were able to show that the wage gap relative to the German reference group can be explained almost entirely by these differences. The major explanatory factors are selectivity in occupations, industries, task levels, and education as well as less labour market experience and firm characteristics. With regard to the differences in pay-offs, it can be seen that there are few significant differences between the native and migrant groups. When foreigners work in the same task levels or occupations, they earn, on average, the same wages as their German counterparts. However, the firm coefficients are higher for foreigners compared to Germans: a result which has been found before. Accordingly, they tend to sort themselves into more firms that pay higher wages, which is also associated with a better distribution in the labour market regions. Furthermore, our results show that from a policy point of view, the most relevant parameters to reduce the wage gap are labour market experience and education. Consequently, the wage gap will shrink over time if migrants stay for a longer period. It can be assumed that this will be accompanied by a degree of economic integration in the sense of Chiswick (1978), according to which migrants improve their language skills and benefit from a better knowledge of the market. Our sensitivity analyses indicate that migrants show a clear catching-up process especially within the first 10 years on the German labour market. In general, our results are very robust, and various changes to the models do not lead to any contradictory statements or findings. Seeing as we consider only employed full-time workers in our analysis, the uncertainty regarding the findings of migrants with greater proximity to the labour market decreases. We infer therefore that our findings are particularly valuable in view of the debate on skilled worker migration. This brings us to the following

3 Foreign Vs. German Wage Differentials in Germany: Does the Home Country Matter?

75

conclusion: As differences in endowments can explain most of the wage gap between the considered groups, we conclude that migrants, on average, earn fair wages given their current endowment levels. For the political debate on the immigration of skilled workers, which is particularly close to the labour market, this means that Germany is an attractive destination for immigration, as migrants generally tend not to be disadvantaged. Possible wage differentials narrow over time, which we can observe particularly among EU-15 migrants, but also foreigners of EU-8 and EU-2 + Balkan countries show similar effects. At the same time, our findings suggest that a potential clustering of migrants in less-favoured occupations might occur. Thus, policy programmes should prevent potential crowding and support the integration process for faster returns on labour market experience.

References Aldashev A, Gernandt J, Thomsen SL (2012) The immigrant-native wage gap in Germany. Jahrbuecher fuer Nationaloekonomie und Statistik 232(5):490–517 Bosseler, Mario (2014): Sorting within and across establishments. The immigrant-native wage differential in Germany. IAB-Discussion Paper 10/2014. URL: http://doku.iab.de/ discussionpapers/2014/dp1014.pdf Blinder AS (1973) Wage discrimination: reduced form and structural estimates. J Hum Resour 8 (4):436–455 Brunow S, Jost O (2019) Wages of migrant and native employees in Germany: new light on an old issue. (IAB-Discussion paper, 10/2019), Nürnberg, 48S Brunow S, Nijkamp P (2016) The impact of a culturally diverse workforce on firms' revenues and productivity: an empirical investigation on Germany. Int Reg Sci Rev 41(1):62–85 Card D, Heining J, Kline P (2013) Workplace heterogeneity and the rise of west German wage inequality. Q J Econ 128(3):967–1015 Chiswick BR (1978) The effect of Americanization on the earnings of foreign-born men. J Polit Econ 86(5):897–921 Fitzenberger B, Osikominu A, Völter R (2005) Imputation rules to improve the education variable in the IAB employment subsample. IAB-FDZ-Methodenreport, Nürnberg Himmler O, Jäckle R (2017) Literacy and the migrant-native wage gap. Rev Income Wealth 64 (3):592–625 Hofer H, Titelbach G, Winter-Ebmer R, Ahammer A (2017) Wage discrimination against immigrants in Austria? Labour 31(2):105–126 Jones FL, Kelley J (1984) Decomposing differences between groups. A cautionary note on measuring discrimination. Sociol Methods Res 12(3):323–343 Lehmer F, Ludsteck J (2011) The immigrant wage gap in Germany: are east Europeans worse off? Int Migr Rev 45(4):872–906 Lehmer F, Ludsteck J (2015) Wage assimilation of foreigners: which factors close the gap? Evidence from Germany. Rev Income Wealth 61(4):677–701 Nanos P, Schluter C (2014) The composition of wage differentials between migrants and natives. Eur Econ Rev, 65:23–44 Oaxaca R (1973) Male-female wage differentials in urban labor markets. Int Econ Rev 14 (3):693–709

76

S. Brunow and O. Jost

Yun M-S (2005) A simple solution to the identification problem in detailed wage decomposition, economic inquiry (2005), 43(4), pp. 766–772, with erratum. Econ Inq 44(1):198 Zibrowius, Michael (2012): Convergence or divergence? Immigrant wage assimilation patterns in Germany. SOEP papers, DIW Berlin, pp. 479–2012

Chapter 4

Bilingualism in the Labour Market Joanna Clifton-Sprigg and Kerry L. Papps

Abstract Previous research has found that among the native-born population, bilingual people earn less in the U.S. labour market. We examine whether a similar pattern exists in the U.K. and attempt to provide an explanation. We find that bilingual men do no worse than monolingual men, but that bilingual women earn significantly less than monolingual women. This is not explained by differences in cultural background, parental education or other family background variables. The result also holds when we control for various degrees of bias in unobserved characteristics. Instead, the result appears to be driven by differences across areas in the prevalence of bilingualism, with the negative earnings effects restricted to bilingual women living in areas with relatively low proportions of English speakers. The negative effects of bilingualism on women are also concentrated among speakers of South Asian languages and relatively uncommon languages.

4.1

Introduction

Bilingualism is increasingly common across the developed world. Although reliable statistics are scarce due to varying definitions of bilingualism, about 6% of students taking the PISA 2009 assessment in the O.E.C.D. countries spoke another language at home (O.E.C.D.: Social Policy Division, Directorate of Employment, Labour and Social Affairs 2012). The phenomenon has triggered a debate about its implications for individuals, the economy and society as a whole. Foreign, as well as European, languages feature in many aspects of the U.K. and E.U. education and labour market policies. The United Nations, UNESCO, the Council of Europe and the E.U. all emphasise the right of an individual to maintain their heritage language and culture and highlight the role of linguistic proficiency in achieving labour market success (European Commission 1995). This reflects a conviction that linguistic ability plays

J. Clifton-Sprigg · K. L. Papps (*) Department of Economics, University of Bath, Bath, UK e-mail: [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 W. Cochrane et al. (eds.), Labor Markets, Migration, and Mobility, New Frontiers in Regional Science: Asian Perspectives 45, https://doi.org/10.1007/978-981-15-9275-1_4

77

78

J. Clifton-Sprigg and K. L. Papps

a role in shaping one’s identity, social integration, employability and productivity, which directly influence the well-being and economic performance of a society as a whole. Within the U.K. there has been concern about a lack of foreign language skills in the labour market at a time when these are becoming more valuable in the global economy (Tinsley 2013). According to the 2011 U.K. Census, 7.7% of the population of England and Wales reported speaking a language other than English as their main language (Office for National Statistics 2011). However, the number of students studying foreign languages at school and university has been falling, as students increasingly choose science, technology, engineering and mathematics subjects instead. Despite this, there is little empirical evidence on the labour market returns to foreign language skills, whether learned at school or at home. Research to date has focused predominantly on the schooling and labour market performance of first- and second-generation immigrants.1 These studies focus on establishing the existence of gaps in labour market performance between native and immigrant populations and analysing factors which contribute to closing these disparities. Knowledge of the host country language has been identified by previous studies as a significant determinant of immigrant labour market outcomes, but few authors have examined the value of speaking an additional language. Given the number of factors that may affect a person’s labour market outcomes, the role of bilingualism is far from clear. Acquisition of more than one language early on in life may be seen as an investment in one’s general human capital (Chiswick and Miller 2018), leading to an increase in cognitive and non-cognitive skills (Carneiro et al. 2013; Bak et al. 2014). The returns to human capital investments in early childhood are particularly high and persistent and are expected to significantly contribute to an adult’s performance in the labour market (Heckman 2008). In this case, one would expect bilinguals to earn more than comparable monolinguals on average. However, bilingualism is likely to be correlated with many other characteristics that affect wages, such as family socio-economic status, cultural background and where a person lives. Many of these are typically unobservable by the researcher. In this paper, we use data from the Understanding Society survey to compare the earnings of U.K.-born individuals who spoke a language other than English at home as children with those of their monolingual counterparts and examine possible explanations for this relationship. We are able to control for a richer set of control variables than most previous studies, allowing us to eliminate the confounding effect of family and cultural background, which may be correlated with language skills. In addition, we use a recently developed technique to adjust for the potential bias

1 For example, Dustmann and Fabbri (2003), Bleakley and Chin (2004), Rooth and Saarela (2007), and Yao and van Ours (2015) consider the role of fluency in the language of the destination country; Blau et al. (2011) and Blau et al. (2013) analyse differences by country of origin, gender and intergenerational transmissions; and Bisin et al. (2011) investigate the influence of ethnicity.

4 Bilingualism in the Labour Market

79

caused by selection on unobservable characteristics. We consider men and women separately, acknowledging that bilingualism is known to have different effects on the learning opportunities of boys and girls (Lee and Hatteberg 2015) and that each gender faces a very different labour market. We focus exclusively on the native population of the U.K. and assess the labour market returns to fluency in languages other than English. The native-born population who are fluent in English and another language (“native bilinguals”) is not comparable to first-generation immigrants. Although many native bilinguals have an immigrant background, they were born in the U.K. Therefore, they have been exposed to the same cultural and institutional environment as their monolingual counterparts. In particular, both groups received education in the U.K. This is particularly useful for two reasons. Firstly, unlike the majority of immigrants, we can assume that all respondents are fluent in English. Thus, our focus is on the gains from the ability to speak a second language, rather than the penalty associated with an insufficient knowledge of English. Secondly, we can eliminate different institutional or educational systems as factors potentially confounding the relationship studied.2 We find that bilingual women have lower earnings than comparable monolingual women, but that there are no significant differences among men. This pattern is not explained by differences in the country of birth of a person’s parents, differences in parental education or other measures of family background. However, the negative effects of bilingualism on earnings among women are found to be concentrated among women who live in areas with a high fraction of non-English speakers. One explanation for this is that there may be limited employment opportunities in such areas. A negative earnings effect is only found among bilingual women who speak South Asian or less-common languages. In the next section, we review the relevant literature on bilingualism in economics and other disciplines. We describe the dataset we use in Sect. 4.3, before presenting our results in Sect. 4.4, starting with estimates of the overall wage effect of being bilingual, before considering heterogeneity between different groups of people. Some concluding comments are given in Sect. 4.5.

2

When analysing a relationship between one’s ability to speak several languages and his/her performance in the labour market, it is highly likely that many other factors (e.g. culture, different quality of education, social norms) that are correlated with bilingualism also affect one’s employability and income. If they are not explicitly accounted for the alleged relationship we are investigating may also capture these other influences, preventing any conclusions about the sole role of language. By comparing individuals born and brought up in the same country, we control for a wide range of such associated factors, which may otherwise confound the relationship.

80

4.2

J. Clifton-Sprigg and K. L. Papps

Background

The role of language in labour market outcomes has long been acknowledged in the economic literature. Most studies focus on the economic performance of first- and second-generation immigrants, who often lack proficiency in the language of the host country. Analyses for first-generation immigrants demonstrate the existence of an earnings gap between natives and immigrants, which closes with the duration of stay in the host country. Some studies (Bleakley and Chin 2004; Chiswick and Miller 1995; Dustmann and van Soest 2001; Yao and van Ours 2015) focus more specifically on linguistic skills and find that they are associated with better labour market performance. There is also evidence of differences in language acquisition, probability of employment and earnings depending on the ethnic origins of immigrants and the strength of their ethnic identity (Dustmann and Fabbri 2003; Bisin et al. 2011). This literature also acknowledges the role of age at arrival and argues that a significant part of the positive effect found among young immigrants is channelled through schooling (Bleakley and Chin 2004). Measurement error inherent in people’s assessments of their linguistic skills and endogeneity in the relationship between language and earnings pose the main estimation challenge, which is often addressed by an instrumental variables approach (Chiswick and Miller 1995; Dustmann and van Soest 2002; Bleakley and Chin 2004). Rooth and Saarela (2007) propose an alternative way of eliminating bias by considering the labour market outcomes of Finns in Sweden. In doing so they are comparing outcomes of immigrants of the same nationality, some of whom are native speakers of Swedish (the host country language) and others are native speakers of Finnish. This allows them to minimise concerns about measurement error in linguistic ability and to control for the role country of origin may play in outcomes. The approach is similar in spirit to ours in that the two groups studied differ solely in terms of language spoken and are otherwise comparable. Their results provide even higher estimates of the positive effect of immigrants’ language proficiency on earnings. The research focusing on native-born individuals, to which our paper contributes, has predominantly focused on second-generation immigrants in the U.S. (Chiswick and Miller 2018; Fry and Lowell 2003) or on native populations in countries with several official languages or dialects (e.g. Carliner (1981) for Canada, Paolo and Raymond (2012) and Rendon (2007) for Catalonia, Chiswick et al. (2000) for Bolivia, Yao and van Ours (2019a, b) for the Netherlands, Henley and Jones (2005) for Wales). These studies find mixed results, with no overall returns found in the U.S. and Welsh cases, negative estimates for the Netherlands and positive associations for Catalonia. Our work is closest in spirit to that of Chiswick and Miller (2018). They investigate the labour market returns to bilingualism among natives using U.S. data, focusing solely on males. They conclude that although overall bilingualism is negatively associated with earnings there is significant heterogeneity within the group of bilinguals and across quantiles of income, with the returns to some languages being positive.

4 Bilingualism in the Labour Market

81

Work in linguistics and cognitive psychology points towards skill enhancement due to language acquisition as a major driver of differences in earnings and employability. Baker (1999) provides an extensive overview of the impacts of bilingualism on cognitive outcomes in children, which typically translate into working life performance. Bilinguals seem to have an advantage in certain thinking dimensions, such as divergent thinking, creativity and metalinguistic awareness (Blumenfeld and Marian 2009), selective attention and inhibitory control (Bialystok et al. 2009). The ability to speak several languages may also delay onset of dementia (Bak et al. 2014). At the same time, however, it has been found that bilinguals may have a slower response time and make more errors in vocabulary tests focused on word retrieval. This may be reflected in speech production and is thought to be related to the processing of two languages at the same time (Bialystok et al. 2009). The research so far has found no correlation between bilingualism and I.Q. (Kaushanskaya and Marian 2007). The demand for particular language skills in the labour market is an alternative explanation for any positive association. Certain languages may be in particular demand, either due to the high frequency of their use locally or due to the nature of the business in which one is employed (Pendakur and Pendakur 2002). For example, an individual who fluently speaks a foreign language may be rewarded by his/her employer if many of the person’s customers speak that language. Conversely, bilingual workers may suffer a wage penalty if discrimination against foreign language speakers occurs. This may take the form of direct pay discrimination against bilinguals in areas where a lot of workers (and potentially a lot of employers) speak other languages. Alternatively, it may be indirect and due to firms choosing not to locate in areas with higher fractions of non-English speakers, resulting in low levels of labour demand. Identifying whether one or more of the aforementioned mechanisms explain any differences in outcomes between bilingual and monolingual individuals is complicated by the fact that language is also highly correlated with culture. Language may influence behaviour (Hicks et al. 2015), including labour force participation (Alesina et al. 2013), through its links to culture and impact on cognitive processes, rather than through an increased skill set or labour market demand. The ability to speak more than one language may, through cultural links, affect labour market prospects of females differently than those of males. For instance, exposure to the language and culture of a country with traditional gender roles may lead to lower labour market participation by females (Gay et al. 2018). In the case of bilinguals, knowledge of a second language is usually a consequence of having strong ties with other countries, typically the countries of birth of a person’s parents or grandparents. This family background may reinforce traditional gender roles or embrace modern attitudes to female labour market participation. For these reasons it is important to differentiate between genders and to control for cultural factors which may be correlated with bilingualism and directly affect labour market outcomes. We do so by running separate regressions for males and females and controlling for mothers’ country of birth dummies in the regressions.

82

J. Clifton-Sprigg and K. L. Papps

Fraction whose main childhood language was not English

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 First

Second

Third

Fourth

Immigrant generation Women

Men

Fig. 4.1 Fraction of non-English speakers by immigrant generation. Note: First generation are those who were born outside the U.K.; second generation are those who were born in the U.K. with at least one foreign-born parent; third generation are those who were born in the U.K. with two U. K.-born parents but at least one foreign-born grandparent; fourth generation are those who were born in the U.K. with U.K.-born parents and grandparents

4.3

Data

We use data from Understanding Society, which is an annual longitudinal study of 40,000 households in the U.K., capturing information about respondents’ demographic characteristics, socio-economic circumstances, attitudes, behaviour and health. Our data come from waves 1 to 5 of the survey, which were conducted between 2009 and 2014. Information on the language a person spoke at home during childhood was collected during wave 2 of Understanding Society. Speaking a language other than English is strongly related to immigration. Figure 4.1 shows the fraction of Understanding Society respondents who spoke another language as a child, by immigrant generation. 69% of first-generation immigrants (those who were born outside the U. K.) spoke another language, with little difference in rates between men and women. However, among the second generation (those who were born in the U.K. but whose parent(s) were born elsewhere), the fraction drops to 29%. Among both the third generation (those born in the U.K. with U.K.-born parents but a grandparent born elsewhere) and fourth generation (where the respondent and his/her parents and grandparents were all born in the U.K.), only 1% spoke a language other than English as a child. Since there is no information in Understanding Society on English proficiency, we exclude the first generation from our analysis, because they may have poor English skills. However, immigrants born in the U.K. should have been educated in

4 Bilingualism in the Labour Market

83

Table 4.1 Descriptive statistics for the estimation sample Variable Gross hourly pay Age A-level education Undergraduate education Postgraduate education Married Number of children Asian Black Other race Mixed race Hours worked Lives in London Second-generation immigrant Mother had high school education Mother had university education Father had high school education Father had university education Number of observations

Women Monolinguals 15.143 40.648 0.114 0.311 0.085 0.484 0.585 0.018 0.025 0.002 0.018 30.037 0.081 0.120 0.446 0.061 0.405 0.076 25,597

Bilinguals 15.031 33.837 0.173 0.332 0.124 0.495 0.778 0.681 0.045 0.007 0.017 29.838 0.328 0.854 0.309 0.054 0.317 0.079 969

Men Monolinguals 15.139 40.638 0.123 0.235 0.097 0.513 0.568 0.024 0.017 0.002 0.013 37.649 0.085 0.116 0.426 0.058 0.426 0.078 20,455

Bilinguals 14.548 33.689 0.152 0.295 0.130 0.563 0.817 0.743 0.025 0.006 0.009 35.846 0.321 0.833 0.282 0.041 0.310 0.110 822

Notes: Observations are weighted by the inverse of the total number of observations for each person in the sample

English and be native speakers. Therefore, an individual is identified as a bilingual in the sample if he/she reported speaking a language other than English at home during childhood. We further focus on individuals aged 16–65 and therefore part of the labour force. This leaves us with a sample of 10,239 female and 8393 male respondents who are observed over the five survey waves, once we drop observations with missing values for the main regression specification used in the next section. In total, we have 26,566 and 21,277 person-year observations on women and men, respectively. We use gross monthly pay, which is a derived variable, explicitly provided in the survey. It measures income from all sources of employment. We adjust for inflation using annual C.P.I. data, so that everything is expressed in 2009 pounds. In addition to Understanding Society data, we merge in data on the proportion of people (aged 3 and over) who speak English in the local area (specifically, the middle-layer super output area or M.S.O.A.) from the 2011 U.K. Census. Some summary statistics can be found in Table 4.1. Bilingual women and men earn slightly less than their monolingual counterparts. However, the two groups differ systematically in terms of their demographic characteristics. The bilingual sample is younger than their monolingual counterparts, has more children and is more likely to be married. Bilinguals also tend to live in areas with much lower fractions of people who speak English.

84

4.4

J. Clifton-Sprigg and K. L. Papps

Analysis

To begin with, we estimate the following specification (which is similar to that used by Chiswick and Miller), separately for employed men and women: ln EARNINGSit ¼ αBILINGUALi þ Xit γ þ uit ,

ð4:1Þ

where ln EARNINGS is the log of monthly gross pay of person i in year t, BILINGUAL is a dummy variable for whether the person spoke a language other than English at home as a child and X is a vector of control variables. We weight by the inverse of the number of observations for each person in the sample. The results of estimating eq. 4.1 are presented in Table 4.2. Initially, we use a similar set of control variables to Chiswick and Miller (2018); specifically, age and age squared, education dummies (completed A-level equivalent, undergraduate degree, postgraduate degree), a dummy for being married, number of children, race dummies (Asian, black, other, mixed), a London dummy and year dummies. As seen in the first and third columns, respectively, bilingualism is associated with a 5.2% reduction in earnings among women but has no significant relationship with earnings among men, in contrast to the findings of Chiswick and Miller. The results for women cannot be interpreted as a causal effect, however, because bilingualism reflects differences in the choices made by the respondents’ parents regarding which language(s) to expose the respondents to when they are young. If the choice to raise a child as bilingual is correlated with other family characteristics that affect the child’s future employment outcomes, this will introduce bias to the estimates. Bilingualism might be more common among immigrants from less developed countries, who would do worse in the U.K. labour market anyway. Furthermore, even within immigrant groups, bilingualism might be more common among less educated families, where the parents may not be able to speak English well and the children might be expected to acquire less human capital than children with educated parents. To test whether bilingualism captures differences in the labour market outcomes of immigrants from different countries, we add fixed effects for a person’s mother’s country of birth to X.3 To control for the influence of parental education, we add dummies for whether the respondent’s mother and father had a high school or university education. In this specification, the effects of bilingualism are estimated from differences in outcomes between bilingual and monolingual people whose parents had the same education level and came from the same country. As seen in the second column of Table 4.2, rather than being wiped out, the coefficient on the bilingual dummy for women becomes stronger and is now significant at the 1% level. The coefficient for men (in the fourth column) remains insignificant. For women, mother’s education, but not father’s education, has a significant positive

3

Very similar results are found when the father’s country of birth is controlled for instead.

4 Bilingualism in the Labour Market

85

Table 4.2 Results for log earnings regressions Variable Bilingual Age Age squared Married Number of children London Hours worked

Women (1) 0.052** (0.025) 0.067*** (0.002) 0.001*** (0.000) 0.048*** (0.009) 0.040*** (0.005) 0.152*** (0.015) 0.007*** (0.000)

Mother had high school education Mother had university education Father had high school education Father had university education Mother’s country of birth fixed effects Observations R-squared

No 26,566 0.156

(2) 0.072*** (0.027) 0.067*** (0.002) 0.001*** (0.000) 0.047*** (0.009) 0.036*** (0.005) 0.149*** (0.015) 0.007*** (0.000) 0.058*** (0.010) 0.071*** (0.019) 0.002 (0.010) 0.010 (0.017) Yes 26,566 0.160

Men (3) 0.017 (0.027) 0.071*** (0.002) 0.001*** (0.000) 0.119*** (0.010) 0.015*** (0.005) 0.160*** (0.015) 0.011*** (0.000)

No 21,277 0.218

(4) 0.008 (0.028) 0.071*** (0.002) 0.001*** (0.000) 0.116*** (0.010) 0.016*** (0.005) 0.157*** (0.016) 0.011*** (0.000) 0.055*** (0.011) 0.044** (0.021) 0.028*** (0.011) 0.024 (0.018) Yes 21,277 0.222

Notes: All regressions include a full set of race (5 categories), education (4 categories) and year (4 categories) dummies Standard errors are presented in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level, respectively Observations are weighted by the inverse of the total number of observations for each person in the sample

effect on earnings. For men, both mother’s education and father’s education (at least as far as high school) increase earnings. There are additional characteristics that might be correlated with whether a person speaks more than one language. For example, socially conservative women may be more likely to speak a second language, but also be likely to earn less. To control for this, we added dummies for whether a person was Christian or of another religion in the first column of Table 4.6. The sample size drops because information on religion is not available for all respondents. For women, being of a non-Christian religion raises earnings by 8%. The wage penalty associated with bilingualism is slightly bigger than in Table 4.2 and remains significant. For men, religion has no effect on earnings and the coefficient on being bilingual remains insignificant.

86

J. Clifton-Sprigg and K. L. Papps

A person’s social group might also be important, given evidence on the importance of ethnic networks in labour market outcomes (Damm 2012). To capture this, we control for the fraction of a person’s friends who are of same ethnic group. Possible answers are “all the same”, “more than half”, “about half” or “less than half”. In the second and fourth columns of Table 4.6, we add a dummy variable for whether a person responded with the third or fourth of these categories. This is associated with significantly lower earnings among both men and women. However, its inclusion makes little difference to the coefficient on bilingualism. To examine the effects bilingualism has at the extensive margin, we repeat the specifications from Table 4.2 using a dummy variable for whether a person is employed as the dependent variable. As shown in Table 4.7, bilingualism is associated with a lower likelihood of a woman being employed and a higher likelihood of a man being employed. This is robust to the inclusion of controls for mother’s country of birth and parents’ education.

4.4.1

Heterogeneity

Next, we consider what factors influence the magnitude of the wage penalty found for bilingual women in Table 4.2. As noted in the previous section, bilinguals can be divided into second-generation immigrants and third- or fourth-generation immigrants, according to whether at least one parent was born in the U.K. In the first and third columns of Table 4.3, we allow the effect of bilingualism to vary according to a Table 4.3 Heterogeneity in the log earnings regressions Variable Bilingual Fraction speaking English in M.S.O.A. Fraction speaking English in M.S.O.A.  bilingual Third/fourth generation

Women (1) 0.451*** (0.124) 0.052 (0.070) 0.463*** (0.147)

Third/fourth generation  bilingual Observations R-squared

23,055 0.153

(2) 0.088*** (0.031)

0.005 (0.021) 0.059 (0.060) 26,566 0.160

Men (3) 0.033 (0.127) 0.259*** (0.070) 0.017 (0.154)

18,416 0.220

(4) 0.030 (0.033)

0.016 (0.022) 0.075 (0.061) 21,277 0.222

Notes: All regressions include age and age squared, hours worked, number of children, married and London dummies and a full set of race (5 categories), education (4 categories), year (4 categories) and mother’s country of birth dummies Standard errors are presented in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level, respectively Observations are weighted by the inverse of the total number of observations for each person in the sample

4 Bilingualism in the Labour Market

87

person’s generation. The earnings penalty for second-generation bilingual women is 8.8%. Third or fourth generation bilingual women earn slightly more, but the difference is insignificant (most likely due to the small number of bilinguals who are third or higher generation). The earnings of bilingual men of either generation are not statistically significantly different from those of monolingual men. The earnings effects associated with bilingualism might also depend on the characteristics of the local labour market in which the person lives. As noted in Sect. 4.2, this may lead to either upward or downward bias in the bilingualism coefficient. If a lot of other people speak the same language in the local area, a bilingual person may have more opportunities in the labour market to exploit his/her language skills. However, if the presence of a lot of non-English-speakers leads to pay discrimination, either directly or by crowding out better-paying jobs because companies choose not to locate in the area, bilingualism might be associated with lower wages. To investigate this possibility, the fraction of people speaking English in the local area was added to the regression in the first and third columns of Table 4.3, alongside the interaction of this variable and the bilingual dummy. For women, the fraction speaking English has no significant effect, but the interaction term does have a significant positive effect. This indicates that bilingual women experience lower pay if they live in areas with relatively few English speakers, but this effect dissipates as the fraction speaking English increases. For women in areas at the tenth percentile of English-speaking fraction (81%), the coefficient on bilingualism is 0.078; whereas for women at the 90th percentile (99%), the coefficient is 0.008 and is not significantly different from zero. For men, neither the uninteracted nor the interacted bilingualism terms are significant. However, the fraction speaking English has a significant positive effect, indicating that all men do better if they live in areas with many English speakers.4

4.4.2

Correction for Selection

The results in Tables 4.2 and 4.6 indicate that the effects of bilingualism on earnings are robust to the inclusion of a variety of controls for family and cultural background. However, it is still possible that being bilingual reflects selection on some unobserved characteristics which also affect the labour market performance of individuals. For example, being bilingual may be a proxy for a person’s preference for a traditional division of labour in the household, even among others of the same

4 A similar pattern of results was found if the fraction of people in the local area speaking the respondent’s own language was used instead.

88

J. Clifton-Sprigg and K. L. Papps

religion, with similar ethnic networks and with parents of the same education level and from the same country. To shed light on the extent to which selection on unobservable characteristics poses a concern, we consider the extent to which it would alter the OLS results. We follow the work of Altonji et al. (2005) and Oster (2019) and make assumptions about the degree of selection on unobservables relative to the degree of selection on observed variables in the regression to produce alternative, bias-adjusted estimates. Altonji et al. (2005) provide a method to correct for selection on unobservables, using the degree of selection on observed variables in the regression as a guide. They, and subsequently Oster (2019), argue that it is plausible that degrees of selection on unobservable and observable variables in a regression are equal (δ ¼ 1) provided that the observable individual characteristics are just a random subset of a greater group of variables important for the outcome. The point estimates based on this assumption constitute an upper bound on OLS. Oster (2019) provides an extension to this work, noting that the unexplained variation in the outcome can be decomposed into two elements—idiosyncratic and one driven by unobserved characteristics. One can thus impose a further upper bound (Rmax) on the R2 from the regression of the dependent variable on all observed and non-idiosyncratic unobserved variables, effectively narrowing the bound on the OLS estimates of interest.5 Both approaches allow also for estimation of the degree of selection on unobservables required to eliminate the treatment effect. We present estimates of α obtained using Oster's method, assuming Rmax ¼ 0.3 and considering a range of values of δ.6 We use the regression specification in the second and fourth columns of Table 4.2. The results can be found in Table 4.4. They indicate that, regardless of the choice of δ, a negative effect of bilingualism is found for women. However, among men, when δ is above 0.3, the bilingual coefficient becomes positive. This implies that selection on unobserved characteristics larger than 30% of the degree of selection on observed characteristics in the regression would be sufficient to mask the negative effect of bilingualism on male earnings found earlier. Hence, the negative earnings effect for women found in Table 4.2 appears to be robust to selection on unobservable characteristics. However, under a relatively small degree of selection on unobservable characteristics, the earnings effect for men might in fact become positive.

5 Note that in Altonji et al. (2005) the assumption was that if all variables were observed the model would be fully explained and Rmax would be equal to 1. 6 The choice of the maximum explanatory power assumed in the regression is justified by the fact that the highest R2 obtained in the regressions here does not exceed 0.25. Therefore, the assumption that the techniques correcting for bias can explain as much as 30% of the variation in the dependent variable is generous and more realistic than assuming 100% of the variation would have been explained.

Assumed δ 0.1 0.083 0.005

0.2 0.096 0.002

0.3 0.112 0.001

0.4 0.131 0.004

0.5 0.154 0.008

0.6 0.185 0.012

0.7 0.226 0.016

0.8 0.284 0.021

0.9 0.371 0.026

1 0.519 0.031

Notes: All regressions include age and age squared, hours worked, number of children, married, London and proud of mother’s country dummies and a full set of race (five categories), education (four categories), year (four categories), mother’s country of birth (28 categories) and parental education (four categories) dummies Observations are weighted by the inverse of the total number of observations for each person in the sample A value of Rmax ¼ 0.3 is assumed

Sample Women Men

Table 4.4 Results for log earnings regressions using Oster’s (2019) approach

4 Bilingualism in the Labour Market 89

90

J. Clifton-Sprigg and K. L. Papps

Table 4.5 Results including separate language groups Variable Welsh or Gaelic Any European language Arabic Chinese South Asian language Any other language Parents’ education fixed effects Mother’s country of birth fixed effects Observations R-squared

Women (1) 0.051 (0.054) 0.074 (0.085) 0.096 (0.161) 0.126 (0.149) 0.068* (0.035) 0.139** (0.058) No No 26,566 0.156

(2) 0.055 (0.056) 0.126 (0.092) 0.164 (0.162) 0.010 (0.190) 0.085** (0.037) 0.179*** (0.062) Yes Yes 26,566 0.160

Men (3) 0.003 (0.055) 0.044 (0.102) 0.096 (0.375) 0.055 (0.125) 0.047 (0.034) 0.071 (0.070) No No 21,277 0.218

(4) 0.023 (0.057) 0.039 (0.116) 0.039 (0.375) 0.150 (0.151) 0.052 (0.036) 0.087 (0.072) Yes Yes 21,277 0.222

Notes: All regressions include age and age squared, hours worked, number of children, married and London dummies and a full set of race (five categories), education (four categories) and year (four categories) dummies Standard errors are presented in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level, respectively Observations are weighted by the inverse of the total number of observations for each person in the sample

4.4.3

Results for Separate Languages

The results presented so far provide evidence of the overall labour market effects of bilingualism. However, it is reasonable to imagine that there will be substantial heterogeneity in this by language group, reflecting the characteristics of different immigrant communities and the different levels of demand for their languages. In Table 4.5, the bilingual dummy is replaced by separate dummies for which language a person spoke as a child. These dummies capture whether a person spoke include a native U.K. language (Welsh or Gaelic), any European language, Arabic, Chinese, a South Asian language (Gujarati, Bengali, Punjabi or Urdu) or any other language. The first and third columns include these dummies alongside the basic control variables used in Eq. 4.1. The results indicate that the overall negative effect of

4 Bilingualism in the Labour Market

91

bilingualism among women is driven by women speaking South Asian and other languages. No significant effects of bilingualism are found for any language among men. Adding controls for family background (in the second and fourth columns) makes little difference to the results and the negative earnings effect associated with speaking a South Asian language or other language persists among women. These results indicate that, for example, among women with well-educated Pakistani-born parents, Urdu speakers fare particularly badly in the labour market, compared to women who speak only English. Given the results in Table 4.3 that the negative effects of bilingualism are concentrated among those who live in areas with relatively low fractions of non-English speakers, the results in Table 4.5 suggest that there may be linguistic “enclaves” where bilingual women do particularly poorly in the labour market. This does not appear to be due to the presence of unobserved person-specific characteristics (as seen in Table 4.4). Instead, it may be due to differential patterns of labour demand or of discrimination across local labour markets.

4.5

Conclusion

In this paper, we have used data from the Understanding Society survey in the U.K. for 2009–2014 to examine the effects bilingualism has on a person’s labour market outcomes. We find no evidence that bilingualism leads to higher earnings, as would be expected if it contributed to a person’s human capital. Instead, bilingualism is associated with lower earnings among females and has no effect on males. The negative effect for women does not appear to be driven by either observable or unobservable differences in a person’s family background. However, neighbourhood effects are found to be important. Specifically, bilingual women do significantly worse if they live in areas with few English speakers, perhaps due to discrimination. There is also significant variation in the returns to speaking specific languages. Among women, speaking South Asian or less common languages is associated with significantly lower earnings, while speakers of other languages do not earn significantly less than monolinguals. Among men, no language has a significant effect on earnings. Acknowledgement The authors would like to thank participants at the IFO Workshop on Labour Market and Social Policy (April 2017), the European Society of Population Economics Annual Conference (June 2017), the CReAM/RWI Workshop on the Economics of Migration (September 2017), the European Economic Association Annual Congress (August 2018), a SOFI Brown Bag seminar, and a University of Western Australia Business School seminar for their helpful suggestions.

92

J. Clifton-Sprigg and K. L. Papps

Appendix

Table 4.6 Robustness tests for log earnings regressions Variable Bilingual Christian Other religion

Women (1) 0.101* (0.054) 0.017 (0.012) 0.079** (0.031)

Half or less of friends of same race Observations R-squared

12,312 0.168

(2) 0.053* (0.029)

0.070*** (0.012) 23,459 0.156

Men (3) 0.023 (0.053) 0.019 (0.012) 0.007 (0.029)

11,543 0.233

(4) 0.020 (0.030)

0.057*** (0.013) 18,031 0.217

Notes: All regressions include age and age squared, hours worked, number of children, married and London dummies and a full set of race (five categories), education (four categories), year (four categories), parents’ education and mother’s country of birth dummies Standard errors are presented in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level, respectively Observations are weighted by the inverse of the total number of observations for each person in the sample

4 Bilingualism in the Labour Market

93

Table 4.7 Results for employment regressions Variable Bilingual Age Age squared Married Number of children London Hours worked

Women (1) 0.028*** (0.006) 0.014*** (0.001) 0.000*** (0.000) 0.001 (0.002) 0.003** (0.001) 0.013*** (0.004) 0.001*** (0.000)

Mother had high school education Mother had university education Father had high school education Father had university education Mother’s country of birth fixed effects Observations R-squared

No 26,566 0.061

(2) 0.028*** (0.006) 0.014*** (0.001) 0.000*** (0.000) 0.001 (0.002) 0.003** (0.001) 0.013*** (0.004) 0.001*** (0.000) 0.001 (0.002) 0.006 (0.005) 0.001 (0.002) 0.003 (0.004) Yes 26,566 0.061

Men (3) 0.032*** (0.007) 0.013*** (0.001) 0.000*** (0.000) 0.005* (0.003) 0.002 (0.001) 0.007 (0.004) 0.002*** (0.000)

No 21,277 0.081

(4) 0.032*** (0.007) 0.013*** (0.001) 0.000*** (0.000) 0.005* (0.003) 0.002 (0.001) 0.008** (0.004) 0.002*** (0.000) 0.003 (0.003) 0.000 (0.006) 0.003 (0.003) 0.006 (0.005) Yes 21,277 0.083

Notes: All regressions include age and age squared, hours worked, number of children, married and London dummies and a full set of race (five categories), education (four categories) and year (four categories) dummies Standard errors are presented in parentheses. *, ** and *** denote significance at the 10%, 5% and 1% level, respectively Observations are weighted by the inverse of the total number of observations for each person in the sample

References Alesina A, Giuliano P, Nunn N (2013) On the origins of gender roles: women and the plough. Q J Econ 128:469–530 Altonji JG et al (2005) Selection on observed and unobserved variables: assessing the effectiveness of Catholic schools. J Polit Econ 12:151–184 Bak TH, Nissan JJ, Allerhand MM, Deary IJ (2014) Does bilingualism influence cognitive ageing? Ann Neurol 75:959–963 Baker C (1999) Foundations of bilingual education and bilingualism. Multilingual Matters, Bristol Bialystok E et al (2009) Bilingual minds. Psychol Sci Publ Inter 10:89–129 Bisin A, Patacchini E, Verdier T, Zenou Y (2011) Ethnic identity and labour market outcomes of immigrants in Europe. Econ Policy 26:57–92

94

J. Clifton-Sprigg and K. L. Papps

Blau FD, Kahn LM, Liu A-H, Papps KL (2013) The transmission of women's fertility, human capital and work orientation across immigrant generations. J Popul Econ 26:405–435 Blau FD, Kahn LM, Papps KL (2011) Gender, source country characteristics and labour market assimilation among immigrants. Rev Econ Stat 98:43–58 Bleakley H, Chin A (2004) Language skills and earnings: evidence from childhood immigrants. Rev Econ Stat 86:481–496 Blumenfeld H, Marian V (2009) Chap. 3: language-cognition interactions during bilingual language development in children. In: Kuzmanovic B, Cuevas A (eds) Recent trends in education. Nova Science Publishers, New York Carliner G (1981) Wage differences by language group and the market for language skills in Canada. J Hum Resour 16:384–399 Carneiro P, Meghir C, Parey M (2013) Maternal education, home environments, and the development of children and adolescents. J Eur Econ Assoc 11:123–160 Chiswick BR, Miller PW (1995) Endogeneity between language and earnings: international analyses. J Labor Econ 13:246–288 Chiswick BR, Miller PW (2018) Do native-born bilinguals in the US earn more? Rev Econ Househ 16:563–583 Chiswick BR, Patrinos H, Hurst M (2000) Indigenous language skills and the labour market in a developing economy: Bolivia. Econ Dev Cult Chang 48:349–367 Damm AP (2012) Ethnic enclaves and immigrant labor market outcomes: quasi-experimental evidence. In: Chiswick BR, Miller PW (eds) Recent developments in the economics of international migration: volume 1: immigration: flows and adjustment. Edward Elgar Publishing, Cheltenham Dustmann C, Fabbri F (2003) Language proficiency and labour market performance of immigrants in the UK. Econ J 113:695–717 Dustmann C, van Soest A (2001) Language fluency and earnings: estimation with misclassified language indicators. Rev Econ Stat 83:663–674 Dustmann C, van Soest A (2002) Language and the earnings of immigrants. Ind Labor Relat Rev 55 European Commission (1995) White paper on education and training. via: https://europa.eu/ documents/comm/white_papers/pdf/com95_590_en.pdf. Accessed 20 May 2020 Fry R, Lowell L (2003) The value of bilingualism in the US labor market. Ind Labor Relat Rev 57:128–143 Gay V, Hicks DL, Santacreu-Vasut E, Shoham A (2018) Decomposing culture: can gendered language influence women's economic engagement? Rev Econ Househ 16(4):879–909 Heckman J (2008) Schools, skills, and synapses. Econ Inq 46:289–324 Henley A, Jones RE (2005) Earnings and linguistic proficiency in a bilingual economy. In: The Manchester school working paper series, 73, pp 300–320 Hicks DL, Santacreu-Vasut E, Shohan A (2015) Does mother tongue make for women’s work? Linguistics, household labor and gender identity. J Econ Behav Organ 110:19–44 Kaushanskaya M, Marian V (2007) Bilingual language processing and interference in bilinguals: evidence from eye tracking and picture naming. Lang Learn 57:119–163 Lee JC, Hatteberg SJ (2015) Bilingualism and status attainment among Latinos. Sociol Q 56 (4):695–722 O.E.C.D.: Social Policy Division, Directorate of Employment, Labour and Social Affairs (2012). C03.6 Percentage of immigrant children and their educational outcomes. via: www.oecd.org/ els/soc/49295179.pdf. Accessed 19 July 2016 Office for National Statistics (2011) Language in England and Wales: 2011. via: https://www.ons. gov.uk/peoplepopulationandcommunity/culturalidentity/language/articles/ languageinenglandandwales/2013-03-04. Accessed 20 May 2020 Oster E (2019) Unobservable selection and coefficient stability: theory and validation. J Bus Econ Statist 37(2):187–204 Paolo AD, Raymond JL (2012) Language knowledge and earnings in Catalonia. J Appl Econ 15:89–118

4 Bilingualism in the Labour Market

95

Pendakur K, Pendakur R (2002) Speaking in tongues: language knowledge as human capital and ethnicity. Int Migr Rev 36:147–178 Rendon S (2007) The Catalan premium: language and employment in Catalonia. J Popul Econ 20:669–686 Rooth D, Saarela J (2007) Native language and immigrant labour market outcomes: an alternative approach to measuring the returns for language skills. J Int Migr Integr 8:207–221 Tinsley, T. (2013): “Languages: The State of the Nation”, Report for British Academy Yao Y, van Ours JC (2015) Language skills and labour market performance of immigrants in the Netherlands. Labour Econ 34:76–85 Yao Y, van Ours JC (2019a) Dialect speech and wages. Econ Lett 177:35–38 Yao Y, van Ours JC (2019b) Dialect speaking and wages among native Dutch speakers. Empirica 46(4):653–668

Chapter 5

The Contrasting Importance of Quality of Life and Quality of Business for Domestic and International Migrants Arthur Grimes, Kate Preston, David Maré, Shaan Badenhorst, and Stuart Donovan

Abstract We examine whether bilateral regional migration flows are driven by the city’s quality of life (QL) or quality of business (QB). The QL and QB measures are constructed using (quality-adjusted) rents and wages in each city. QL and QB reflect the willingness to pay of households and firms, respectively, for local amenities. The measures are constructed for 31 urban areas in New Zealand using five-yearly census data covering 1986–2013. We adopt a gravity model of regional migration—augmented by destination and origin QL and QB—to model bilateral flows of workingage migrants (post-tertiary education and pre-retirement age). We also model flows between urban and rural areas and flows for the urban areas to and from overseas locations. We find different attractors for international versus domestic migrants according to the type of city amenity. International migrants are more attracted to cities with productive amenities, whereas domestic migrants are more attracted to places with consumption amenities. Thus, in deciding on the type of city amenity to enhance, city officials implicitly choose the type of migrant they attract as well as the type of city that may result.

A. Grimes (*) Motu Economic and Public Policy Research, Wellington, New Zealand Victoria University of Wellington, Wellington, New Zealand e-mail: [email protected] K. Preston Ministry of Justice, Wellington, New Zealand D. Maré Motu Economic and Public Policy Research, Wellington, New Zealand University of Waikato, Hamilton, New Zealand S. Badenhorst Motu Economic and Public Policy Research, Wellington, New Zealand S. Donovan Vrije Universiteit, Amsterdam, The Netherlands © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 W. Cochrane et al. (eds.), Labor Markets, Migration, and Mobility, New Frontiers in Regional Science: Asian Perspectives 45, https://doi.org/10.1007/978-981-15-9275-1_5

97

98

A. Grimes et al.

Keywords Migration · Amenities · Quality of life · Quality of business · Gravity model

5.1

Introduction

We test the importance of local quality of life (QL) and local quality of business (QB) in driving regional migration flows. Often, studies of regional migration focus on flows of domestic migrants,1 while studies of international migration focus on flows of migrants between countries.2 We examine flows of migrants to, from, and within New Zealand. We distinguish between the location choices of (1) domestic residents who relocate within New Zealand, (2) domestic residents who relocate to another country, and (3) international immigrants who choose a specific location within New Zealand. Domestic residents are defined here as those people who were in a New Zealand location 5 years prior (regardless of their official immigration or citizenship status), while international migrants are defined as those who were not in New Zealand 5 years prior (regardless of their official immigration or citizenship status). The distinction between domestic and international migrants proves to be important in assessing the effects of a location’s quality of life and its quality of business on migrants’ location choices.3 The QL and QB measures that we adopt are based on those of Gabriel and Rosenthal (2004) and Chen and Rosenthal (2008). They reflect the willingness to pay by workers and firms, respectively, for a location’s (consumption and productive) amenities. For a given location, QL is a function of local quality-adjusted rents minus wages, whereas QB is a function of rents plus wages. The intuition underpinning these measures is that places with relatively high rents and low wages must have attractive consumption amenities (high QL), otherwise people would not be willing to live there at those prices. Similarly, firms that choose to locate in places that have high rents and high wages must regard those places as having offsetting productive amenities (high QB), otherwise they would choose to relocate elsewhere. We use a gravity model of regional migration—augmented by destination and origin QL and QB—to model bilateral flows of working-age migrants (post-tertiary education and pre-retirement age). Our data covers the migration flows between the 31 main and secondary urban areas in New Zealand, derived from the 1986–2013 censuses. The QL and QB measures are derived from wage and rent data constructed for each urban area for each census wave. We incorporate migration flows between urban areas and rural New Zealand and also between urban areas and international

1

Biagi and Dotzel (2018) survey interregional migration models. Ariu (2018) surveys international migration models. 3 In Sect. 5.4, we discuss implications for our results of treating new migrants and returning New Zealanders as separate categories of migrants. 2

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

99

locations. Thus, we bring together two well-grounded models from the urban economics literature—the gravity model of migration and the amenity values of cities—within a single modelling framework. Our work builds on prior studies of regional within-country migration (e.g. Sjaastad 1962; Harris and Todaro 1970; Stark and Bloom 1985) which focus on the importance of location-specific factors. Some of these prior studies emphasise the importance of both pecuniary and non-pecuniary factors in determining residents’ location choice.4 Chen and Rosenthal (2008) explicitly model regional within-country migration based on QL and QB but do not analyse the impact of these influences on international migrants to or from the USA. A second set of related studies examines the choice of regional location for new migrants to a country (e.g. Bartel 1989; Bauer et al. 2007; Epstein 2008; Lichter and Johnson 2006; Munshi 2003; White 1998). Within New Zealand, studies by Maré et al. (2007), Maré et al. (2016), and Smart et al. (2018) have modelled location choices of new migrants to the country. A number of these studies incorporate both labour market variables and non-pecuniary variables as determinants of migrants’ location choices in their analysis. For instance, Smart et al. (2018) find that international migrants to New Zealand are attracted both to areas in which they can earn high wages and to areas with a relatively high proportion of migrants from their origin country. The influence of high wages is consistent with places that have high quality of business being attractive to international migrants while the origin country influence is consistent with a high quality of life being a determinant of migration choices. However, none of these cited studies models international migration explicitly within a framework that incorporates theoretically derived measures of both quality of life and business in the migration model. Furthermore, it is rare to find a study that incorporates both international and domestic migrants’ location choices within the same model. A recent contribution that does incorporate both types of migrants is that of Maré and Poot (2019); however, that analysis does not incorporate explicit measures of QL and QB as determinants of migrants’ location choices. By including international–urban and rural–urban flows into our gravity model of regional migration, we are able to observe differences in the importance of QL, QB, and other factors for migrants coming and going between urban, rural, and international destinations.5 We find evidence of different attractors for international versus domestic migrants according to the type of city amenity. International migrants are more attracted to cities that are based on productive amenities (QB), whereas domestic migrants are more attracted to places with consumption amenities (QL). This important difference in attractors for different types of migrants has frequently

4

For instance, after discussing monetary returns to regional migration, Sjaastad (1962) states (p. 86): “In addition, there will be a non-monetary component, again positive or negative, reflecting his preference for that place as compared to his former residence”. 5 Formally, we treat migrant groups as being homogeneous within the relevant group, but as heterogeneous between groups. As we discuss subsequently, some heterogeneity within groups and different constraints facing different groups are both likely to influence the empirical findings.

100

A. Grimes et al.

been overlooked by researchers and city officials alike. Indeed, in deciding on the type of city amenity to enhance (e.g. a port that facilitates business or a concert hall that facilitates consumption), city officials are implicitly choosing the type of migrant they attract as well as the type of city that may result.

5.2

Model

The gravity model of migration treats migration flows between two locations as increasing in the size of each of their populations and decreasing with the distance between them. We estimate a gravity model of migration, where Mijt represents the migration flow between origin location i and destination location j from time t
1 to time t.6 The basic model is set out in Eq. (5.1): ln M ijt ¼ α þ β1 ln Pit
1 þ β2 ln Pjt
1 þ β3 QLit
1 þ β4 QLjt
1 þ β5 QBit
1 þ β6 QBjt
1 þ β7 Cij þ γ t þ δi þ μ j þ εijt

ð5:1Þ

where Pit
1 (Pjt
1) is the population of location i ( j) at time t
1, and QLit
1 and QBit
1 (QLjt
1 and QBjt
1) refer to the quality of life measure and quality of business measure for location i ( j), respectively, at t
1. These variables are lagged to incorporate their values at the beginning of the migration period, reducing the risk of reverse causality. We control for the cost of moving from location i to location j, Cij,7 and include time fixed effects, γ t, origin fixed effects, δi, and destination fixed effects μj. The β terms represent the parameters to be estimated and εijt is a random error term clustered at the origin and destination location pair level. To interpret results, note that a one standard deviation increase in either QL or QB will result in (approximately) a 100*β% increase in the migration flow.8 From the standard gravity model, we expect to find positive coefficients on both origin and destination population. Moving costs, Cij, are captured by an indicator for whether locations i and j are located on the same island plus separate variables measuring the distance between the two locations for those on the same island and those on different islands.9 Including the distance measures separately allows for For our OLS estimates, we add one to the migration flow since Mijt enters the equation in logarithmic form; this enables us to include pairs of locations between which there is no migration in the estimation sample. This adjustment is not needed for our Poisson and negative binomial regressions. 7 Ideally one would control for time-specific moving costs but these have not changed materially over our sample period. Estimates of effects of time-varying travel costs on migration show only minor differences (Poot et al. 2016). 8 The exact estimated migration increase from a one standard deviation change in QL or QB will be 100*(eβ
1)%. 9 All of the urban areas in our data are located on either of New Zealand’s two major islands: North Island and South Island. New Zealand (with a population of 4.24 million in March 2013) has a land area of 268,000 km2 which is similar to that of the United Kingdom (242,000 km2). 6

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

101

different costs associated with distance when moving to another urban area on the same island versus moving to an urban area on a different island. We expect greater migration between urban areas on the same island (reflecting lower moving costs) with reduced migration between locations that are further apart (reflecting higher moving costs). We hypothesise that there will be greater migration towards locations with relatively high QL and QB since we expect that some groups will be attracted to places that are good to reside in and others will be attracted to places that are good for jobs and income. It is less clear whether higher quality places will also retain their populations. Their high levels of productive and consumption amenities will have a direct impact in reducing emigration from those locations. However, there may also be two other effects that are relevant. The first is a selection effect. People who value consumption amenities highly are more likely to be already located in medium to high QL areas, and they may then move from such an area to an even higher QL area. Conversely, those people who do not value consumption amenities highly are more likely to be already located in low QL places and to stay there. The second effect is a life-course effect. Some people, for instance, may choose to locate in a high QL location when young and then move to a high QB location to earn more later in life.10 Each of these effects will be reflected in the origin QL coefficient (and similarly for QB if relevant) which is, therefore, of ambiguous sign. For this reason, we place more emphasis on the destination coefficients than the origin coefficients in interpreting results. Model (1) is an unrestricted model in which the relative influences of origin and destination variables (QL and QB) can differ. We also report results from a restricted version of this model in which we set β3¼
β4 and β5¼
β6 so that it is the difference between origin and destination variables that influences migration choices. We estimate both these model variants as a function solely of the variables shown in (1), and also with a vector of added amenity variables. We do so in case our estimates of QL and QB impacts are reflecting a specific amenity rather than the aggregated value of amenities as summarised by the QL and QB measures. The chosen amenities are those used by Preston et al. (2018) to estimate determinants of QL and QB across a larger range of urban areas in New Zealand. The amenities include: rainfall, sun hours, wind strength, proximity to the sea or a lake, and shares of employment in each of accommodation/food/recreation services, education, health, land transport, and air transport. Following the approach of Poot et al. (2016), we extend model (1) to incorporate international–urban and rural–urban migration flows. This extended model forms the primary focus of our analysis. Because international and rural locations are not confined to a specific location, we do not observe the distance variables for them, nor can we include their origin population, QL and QB measures (or separate

10

Grimes et al. (2017) provide an explicit model showing that individuals with a high rate of time preference will tend to move from a high consumption amenity area to a high income area over their lifetime, while those with a low rate of time preference will tend to move in the opposite direction.

102

A. Grimes et al.

amenity variables for them). Thus we do not estimate the restricted model for origin and destination QL and QB effects, nor do we estimate equations with separate amenity variables for this model. In order to include rural and international flows in our model, we define the following dummy variables for each type of migration flow: • UtoUij ¼ 1 if and only if the origin i and destination j are both urban areas and 0 otherwise; • UtoRij ¼ 1 if and only if the origin i is an urban area and the destination j is rural and 0 otherwise; • UtoWij ¼ 1 if and only if the origin i is an urban area and the destination j is overseas and 0 otherwise; • RtoUij ¼ 1 if and only if the origin i is rural and the destination j is an urban area and 0 otherwise; • WtoUij ¼ 1 if and only if the origin i is overseas and the destination j is an urban area and 0 otherwise. We then define our gravity model with rural and international–urban migration flows as:
lnM ijt ¼ δ0 þ UtoU ij α1 Oit 2 1 þ β1 Djt 2 1 þ β4 C ij þ UtoRij ðδ1 þ α2 Oit 2 1 Þ

þUtoW ij ðδ2 þ α3 Oit 2 1 Þ þ RtoU ij δ3 þ β2 Djt 2 1 þ WtoU ij δ4 þ β3 Djt 2 1 þγ t þ θi þ μ j þ εijt ð5:2Þ where Oit 2 1 is a vector of origin characteristics in the previous wave (lnPit
1, QLit
1, QBit
1) and Djt 2 1 is a vector of destination characteristics in the previous wave (lnPjt
1, QLjt
1QBjt
1).11 We estimate our models with and without origin and destination fixed effects. When we do not include these fixed effects, we can observe the impact of the (timeinvariant) distance (and amenity) variables on migration and we can assess the raw association between migration flows and cities’ QL and QB. However, omitted (unobserved) variable bias is potentially a problem with these estimates, so our preferred estimates include the fixed effects which account for the impact of timeinvariant characteristics of each place on migration flows. There remains some risk that time-varying omitted variables could be correlated with our time-varying explanatory variables (population, QL, and QB). Hypothetically, for instance, travel connections may have changed over time and these may be reflected in travel costs as well as in population and amenity variables. The prior results of Poot et al. (2016) suggest that any such changes have not been substantive enough to materially alter our estimates.

11

Note that the UtoUij dummy is omitted as a stand-alone variable to avoid perfect multicollinearity.

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

103

Another risk to our modelling strategy is measurement error, especially in QL and QB. An earlier study (Preston et al. 2018) found QL and QB measures to be more volatile (across censuses) for smaller localities, reflecting thin rental and labour markets. For this reason, we restrict our attention to the 31 main and secondary urban areas in the country.

5.3

Data

We use census data for population movements. We focus on the main and secondary urban areas as defined by Statistics NZ using 2013 boundaries. Where urban areas are geographically contiguous, we combine them into a single urban area, leaving us with 31 urban areas.12 Our data span six censuses from 1986 to 2013, held every five years (except for the 2013 census, which was delayed by 2 years due to the February 2011 Christchurch earthquake). The data are limited to migration flows of usual residents aged 30–59 years of age in year t (i.e. 25–54 five years previously) and their movements relative to 5 years earlier. The age range is chosen so that we focus on migration flows of the working-age population, thus we (intentionally) ignore movements of people who are of typical student and retirement ages. In our formal modelling, we do not distinguish between international migrants new to New Zealand and those who are New Zealanders returning home; subsequently we discuss implications for our results when we split international migrants into these two categories. Information on migration in the New Zealand Census is obtained from a question on current place of residence and that of 5 years ago. We define bilateral migration flows as the counts of census respondents aged 30–59 in each destination location, who were resident in the origin location 5 years ago. We have no information on any intervening location choices, so we effectively assume that the migration was direct. The migration flows between the 31 urban areas form 930 origin–destination pairs. These pairs are observed over six censuses, giving rise to a total of 5580 observations. All migration flows between the urban areas in our data and any other part of New Zealand are coded as urban–rural migration flows. Since flows of individuals from each urban area to overseas are not observed in the census, migration flows between the urban areas in our data and overseas are imputed as the residual change in the population of the urban area after accounting for immigration, internal migration, and observed registrations of deaths.13 Adding in urban–

12 Urban areas that we combine are Northern Auckland Zone, Western Auckland Zone, Central Auckland Zone, and Southern Auckland Zone (into Auckland); Hamilton Zone, Cambridge Zone, and Te Awamatu Zone (into Hamilton); Wellington Zone, Lower Hutt Zone, Upper Hutt Zone, and Porirua Zone (into Wellington); Napier Zone and Hastings Zone (into Napier-Hastings). 13 The resulting numbers contain some error due to census undercounting, etc.

104

A. Grimes et al.

international and urban–rural flows results in 6324 observations (i.e. an additional 744 destination–origin pairs). To ensure confidentiality, the migration data derived from the census is randomly rounded to base three,14 this rounding occurs after any positive flow that is less than six is suppressed. Of our 6324 potential observations, 1382 are suppressed while in a further 479 cases we know that the true bilateral flow was zero for that inter-censal period. We treat all suppressed observations as having a flow of two.15 Appendix Table 5.3 details the distribution of the aggregated and disaggregated migration flows. The adjustments for nil or low flows clearly have some arbitrariness attached to them. For this reason, the estimates reported in the main text exclude all observations with true zero or suppressed flows. In case these exclusions cause problems of selection bias, we report estimates in the Appendix in which all observations are included. Results are stable across the two samples although estimates are less contaminated by noise in our main set of results. For the population variable we use counts of the usually resident population aged 30–59, to be consistent with the migrating sample. The population of this age group will be strongly correlated with total population, so in the regression will likely pick up the attractiveness of total city size. Distance is measured as kilometres between the city-centre of each urban area in 2013, obtained using Google Maps (entered as separate variables for urban areas on the same or different islands).16 Our QL and QB measures for each urban area and time period are based on those used in Preston et al. (2018) and in Maré and Poot (2019). Briefly, these models are based on the spatial equilibrium insights of Rosen (1979) and Roback (1982, 1988) as extended by Gabriel and Rosenthal (2004) and Chen and Rosenthal (2008). The quality of business and quality of life in location i are calculated as: QBit ¼

γ ln ðr it Þ þ ln ðwit Þ 1
γ

QLit ¼ α ln ðr it Þ
ln ðwit Þ

ð5:3Þ ð5:4Þ

where ln(rit) is the quality-adjusted rent premium in location i at time t, ln(wit) is the quality-adjusted wage premium in location i at time t, γ (1
γ) is the coefficient on

14 For instance, a migration flow of 58 is reported as 57 with probability 2/3 and as 60 with probability 1/3; a flow of 59 is reported as 57 with probability 1/3 and as 60 with probability of 2/3; a flow of 60 is reported as 60; and similarly for flows of 61 and 62 (where 63 replaces 57 as the alternative possibility). 15 As with all other observations, one is then added to this value for the OLS estimates. Thus lnMijt ¼ 0 for true zero flows, lnMijt  1.1 for suppressed flows, and lnMijt  2.0 for the lowest reported flows (of six). 16 The distance information between urban areas was provided by the authors of Poot et al. (2016). For urban areas which we combined because they are contiguous, we took the average of the distance between each of the combined urban areas and the other location.

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

105

land (labour) in the representative firm’s Cobb–Douglas production function, and α is the coefficient on housing in the representative consumer’s utility function.17 Based on aggregate data, we set γ ¼ 0.1 and α ¼ 0.2. Firms that operate in a location with highly productive amenities (i.e. in places with a high quality of business) can afford to pay a combination of high rents and high wages given their higher productivity. Individuals who live in a location with high consumption amenities (i.e. in places with a high quality of life) will be prepared to pay higher rents relative to wages than elsewhere since they benefit from the non-pecuniary amenities in that location. Thus the QL and QB measures reflect the value that households (for QL) and firms (for QB) place on local amenities, and rents will be bid up to reflect higher productive and consumption amenities. Each of QL and QB has been normalised to mean zero and standard deviation one across the sample of 130 urban locations listed in Preston et al. (2018). That study showed that QB tended to be high in larger places so the bulk of our 31 (“large”) urban areas has QB > 0. Similarly, smaller places tend to have higher quality of life so a majority of our areas has QL < 0. In constructing QL and QB for this study we have made one conceptual change relative to the measures in Preston et al. (2018). The prior studies defined rents and wages as those paid and earned by individuals who resided in a specific location. This meant if an individual resided in location A and worked in a separate location B that the wage earned in B would be attributed to location A. In practice, some of our (smaller) locations are close enough to larger cities to enable commuting, and this approach may bias upwards the wages attributed to these smaller locations (resulting in an upwardly biased estimate for QB and a downwardly biased estimate of QL in those towns). For the measures used in the current study we instead define wages for a location to be the wages earned by people who work in that location. In equilibrium, the person’s wage in a town near a main city will equal the wage in the city less transport costs; thus for the individual the wage earned in the home location is the relevant (after transport cost) income indicator. The resulting QL and QB measures for our 31 urban areas in 2013 are displayed in Fig. 5.1. Figs. 5.2 and 5.3 show the 1986 and 2013 values for QL and QB, respectively, indicating movements in quality of life and quality of business over our sample period. In each of the figures, the size of circle is proportional to the urban area’s 2013 population. Appendix Tables 5.4 and 5.5 provide the QL and QB values, respectively, for each census year for each location. We see from the three figures and the Appendix tables that New Zealand’s largest city (Auckland) has the highest quality of business, closely followed by the capital city (Wellington) and then by the next two largest cities (Christchurch and Hamilton). The high QB in larger urban areas is consistent with agglomeration economies

17 Rents and wages are quality-adjusted at each census date. Rents are quality-adjusted by regressing actual rents on the number of rooms, number of bedrooms, dwelling type, and available heating types. Wages are quality-adjusted by regressing actual wages on age, gender, ethnicity, industry, birthplace, religion, and qualifications.

106 Fig. 5.1 2013 QL and QB measures for 31 urban areas

Fig. 5.2 1986 and 2013 QL measures

Fig. 5.3 1986 and 2013 QB measures

A. Grimes et al.

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

107

reported in other studies (e.g. Maré and Graham 2013). Smaller places, on average, have a higher quality of life, with Queenstown (a popular tourist resort) being the most favoured area in this respect; however, some smaller urban areas (e.g. Tokoroa) have low QL. Figs. 5.2 and 5.3 reveal considerable persistence in both QL and QB from the start to the end of the sample period, but also indicate cases of substantial changes in fortunes over time. For instance, Ashburton has moved substantially upwards as a place for business, while Levin has moved substantially downwards; Auckland’s quality of life has declined substantially over the sample period. Thus, we see considerable temporal as well as spatial variation in our QL and QB measures over the sample which enables us to test the influence of these variables on migrants’ location choices even after accounting for destination and origin fixed effects.

5.4

Gravity Model Results

We begin by using Eq. (5.1) to model bilateral urban to urban flows for the 31 urban areas over the six census waves using OLS regressions. This specification does not include flows to and from rural and international locations. Table 5.1 reports the results for eight specifications, each based on Eq. (5.1), for the sample that excludes zero and suppressed flows. Appendix Table 5.6 reports the results for the full sample that includes these flows. In each case the cost vector is proxied by the three distance-related variables described in Sect. 5.2. Time fixed effects and a constant are included in all equations (but not reported). The first four columns (in each of Tables 5.1 and 5.6) do not include separate amenity variables, which are added for the final four columns. Even numbered columns include origin and destination fixed effects, while odd numbered columns omit these effects. Columns 1, 2, 5, and 6 include unrestricted destination and origin QL and QB, while the remaining columns restrict impacts of these variables to be related to the difference between their destination and origin values.18 Our analysis of results concentrates on the estimates shown in Table 5.1, but the full sample estimates (Table 5.6) show similar patterns. Prior to analysing the effects of our focal variables (QL and QB) we note that the cost variables all have the expected signs (and are significant in 22 of 24 cases); thus people are more likely to migrate to places in the same island and that are closer to their origin location. Population has positive effects on both arrivals and departures so that people both tend to leave large places and to migrate to large places. Once origin and destination fixed effects are added, we find that the impact of origin 18 We explored a further specification based on Eq. (5.1) that was estimated just for emigrants from Christchurch (one of New Zealand’s three largest cities) that suffered devastating earthquakes between 2006 and 2013. This specification explored whether estimated patterns of emigration from Christchurch changed following the (exogenous) earthquakes. We found no evidence of a change in the emigration pattern from Christchurch in relation to QL and QB following the earthquakes.

108

A. Grimes et al.

population is materially larger than that of destination population indicating that (domestic) residents are, on balance, tending to leave the larger cities for smaller places. Coefficients on the time fixed effects (which are not reported) show a tendency to become increasingly negative through the sample period (relative to a zero base in 1986) implying, ceteris paribus, that inter-urban migration flows have tended to reduce over time.19 Turning to quality of life, we see that destination QL is, in all (unrestricted) cases, a strong drawcard for migrants. We also observe that migrants tend to leave places with high QL, consistent with a selection effect or a life-course pattern (at least for those with a high rate of time preference). In each case, the destination QL effect is greater than the origin effect so that people, on balance, leave places that are nice to live in and move to places that are even more attractive. Once the coefficients are restricted, this effect remains significant for the raw results without other amenities (column 3) but not with other controls added. With respect to quality of business, all unrestricted equations show that people tend to migrate out of cities with high QB. In addition, once origin and destination fixed effects are included, there is little evidence that they are attracted to places where quality of business is high. In the restricted equations, migrants are seen either to be indifferent to QB or, on balance, to leave places with high QB. Thus, locations with a high quality of business do not appear to be attractive to domestic (working age) migrants. Instead, the buoyant economic conditions in cities that are good for business may afford existing residents in those cities the incomes and the capital (through high house prices and rents associated with both high QL and high QB) to leave those locations for more pleasant places in which to live.20 The results in Table 5.1 apply only to inter-urban flows and so do not include inflows and outflows of urban locations with rural and international locations. We gain deeper insights by extending the analysis to these additional flows, based on Eq. (5.2). OLS results are reported in Table 5.2 for the sample that excludes zero and suppressed flows. Table 5.7 in the Appendix reports results for which these flows are included. The first two columns of Table 5.7 report OLS results (with and without location fixed effects); column 3 reports Poisson regression estimates and column 4 presents negative binomial estimates (each incorporating location fixed effects).21 We cannot estimate the restricted format for this specification since we do not

19

This pattern is the opposite of what we might expect if it were the case that reductions in transport costs through the period had boosted migration flows. 20 We note that the attractiveness of (expensive) high QB locations may be different for people at the start of their working careers (e.g. those aged under 25 years) and this group is (intentionally) omitted from our sample. Grimes et al. (2020) examine location choices of tertiary graduates with respect to QL and QB across potential destinations. 21 The Poisson and negative binomial regressions do not require us to add one to migration flows to enable the zero flows to be included; however, we still need to impose an arbitrary flow (assumed to be two) for suppressed flows, which inevitably results in some inaccuracy. The negative binomial results are preferred to the Poisson regression results in this case since the data displays overdispersion contrary to the assumptions of the Poisson approach.

0.761*** (0.022) 0.738*** (0.022) 1.593*** (0.473)
0.737*** (0.031)
0.524*** (0.063) 3740 0.843 No No

1 0.440*** (0.038) 0.134*** (0.034) 0.155*** (0.038) 0.097** (0.035)

0.286*** (0.070) 1.004*** (0.091) 1.869*** (0.471)
0.855*** (0.026)
0.615*** (0.066) 3740 0.904 No Yes

2 0.128*** (0.024) 0.049* (0.023) 0.079** (0.024) 0.114*** (0.024) 0.133*** (0.033) 0.014 (0.032) 0.769*** (0.023) 0.745*** (0.024) 0.825 (0.457)
0.686*** (0.034)
0.584*** (0.061) 3740 0.821 No No

3

0.022 (0.018)
0.033* (0.016) 0.388*** (0.065) 1.103*** (0.086) 1.864*** (0.472)
0.854*** (0.026)
0.615*** (0.066) 3740 0.903 No Yes

4

0.762*** (0.024) 0.742*** (0.023) 1.593*** (0.469)
0.739*** (0.030)
0.527*** (0.063) 3740 0.849 Yes No

5 0.350*** (0.039) 0.117*** (0.035) 0.245*** (0.037) 0.111** (0.036)

0.164* (0.073) 1.139*** (0.083) 1.861*** (0.472)
0.856*** (0.026)
0.617*** (0.066) 3740 0.905 Yes Yes

6 0.117*** (0.025) 0.032 (0.023) 0.092*** (0.023) 0.132*** (0.023) 0.042 (0.027)
0.002 (0.028) 0.770*** (0.023) 0.748*** (0.023) 0.822 (0.455)
0.688*** (0.033)
0.587*** (0.062) 3740 0.828 Yes No

7

0.01 (0.018)
0.051** (0.016) 0.267*** (0.069) 1.237*** (0.078) 1.856*** (0.472)
0.855*** (0.026)
0.617*** (0.066) 3740 0.904 Yes Yes

8

The dependent variable in all models is the natural logarithm of the migration flow (plus one) between the destination and origin location, i.e. the population aged 30–59 in each destination urban area that was usually resident in the origin city 5 years ago. An observation is an origin–destination pair. The sample includes 31 (continued)

N R2 Added amenity variables Destination and origin fixed effects

Distance if different island (ln)

Distance if same island (ln)

Same island dummy

Origin population (lnPit
1)

Destination population (lnPjt
1)

Difference QB (QBjt
1 - QBit
1)

Difference QL (QLjt
1 - QLit
1)

Origin QB (QBit
1)

Origin QL (QLit
1)

Destination QB (QBjt
1)

Destination QL (QLjt
1)

Table 5.1 Gravity model: Inter-urban migration (zero and suppressed flows dropped)

5 The Contrasting Importance of Quality of Life and Quality of Business for. . . 109

cities (930 destination–origin pairs) over six censuses from 1986 to 2013. Distance is the 2013 driving distance between the origin and destination cities in km. All models include year fixed effects and a constant not shown. Added amenity variables include shares of employment in each of accommodation/food/ recreation services, education, health, land transport, and air transport in columns 5–8, plus rainfall, sun hours, wind strength, and proximity to the sea or a lake in columns 5 and 7. Estimation is by OLS. Standard errors clustered by origin–destination location pair in parentheses; *p < 0.1, **p < 0.05, ***p < 0.01

110 A. Grimes et al.

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

111

observe QL and QB for rural and international observations. We also do not observe amenities for these locations and so drop the specifications with added amenities (noting that the Table 5.1 results were not materially impacted by the inclusion or exclusion of these variables). In Tables 5.2 and 5.7, prefixes before each of the QL, QB, and population variables indicate whether the estimated coefficient refers to flows that are urban to urban (U-U), rural to urban (R-U), world (international) to urban (W-U), urban to rural (U-R), or world to urban (W-U). In the discussion that follows, we concentrate primarily on the results with origin and destination fixed effects included (i.e. column 2 of Table 5.2). These results are similar to each of the OLS, Poisson, and negative binomial regression results (with location fixed effects added) that are reported in Table 5.7 for each of our focal variables. The cost variables, which are applicable only to urban to urban flows, again indicate that people are more likely to migrate to locations in the same island and are less likely to migrate to distant places. Origin and destination population in each case has a positive effect on migration flows both for intra-country and inter-country migration. It is noteworthy that the largest impact of destination population on migration flows is on flows of international migrants to urban areas (W-U), as opposed to flows of urban to urban (U-U) or rural to urban (R-U) migrants. Thus, consistent with the gateway city findings of Smart et al. (2018), international migrants to New Zealand are attracted to the larger population centres. Migration from rural to urban areas (R-U) is negatively impacted by a destination’s QL and QB (albeit not significantly so for the latter). House prices (and rents) are low in rural areas compared with those in places with high quality of life and quality of business. The negative impact of destination QL and QB on rural flows is consistent with housing in these more attractive urban areas being out of reach of many potential migrants from rural locations. Consistent with this hypothesis, we see some evidence (at the 10% significance level) of urban residents from high QB areas (which are likely to have high house prices) migrating to rural locations. Urban to urban migration is again affected positively by both origin QL and by destination QL. Thus, there appears to be an interchange of domestic residents between urban areas that are valued highly as places to live. However, the same pattern does not extend to the impacts of QB. Here we see that destination QB is not an attractor for domestic urban (or rural) migrants, while high origin QB boosts migration to other urban areas. A one standard deviation increase in QB in an origin location increases migration out of that location to other urban areas by approximately 20%. Consistent with the results in Table 5.1, therefore, a city that boosts its attractiveness to business also reduces its retention of domestic residents. This pattern is consistent with places that have high QB having high house prices, so enabling or incentivising existing residents to move elsewhere. Migration responses are very different for international migrants to urban locations (i.e. W-U). While international migrants tend to move to locations that have high QL (column 1 of Table 5.2), a location that further improves its QL does not attract additional international migrants (i.e. column 2 which incorporates location fixed effects). By contrast, destination QB is a highly significant attractor (both with

112

A. Grimes et al.

Table 5.2 Gravity Model: Inter-Location Migration (zero and suppressed flows dropped) U-U: Destination QL (QLjt
1) R-U: Destination QL (QLjt
1) W-U: Destination QL (QLjt
1) U-U: Destination QB (QBjt
1) R-U: Destination QB (QBjt
1) W-U: Destination QB (QBjt
1) U-U: Origin QL (QLit
1) U-R: Origin QL (QLit
1) U-W: Origin QL (QLit
1) U-U: Origin QB (QBit
1) U-R: Origin QB (QBit
1) U-W: Origin QB (QBit
1) U-U: Destination population (lnPjt
1) R-U: Destination population (lnPjt
1) W-U: Destination population (lnPjt
1) U-U: Origin population (lnPjt
1) U-R: Origin population (lnPjt
1) U-W: Origin population (lnPjt
1) Same island dummy Distance if same island (ln) Distance if different island (ln) N R2 Destination and origin fixed effects

1 0.448*** (0.037) 0.119 (0.077) 0.362** (0.120) 0.138*** (0.033) 0.017 (0.053) 0.473*** (0.134) 0.161*** (0.037) 0.125** (0.046) 0.368 (0.189) 0.101** (0.034) 0.172*** (0.049) 0.571*** (0.155) 0.758*** (0.022) 0.664*** (0.034) 1.046*** (0.077) 0.735*** (0.022) 0.734*** (0.027) 1.013*** (0.098) 1.613*** (0.474)
0.737*** (0.031)
0.522*** (0.063) 4463 0.915 No

2 0.119*** (0.027)
0.259** (0.085)
0.016 (0.079) 0.019 (0.025)
0.122 (0.095) 0.334*** (0.083) 0.153*** (0.026) 0.035 (0.088) 0.263 (0.136) 0.183*** (0.025) 0.218* (0.097) 0.598*** (0.097) 0.432*** (0.080) 0.331*** (0.098) 0.714*** (0.085) 0.657*** (0.086) 0.647*** (0.107) 0.933*** (0.104) 1.694*** (0.441)
0.846*** (0.027)
0.631*** (0.061) 4463 0.943 Yes

The dependent variable in all models is the natural logarithm of the migration flow (plus one) between the destination and origin location, i.e. the population aged 30–59 in each destination urban (continued)

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

113

area that was usually resident in the origin city 5 years ago. An observation is an origin–destination pair. The sample includes 31 cities, plus “rural” (i.e. all New Zealand locations other than the 31 urban areas) and world (1054 destination–origin pairs) over six censuses from 1986 to 2013. Distance is the 2013 driving distance between the origin and destination cities in km. All models include year fixed effects, dummies for rural to urban, world to urban, urban to rural, and urban to world, and a constant not shown. Estimation is by OLS. Standard errors clustered by origin– destination location pair in parentheses; *p < 0.1, **p < 0.05, ***p < 0.01

and without fixed effects). A one standard deviation increase in QB in a destination location increases international migration into that location by approximately one-third. Notably, the attractiveness of QB holds even though we control for the impacts on migration flows of population in the destination location—i.e. after controlling for the gateway city phenomenon. As Smart et al. (2018) discuss, New Zealand’s immigration policies prioritise migrants with marketable skills, so it is likely that these people are attracted to places with a high quality of business (as well as to places with a large population). Our definition of international migrants includes both returning New Zealanders and migrants who are new to New Zealand. We check whether this aggregation may have influenced our results by splitting new migrants and returning New Zealanders in 2013 and observing whether they have different location patterns with respect to QL and QB. We form the ratio of new migrants to returning New Zealanders (RATIO) for each city and examine whether this ratio is correlated with each of QL and QB in 2013. If there is zero correlation, we can conclude that the relevant amenities act as similar drawcards to new migrants and returning New Zealanders, whereas if the correlation is positive [negative] we can conclude that the amenities act as a stronger [weaker] drawcard for new migrants relative to returning New Zealanders. The correlation of RATIO with QL is 0.17, which is not significantly different from zero at the 10% level. This result, combined with our prior finding that changes in QL are not a drawcard for international migrants in general, indicates that neither new migrants nor returning New Zealanders are strongly influenced in their location choice by consumption amenities. By contrast, the correlation coefficient of RATIO with QB is 0.46, which is significantly different from zero at the 1% level. This finding implies that new migrants are more attracted than returning New Zealanders to places with highly productive amenities, which tend to be the large cities. Another interpretation of this result is that returning New Zealanders, like domestic residents, are less likely to be attracted to highly productive places. If anything, therefore, our results in Table 5.2 understate the attraction of productive amenities for new international migrants. The attractiveness of places with high (and improving) quality of business to international migrants is consistent with the housing channel affecting domestic residents’ migration patterns. A number of studies find that immigration has a positive impact on house prices in New Zealand either at the aggregate level (Coleman and Landon-Lane 2007; McDonald 2013) or at the local level (Stillman

114

A. Grimes et al.

and Maré 2008).22 A location that improves its quality of business attracts international migrants which, consistent with these prior studies, raises house prices, and this both crowds out rural residents from migrating to these locations and encourages urban residents of these locations to move elsewhere; i.e. to move to locations with lower QB accompanied by lower housing costs. Thus we see a migration system in which changes to amenities that boost business productivity affect the migration patterns of both international and domestic migrants; international migrants increase their presence in high QB locations at the same time as domestic residents reduce their presence in these locations.

5.5

Conclusions

Local policy-makers and planners make decisions over local amenities that affect the attractiveness of their localities. They may choose to improve consumption amenities such as libraries, parks, cycle-ways, cultural facilities, skate parks, etc., and so improve the quality of life for residents in a particular place. Alternatively, or in addition, they may choose to improve productive amenities such as railyards, ports, and high speed Internet services that make a city a better place for firms to locate in and do business. Coleman and Grimes (2010) show that, provided the benefits outweigh the costs of such decisions, the locality will improve its attractiveness to migrants from other areas and this will place pressure on local property prices. We show that these aggregate effects represent only a portion of the full story. An improvement in a destination’s quality of life will attract domestic migrants while at the same time incentivising some existing residents to move elsewhere, so there is an exchange of domestic residents. There is no discernible effect of an improvement of QL on the city’s attractiveness for international migrants. An improvement in a locality’s quality of business, by contrast, attracts international migrants—especially new migrants—but not domestic migrants, while existing residents are again incentivised to shift elsewhere. A one standard deviation increase in a location’s quality of business is estimated to increase international migration into that location by approximately one-third, while raising domestic residents’ migration out of that location by approximately one-fifth. These patterns are similar to the localised patterns that we see in the literature on effects of gentrification (e.g. Van Criekingen 2009; Hochstenbach and van Gent 2015) in which gentrification pushes up local property values and encourages an exchange of residents to and from the area. A key difference with our analysis is that as well as the exchange of domestic residents when QL changes we see an exchange of international migrants for domestic residents when QB changes.

Stillman and Maré find that returning (New Zealand-born) migrants have a greater effect on house prices than do foreign-born migrants.

22

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

115

Our results indicate that city investments to improve amenities will not be neutral with regard to the demographic composition of a city. In particular, a drive to improve amenities that raise firms’ productivity is likely to change the composition of the city’s population towards a greater proportion of international migrants. Based on prior studies of the effects of migration on housing markets, the mechanism by which this demographic switch occurs is likely to be through a rise in house prices consequent on the rise in international migration to the city. The house price pressures enable existing homeowners (and incentivise existing renters) to leave the city in favour of places with lower housing costs. Thus policy-makers should be aware that productivity-oriented amenity investments make a city more attractive in aggregate but the disaggregated effects on location choice will differ between existing residents of the city, domestic residents elsewhere in the country, and potential international migrants. Acknowledgements We thank the MBIE-funded National Science Challenge 11: Building Better Homes, Towns and Cities for enabling this work. Disclaimer Access to the data used in this study was provided by Statistics New Zealand (SNZ) under conditions designed to give effect to the security and confidentiality provisions of the Statistics Act 1975. All frequency counts using Census data were subject to base three rounding in accordance with SNZ’s release policy for census data.

Appendix Table 5.3 Distribution of migration flows and usually resident population

Migration flow Mijt: All Mijt: U-U Mijt: U-R Mijt: U-W Mijt: R-U Mijt: W-U Populationit

% zero observations 7.6% 8.6% 0.0% 0.0% 0.0% 0.0%

% suppressed observations 21.9% 24.4% 0.0% 11.3% 0.0% 0.0%

Excluding zero and suppressed flows Mean Min Max flow flow flow 525 6 75,168 110 6 4185 1715 207 18,648 5061 12 66,639 1282 240 8394 2918 54 75,168 Mean Min Max 35,850 1494 532,437

116

A. Grimes et al.

Table 5.4 Quality of life (QL) by urban area and census year Urban area Ashburton Auckland Blenheim Christchurch Dunedin Feilding Gisborne Greymouth Hamilton Hawera Invercargill Kapiti Levin Masterton Napier-Hastings Nelson New Plymouth Oamaru Palmerston North Pukekohe Queenstown Rangiora Rotorua Taupo Tauranga Timaru Tokoroa Wanganui Wellington Whakatane Whangarei

Year 1986 0.032
0.163 0.941 0.185 0.041
0.130 0.305
0.439
0.055
0.707
0.079 0.446
0.668
0.237
0.046 0.803
0.039
0.180 0.143
0.222 1.454 0.426 0.047
0.378 0.353 0.055
1.335
0.111
0.584
0.045 0.185

1991
0.040
0.684 0.518
0.079 0.385
0.689 0.146 0.002
0.547
1.492
0.722 0.205
0.348
0.139
0.024 0.656
0.294
0.050
0.152
0.235 0.586 0.061
0.611
0.368
0.021
0.100
1.946
0.508
1.001
0.134
0.382

1996
0.365
0.793 0.707
0.117 0.442
0.352 0.456 0.010
0.495
0.920
0.764 0.725
0.302
0.266 0.517 0.732
0.509 0.345
0.503
0.293 0.298 0.244
0.394
0.411 0.197
0.340
2.159
0.415
1.449 0.242
0.201

2001
0.550
1.589 0.538
0.278 0.229
0.600
0.267 0.004
0.876
1.971
0.658 0.512
0.292 0.066 0.114 0.709
0.870 0.060
0.602
0.399 0.290
0.067
0.644
0.043
0.007
0.615
2.324
0.519
1.841
0.282
0.679

2006
0.563
1.594 0.423
0.321 0.196
0.404
0.588 0.084
0.901
1.323
0.318 0.600
0.173
0.162
0.058 0.503
0.657 0.244
0.733
0.747 0.133 0.197
0.730
0.300 0.037
0.475
2.147
0.512
1.803
0.196
0.852

2013
0.831
1.117 0.483
0.368 0.111
0.026 0.112
0.401
0.839
1.469
0.512 0.484 0.142 0.128 0.221 0.797
0.695
0.026
0.513
0.171 1.094 0.265
0.501
0.146 0.112
0.774
2.578
0.447
1.385 0.100
0.532

Quality of life (QL) values are standardised to have a mean of zero and a standard deviation of one across 130 New Zealand urban areas

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

117

Table 5.5 Quality of business (QB) by urban area and census year Urban area Ashburton Auckland Blenheim Christchurch Dunedin Feilding Gisborne Greymouth Hamilton Hawera Invercargill Kapiti Levin Masterton Napier-Hastings Nelson New Plymouth Oamaru Palmerston North Pukekohe Queenstown Rangiora Rotorua Taupo Tauranga Timaru Tokoroa Wanganui Wellington Whakatane Whangarei

Year 1986
0.033 2.551 0.140 1.147 0.722 0.684 0.341 0.225 1.283 0.453 0.795 0.561 0.955 0.269 1.007 0.532 1.398
0.236 1.281 0.643 1.168 0.128 1.469 1.573 0.978 0.468 2.025 0.842 2.627 0.917 1.425

1991
0.248 3.418 0.226 1.481 0.887 1.316 0.333 0.191 1.607 0.667 0.555 1.411 0.909 0.135 0.757 0.710 1.204
0.326 1.709 1.137 0.744 0.510 1.601 1.470 1.242 0.170 1.919 0.825 3.556 0.836 1.152

1996 0.260 3.180
0.226 1.417 0.566 0.777 0.308 0.308 1.518 0.795 0.689 0.478 0.392 0.087 0.338 0.708 1.214
0.589 1.508 1.406 2.042 0.337 1.348 1.228 1.123 0.286 1.759 0.533 3.160 0.423 0.975

2001 0.155 3.569
0.022 1.284 0.398 0.576 0.133
0.074 1.717 1.435 0.141 1.001 0.149
0.078 0.352 0.505 0.885
0.685 1.096 1.633 1.764 0.803 1.449 1.070 1.196 0.150 1.725
0.032 3.322 0.710 1.028

2006 0.610 3.324 0.333 1.344 0.552 0.265 0.281
0.198 1.609 0.640 0.202 0.446
0.300 0.067 0.513 0.622 0.868
0.810 0.895 1.900 2.357 0.745 1.131 1.152 1.089
0.005 0.847
0.295 2.832 0.600 1.070

2013 1.261 2.691 0.087 1.402 0.466 0.055
0.303 0.439 1.379 0.900 0.314 0.539
0.524
0.015 0.167 0.300 1.139
0.256 0.630 1.267 0.911 0.796 0.698 0.783 0.831 0.379 1.006
0.486 2.517 0.123 0.728

Quality of business (QB) values are standardised to have a mean of zero and a standard deviation of one across 130 New Zealand urban areas

0.847*** (0.018) 0.827*** (0.019) 2.188*** (0.446)
0.875*** (0.028)
0.569*** (0.061) 5580 0.858 No No

0.460*** (0.031) 0.154*** (0.027) 0.209*** (0.034) 0.121*** (0.030)

1

0.342*** (0.082) 0.887*** (0.091) 1.627*** (0.461)
0.964*** (0.028)
0.757*** (0.066) 5580 0.893 No Yes

0.162*** (0.031) 0.071* (0.028) 0.072* (0.031) 0.137*** (0.028)

2

0.126*** (0.029) 0.016 (0.026) 0.873*** (0.020) 0.853*** (0.021) 1.491** (0.468)
0.830*** (0.031)
0.625*** (0.065) 5580 0.839 No No

3

0.045* (0.023)
0.033 (0.020) 0.458*** (0.076) 1.003*** (0.087) 1.627*** (0.461)
0.964*** (0.028)
0.757*** (0.066) 5580 0.893 No Yes

4

0.849*** (0.021) 0.825*** (0.019) 2.189*** (0.448)
0.875*** (0.027)
0.570*** (0.062) 5580 0.861 Yes No

0.405*** (0.034) 0.141*** (0.031) 0.265*** (0.034) 0.135*** (0.031)

5

0.192* (0.087) 1.036*** (0.091) 1.627*** (0.461)
0.964*** (0.028)
0.757*** (0.066) 5580 0.894 Yes Yes

0.155*** (0.031) 0.058* (0.029) 0.078* (0.031) 0.150*** (0.028)

6

0.070** (0.027) 0.003 (0.025) 0.875*** (0.021) 0.850*** (0.019) 1.491** (0.470)
0.830*** (0.030)
0.625*** (0.065) 5580 0.842 Yes No

7

0.039 (0.023)
0.046* (0.021) 0.309*** (0.080) 1.153*** (0.087) 1.627*** (0.461)
0.964*** (0.028)
0.757*** (0.066) 5580 0.893 Yes Yes

8

The dependent variable in all models is the natural logarithm of the migration flow (plus one) between the destination and origin location, i.e. the population aged 30–59 in each destination urban area that was usually resident in the origin city 5 years ago. An observation is an origin–destination pair. The sample includes 31 cities (930 destination– origin pairs) over six censuses from 1986 to 2013. Distance is the 2013 driving distance between the origin and destination cities in km. All models include year fixed effects and a constant not shown. Added amenity variables include shares of employment in each of accommodation/food/recreation services, education, health, land transport, and air transport in Eqs. 5–8, plus rainfall, sunhours, wind strength, and proximity to the sea or a lake in Eqs. 5 and 7. Estimation is by OLS. Standard errors clustered by origin– destination location pair in parentheses; *p < 0.1, **p < 0.05, ***p < 0.01

N R2 Added amenity variables Destination and origin fixed effects

Distance if different island (ln)

Distance if same island (ln)

Same island dummy

Origin population (lnPit
1)

Destination population (lnPjt
1)

Difference QB (QBjt
1 - QBit
1)

Difference QL (QLjt
1 - QLit
1)

Origin QB (QBit
1)

Origin QL (QLit
1)

Destination QB (QBjt
1)

Destination QL (QLjt
1)

Table 5.6 Gravity model: Inter-urban migration (full sample) 118 A. Grimes et al.

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

119

Table 5.7 Gravity model: Inter-location migration (full sample) U-U: Destination QL (QLjt
1) R-U: Destination QL (QLjt
1) W-U: Destination QL (QLjt
1) U-U: Destination QB (QBjt
1) R-U: Destination QB (QBjt
1) W-U: Destination QB (QBjt
1) U-U: Origin QL (QLit
1) U-R: Origin QL (QLit
1) U-W: Origin QL (QLit
1) U-U: Origin QB (QBit
1) U-R: Origin QB (QBit
1) U-W: Origin QB (QBit
1) U-U: Destination population (lnPjt
1) R-U: Destination population (lnPjt
1) W-U: Destination population (lnPjt
1) U-U: Origin population (lnPjt
1) U-R: Origin population (lnPjt
1) U-W: Origin population (lnPjt
1) Same island dummy Distance if same island (ln) Distance if different island (ln) N R2 Destination and origin fixed effects

1: OLS 0.451*** (0.031) 0.096 (0.077) 0.339** (0.125) 0.148*** (0.027) 0.005 (0.055) 0.462** (0.142) 0.200*** (0.034) 0.103* (0.048) 0.270 (0.338) 0.115*** (0.030) 0.160** (0.054) 0.996** (0.303) 0.849*** (0.018) 0.671*** (0.036) 1.053*** (0.081) 0.829*** (0.019) 0.741*** (0.030) 1.046*** (0.171) 2.177*** (0.447)
0.874*** (0.028)
0.570*** (0.061) 6324 0.896 No

2: OLS 0.081* (0.038)
0.288** (0.090)
0.045 (0.093)
0.034 (0.039)
0.173 (0.097) 0.284** (0.095) 0.082* (0.034)
0.056 (0.095) 0.112 (0.266) 0.163*** (0.031) 0.190 (0.101) 1.026*** (0.249) 0.700*** (0.121) 0.520*** (0.132) 0.902*** (0.127) 0.661*** (0.091) 0.581*** (0.111) 0.886*** (0.149) 1.387** (0.453)
0.959*** (0.027)
0.786*** (0.065) 6324 0.918 Yes

3: Poisson 0.14 (0.17)
0.19 (0.18)
0.05 (0.15) 0.05 (0.07)
0.11 (0.09) 0.34*** (0.07) 0.22* (0.13) 0.15 (0.16) 0.29** (0.15) 0.12* (0.07) 0.12 (0.09) 0.27*** (0.05) 0.84*** (0.29) 0.59** (0.28) 0.98*** (0.28) 0.44** (0.21) 0.34 (0.24) 0.62*** (0.21) 0.97 (1.02)
0.77*** (0.09)
0.65*** (0.15) 6324 0.974 Yes

4: Neg Bin 0.13 (0.08)
0.32*** (0.12)
0.06 (0.09) 0.01 (0.08)
0.09 (0.11) 0.32*** (0.08) 0.14** (0.07) 0.02 (0.10) 0.19* (0.10) 0.20*** (0.07) 0.29*** (0.11) 0.74*** (0.11) 0.59** (0.25) 0.33 (0.26) 0.71*** (0.25) 0.83*** (0.26) 0.65** (0.26) 0.95*** (0.25) 1.42 (1.06)
1.02*** (0.05)
0.85*** (0.18) 6324 0.787 Yes

The dependent variable in the OLS models is the natural logarithm of the migration flow (plus one) between the destination and origin location, i.e. the population aged 30–59 in each destination urban (continued)

120

A. Grimes et al.

area that was usually resident in the origin city 5 years ago. The dependent variable in the Poisson and negative binomial (Neg Bin) models is the migration flow. An observation is an origin– destination pair. The sample includes 31 cities, plus “rural” (i.e. all New Zealand locations other than the 31 urban areas) and world (1054 destination–origin pairs) over six censuses from 1986 to 2013. Distance is the 2013 driving distance between the origin and destination cities in km. All models include year fixed effects, dummies for rural to urban, world to urban, urban to rural, and urban to world, and a constant not shown. Standard errors clustered by origin–destination location pair in parentheses; *p < 0.1, **p < 0.05, ***p < 0.01

References Ariu A (2018) Determinants and consequences of international migration. In: Biagi B, Faggian A, Rajbhandari I, Venhorst V (eds) New frontiers in interregional migration research, advances in spatial science. Springer, Cham, pp 49–60 Bartel A (1989) Where do the new U.S. immigrants live? J Labor Econ 7(4):371–391 Bauer T, Epstein G, Gang I (2007) The influence of stocks and flows on migrants’ location choices. Res Labor Econ 26(6):199–229 Biagi B, Dotzel K (2018) Theoretical advances on interregional migration models. In: Biagi B, Faggian A, Rajbhandari I, Venhorst V (eds) New frontiers in interregional migration research, advances in spatial science. Springer, Cham, pp 21–47 Chen Y, Rosenthal S (2008) Local amenities and life-cycle migration: Do people move for jobs or fun? J Urban Econ 64(3):519–537 Coleman A, Grimes A (2010) Betterment taxes, capital gains and benefit cost ratios. Econ Lett 109 (1):54–56 Coleman A, Landon-Lane J. 2007. Housing markets and migration in New Zealand, 1962–2006. Discussion Paper DP2007/12. Wellington: Reserve Bank of New Zealand Epstein G (2008) Herd and network effects in migration decision-making. J Ethn Migr Stud 34 (4):567–583 Gabriel S, Rosenthal S (2004) Quality of the business environment versus quality of life: Do firms and households like the same cities? Rev Econ Stat 86(1):438–444 Grimes A, Badenhorst S, Maré D, Poot J, Sin I (2020) Hometown and whānau, or big city and millennials? In: The economic geography of graduate destination choices in New Zealand. Motu Working Paper 20–04. Motu, Wellington Grimes A, Ormsby J, Preston K (2017) Wages, wellbeing and location: Slaving away in Sydney or cruising on the Gold Coast? Motu Working Paper 17–07. Motu, Wellington Harris J, Todaro M (1970) Migration, unemployment and development: a two-sector analysis. Am Econ Rev 60(1):126–142 Hochstenbach C, van Gent W (2015) An anatomy of gentrification processes: variegating causes of neighbourhood change. Environ Plan A 47:1480–1501 Lichter D, Johnson K (2006) Emerging rural settlement patterns and the geographic redistribution of America’s new immigrants. Rural Sociol 71(1):109–131 Maré D, Graham D (2013) Agglomeration elasticities and firm heterogeneity. J Urban Econ 75:44–56 Maré D, Morten M, Stillman S (2007) Settlement patterns and the geographic mobility of recent migrants to New Zealand. N Z Econ Pap 41:163–195 Maré D, Pinkerton R, Poot J (2016) Residential assimilation of immigrants: a cohort approach. Migration Studies 4(3):373–401 Maré D, Poot J (2019) Valuing birthplace diversity. Motu Working Paper (forthcoming). Motu Economic & Public Policy Research, Wellington McDonald C (2013) Migration and the housing market. Discussion Paper AN2013/10. Reserve Bank of New Zealand, Wellington

5 The Contrasting Importance of Quality of Life and Quality of Business for. . .

121

Munshi K (2003) Networks in the modern economy: Mexican migrants in the U.S. labor market. Q J Econ 118(2):549–599 Poot J, Alimi O, Cameron M, Maré D (2016) The gravity model of migration: the successful comeback of an ageing superstar in regional science. InvestigReg J Reg Res 36:63–86 Preston K, Maré D, Grimes A, Donovan S (2018) Amenities and the attractiveness of New Zealand cities. In: Motu Working Paper 18–14. Motu, Wellington Roback J (1982) Wages, rents, and the quality of life. J Polit Econ 90(6):1257–1278 Roback J (1988) Wages, rents, and amenities: differences among workers and regions. Econ Inq 26 (1):23–41 Rosen S (1979) Wage-based indexes of urban quality of life. In: Mieszkowsji P, Straszheim M (eds) Current issues in urban economics. Johns Hopkins University Press, Baltimore, pp 74–104 Sjaastad L (1962) The costs and returns to human migration. J Polit Econ 70:80–93 Smart C, Grimes A, Townsend W (2018) Ethnic and economic determinants of migrant location choice. In: Biagi B, Faggian A, Rajbhandari I, Venhorst V (eds) New frontiers in interregional migration research, advances in spatial science. Springer, Cham, pp 181–204 Stark O, Bloom D (1985) The new economics of labor migration. Am Econ Rev 75:173–178 Stillman S, Maré D (2008) Housing markets and migration: evidence from New Zealand. Department of Labour, Wellington Van Criekingen M (2009) Moving in/out of Brussels’ historical core in the early 2000s: migration and the effects of gentrification. Urban Stud 46(4):825–848 White P (1998) The settlement patterns of developed world migrants in London. Urban Stud 35 (10):1725–1744

Chapter 6

Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy Initiatives Malachy Buck and Peter Batey

Abstract The focus in this chapter is on the relationship between migration and neighborhood change. The chapter reviews the relationship between urban deprivation, residential mobility patterns, and urban regeneration policy, commenting on theoretical concepts and empirical findings in earlier studies. It makes extensive use of small area statistics from the UK Population Census, linking together migration data at the most detailed spatial level, Output Areas, with a geodemographic classification system. This provides some interesting insights about the structure of residential mobility in urban areas. The chapter proceeds to examine the impact of urban policy in the Merseyside region, as part of the EU Objective One programme. It uses a matched comparison method to examine the impact of the Pathways areabased initiative upon migration flows to and from Pathways Areas. Its key finding is that the Pathways programme had a significant impact upon residential stability.

6.1

Introduction

Throughout much of the last 40 years, poverty and disadvantage have intensified geographically in British cities. The response from Government has been to develop and implement urban policy aimed at regenerating those areas most badly affected. Much of the focus has been on neighborhoods within inner-city areas and on a range of place- and people-based policy initiatives. A growing body of applied research has been carried out aimed at assessing the impacts of these initiatives, in the hope of establishing “what works?” and therefore informing subsequent policy interventions (Lupton et al. 2016; Hughes and Lupton 2018). As part of this research, much attention has been paid to questions about the complex relationships between migration, urban deprivation, social mobility, the housing market, and neighborhood change. This work has been aided by the

M. Buck · P. Batey (*) University of Liverpool, Liverpool, UK e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 W. Cochrane et al. (eds.), Labor Markets, Migration, and Mobility, New Frontiers in Regional Science: Asian Perspectives 45, https://doi.org/10.1007/978-981-15-9275-1_6

123

124

M. Buck and P. Batey

increasing availability of small area statistics, notably from the decennial Population Census. The 2001 Census1 in particular opened up the possibility of linking migration data to area typologies defined by the use of census-based geodemographic classifications. The focus in this chapter is on exploring the detailed relationship between migration and neighborhood change, with particular reference to the region of Merseyside in north-west England which has been the subject of numerous urban policy interventions in response to long-standing severe social and economic problems. We explore the example of the Pathways area-based policy initiative. The chapter analyzes the migration flows into and out of different neighborhood types in the UK as a whole, identifying flows that are greater than, or less than, expected. From this, there emerges a clear picture of the structure of migration flows. Migration flows are examined between geodemographic area types. A matched comparison method is then used to compare Pathways and Pathwayslike areas sharing very similar social and economic characteristics to Pathways Areas but without the same designation and targeted resources. An assessment is made of whether this community development component of Merseyside’s Objective One programme, and the resources this brought, made a positive impact upon the population of the Merseyside region’s most deprived areas. It covers the period of the first phase of Objective One funding, from 1994 to 1999. Forming the cornerstone of regeneration programmes in the 1990s and early 2000s were area-based initiatives (ABIs) targeted at neighborhoods viewed as having the greatest needs. ABIs generally involved a combination of place-focused and people-focused initiatives, aimed at improving neighborhoods while at the same time enhancing the life chances of people living there. Cole et al. (2007, p. 5) identified a tension between these two objectives. On the one hand: Regenerating a neighborhood should make (an) area more attractive to existing residents. Fewer residents will want to leave the area and those who do will be replaced more rapidly. Community stability and cohesion will improve.

while on the other hand Improving life chances through education, health promotion, training, job mentoring, etc may help prospects and material circumstances of local residents. More may want, and be able, to leave the area. Out-movers may be replaced by more disadvantaged households. The neighborhood will become more deprived.

This process of neighborhood change was characterized as a moving escalator by Cole et al. (2007) in a comprehensive evaluation of one of the biggest regeneration programmes, the New Deal for Communities. In Fig. 6.1, for the purpose of the present study, four types of migration are shown. On the up-bound escalator, there are migrants from less affluent neighborhoods moving into regenerated neighborhoods (1) and migrants moving from

1 Sadly this proved to be a one-off. The migration flows from the 2011 Census are not available in the same level of detail, ostensibly for reasons of confidentiality.

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

125

Fig. 6.1 The moving escalator analogy

regenerated neighborhoods into more affluent neighborhoods (2). On the downbound escalator, there are migrants from more affluent neighborhoods moving into regenerated neighborhoods (3) and migrants moving from regenerated neighborhoods moving into less affluent neighborhoods (4). Not shown here is a fifth type of migrant whose (horizontal) move is within a regenerated neighborhood.

6.1.1

Research Aim and Questions

The moving escalator will serve as a framework for the present research, the aim of which is to develop and apply a method for assessing the degree to which regeneration activity has affected migration to, and from, deprived urban neighborhoods. In pursuit of this aim, four research questions will be addressed: • What evidence is there of a relationship between urban deprivation, migration, and regeneration policy? • What potential is there for combining geodemographic classifications with small area census migration data in order to evaluate the impact of urban policy? • How can this approach be applied to assess the effect regeneration funding has upon migration to and from particular targeted neighborhoods? • Does this evidence support the notion of a moving escalator in neighborhood renewal, involving the “export” of affluence and the “import” of poverty?

126

6.2 6.2.1

M. Buck and P. Batey

The Relationship Between Urban Deprivation, Migration, and Regeneration Policy Migration and Neighborhood Change

ABIs may be seen as an attempt by policymakers to alter the endogenous factors (e.g. employment opportunities and housing stock) which might then help bring about positive change (or at least arrest decline) in a deprived neighborhood. However, many factors interact to influence the trajectory of neighborhood change (Hincks 2017), one of which is migration, which may accelerate the polarization of the neighborhood (Delmelle 2015) or displace existing residents in a “gentrifying” neighborhood (Robson et al. 2008). Considerable effort has gone into understanding the effects of migration, particularly the effect upon deprived neighborhoods. Robson et al. (2008) explored the differences in deprivation of the origin and destination of migrants within deprived areas in England. This led them to create a four-fold typology of residential moves to and from deprived neighborhoods, based upon their function within a local housing market: • Improver/gentrifier: wealthier households moving into an area, with possible displacement effects; • Escalator: households moving from poorer areas up the housing ladder; • Transit: young households looking for cheaper starter housing and moving on to better areas; and • Isolate: poor households experiencing a degree of entrapment in deprived areas. Hughes and Lupton (2016) used the typology to understand the role that migration can play in “inclusive growth” within city-regions. Migration between Escalator and Improver areas was seen to drive change. In the first of these neighborhood types, out-migrants move to less deprived areas and are replaced by in-migrants from more deprived areas. In the second type, in-migrants tend to move from more affluent areas, while out-migrants move to more deprived areas. In these areas, there is the risk of displacement as housing costs rise (Hughes and Lupton 2016; Robson et al. 2008). Instead greater challenges are felt in Transit areas, where migrants mostly move from and move to areas which are similarly or less deprived. This means there is a limited period in which migrants can have an influence upon relative deprivation (Robson et al. 2008; Hughes and Lupton 2016). Isolate areas are a particular challenge for policymakers since migration flows in both directions are from similar or more deprived neighborhoods, which then limits the opportunity for change through migration. Hincks (2017) found that in such neighborhoods the population is likely to be trapped in poverty and especially vulnerable to macroeconomic trends, though the precise effect on neighborhood change depends on the demographic structure. Rae et al. (2016) extended the analysis of residential mobility to cover the connections with labor markets, finding that, despite the closeness of primary

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

127

employment areas to many Isolate neighborhoods, there was frequently only limited engagement with local labor markets.

6.2.2

Residential Mobility Patterns

The decision to move can be conceptualized in many ways. One theory is that households decide by comparing the utility of their current dwelling with that of other possibilities (Tu and Goldfinch 1996; Quigley and Weinberg 1977). Another, e.g. Marsh and Gibb (2011), holds that the sheer complexity involved in the decision to move leads to householders satisficing in their decision-making. The role of triggers, such as dissatisfaction with their current home or neighborhood, or lifecourse events, e.g. birth of a child or changing employment is also thought to be key (Clark et al. 1984; Rudel 1987; Rabe and Taylor 2010). Decisions may be enforced through an inability to meet housing costs (Preece et al. 2020) and the role of social and family ties may lead to a decision to move or stay put (Hickman 2010; Cole 2013; Lee et al. 1994). However, financial resources constrain nearly all households’ decision-making. This point is underlined by Kearns and Parkes (2003), who conclude that mobility patterns in deprived areas are not the result of a certain “cultural outlook,” but simply reflect limited financial resources to realize an aspired move. Over time these constraints have driven the process of residualization in many deprived areas of the UK (Rae et al. 2016), often accelerating the decline of deprived, undesirable areas, while driving improvement in more desirable areas (Van Ham et al. 2012). The ability to discriminate between neighborhoods, rather than being constrained to certain neighborhoods, is thought to be an important factor in enabling social mobility, given the influence of “neighborhood effects” (Lupton 2003; Van Ham et al. 2012). This is reflected in research in understanding the ability of households to do so. Clark et al. (2014) found the financial constraints on households created a “sorting effect,” where the greatest difficulty in “upward mobility” was found in the most deprived neighborhoods, with an increasing ability to realize a move as deprivation decreased. Even in the most deprived neighborhoods, education and homeownership both contributed positively to the ability to make an upward move, while the tenure within in a social-rented home had a negative effect. Bailey and Livingston (2008) found that migration patterns across England and Scotland were reinforcing existing patterns of segregation. For example, they found that more deprived areas experienced net out-migration of those with higher educational qualifications, while less deprived areas attracted those with such qualifications, although the net effect of this pattern was estimated to increase the proportion of those with a lower qualification in deprived areas by just 0.11%. They also found that, in total, 50% of migrants within deprived area moved between deprived and non-deprived neighborhoods, with flows in both directions, emphasizing that these areas were not as disconnected as previously suggested.

128

6.2.3

M. Buck and P. Batey

The Moving Escalator

We have seen earlier that, once the financial constraints are removed, individuals will realize a move out of a neighborhood with which they are dissatisfied. New Deal for Communities, as one of the biggest regeneration programmes in Britain, focused upon tackling a culture of worklessness (HM Treasury 2003). For those receiving employment training, this could present an opportunity to move to a less deprived area following a transition into labor markets, but with the risk of being “replaced” by those worse off, hence the concept of a moving escalator, as described earlier. It could mean the benefits of the scheme “leaking away,” leaving behind a concentration of those who are hardest to help (Bailey and Livingston 2008; Cole et al. 2007). The idea that “those who can, move out” (Social Exclusion Unit 2001) had previously received support from Cheshire et al. (1998) in the Harlesden City Challenge evaluation. This found that in-movers were more likely to be unemployed than out-movers, meaning by the end of the programme the area’s unemployment rate was higher than when the programme began (Cheshire et al. 2003). Cole et al. (2007) used 2002 Household Survey data to study the characteristics of those moving into and out of 39 New Deal Communities (NDC) in many British towns and cities. The studies found that those moving out of NDC areas were more likely to be in employment, with higher educational qualifications and seeking to enter the owner-occupied sector. Whereas those who choose to move into the same areas are likely to have a lower income, be unemployed and seeking a move into rented accommodation. CRESR (2005) identified these patterns as a key challenge to NDC success, given that the partnerships were dealing with increasingly deprived populations. When Cole et al. (2007) attempted to understand the factors which led to residents moving away from NDC areas, they found inconclusive evidence that people-based NDC interventions encouraged residents to move since out-movers and stayers participated in initiatives equally. Significantly they found no direct evidence that such programmes lead to employment, which Lawless and Pearson (2012) highlight in their critique of the “moving escalator” concept. While that it had been assumed that NDC was central to support residents into employment, in fact the focus of such programmes was upon those who had the greatest challenges in accessing labor markets, and there was little evidence that this moved significant numbers into employment (CLG 2009). Furthermore, Lawless and Pearson (2012) argued that instead the demographic structure of the area strongly determined mobility patterns, with 72% of the variance in mobility explained by the proportion of 16–34 year-olds. CLG (2009) data, cited by Lawless (2011) also indicates that migration into a NDC area was often driven by citizens from the 2004 EU Accession states, and they were typically less disadvantaged than existing residents. Using employment records from UK Department of Work and Pensions, Holden and Frankal (2012) explored these questions in the context of Greater Manchester, finding that out-migration of those gaining employment was not a key factor in persistent levels of worklessness. These findings were supported by Barnes et al.

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

129

(2011) at a national scale. The same authors (Holden and Frankal 2012) found instead that within deprived neighborhoods approximately a third of those gaining employment moved out of the neighborhood and, of those who did leave, only a slighter greater percentage moved to a neighborhood with lesser deprivation.

6.2.4

Geodemographic Classification Systems

Geodemographic classification systems have been employed by a number of researchers to understand migration patterns at a range of scales, for example, Duke-Williams’ (2010) findings broadly align with the view that migration is largely between similar areas. He used the Office of National Statistics OAC geodemographic system to find the greatest number of in-migrants moved from neighborhoods in the same geodemographic groups, linking this to strong spatial auto-correlation of geodemographic groups. Others have used migration, e.g. Dennett and Stillwell (2011) or travel to work data, e.g. Martin et al. (2018) to construct geodemographic systems. Another area where these systems have been used is the assessment of targeting of ABIs. Batey and Brown (2007) used the same geodemographic system as this study (People and Places) to assess the spatial targeting of ABIs. They concluded that this system had considerable value in identifying neighborhoods which were either wrongly targeted, or missed from targeting. In an effort to develop a methodology to tackle this issue, Batey et al. (2008) used a comparative approach to compare the value of the Index of Multiple Deprivation (IMD) and People and Places for the spatial targeting of interventions. They found that targeting of a specific group, e.g. the 10% most deprived neighborhoods, was highly dependent upon the approach taken. Instead, a hybrid system, which synthesizes IMD data at a finer spatial scale, led to a greater proportion of the target group living in areas with a spatially targeted intervention.

6.3

The Research Method

The research method used here brings together a geodemographic classification of residential neighborhoods and detailed migration flow data from the 2001 Population Census. In the geodemographic system neighborhoods throughout the UK are classified on the basis of 84 census variables measuring demographic, social, and economic characteristics using the multivariate statistical method of cluster analysis. The particular system is known as People and Places (P + P) and was developed jointly by researchers at the University of Liverpool and Beacon Dodsworth, a consulting company specializing in spatial data analysis. The classification makes use of Output Areas, the most detailed level of census geography. Some 220,000 Output Areas,

M. Buck and P. Batey

Out-migration: P+P Trees in Affluence Order

130

1 1 2 3 4 5 . . . . 13 Σ xi

In-migration: P+P Trees in Affluence Order 2 3 4 5 . . . .

13

Σ xj

xij

ΣΣ xij

Fig. 6.2 Studying the relationship between migration and affluence

each containing an average of 100 households (250 people), are classified. Three levels of neighborhood types are available: 156 Leaves, 40 Branches, and 13 Trees, all of which are capable of being ranked according to affluence. In the present study, the focus will be upon Trees and Branches, two of the three classification levels. The 2001 Population Census provides small area (Output Area) migration data capturing changes of address in the 12-month period 2000–2001 for the whole of the UK. Inter-Output Area migration flows enable the identification of the neighborhood type at both origin and destination. In the present study, this makes it possible to generate a 13
13 inter-residential type matrix that can be used to analyze key characteristics of migration. To study the relationship between migration and affluence, residential types are ranked in affluence order. This property is used in constructing a matrix of migration flows (Fig. 6.2): Each cell in this matrix represents the migration flow, xij, from one neighborhood type (i) to another ( j). In practice the matrix would have enough rows and columns to accommodate all 13 P + P (People and Places) Trees. Instead of presenting absolute migration flows, the matrix will be modified to show standardized flows. The standardization is based on a comparison of observed (O) and expected (E) flows:
Oij  E ij =E ij , where Eij is the flow expected between area types i and j, proportionate to the total size of the flows originating from, and terminating at, the relevant neighborhood types. Here:
E ij ¼ Σ xi  Σ x j =ΣΣ xij

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

131

Fig. 6.3 Standardized geodemographic migration matrix for P + P Tree neighborhood types

The result of standardization is to create a series of positive values when the observed flow is greater than the expected flow, and negative values when the observed flow falls short of that expected. It is therefore possible to compare one area type with another in order to establish differences in the patterns of inward and outward flows. Without standardization, these differences could largely be attributed to variations in the “trip ends”: the flow between two relatively small area types would almost inevitably be less than that between two larger area types. The standardized matrix provides valuable information about the extent to which migration takes place up and down the affluence “ladder.” Entries above the main diagonal show flows where the movement is upward, and entries below the diagonal show flows where the movement is downward. Fig. 6.3 shows the standardized migration matrix for the 13 neighborhood types at the P + P Tree level, labelled A to M, in affluence order where A is highest. Where a standardized flow is +0.25 or more, it is shaded red, and if it is 0.25 or less, it is shaded blue. Clear patterns emerge: least affluent neighborhood types interact with other least affluent neighborhood types in the bottom right-hand quadrant, while most affluent neighborhood types show a similar tendency to interact with other most affluent neighborhood types in the top left-hand quadrant. There is also an understandable tendency for diagonal entries to display high positive scores, a reflection of a high level of moves within a neighborhood type. Figure 6.4 provides a more illuminating way of presenting these migration flows. Here the 13 neighborhood types (“Trees”) have been re-arranged in groupings depending on the same scores and thresholds adopted in Fig. 6.3. Three main features are apparent here: (i) an underlying pattern of migration to more affluent neighborhood types, as shown by the strong central vertical axis with an upward trajectory; (ii) three migration sub-systems: affluent (A, B, C, D, and F); deprived (J, K, L, and M) and metropolitan (E and I); and an outlier (H) representing new starters in the housing market2. Overall, it can be seen that this pattern broadly supports the notion of a moving escalator.

2

Neighborhood types are shown by a letter and a label. Labels of this kind are common in applications of geodemographics and provide, with varying degrees of success, a succinct descriptor.

132

M. Buck and P. Batey

Fig. 6.4 Standardized geodemographic migration diagram for P + P Tree neighborhood types

6.4

Application: EU Merseyside Objective 1 Pathways Areas

Combining geodemographics and migration data potentially enables us to answer questions like: • Are residents in targeted neighborhoods more or less likely to move than their counterparts elsewhere? • Do targeted neighborhoods export population to more affluent areas and import population from less affluent areas?

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

133

First, however, we need to extend the research method and to focus ideas it will be helpful to examine a specific example of an area-based initiative: Merseyside Objective One Pathways. In 1993 the lagging city region of Merseyside in North West England was designated an Objective One region, making it eligible for the highest level of European Union (EU) Structural Funds. This followed decades of industrial decline, mounting social problems and widespread urban deprivation. The first “post-industrial” Objective One region, Merseyside submitted a plan for approval by the European Commission. The plan, known as a Single Programming Document, was intended to guide the implementation of a comprehensive programme of economic development funded through the European Regional Development Fund (ERDF) and community development supported by the European Social Fund (ESF) (Evans 2002). Although initially the programme extended over a six-year period (1994–1999), there was the prospect of further tranches of EU funding as long as Merseyside retained its Objective One status. The programme was organized as five so-called drivers, one of which is relevant here: Action for the People of Merseyside. This driver identified some 38 deprived neighborhoods3—Pathways Areas—in Merseyside and accounting for 35% of the region’s population. These areas would receive extra support through a package of people- and place-based regeneration measures including lifetime training in growth sectors; equal opportunities measures; improving access to jobs and training for those with special needs; improved education, training and employment services; improving access to work via public transport; and treating derelict, contaminated and neglected land (Evans 2002). Figure 6.5 shows the Pathways Areas and how these were distributed across the five local authority areas making up the region. Pathways Areas continued throughout the first period of Objective One support and also the second period, 2000–2006. We are interested in movement into, and from, these Pathways Areas as well as to and from Pathways-like Areas, which, although sharing very similar social and economic characteristics, have not been targeted as Pathways Areas. To do this we follow a 6-stage matched comparison method: 1. Identify within Merseyside the Output Areas that together make up Pathways Areas. 2. Create a new “notional” cluster by aggregating these Output Areas. Locate the notional cluster’s centroid in n-dimensional space using the component loadings for the first n principal components calculated as a preliminary to the cluster analysis that created People and Places. 3. Calculate the distance in n-dimensional space of all Output Areas from this centroid. Starting with the Output Areas furthest away from this centroid, progressively eliminate Output Areas in order to develop a tighter cluster with fewer 3

Although all Pathways Areas can be considered to be deprived, there was some inconsistency in their definition. This was primarily for political reasons to ensure support from all five of the local authorities in the Merseyside region.

134

M. Buck and P. Batey

Fig. 6.5 Pathways areas in Merseyside

outlying Output Areas. Every so often in this process, re-compute the notional cluster centroid. Continue re-computing the centroid until its location in n-dimensional space stabilizes. 4. Extract a sample of Output Areas that satisfy this selection criterion and are located outside Pathways Areas (Pathways-like Areas). This sample is to be drawn from Output Areas that make up the group of local authorities that together contribute m% of the migration to and from Merseyside. 5. Extract a second sample of Output Areas (from the same set of local authorities) that fail to satisfy the selection criterion (other areas—distinguishing between less and more affluent areas). 6. Analyze the migration flows within and between these four types of areas: Pathways Areas; Pathways-like Areas; less affluent areas; and more affluent areas. The results are shown in a sequence of figures. Figure 6.6 presents the migration flows for the working population of Merseyside in an un-standardized form. It underlines the importance of residential mobility within each of the four categories. These flows are shown in yellow. Some 60% of all flows are of this intra-group type. Figure 6.7 defines the five basic migration rates as introduced in Fig. 6.1. It also defines the gross turnover rate, a measure of total migration activity and therefore a means of comparing residential stability with and without the impact of the ABI, in this case Pathways.

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

Fig. 6.6 Migration flows in the Merseyside region 2000–2001: working age population

1. 2. 3. 4. 5. 6.

Upwardly-mobile in-migration rate: 1/x Upwardly-mobile out-migration rate: 2/x Downwardly-mobile in-migration rate: 3/x Downwardly-mobile out-migration rate: 4/x Horizontally-mobile migration rate: 5/x Gross turnover rate:(1+2+3+4+5)/x where x is the population at in 2001 and all rates are expressed as per 10,000 population

Fig. 6.7 Defining migration rates: see also Fig. 6.1

Fig. 6.8 Moving escalator for Pathways Areas

135

136

M. Buck and P. Batey

Fig. 6.9 Moving escalator for Pathways-like Areas

Fig. 6.10 Moving escalator showing the difference between Pathways and Pathways-like areas

Figures 6.8 and 6.9 present the moving escalator for Pathways and Pathways Areas, respectively, in the same format as Fig. 6.1. Figure 6.10 shows the difference between Figs. 6.8 and 6.9 and therefore provides a concise measure of the impact of the Pathways ABI. Finally, Fig. 6.11 summarizes the results, enabling Pathways Areas migration rates to be compared with those for Pathways-like Areas. Figure 6.6 gives a clear indication of the relative importance of residential mobility within the four categories of neighborhood. In fact, some 60% of flows are of this kind. They are shown in yellow. Figure 6.8 shows that, for Pathways Areas, out-migration and in-migration are almost exactly the same, that those moving out are much more likely to move to more affluent neighborhoods than less affluent ones, and that those moving into Pathways Areas are more likely to come from more affluent neighborhoods than less affluent ones. Figure 6.9 shows that for Pathways-like Areas there is a greater propensity to move out than in, and that those moving in are much more likely to be from more affluent neighborhoods than from less affluent ones. Figure 6.10, which measures the differences between the migration rates for Pathways and Pathways-like Areas, shows very clearly the impact of the Pathways ABI, suggesting that the extra

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

137

Fig. 6.11 Pathways migration: highlighting the role of other neighborhood types

resources that come with Pathways status are slowing down the escalator, stemming the “leakage” of population. In addressing the question about importing poverty and exporting affluence, we find sharply contrasting evidence for Pathways and Pathways-like neighborhoods. Figure 6.8 shows very little net exporting of affluence (394–375) ¼ +19 and of net importing of poverty (36–28) ¼ +8, while Fig. 6.9 shows the equivalent information for Pathways-like neighborhoods: (560–459) ¼ +101 and (145–150) ¼ 5. Figure 6.11 shows migration flows to and from Pathways and Pathways-like Areas using four neighborhood types differentiated according to affluence level. These four types are based on aggregated blocks of 40, affluence-ranked, People and Places Branches (1–20, 21–30, 31–36, 37–40). Figure 6.11 shows that two-thirds of migration is found to take place within Pathways Areas and/or Pathways-like Areas. The profile of neighborhoods contributing migrants to Pathways and Pathways-like Areas is very similar. People migrating from Pathways Areas are more likely to move to neighborhoods that are slightly more affluent than Pathways Areas. Those migrating from Pathways-like Areas move to a more diverse range of neighborhoods. In Fig. 6.12 it can be seen that in Pathways Areas, the majority of migration (69%) occurs within and between the two least affluent groupings of neighborhood types. Almost half of migrants (48%) in Pathways Areas remain within the same neighborhood affluence category when they move. For those who do shift affluence category, Fig. 6.13 shows that there is a clear pattern of upward movement, from the less affluent Pathways neighborhoods to Pathways Areas that are more affluent; a much less pronounced pattern is found in Pathways-like Areas. Finally, and arguably most importantly, we see from Fig. 6.14 further strong confirmation of the greater community stability and cohesion attributable to the

138

M. Buck and P. Batey

Fig. 6.12 Gross migration within Pathways Areas

Fig. 6.13 Net migration within Pathways and Pathways-like Areas

Objective One Pathways as measured by the gross turnover rate: 2381 for Pathways compared with 3431 for Pathways-like Areas.

6.5

Conclusions

This chapter has argued that the study of changing migration patterns is a crucial step in evaluating the effectiveness of any area-based regeneration policy. It has shown how, by combining small area census data on migration with a geodemographic classification of residential neighborhoods, much can be learnt about the structure of migration patterns. It has explored the concept of a moving escalator as a way of characterizing residential mobility to and from deprived neighborhoods. In

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

139

Fig. 6.14 Migration rates per 10,000 population compared

particular, it has demonstrated the value of this concept in making detailed comparisons between neighborhoods subject to different policy interventions. A new method, based on a matched comparison approach, was used to assess the impact of a specific area-based regeneration initiative forming part of the EU Objective One programme for Merseyside. It showed quite clearly that the initiative had a positive impact upon residential stability in the targeted neighborhoods. The study as presented here had the advantage that migration data was available at the most detailed spatial level, the Output Area level. This was particularly helpful in linking with the geodemographic classification system. However, the same degree of detail is not available in earlier censuses before or since the 2001 Census. A further consideration is the limited amount of information provided about the migration flows. Other than origin, destination and broad age categories, there is no other information that tells us who is doing the migrating. Because there can be considerable heterogeneity within Output Areas, and certain groups may have a greater propensity to migrate, limits the conclusions that can be drawn about the effects of migration. Acknowledgments The authors wish to acknowledge the contribution of Peter Brown, Simon Whalley and Simon Pemberton to the early stages of the project reported here.

References Bailey N, Livingston M (2008) Selective migration and neighbourhood deprivation: evidence from 2001 census migration data for England and Scotland. Urban Stud 45(4):943–961 Barnes H, Garratt E, McLennan D, Noble M (2011) Understanding the worklessness dynamics and characteristics of deprived areas. Department for Work and Pensions, London

140

M. Buck and P. Batey

Batey P, Brown P (2007) The spatial targeting of urban policy initiatives: a geodemographic assessment tool. Environ Plan A 39(11):2774–2793 Batey P, Brown P, Pemberton S (2008) Methods for the spatial targeting of urban policy in the UK: a comparative analysis. Appl Spat Anal Policy 1(2):117–132 Cheshire P, Flynn N, Jones DA (1998) Harlesden City challenge: final evaluation. LSE, London Cheshire P, Monastiriotis V, Sheppard S (2003) Income inequality and residential segregation: labor market sorting and the demand for positional goods. In: Martin R, Morrison PS (eds) Geographies of labor market inequality. Routledge, London, pp 91–117 Clark WA, Deurloo MC, Dieleman FM (1984) Housing consumption and residential mobility. Ann Assoc Am Geogr 74(1):29–43 Clark WA, Van Ham M, Coulter R (2014) Spatial mobility and social outcomes. J Housing Built Environ 29(4):699–727 Cole I (2013) Whose place? Whose history? Contrasting narratives and experiences of neighborhood change and housing renewal. Hous Theory Soc 30(1):65–83 Cole I, Lawless P, Manning J, Wilson I (2007) The moving escalator? Patterns of residential mobility in the New Deal for Communities areas: research report 32. DCLG, London Communities and Local Government (CLG) (2009) Understanding and tackling Worklessness volume 2: neighborhood-level problems, interventions and outcomes. CLG, London CRESR (2005) Research report 17 New Deal for Communities 2001–2005: an interim evaluation. ODPM, London Delmelle EC (2015) Five decades of neighborhood classifications and their transitions: a comparison of four US cities, 1970–2010. Appl Geogr 57:1–11 Dennett A, Stillwell J (2011) A new area classification for understanding internal migration in Britain. Popul Trends 145(1):146–171 Duke-Williams O (2010) Mapping the geodemographic classifications of migrants’ origins and destinations. J Maps 6(1):360–369 Evans R (2002) The Merseyside objective one Programme: exemplar of coherent city-regional planning and governance or cautionary tale? Eur Plan Stud 10(4):496–517 Hickman P (2010) Understanding residential mobility and immobility in challenging neighborhoods. Centre for Regional Economic and Social Research, Sheffield Hincks S (2017) Deprived neighborhoods in transition: divergent pathways of change in the Greater Manchester city-region. Urban Stud 54(4):1038–1061 HM Treasury (2003) Full employment in every region. HMSO, London Holden J, Frankal B (2012) A new perspective on the success of public sector worklessness interventions in the UK’s most deprived areas. Local Econ 27(5–6):610–619 Hughes C, Lupton R (2016) Residential and labor market connection of deprived neighborhood in Greater Manchester and Leeds City Region. Joseph Rowntree Foundation, York Hughes C, Lupton R (2018) Understanding changes in Greater Manchester’s deprived neighborhood 2004–2015 using a typology of residential mobility. Joseph Rowntree Foundation, York Kearns A, Parkes A (2003) Living in and leaving poor neighborhood conditions in England. Hous Stud 18(6):827–851 Lawless P (2011) Understanding the scale and nature of outcome change in area-regeneration programmes: evidence from the new Deal for communities Programme in England. Environ Plan C Govern Policy 29(3):520–532 Lawless P, Pearson S (2012) Outcomes from community engagement in urban regeneration: evidence from England’s new deal for communities programme. Plan Theory Pract 13 (4):509–527 Lee B, Oropesa RS, Kana JW (1994) Neighborhood context and mobility. Demography 31:249–270 Lupton R (2003) Poverty street: the dynamics of neighborhood decline and renewal. The Policy Press, Bristol Lupton R, Rafferty A, Hughes C (2016) Inclusive growth: opportunities and challenges for Greater Manchester. Joseph Rowntree Foundation, York

6 Migration, Neighborhood Change, and the Impact of Area-Based Urban Policy. . .

141

Marsh A, Gibb K (2011) Uncertainty, expectations and behavioral aspects of housing market choices. Hous Theory Soc 28(3):215–235 Martin D, Gale C, Cockings S, Harfoot A (2018) Origin-destination geodemographics for analysis of travel to work flows. Comput Environ Urban Syst 67:68–79 Preece J, Crawford J, McKee K, Flint J, Robinson D (2020) Understanding changing housing aspirations: a review of the evidence. Hous Stud 35(1):87–106 Quigley JM, Weinberg DH (1977) Intra-urban residential mobility: a review and synthesis. Int Reg Sci Rev 2(1):41–66 Rabe B, Taylor M (2010) Residential mobility, quality of neighbourhood and life course events. J R Stat Soc A Stat Soc 173(3):531–555 Rae A, Hamilton R, Crisp R, Powell R (2016) Overcoming deprivation and disconnection in UK cities. Joseph Rowntree Foundation, York Robson B, Lymperopoulou K, Rae A (2008) People on the move: exploring the functional roles of deprived neighborhoods. Environ Plan A 40(11):2693–2714 Rudel TK (1987) Housing price inflation, family growth, and the move from rented to owner occupied housing. Urban Stud 24(4):258–267 Social Exclusion Unit (SEU) (2001) A new commitment to neighborhood renewal: national strategy action plan. Social Exclusion Unit, London Tu Y, Goldfinch J (1996) A two-stage housing choice forecasting model. Urban Stud 33 (3):517–537 Van Ham M, Manley D, Bailey N, Simpson L, Maclennan D (2012) Neighborhood effects research: new perspectives. In: Van Ham M, Manley D, Bailey N, Simpson L, Maclennan D (eds) Neighborhood effects research: new perspectives. Springer, Dordrecht, pp 1–21

Chapter 7

Outmigration and Remittances as Facilitating Conditions for Economic Transition in Romania Cristian Incaltarau, Liviu George Maha, Karima Kourtit, and Peter Nijkamp

Abstract Cross-border migration is a phenomenon of all times, because people seek to improve their economic fortune—and that of their relatives or community— through labour mobility. There is an extant literature on the backgrounds and rationality of migration decisions. Research has extensively addressed both the specific and the wider impacts of a migration influx into a destination or host country. The impacts of outmigration and remittances on the source country have received less attention. The present paper examines whether outmigration from (national or regional) labour markets—and the related remittances—have exerted an impact on the Romanian political-economic transition in the post-communist period, especially after the country’s access to the EU. Thus, the system-wide migration impacts transmitted through both remittances and labour market mechanisms are analysed for Romania, using a long-term time series (1991–2015), while accounting for endogeneity of migration and remittances. The post-communist period is split up into three characteristic stages, while the difference in the magnitude of these effects for each stage is tested. Our findings show that migration has exerted an overall net positive effect on economic growth, except for the second stage when the Romanian economy moved from a shortage of jobs to a shortage of labour. The positive effects were due to a reduction in the social security burden, given the prevailing high unemployment rate. Remittances from migration were also

C. Incaltarau (*) Centre for European Studies, Alexandru Ioan Cuza University of Iasi, Iasi, Romania e-mail: [email protected] L. G. Maha Faculty of Economics and Business Administration, Alexandru Ioan Cuza University of Iasi, Iasi, Romania e-mail: [email protected] K. Kourtit · P. Nijkamp Centre for European Studies, Alexandru Ioan Cuza University of Iasi, Iasi, Romania Open University, Heerlen, The Netherlands © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 W. Cochrane et al. (eds.), Labor Markets, Migration, and Mobility, New Frontiers in Regional Science: Asian Perspectives 45, https://doi.org/10.1007/978-981-15-9275-1_7

143

144

C. Incaltarau et al.

shown to enhance growth, with a stronger effect during the second stage, which may have offset the negative effects of outmigration. The lesson for Romanian policymaking is that migration is not sufficient for economic development, as home country conditions need to continuously improve for unleashing the full development potential of outmigration. Keywords Migration · Remittances · Economic growth · Labour market · Postcommunist transition · Romania

7.1

Aims and Scope

We live nowadays in the ‘age of migration’: a significant share of the world population is ‘on the move’. Clearly, migration is a common phenomenon in the history of mankind, but nowadays the cross-border migration flows are rapidly rising, both absolutely and relatively. Migration is the result of attraction and repulsion forces of all kind (e.g. economic, social, environmental, political) and is therefore difficult to manage and control in an open and globally interconnected system of nations. The complex nature of migration flows and their drivers have prompted many heated debates in both academic and policy circles. In recent years, we have witnessed a rising interest in the empirical assessment of the manifold consequences of heterogeneous migration, in both the destination countries (or regions) and the countries (or regions) of origin. This scientific approach is called ‘migration impact assessment’ (MIA), an exercise in which Jacques Poot has played a critical role (see Nijkamp et al. 2012). MIA is defined as ‘the integrated application of scientific tools to trace the broad socio-economic impacts of crossborder migration and related policies’ (p. 4). In subsequent studies (see Nijkamp et al. 2015) various dedicated applied investigations into the diversity effects of migrants are mapped out, while in a recent publication the combined migrationageing impacts on spatial labour markets are addressed (see Stough et al. 2018). Comprehensive overviews on the economics and economic geography of migration can be found in Chiswick and Miller (2015) and Kourtit et al. (2020). There is no doubt that cross-border migration still is—and probably will remain—a controversial issue because of the many (positive and negative) effects it generates at various levels among both sending and receiving communities. The general opinion on the consequences of migration for the emigrants themselves—but also on the effects induced in the origin and destination countries or regions—have swung like a ‘pendulum’ between optimism and pessimism over the twentieth century (de Haas 2012). The negative attitude towards migration started regaining ground along with the recent economic crisis (Gamlen 2014), claiming that optimism was inflated just as the credit bubble, and that neither migration nor remittances can substitute development. Overall, both optimistic and pessimistic perspectives may actually be simultaneously right, as claimed by the ‘pluralist’ views on migration which take for granted

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

145

both benefits and negative effects induced by migration. As migration impacts the source country’s economy through various channels, the overall net effect may be either positive or negative, depending on the home country’s existing conditions, the amplitude of migration flows, but also the stage of the migration cycle. In the present study, we will address in particular the effects on sending countries or regions. On the one hand, according to neoclassical theory, migration leads to growth through a quantitative relocation of workers, which induces changes in the capital–labour ratio. Thus, migration flows are decreasing the stock of labour relative to that of capital in usually less developed home countries, leading to an increase in income per capita (Barro and Sala-i-Martin 2004). Even though neoclassical models did not often refer to remittances, these flows may have an important role for capital accumulation in source countries (Barajas et al. 2009). Furthermore, the increase of capital accumulation may be also indirectly supported by remittances, through a decreased cost of capital investment due to raising the credit ranking of investors, thus inducing additional new investments. Also, remittances may support macroeconomic stability, which makes investments more attractive. One the other hand, beside the quantitative effect described by neoclassical thinking, the composition effect, given by the skill structure of migrant flows, may act in an opposite way, by undermining growth in home countries (Shioji 2001). If migrants are more productive than stayers—which is a plausible assumption given that migration has been shown to be a selective process especially during early stages—migration may be detrimental to growth in origin areas (Fratesi and Riggi 2007). Beside the negative effect induced by the composition effect in origin countries, the negative effect of migration may be strengthened by remittances, as these inflows may reduce labour force participation (Acosta 2007; Jawaid and Raza 2016; Chami et al. 2003; Naiditch and Vranceanu 2009). Overall, the current literature describes diverse mechanisms acting behind migration and remittances, showing how these can act in conflicting ways, while the net effect highly depends on their relative magnitude. Given the different mechanisms migration enacts and the different opinions gravitating around this topic, our study investigates the Romanian case and estimates the impact of migration, transmitted through both labour market and remittances channels, by using long time-series data from 1991 to 2015. The Romanian case indeed represents a fertile ground for analysing migration impacts (as advocated by Nijkamp et al. 2012), as it faced substantial emigration flows during the postcommunist period. The changes driven by the fall of communism at the end of 1989 forced population to look for better opportunities abroad; almost half of the Romanian households have at least one member working abroad (Dincu et al. 2015). The Romanian National Institute for Statistics (RNIS) (2014) estimated that the number of outflows during 1989–2012 exceeds 10% of population (over 2.4 million people). The World Bank (WB) estimates are even higher, showing that in 2013 the stock of emigrants was reaching over 17% of population (3.4 million emigrants) (WB 2016). In addition, the high emigration rates also triggered large remittance inflows, making the Romanian case a relevant one for analysing the economic impact of migration driven through remittances channels. According to both the

146

C. Incaltarau et al.

National Bank of Romania (NBR) and the WB, these transfers reached at least 5% of GDP before the recent economic crisis. And finally, the road from the communist regime to EU accession has driven major structural changes in the Romanian economy, with important effects on the labour market. The labour market went from a ‘stagnant pool’ of unemployment during the 1990s (Earle and Pauna 1998) to a labour shortage before the recent economic crisis (Cindrea 2007; Voicu et al. 2008), which makes the overall net impact of migration ambiguous, while leaving room for further research. Therefore, our study aims to estimate the impact of outmigration in Romania during the post-communist period (1991–2015) by specifically addressing the following effects relating to migration: (1) Did migration influence economic growth in Romania? (2) Were remittances a relevant migration channel for fostering economic growth? (3) How did migration and remittances affect the labour market? (4) Did the intensity of these effects differ by migration stage? Our study introduces several amendments in the already existing literature. First, although there are several studies using large-scale surveys for analysing the impact of migration at microeconomic level (Sandu 2006b; Dincu et al. 2015; Stoiciu et al. 2011), to our knowledge there is no other study estimating the growth effect of migration at macroeconomic level by using Romanian data. There are only a few studies estimating the effects transmitted through remittances on consumption, investments and trade (Litan 2009; Blouchoutzi and Nikas 2010; Chirila and Chirila 2017; Incaltarau and Maha 2012), but—given they do not directly focus on growth and that there are multi-faceted mechanisms acting in conflicting directions—their relevance for the growth hypothesis is limited. Second, our study brings the current insights further, by separately estimating the emigration impact transmitted through remittances, as well as through labour market mechanisms. Therefore, after directly estimating the growth effect of outmigration, our study also addresses its impact on unemployment (as a proxy for labour market frictions). Additionally, while the direct impact of remittances on growth is estimated, our study also estimates the net remittances’ impact on unemployment, in order to analyse for the moral hazard scenario which may discourage labour force participation, against the production– multiplier effect, which may raise labour demand and thus reduce unemployment. Third, our study distinguishes three distinct post-communist stages, by taking into account the interaction between emigration, remittances and labour market, and attempts to estimate the differences in magnitude of these effects by stage. And fourthly, our study is not only using long time-series data from 1991 to 2015, but also the estimation method accounts for the possible endogeneity of outmigration and remittances, by using an instrumental estimation technique. The remainder of the paper is structured as follows. Section 7.2 presents a brief literature review on the effects induced by outmigration through remittances and the labour market in the origin country, while Sect. 7.3 describes three distinct stages in terms of migration, remittances and labour market outcomes. Next, Sect. 7.4 presents the methodology and data used for estimating the impact of migration on the Romanian economy. Section 7.5 points out the estimation results and discusses the main findings according to our hypothesis, while Sect. 7.6 presents our conclusions.

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

7.2

147

Economic Impacts of Migration on the Source Country: An Empirical Review

International migration of labour implies the movement of production factors (labour and, maybe, capital owned by emigrants) from one country to another. In this way, the relative quantities of factors from the two economies change, and, therefore, the production levels based on using these inputs and the wage levels may change as well. On the one hand, if migration increases productivity in the origin country, by raising capital intensity relative to labour, this may lead to an increase in per capita income, as claimed by neoclassical theory (Barro and Sala-i-Martin 2004). On the other hand, migration may also undermine productivity. First, as human capital accumulation is one of the main economic growth drivers (Lucas 1988), if migrants are more educated and more productive than the workers left behind, this may cause a loss in productivity. Many of studies have evidenced that the negative composition effect given by the loss in human capital may dominate the positive quantity effect driven by the capital–labour ratio increase (Shioji 2001; Fratesi and Riggi 2007; Kubis and Schneider 2016). But, despite that migration may drain origin areas of highly skilled population through the composition effect, migration may also enact an opposite ‘wage premium’ effect, which encourages brain gain. As the experience gained abroad ensures a ‘wage premium’ for emigrants when they return to origin countries, this can further encourage human capital accumulation and thus lead to an overall increase in human development level, because not all of them will actually emigrate and, even so, some of them are going to return home (Ambrosini et al. 2015). Second, a negative effect on productivity may be also caused by the upward pressure on domestic wages as a result of migration, which generates a workforce reduction (Mishra 2015; Elsner 2013). This upward pressure may be tempered by a concomitant reduction in labour demand, given that migrants are also moving their consumption in destination countries. Moreover, the higher incomes earned in the destination country allows them to send remittances to family members left behind, offsetting the initial decrease in demand for goods and services in the origin country caused by their departure. Various studies at both microeconomic and macroeconomic levels evidenced that remittances can indeed stimulate consumption, being used for the basic needs of the population in origin areas (Duval and Wolff 2010; Sandu 2006b; Blouchoutzi and Nikas 2010; Grigoras 2006). But even if remittances are mainly directed to consumption, economic growth can still be indirectly stimulated by the increased demand for goods and services, which will further be transmitted to their suppliers, resulting in a Keynesian ‘multiplier effect’. Its amplitude generally depends on the migrant households’ propensity to consume at the ability of local producers to increase production, but also on the economic links of the origin area of migrants to other areas (capacity of transmission) (Katseli et al. 2006). But besides stimulating consumption, remittances can also directly support economic growth through investments. Remittances can enable households to overcome

148

C. Incaltarau et al.

capital and risk constraints, therefore facilitating investments (Acosta 2007; Incaltarau and Maha 2012; Kireyev 2006; Lartey 2013; Litan 2009; Senbeta 2013; Woodruff and Zenteno 2007; Giuliano and Ruiz-Arranz 2009). In addition to providing capital, the skills acquired overseas may also matter for explaining entrepreneurship (McCormick and Wahba 2003). Moreover, migration raises awareness about the importance of risk for succeeding in life, and the percentage of people who are entrepreneurs is higher among households with migration experience than among households without migration experience (Sandu 2010; Toth and Toth 2006). Nevertheless, remittances may induce further positive effects on growth through human capital accumulation (Edwards and Ureta 2003). Despite all the growth supporting mechanisms caused by remittances, studies estimating the direct impact of remittances on economic growth brought out ambiguous results. For example, some studies have found an important contribution of remittances to economic growth in origin countries (Catrinescu et al. 2009; Giuliano and Ruiz-Arranz 2009; Jawaid and Raza 2016; Lartey 2013; Azam 2015; Tehseen Jawaid and Raza 2012; Kumar et al. 2018; Incaltarau and Pascariu 2019), but others have shown that the impact is negligible or insignificant (Ruiz et al. 2009; Jongwanich 2007) or even negative (Barajas et al. 2009; Chami et al. 2003; Jawaid and Raza 2016; Tehseen Jawaid and Raza 2012). The different results evidenced in the literature reflect the impact of remittances on growth is transmitted through various mechanisms acting at the same time, but sometimes in opposite directions, which causes a different net effect. One of the negative effects induced by remittances is driven by the ‘moral hazard problem’ (Acosta 2007; Jawaid and Raza 2016; Chami et al. 2003): considering that remittances may increase the reservation wage, remittance inflows may encourage voluntary unemployment and thus reduce the existing labour supply. Additionally, the institutional environment can play a key role in helping remittances to be channelled more efficiently for enhancing economic growth (Catrinescu et al. 2009). They need to minimise other potential negative effects of remittances which may erode economic growth. For example, although remittances are an important source of foreign currency for source countries, a risk of neglecting the trade deficit may emerge over time, by increasing imports to a greater extent than exports, hoping that the deficit will be covered by remittances (Blouchoutzi and Nikas 2010; Chami et al. 2003). Nevertheless, large remittance inflows may result in real exchange rate appreciation, which may further deteriorate the trade balance (Amuedo-Dorantes and Pozo 2004; Acosta et al. 2009). Such developments may trigger further imbalances, thus undermining economic growth in the source economy. Given the various effects induced by migration on source economies, the next section will take a closer look at Romania’s case, by distinguishing three distinct stages according to the migration evolution during its economic transition.

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

7.3

149

Stages and Magnitude of the Romanian Emigration in Post-Communist Times

Migration has been one of the most attractive survival strategies for the Romanian population after the communist regime collapsed. There were certainly more factors pushing population into mass labour migration. First, the last decade of the communist regime has worsened the living standards, therefore, intensifying the migration aspiration within the Romanian society (Horvath 2012). Feeding up population discontent was not only related to the strict control of movements, but also to the fact that the regime failed to provide sufficient resources to the now adult generation born during the baby-boom in the 1960s by limiting their access to houses and higher education. Second, the severe transition following communism was to exacerbate the already existing burden and encourage emigration even more. And thirdly, the persistence of development disparities (especially between urban and rural areas)— despite more than a decade of growth—has kept outmigration high. Overall, the lack of employment and low incomes, which are not sufficient to cover the daily living costs, are the main drivers of Romanian emigration (Cucuruzan 2009; Dincu et al. 2015; Holland et al. 2011; Lazaroiu and Alexandru 2008; Mara 2012; Sandu 2006b; Stoiciu et al. 2011; Sandu 2006a). Thus, while large emigration flows emerged during the post-communist period, more types of migration outflows became visible (Ambrosini et al. 2015; Sandu 2006b; BaldwinEdwards 2008), namely: migration outflows with positive selectivity, which include a large number of young people who left towards old immigrant destinations (USA, Canada, Australia) in order to improve their level of education; migration outflows with neutral selectivity (located between the upper and the lower ends of the distribution of skills), that were heading towards the European continent (Germany, Austria, France); migration flows with negative selectivity that emerged in the late 1990s and the early 2000s’, mainly towards Spain and Italy. The cultural similarities eased the rapid enlargement of Romanian diaspora in these two destination countries, while the migration networks kept feeding them, allowing more and more unqualified people to join them. Despite the high migration outflows, Romania is still one of leading working-age population providers in the EU28 (second after Poland), as in 2018 it accounted for over 16% of all labour movers from EU-28/ EFTA1 (EC 2020). In 2018, Romania was shown to have the largest stock of working-age (20–64 years) emigrants (2.5 million) at EU28 level, with Italy, Spain, Germany and United Kingdom as the main destinations2 (EC 2020). Giving the large migration outflows experienced by Romania, we will take a closer look at its specificities in order to explain the labour market and migration 1 The statistics refers to working age of movers (20–64 year) from EU and European Free Trade Area (EFTA) countries who are residing in other than their country of citizenship. In 2018 the Romanian outflows accounted for 158 from a total of 971 thousands at EU28 level. 2 The largest stocks of Romanian working-age population were shown to be in Italy (931 thousands), Spain (559 thousands), Germany (370 thousands) and United Kingdom (345 thousands).

150

C. Incaltarau et al.

interactions triggered by the broader economic outcomes (which were affected by important developments, such as the economic transition, EU accession or the worldwide economic crisis), while marking three distinct stages during the postcommunist period.

7.3.1

Stage 1: The ‘Stagnant Pool’ of the Slow Transition

The first stage was deeply marked by a slow and hesitant economic transition from a socialist to a capitalist economic system. The reforms and structural adjustments have driven high economic volatility during the first post-communist decade. Beside the stability programme enforced by the Central Bank between 1994 and 1996, the overall lack of deep restructuring has been transposed in successive economic downturns throughout the 1990s. The post-communist political regime embodied many people who were part of the communist party. They knew the system quite well and they focused on using public resources for their own benefit. Thus, reforms were only partially accomplished, while bureaucracy was politicised to strengthen party control of the state, which overspread corruption. The few privatisations performed have only targeted SMEs, while avoiding Romanian economy’s ‘giants’, which were still receiving large subsidies from the Romanian government despite being highly inefficient. By the time of the political elections of 1996, the government had only privatised 12% of the assets owned by the state (Gallagher 2004). The formal and informal barriers were still restraining trade, while the banking system (still dominated by state banks) was still inaccessible to the private sector (Carothers 1998). The new political coalition elected in 1996 engaged in initiating the second period of reforms (Scrieciu and Winker 2002). A floating exchange rate was adopted and the direct subsidies and credits in agriculture were prohibited. These measures were harsh and caused a slump in industry employment as a result of closing down unprofitable state companies, but also because of the drop in demand (caused by the explosive increase of prices, following price liberalisation), which also constrained business activity. Private sector development was also undermined by the frequent changes of regulations, and by the high level of bureaucracy, corruption and taxes (Neef 2002). Therefore, the private sector was not able to absorb the abundant labour as a result of the mass lay-offs. Although unemployment did not reach alarming levels (it stayed at around 10% most of the 1990s, reaching its peak of 11.2% in 1999; see Fig. 7.1), a very low fraction of the unemployed managed to reintegrate in the labour market, and unemployment in Romania becomes more like a ‘stagnant pool’ (Earle and Pauna 1996, 1998). Moreover, the population was not yet accustomed to the labour market competition, as during the communist period they had not faced unemployment. Additionally, the dominant labour market policies were passive-oriented, with little impact on job creation (Earle and Pauna 1998). Given the unfavourable context and the ‘stagnant pool’ of labour, the population had only few solutions to choose from in order to combat unemployment. First, since the living standards have drastically

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

151

Fig. 7.1 Romanian temporary outmigration and unemployment during the post-communist period, 1990–2016. Note: Temporary emigrants (outmigration) are those who emigrated abroad for a period of at least 12 months. The data are reported as number of emigrants per 1000 working-age population. Source: data retrieved from Sandu (2006a) and completed with the estimates of the National Institute for Statistics in Romania starting with 2007

deteriorated (the high inflation considerably reduced the real incomes of population), whereas urban labour became abundant, and more and more people decided to return to rural areas (after getting back the lands that the communist regime had confiscated, according to the Act in 1991),3 where living costs were lower, while supplementing the incomes they earned from informal agricultural activities with the social benefits they were receiving (in 1997 rural–urban net migration even turned negative) (Albu and Nicolae 2003). Thus, the employment share in agriculture overreached the industry share in 1992, while still continuing to grow by the early 2000s. The informal economy has also flourished along with subsistence farming, becoming an important safety valve for the poor (in 1998 the subsistence farming accounted for almost half of the informal economy) (Neef 2002). Second, easing retirement had also contributed to escape from unemployment (for instance, Decree-Law no. 60 of 07.02.1990 allowed early retirement, based on permissive justification, as if the employee exhibits poor performance). Third, starting with the

3 According to the Land Fund Law no. 18/1991 available at: http: //www.cdep.ro/pls/legis/legis_ pck.htp_act_text?idt¼7996.

152

C. Incaltarau et al.

Fig. 7.2 Real GDP and real average net monthly wage evolution in Romania, 1990–2016. Note: The evolution of real GDP and real monthly net wages in Romania is displayed as indexes, where their values start from 100 in 1990. Source: own representation using data provided by the RNIS

mid-90s, migration has begun to become a reliable solution for escaping the ‘stagnant pool’. Still, beside the high ethnic migration flows that occurred in the early 1990s (especially German and Hungarian minorities), Romanian migration was still at an exploratory stage, with low emigration rates (migration rates stayed below 5 emigrants per 1000 population aged 15–64; see Fig. 7.2) and a high selectivity of outflows. Given the low emigration rates, remittance inflows also stayed at low levels during this stage (see Fig. 7.3), with an average of around 0.5 bill. USD per year.

7.3.2

Stage 2: From Shortage of Jobs to Shortage of Labour—The Boom and the Mass Emigration Drain Out the Pool

Unlike the first stage, starting with the early 2000s, the Romanian authorities showed more commitment to ensuring that the economy would finally earn the functional market economy status following the Copenhagen accession criteria (the year 2002 may mark the beginning of the second stage, as it was the year when procedures for

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

153

Fig. 7.3 The evolution of remittance inflows sent by Romanian emigrants, 1990–2016. Note: Remittances NBR (%GDP) and remittances WB (%GDP) are displayed according to the right axis. Remittances NBR (mln USD) and remittances WB (mln USD) are represented on the left axis. Source: own processing on data collected from the National Bank of Romania (NBR—converted into US dollars using the average annual reference currency) and World Bank (WB)

accessing the Schengen Area were simplified, allowing migration to upsurge, while a consistent economic growth rate for the second consecutive year finally confirmed that a certain level of macroeconomic stability had been reached). The European Council’s decision, during the 1999 meeting in Helsinki, to start negotiations with Romania had managed to draw together political interests, in order to follow the unique goal of EU accession. Furthermore, the European Council decision acted as a guarantee for the future authorities’ commitment in trying to consolidate its institutions in order to guarantee democracy, but also to develop a functional market economy, just as the membership criteria were requiring. This helped Romania earn more credit from international economic partners, thus attracting large FDI inflows. Preserving macroeconomic stability and easing business development have fostered private sector development, which absorbed more and more labour (according to the RNIS, the share of employed in the private sector grew from 49% in 1996, to 69% in 2002 and up to 80% in 2008). Between 2002 and 2008, the Romanian economy has grown at an average rate of 11.5% per year in real terms. In 2003, after 13 years, Romania finally managed to overreach its GDP value in 1990 for consecutive years (the 1990 level of GDP was also overreached in 1996, before

154

C. Incaltarau et al.

the 1997–1998 recession; see Fig. 7.2), proving that the delay in reforms made the transition far longer, with dreadful effects on living standards. Further on, while the economic upturn for several years in a row continuously increased labour demand, emigration was also draining the ‘stagnant pool’ for labour. Thus, after lifting up visas for entering the Schengen Area (2002) for Romanian citizens, outmigration rates began to rise sharply, from 9.4 emigrants per 1000 inhabitants aged between 15 and 64 years in 2002, to 21.5 in 2008 (after reaching the highest level in 2007–37.6, see Fig. 7.1). The migration networks, which had been consolidated for years, have also added to the outflows enlargement by decreasing their selectivity (Sandu and Alexandru 2009; Elrick and Ciobanu 2009). While emigration was tending to focus on the same destinations, particularly to Italy or Spain (Baldwin-Edwards 2008; Sandu 2006b), the costs and risks related to migration were becoming lower and lower, convincing more people to look for better opportunities abroad. The higher earnings gained abroad were attracting more and more labour from both urban and rural areas, continuing to shrink labour supply and causing companies in the Romanian economy to face recruitment difficulties (AGS 2008; Cindrea 2007; Voicu et al. 2008). The new economic boom during the 2000s could not reverse the ‘reruralisation’ (according to the RNIS, since the second half of the 1990s urban–rural and urban–urban have become the leading migration flows), given that the population preferred to go abroad, rather than return to the urban areas (Sandu et al. 2004). Therefore, unlike the previous stage, when labour was abundant, a labour supply shortage emerged, pressuring up wages. Therefore, the average monthly net wage doubled between 2002 and 2008. The unemployment rate dropped from 8.4 in 2002 to 4.4 in 2008, confirming that the stagnant pool of unemployed was already drained by the consecutive years of emigration and economic growth. Given the labour deficit, Romanian authorities started to react by easing the access of immigrant labour (since 2007, the Romanian government finally allowed immigrants to be paid with the minimum wage, as before they had to be paid with the medium wage at least).4 Furthermore, Romanian authorities attempted to attract Romanian emigrants from abroad by designing an action plan for encouraging the return of Romanian citizens working abroad.5 Along with the high emigration rates, the remittance inflows surged as well, overreaching 5% of the GDP all over the 2004–2008 period (according to the NBR estimates). The peak was reached in 2008, when remittances amounted for 11.4 billion USD, as estimated by the NBR and 9.3 billion USD

4 See the Emergency Ordinance no. 55 in 20 June 2007 for the establishment of the Romanian Immigration Office through the reorganization of the Authority for Aliens and the National Office for Refugees, but also for amending and supplementing certain normative acts, available at: http:// www.mmuncii.ro/pub/imagemanager/images/file/Legislatie/ORDONANTE-DE-GUVERN/ OUG55-2007_act.pdf. 5 See the Emergency Ordinance no. 187 of 20 February 2008 concerning the approval of the Action plan for encouraging the return of Romanian citizens working abroad, available at http://www. mmuncii.ro/pub/imagemanager/images/file/Legislatie/HOTARARI-DE-GUVERN/HG187-2008. pdf.

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

155

according to the World Bank (see Fig. 7.3). Given their potential in fostering growth, as we have argued in the previous sections of this paper, they may have offset the negative effects caused by emigration.

7.3.3

Stage 3: The International Downturn Is Partially Refilling the Pool

The labour shortage that emerged during the previous stage did not last for long, as the recent economic crisis in the early 2000s forced the Romanian economy to cool down as well. In 2009, Romania entered in economic recession, marking the beginning of a new stage in its economic development. Between 2009 and 2011, the GDP decreased by 9% and real wages by 7% (see Fig. 7.3). Unemployment also went up from 4.4 in 2008, to 7% in 2010. Amid the economic crisis, in order to protect Romanian labour, the Government raised the minimum wage level for immigrants to the average gross wage.6 The international diffusion of the current crisis also tempered Romanian emigration, because it left emigrants no choice in terms of destinations, by affecting most of the EU members. Moreover, since during economic downturns job insecurity for immigrants becomes even higher than for the resident population (Dumont and Garcon 2010), the economic position of immigrants becomes more vulnerable. Thus, leaving abroad suddenly became unattractive and the temporary emigration rate dropped from 21.5% in 2008 to 11.9% in 2013. Nevertheless, even if emigration decreased during the recession, we did not see a mass return to Romania, not only because of the welfare differences between the more developed destinations and the Romanian welfare state, but also because they considered they had higher chances of getting a job abroad than dealing with the harmed economy back home (Stanculescu et al. 2010). The increased uncertainty in destination countries, because of the high unemployment and the reduction in working hours, caused them more concerns, thus forcing them to increase savings at the expense of transfers to Romania. After the hike in 2008, remittance inflows have faced the first major reduction in the postcommunist period (Stanculescu et al. 2010), dropping to just 3.5% of GDP in 2010 and 2.8 in 2015 according to NBR estimates (see Fig. 7.3). The recent evolutions are indicating that the Romanian economy seems to get back on its growth tracks and may be heading to a new stage. The GDP grew by a yearly average of 4.6% between 2014 and 2015, getting really close to its peak in 2008. The wages have already equalled the peak, increasing by an average of 5.7%

6

See Law 157/11 July 2011, available at: http://www.mmuncii.ro/j33/images/Documente/ Legislatie/LEGE157din11iul2011.pdf. However, given the labour shortage, this measure was cancelled by the Law 247/2018 which reduced the required salary for migrant workers from the average gross wage to the minimum gross wage: https://ec.europa.eu/migrant-integration/? action¼media.download&uuid¼F40D5BCD-A311-1522-B3B4AAAB27705D70.

156

C. Incaltarau et al.

during the same period. Unemployment is already low, indicating that a shortage in labour supply may soon occur. Whereas for the current situation, emigration is still low (although on an increasing trend from 11.9 in 2013 to 14.5 in 2015—Fig. 7.1), it may upsurge again if the Euro zone economy recovers from its currently sluggish trend, reinforcing pull mechanisms. Also, despite the large economic growth rates and the Cohesion Policy instruments Romania benefited, regional disparities have increased (Goschin 2017), leaving lagging regions behind, which may further push migration up.

7.4

Methodology and Data Used

As the empirical evidence revealed so far, there are various mechanisms through which outmigration supports economic growth. Our study focuses on two of the main transmission channels, namely remittances and labour market. To the best of our knowledge, there is no other econometric analysis used for estimating the direct impact of migration on growth in the Romanian case, nor the direct impact of remittances on growth during such a long-term period. Furthermore, our estimation strategy has several main strengths. First, it allows us to address the specificity of the Romanian case, by capturing the different intensity of migration and remittances impact during the stages previously defined. Second, it controls for any possible endogeneity issues faced by variables referring to migration and remittances (Abdih et al. 2012; Catrinescu et al. 2009; Atoyan et al. 2016; Adams and Page 2005) by estimating an instrumental variables regression. We are well aware that both migration and remittances could be endogenous because of the simultaneity with the dependent GDP per capita and the unemployment rate. While even the early theoretical approaches are claiming that migration is driven by low development levels and a lack of jobs (Massey et al. 1993), remittances are also related to the economic conditions in the origin countries (Blue 2004; Bettin et al. 2017). Third, focusing on Romania’s case brings the advantage of selecting the most relevant indicators for our primary variables of interest. For example, we take advantage of the time series relating to temporary emigration (provided by NISR and Sandu 2006a, b) which are more appropriate for capturing the real impact of migration, compared to the underestimated permanent flows. Additionally, the data used are provided by NBR, which are more relevant than the more common WB data, because they cover a longer period of time, but also because the series from the WB database displays a break in 2004, which affects their uniformity. Although both remittances and migration data may still be underestimated due to the alternative informal channels used for sending money back home in order to avoid further financial agencies’ fees (Lazaroiu and Alexandru 2008; Mara 2012) or due to unregistered Romanian emigrants that have been illegally working in the destination countries (Sandu 2006b; Mara 2012), they are still closer to the real values.

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

157

The basic specification of the model we have used for estimating the impact of migration on economic growth is based on the human capital augmented version of the standard Solow model (Mankiw et al. 1992): lnYt ¼ β0 þ β1 lnYt1 þ β2 Migrationt þ β3 Human capital þ β4 X t þ εt

ð7:1Þ

where lnYt is the natural log of GDP per capita in year t, lnYt-1 is the initial log of GDP per capita, migrationt the temporary emigration rates, Xt is a set of control variables and εt is the error term. Additional interaction terms were also included in order to find evidence on the differences across migration stages, as—considering the slow transition period and the recent economic crisis effects—we expect a different impact intensity during the second stage defined in the previous section. Thus, specific interaction dummy variables were included for both migration and remittances for the second stage (2002–2008). Further, we use a similar model in order to estimate the impact of remittances on growth: lnYt ¼ β0 0 þ β1 ’ lnYt1 þ β2 ’ Remittancest þ β3 ’ Human capital þ β4 ’ X ’ t þ εt

ð7:2Þ

where Remittancest refers to the remittance inflows as a share to GDP, and X’t is a set of control variables, with all other notations being the same as already described for Eq. (7.1). Next, models (7.3) and (7.4) were designed to estimate the effect of migration and remittances on unemployment following the analytical framework suggested by Jackman (2014): Unemploymentt ¼ α0 þ α1 Migrationt þ α2 Z t þ εt

ð7:3Þ

where Unemploymentt is the unemployment rate in year t, and Zt is a set of control variables, while all other notations are the same as already described for Eq. (7.1). We estimate the effect of remittances on unemployment by dropping the migration variable while instead including remittances and a set of control variables Zt’: Unemplt ¼ α0 ’ þ α1 ’ Remittancest þ α2 ’ Z t ’ þ εt

ð7:4Þ

A dynamic model specification was initially tried for the Eqs. (7.3) and (7.4), but the initial unemployment (Ut-1) did not prove to be statistically significant, so we finally decided to drop it from our model. Interaction terms were also included as described for Eq. (7.1). The models are also controlling for other drivers of growth and unemployment (see Tables 7.1 and 7.2 in Appendix 1 for more details regarding variables and sources used). Following the neoclassical Solow model (Solow 1956; Mankiw et al. 1992), our models control for the existing labour force, physical capital and human

158

C. Incaltarau et al.

capital. For labour force we used age dependency, as it is a good measure of demographic burden, while for physical capital we relied on the share of gross fixed capital formation to GDP. Similar measurements were also used in other migration and remittances related studies (Tehseen Jawaid and Raza 2012; Sikder and Higgins 2016; Catrinescu et al. 2009). For measuring human capital endowment, we have relied on the number of graduated students per 1000 resident population. Comparable proxies, computed using the share of students/graduated students, were also used in other studies evidencing the migration impact on growth (Vakulenko 2016; Borozan 2017). Trade openness is also accounted by our model, as it promotes economic development by more efficient economic activities. This proxy is also largely used in existing growth studies (Lartey 2013; Azam 2015; Le 2009). Also, knowing that the government can improve the economic environment through its public spending (Kelly 1997; Devarajan et al. 1996; Wu et al. 2010), public expenditure related to economic actions was also included. Nevertheless, because infrastructure quality was evidenced to be a significant driver of growth (Cosci and Mirra 2017), this was also accounted for by our model, by controlling for the share of modernised roads. The time series has been checked for stationarity before being used in the models. Table 7.3 in Appendix 2 shows the empirical results for the different unit-root tests performed, thereby providing evidence for the stationarity of series after being detrended. In addition, the combination of the variables as shown in models (7.1)– (7.4) yields stationary residuals (see Table 7.4). Consequently, regression results will be non-spurious and the coefficients from this regression are the long-run multiplier. For the dynamic growth models (7.1) and (7.2), which are estimating the impact of migration and remittances on growth, we have used the Prais and Winsten (1954) transformed regression estimator, because it is consistent with first-order serial correlation, while the standard errors are robust to heteroskedasticity. Models (7.3) and (7.4), which did not prove to be autoregressive, have been estimated using an OLS method. All models have been also run using the instrumental7 GMM method in order to control for endogeneity of migration and remittances. The validity of the overidentifying restrictions was tested by the Hansen’s J statistic χ2 test (Hansen 1982). Also, as the weak instruments are claimed to be a common problem in growth studies (Kraay 2015), the validity of instruments was tested by using the weakiv module (Finlay et al. 2013) available in Stata.

7 More details about instruments used are given as a note to every estimation table. Lags of endogenous regressors were used as internal instruments, which are commonly used for instrumental variables. External instruments were also included, namely lags in wages for the migration models and lags in migration for the remittance models.

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

7.5

159

Discussion and Findings

Our findings are displayed in Tables 7.5, 7.6, 7.7 and 7.8 in Appendix 3. While the first two tables show the estimation results with detrended series, Tables 7.7 and 7.8 report results for the series in levels. Tables 7.5 and 7.7 display the estimation results of the migration impact on growth and unemployment, while Tables 7.6 and 7.8 show the results of the remittances impact estimations. The additional postestimation tests computed for the GMM estimations confirmed the validity of our models. While the Hansen test shows that the models are correctly specified, the validity of the instruments hypothesis could not be rejected by most of the tests performed by using the weakiv package in Stata. The findings in Table 7.5 confirm the main hypothesis of our paper, which follows the optimistic view on migration, namely that migration had an overall positive net effect on economic growth. This effect was also confirmed when migration was treated as endogenous [model (1d) in Tables 7.5 and 7.7]. Therefore, we found evidence for a dominant neoclassical effect over the composition effect. As we have expected, there is no evidence for a negative impact of migration transmitted through the labour market, because the high unemployment persisted all over the 1990s (first stage), making labour supply look like a ‘stagnant pool’. Also, the economic crisis of the late 2000s (during the third stage) slowed down economic activity and raised unemployment level again, limiting the impact of migration. Nevertheless, the role played by migration in draining the abundant labour supply and reduce unemployment level is evidenced in the second half of Tables 7.5 and 7.7) [Eqs. (2a)–(2d)].8 Therefore, our findings demonstrate that emigration in Romania did not cause significant loses in terms of productivity, but, in addition, it reduced the pressure of social spending by reducing unemployment. Without the migration safety valve, some of the workers would have stayed jobless, being dependent on the social welfare system. The social benefits were becoming a real burden because they represented the core of the passive labour market policies implemented in Romania during the 1990s (Earle and Pauna 1998), as well as during the recent economic crisis (Incaltarau and Maha 2014). Our estimations test for a different dynamic pattern during the second stage, by using dummy intersection regressors for the 2002–2008 period. As expected, our hypothesis claiming a different intensity of the migration impact was confirmed, since the results pointed out that, during the second stage, the impact of migration turned slightly negative (the overall effect is given by the cumulation of both main effects and interactions which gives a low negative coefficient—Eqs. (1b)–(1c) in Tables 7.5 and 7.7). Because neither the interaction term nor the main effect did turn statistically significant for the 2002–2008 period, we shortened the period, knowing that the mechanisms described above became even more intense; the results showed a weak, but significant, negative effect during the 2004–2008 period. While 8 It is interesting to note that, while models (1a)–(1d) display better the AIC and BIC scores with the series in levels (Table 7.7), models (2a)–(2d) display better scores with a detrended series.

160

C. Incaltarau et al.

formalities for accessing the Schengen Area were simplified in 2002, which made migration ‘explode’ given that the Romanian economy has already been booming for several years in a row, raising the labour demand, a labour supply shortage started to occur during the second stage (Voicu et al. 2008; AGS 2008), affecting productivity, although the overall net effect during the three stages was shown to be positive. Results showed a similar dynamic when estimating the migration impact on unemployment. Even though migration did prove to have a significant impact in reducing unemployment over the three stages, during the 2002–2008 period the interaction term did not turn statistically significant. Nevertheless, it showed a significantly lower impact for the shorter 2004–2008 period [Eqs. (2b)–(2c) in Table 7.5].9 The reason for the weaker effect may be related to the fact that during this stage unemployment was already at low levels, due to the flourishing economy, which was officially recognised as a functioning market economy by the European Commission (2004). The estimations in both Tables 7.6 and 7.8 reinforce our hypothesis which claims remittances to be a leading mechanism for transmitting the positive effects of migration. This finding is also confirmed when remittances are treated as endogenous—Eq. (3c). The results are proving that the positive effects of remittances are dominating the negative ones. The last three equations in Table 7.6 (4a, b, c) are testing for the above-mentioned moral hazard problem, as receiving of remittances may cause a reduction in number of hours worked or raise voluntary unemployment (Chami et al. 2003; Naiditch and Vranceanu 2009; Acosta 2007). The results are clearly evidencing that remittances have a negative impact on unemployment (Eq. (4a)), and are confirmed when controlling for remittances endogeneity (Eq. (4c)), infirming the moral hazard hypothesis. Differentiating between migration stages proves once again that the impact of remittances in fostering growth was stronger during the 2002–2008 period, as demonstrated by the interaction regressor (Eq. (3b)). Therefore, our hypothesis claiming a different intensity of impact during the second stage—given the raise in remittance inflows due to the migration upsurge after 2002—was confirmed as well. These expectations were also confirmed when estimating the remittances impact on unemployment, showing that the negative effect of remittances on unemployment was stronger, although this only turned statistically significant when using series in levels [Eq. (4b) in Table 7.8] .10 The Keynesian mechanism of stimulating aggregate demand may be one of the main mechanisms by which remittances fostered economic growth and reduced unemployment. The harsh transition process in Romania has initially driven a large share of the population into poverty. The absolute poverty rate rose 6.3 times

9

Estimations with detrended series display better the AIC and BIC scores. However, this result is not confirmed by the models using the series in levels [Eqs. (2b)–(2c) in Table 7.7]. 10 However, the result is not confirmed by the models using detrended series which report better AIC and BIC scores [Eq. (4b) in Table 7.6].

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

161

(from 5.7% in 1990 to 57.9% in 2000), and fell back to the 1990 level only after 18 years, in 2009 (Zamfir et al. 2010). Thus, remittances supplemented household income for living expenses, daily consumption (including honouring debts), housing investment, but also, to a lesser extent, for opening businesses (Grigoras 2006; Lazaroiu and Alexandru 2008; Mara 2012; Stoiciu et al. 2011; Dincu et al. 2015). Other econometric studies have already evidenced a stronger contribution of remittances to investments (Incaltarau and Maha 2012; Litan 2009; Blouchoutzi and Nikas 2010). Overall, our findings prove that the overall net impact of migration during the post-communist period was positive, as the enacted quantity effect, driven by neoclassical mechanisms, dominates the other negative effects. Its intensity was not uniformly distributed during the three stages, since our results evidence a weaker negative effect during the second stage. Nevertheless, considering the overall positive net effect, it is plausible that they have been offset by the positive impact of migration during the other two stages. Remittances represent one of the main mechanisms contributing to the positive effects induced by migration. Their positive effects were stronger during the second stage, when they managed to compensate for the negative effects of migration. However, we need to be aware that whilst the impact of migration in source countries’ economies is transmitted through various channels, it is hard to account for all of them in the same model. While positive effects may occur due to remittances and a decreased pressure on social protection budgets, negative effects may be related to the brain drain and a labour supply shortage that may hinder competitiveness (e.g. OECD 2013; Kaczmarczyk and Okólski 2008; Atoyan et al. 2016). Our study focuses on labour market and remittances, as main transmission mechanisms, and argues for a net positive impact of migration. As already argued in the previous sections, our findings are also related to the particularity of the Romanian case, which smoothed some of the negative migration effects. First, unlike the other Central and Eastern European countries in the EU, Romania went through a prolonged transition, which slowed down the post-communist economic recovery and thus tempered the labour supply deficit caused by migration. Second, besides the slow transition, Romania (and Bulgaria) accessed the EU later than the other states in Central and Eastern Europe, which delayed the outmigration peak and thus reduced labour supply shortage that occurred up to the Great Recession. Third, being one of the least urbanised countries in the EU (with an urbanisation level of somewhere around 55%), rural areas have substantially contributed to outmigration flows. While being largely involved in subsistence farming activities, their migration has affected to a lesser extent the labour market, while generating substantial remittances inflows. Looking at the other explanatory variables, investments, human capital and public expenditure proved to be important economic growth drivers. Investments were confirmed as one of the leading factors triggering economic growth and diminished unemployment. They increased from 14% of GDP in 1991 to 38% in 2008, although the crisis forced them to drop to almost 26%. These were excluded from the models analysing the remittances impact in order to avoid endogeneity issues, as some of the remittances were also heading investments purposes.

162

C. Incaltarau et al.

As expected, age dependency displayed a significant negative impact on growth due to its pressure on productive population, while being positively associated with unemployment. Human capital endowment was also confirmed to be an important growth driver, but only in the equations measuring the remittances impact. Treating human capital as endogenous, as migration also induces changes in terms of human capital stock, showed that it might have exerted a negative impact on growth. Such findings are similar to those from other studies which showed that migration might have a negative effect on enrolment if the expected returns from education are low (Mastrorillo and Fagiolo 2015). However, considering that a negative result was only significant in one of the equations [Eq. (1c) in Table 7.5] and that the human capital positive contribution to growth was strongly confirmed in remittances models, caution should be exercised in drawing strong unambiguous conclusions. Beside investments and human capital, public expenditure on economic activities also proved to be positively associated with economic growth and negatively with unemployment. This appears to be also the case for infrastructure quality, as the network of roads is of course crucial for economic activity development. Unlike public expenditure and infrastructure quality, the negative association between trade openness and economic growth is related to the persistence of a negative trade balance all over the post-communist period. This became larger and larger, overreaching 10% of GDP during the 2003–2008 period, while still exceeding 5% after the crisis. The gradual liberalisation, as a result of EU accession agreements, has also affected Romania’s commercial activity. The Romanian companies had to face the more competitive companies in the EU Single market. Consequently, imports continued to increase, while domestic companies were continuously improving their efficiency.

7.6

Conclusions

This study investigates the impact of migration on economic growth in Romania by focusing on two leading mechanisms through which migration uses to transmit its effects, namely remittances and labour market. The Romanian case is a fertile ground for analysing the migration impact, as it faced substantial outmigration flows during the post-communist period, which triggered large remittance inflows as well. Furthermore, the bumpy transition road from the communist regime to EU accession has induced major structural changes in the Romanian economy, with mixed effects on the labour market, which makes the net migration impact unclear. The paper brings some important theoretical and methodological contributions in the Romanian case. Theoretically, it defines three distinct post-communist stages according to the interaction between migration, remittances and labour market outcomes and addresses differences in the magnitude of these effects by stage. Methodologically, it is the first study to employ a long-term analysis from 1991 to 2015, while accounting for endogeneity of migration and remittances, by taking advantage

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

163

of the more reliable data relating to temporary emigration and remittance inflows provided by the RNIS and RNB. The paper argues for the delimitation of three distinct stages regarding the evolution of migration and labour market during the post-communist period in Romania, which are marked by two cut-off points, namely the Schengen Area access simplification for Romanian citizens and the economic crisis. The first stage encloses the early post-communist transition period, characterised by high economic instability, which triggered low exploratory migration flows and low remittance inflows. At the same time, the mass lay-offs triggered by privatisation of state-owned companies caused an abundant labour supply, which made unemployment look more like a ‘stagnant pool’. The perspective of accessing the EU brought a simplification of Schengen Area access for Romanian citizens, which allowed Romanian migration flows to upsurge during the second stage. Along with international migration, the ‘reruralisation’, as well as the labour demand driven by economic growth, ‘drained out the unemployment pool’, and the labour market went from an abundant labour to a labour shortage. The remittance inflows boomed as well, contributing to the modernisation of the Romanian society. Later on, the economic crisis of the late 2000s marked the beginning of the third stage, by slowing down economic activity in Romania, which reduced labour demand pressure and started to partially ‘refill the unemployment pool’. The crisis has also tempered migration, as it also reduced the attractiveness of destination countries, and thus halving remittance inflows. Our empirical results bring evidence for the neoclassical optimistic view on migration during the post-communist period, proving that outmigration had an overall positive net effect on growth during the whole three stages. These positive effects were transmitted through a significant reduction of unemployment, which reduced the burden of the social protection system, but also through the remittances channel that proved to be a relevant migration mechanism for fostering economic growth. The results showed no evidence for a moral hazard problem induced by remittance inflows (Acosta 2007; Chami et al. 2003), as they significantly reduced unemployment, through a dominant effect of stimulating production level. The intersection dummy variables evidenced that the migration intensity impact was not uniformly displayed during the three stages. Despite an overall net positive effect of migration on growth, during the second stage this turns slightly negative, as the labour shortage which occurred induced an upward pressure on real wage levels, affecting productivity level. Nevertheless, during the second stage, the growth stimulating effects of remittances also proved to be stronger, stimulated by the increasing remittance inflows which may have offset the negative effects of migration. In policy terms, Romania needs to keep itself on the reform track, aiming to converge to the older EU members’ political and economic models. Otherwise, if it keeps staying unattractive compared to other EU countries, emigration has its own self-sustaining mechanisms and may keep undermining development. De Haas (2012) warns against the risk that, under an optimistic view on migration, migration policies distract attention from other important issues in origin countries, hoping that

164

C. Incaltarau et al.

beneficial effects of migration will resolve them as well. But migration per se is not sufficient for development. As we have shown in the paper, in Romania, the positive impact of migration was due to a specific transition context, characterised by economic instability and abundant labour. Later on, while the labour shortage occurred, migration was also shown to negatively affect the Romanian economy until the recent economic crisis. Given such outcomes, the economic upturns during the last years may reinstall it again. There are several policy lessons to be drawn for our analysis. First, Romanian policy-makers should focus on designing a framework favourable to development (Careja 2013) by improving tangible (e.g. improving infrastructure quality; ensuring an equitable welfare system) and intangible goods (e.g. ensuring economic and political stability; fighting against corruption; ensuring a fair judicial system and the rule of law). The EU Cohesion Policy support is a valuable asset which can help Romania to trigger development and generally improve the standard of living. Second, policy-makers need to focus on using migration as a resource for development. In this sense, they can design friendly policy measures to attract large inflows of remittances and a better leveraging towards investment rather than consumption. Policies oriented towards improving the ease of doing business may be such an example, like cutting the ‘red tape’, reducing the fiscal burden on increasing transparency in the decision-making process in central and local public administration, while facilitating the dialogue with entrepreneurs in order to understand and address their needs. Such measures may both discourage capital outflows and stimulate capital inflows, including investments driven by remittances. Besides capital, the experience gained abroad may also be an important asset for the Romanian economy. Supporting returnees through bilateral cooperation agreements may also ease their reintegration and help them capitalise the accumulated skills abroad (Andren and Roman 2016). A third set of policy interventions aimed at preventing or mitigating the negative effects of migration should also be considered. For example, facilitating vulnerable groups integration (e.g. minorities), as they are more prone to emigrate (Duval and Wolff 2016), may avoid labour force depletion. Also, policy measures for increasing labour force participation and for increasing utilisation of the remaining workforce may prevent a future labour market crisis (Atoyan et al. 2016). Additionally, particular attention should be given to the negative effects on children left at home (Botezat and Pfeiffer 2020), as parents’ emigration was shown to have a negative effect on their school performance, and make them more exposed to mental and physical health problems. Acknowledgments This work was supported by a grant of Ministry of Research and Innovation, CNCS–UEFISCDI, project number PN-III-P4-ID-PCCF-2016-0166, within PNCDI III project ‘ReGrowEU—Advancing ground-breaking research in regional growth and development theories, through a resilience approach: towards a convergent, balanced and sustainable European Union’.

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

165

Appendix 1: Data Table 7.1 Variable definitions and sources Variable Age dependency D2002–2008 D2004–2008

Human capital Infrastructure quality Investments

Definition Population aged 0–19 and 60 and more to pop. Aged 20–59 Dummy variable taking 1 value during 2002–2008 (stage 2) or 0 otherwise Dummy variable taking 1 value during 2004–2008 (stage 2) or 0 otherwise School-age population enrolment rate (% total population) Share of modernised roads

Ln(Y)

Gross fixed capital formation as share of GDP Logarithm of real GDP per capita

Ln(wage)

Real average net wage

Migration

Temporary emigrants per 1000 inhabitants aged 15–64 years (temporary emigrants are defined as persons who emigrated abroad for a period of at least 12 months) Sum of imports (CIF) and exports (FOB) to GDP Public expenditure related to economic actions to GDP Real remittance inflows as a share of GDP (encloses both private transfers and compensation of employees) Unemployment rate

Openness Public expenditure Remittances

Unemployment

Source National Institute for Statistics in Romania Own calculations Own calculations

National Institute for Statistics in Romania National Institute for Statistics in Romania National Institute for Statistics in Romania National Institute for Statistics in Romania National Institute for Statistics in Romania Data gained from Sandu (2006b) and completed with the estimates of the National Institute for Statistics in Romania starting with 2007 National Institute for Statistics in Romania National Institute for Statistics in Romania National Bank of Romania

National Institute for Statistics in Romania

Table 7.2 Descriptive statistics variable Age dependency Human capital Infrastructure quality Investments Y Wage Migration Openness Public expenditure Remittances Unemployment

Obs. 25 25 25 25 25 25 25 25 25 25 25

Mean 81.97 4.81 27.37 23.49 52,671.64 2928.08 12.05 61.56 6.10 3.02 7.28

SD 4.56 2.72 4.20 5.24 21,289.66 938.59 9.20 11.12 2.52 1.76 2.41

Min. 77.30 1.29 23.22 14.38 29,447.20 1910.96 2.30 36.55 3.27 0.24 3.00

Max. 90.60 11.02 38.00 38.40 89,868.09 4658.17 37.65 75.18 12.51 6.27 11.80

166

C. Incaltarau et al.

Appendix 2: Stationarity and Cointegration Tests Table 7.3 Unit root and stationarity tests on detrended series Variable lnYd

Unemploymentd

Migrationd

Remittancesd

Age dependencyd

Human capitald

Investmentsd

Opennessd

Public expenditured

Infrastructure qualityd

PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS

At level with constant

Optimal lag

2.470 2.764* 2.238* 0.16 4.282*** 4.438*** 2.295* 0.103 1.848 2.475 2.220* 0.14 0.983 1.681 0.946 0.209 0.304 1.420 0.746 0.223 1.442 2.346 1.394 0.124 2.440 2.578* 2.245* 0.103 3.534** 3.445** 1.723 0.127 2.826* 2.859* 2.578* 0.0847 0.278 2.297 3.892*** 0.243

2 1 2 2 2 0 3 2 2 4 4 2 2 2 1 2 2 2 1 2 2 3 1 2 2 1 1 2 2 0 1 2 2 2 1 2 2 6 6 2

Notes: PP Phillips–Perron unit-root test critical values, ADF the Augmented Dickey–Fuller unit-root test critical values, DFGLS Dickey–Fuller GLS critical values, KPSS Kwiatkowski–Phillips–Schmidt–Shin test for stationarity critical values. Whereas PP, ADF and DFGLS tests H0: time series has a unit root, for KPSS H0: time series is level stationary. Significance levels: * p < 0.1, ** p < 0.05, *** p < 0.01

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

167

Table 7.4 Unit root and stationarity tests on model residuals Variable Models with detrended series (Tables 7.5 and 7.6) 1(a)

2(a)

3(a)

4(a)

Models with series in level (Tables 7.7 and 7.8) 1(a)

2(a)

3(a)

4(a)

At level with constant

Optimal lag

PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS

3.937*** 4.590*** 3.912*** 0.117 5.318*** 5.238*** 1.539 0.296 3.119** 3.115** 1.863 0.486** 5.400*** 5.398*** 2.341* 0.0664

2 2 1 2 2 0 1 2 2 0 1 2 2 0 1 2

PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS PP ADF DFGLS KPSS

4.324*** 5.214*** 3.766*** 0.0419 5.172*** 7.131*** 1.708 0.0583 4.832*** 5.073*** 3.206*** 0.0417 5.426*** 6.382*** 0.910 0.0576

2 2 1 2 2 2 5 2 2 2 1 2 2 2 4 2

Notes: PP Phillips–Perron unit-root test critical values, ADF the Augmented Dickey–Fuller unitroot test critical values, DFGLS Dickey–Fuller GLS critical values, KPSS Kwiatkowski–Phillips– Schmidt–Shin test for stationarity critical values. PP, ADF and DFGLS tests H0: time series has a unit root, while KPSS tests H0: time series is level stationary. Significance levels: * p < 0.1, ** p < 0.05, *** p < 0.01

Openness

Investments

Human capital

Age dependency

T2004–2008

T2002–2008

Migration*T2004–2008

Migration*T2002–2008

Migration

lnYt-1

0.0294** (0.0138) 0.0233 (0.0160) 0.0217*** (0.00426) 0.00691*

0.0326* (0.0176) 0.0249 (0.0292) 0.0225** (0.00992) 0.00695

0.0230 (0.0653)

Prais–Winsten OLS (1a) (1b) lnY lnY 0.534*** 0.523*** (0.117) (0.145) 0.00459 0.0101 (0.00279) (0.0128) 0.00680 (0.0113)

0.0965* (0.0468) 0.0298* (0.0161) 0.00851 (0.0232) 0.0173** (0.00709) 0.00538

0.0134** (0.00603)

(1c) lnY 0.460*** (0.124) 0.0126* (0.00688)

0.0406*** (0.0113) 0.0544*** (0.00981) 0.0279*** (0.00273) 0.0106***

GMM (1d) lnY 0.638*** (0.0706) 0.00424* (0.00218)

1.441** (0.549) 0.278 (0.268) 0.306*** (0.0930) 0.0795

0.285** (0.107)

OLS (2a) Unempl

1.690*** (0.539) 0.642 (0.516) 0.426** (0.149) 0.103

0.534 (1.596)

0.613** (0.260) 0.376 (0.233)

(2b) Unempl

2.001 (1.435) 1.639*** (0.529) 0.369 (0.331) 0.338** (0.130) 0.0703

0.389** (0.182)

0.552** (0.196)

(2c) Unempl

1.507*** (0.433) 0.813 (0.701) 0.00594 (0.252) 0.0673

0.462*** (0.104)

GMM (2d) Unempl

Table 7.5 Estimation results of the long-term impact of migration on GDP per capita and unemployment in Romania (1991–2015, detrended series)

Appendix 3: Estimation Results

168 C. Incaltarau et al.

(0.00345) 0.0196*** (0.00664) 0.0374* (0.0202) 0.0160 (0.0134) 25 0.894 0.840 71.33 60.36

(0.00413) 0.0164* (0.00823) 0.0478** (0.0185) 0.00438 (0.0356) 25 0.893 0.817 68.95 55.55

(0.00362) 0.0100 (0.00748) 0.0606** (0.0209) 0.00280 (0.0163) 25 0.867 0.772 72.39 58.99 0.417 0.242 0.155 0.544 0.087 0.327 0.000 2

(0.00172) 0.0244*** (0.00472) 0.0184 (0.0204) 0.0241*** (0.0104) 24

(0.0764) 0.410** (0.186) 2.495** (0.960) 0.129 (0.338) 25 0.692 0.565 92.26 102.0

(0.0816) 0.313 (0.193) 2.884** (1.042) 0.904 (0.777) 25 0.752 0.603 90.84 103.0

(0.0744) 0.249 (0.187) 3.016*** (0.998) 0.730 (0.554) 25 0.760 0.615 90.05 102.2 0.226 0.678 0.278 0.739 0.072 0.064 0.000 2

(0.103) 0.0754 (0.176) 3.768*** (0.702) 0.501 (0.329) 25

Notes: Robust standard errors are given in parentheses. Significance levels: * p < 0.1, ** p < 0.05, *** p < 0.01. In GMM estimations, migration and human capital variables were treated as endogenous. While both first and second lags of migration, human capital and net wage were used as instruments in the GDP per capita models, first lags were included in the unemployment models. This explains the lower number of observations is the migration models. Just Instruments validity tests results are also displayed in the last part of the table (weakiv command in Stata). AIC and BIC stand for the Akaike Information Criterion and Bayesian Information Criterion. Source: authors’ estimations

Observations R2 Adj. R2 AIC BIC Hansen test CLR AR K J Rank Wald Endogenous regressors

Constant

Infrastructure quality

Public expenditure

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . . 169

Observations R2 Adj. R2 AIC

Constant

Infrastructure quality

Public expenditure

Openness

Human capital

Age dependency

T2002–2008

Remittances*T2002–2008

Remittances

lnYt-1

0.00114 (0.0132) 0.0284** (0.00984) 0.00180 (0.00376) 0.00660 (0.00712) 0.0519* (0.0248) 0.00549 (0.0188) 25 0.753 0.651 61.47

Prais–Winsten OLS (3a) lnY 0.496*** (0.154) 0.0487** (0.0196) (3b) lnY 0.545*** (0.145) 0.0586 (0.0436) 0.0986* (0.0545) 0.0398 (0.0769) 0.0122 (0.0141) 0.0328** (0.0135) 0.000797 (0.00304) 0.00289 (0.00665) 0.0411* (0.0215) 0.0646** (0.0301) 25 0.861 0.777 63.16 0.0168 (0.0151) 0.0340*** (0.0112) 0.00363 (0.00337) 0.0129* (0.00656) 0.106*** (0.0393) 0.000524 (0.0121) 24 0.816 0.735

GMM (3c) lnY 0.498*** (0.111) 0.0845*** (0.0254)

0.836* (0.468) 0.714*** (0.227) 0.0114 (0.0766) 0.284 (0.201) 3.157** (1.102) 0.314 (0.324) 25 0.646 0.528 93.74

2.210*** (0.674)

OLS (4a) Unempl

0.871 (1.419) 0.966 (1.317) 1.490 (1.913) 1.010** (0.468) 0.809*** (0.202) 0.00391 (0.0778) 0.167 (0.193) 3.176** (1.368) 0.514 (0.622) 25 0.694 0.542 94.05

(4b) Unempl

0.479 (0.370) 0.685*** (0.155) 0.0572 (0.0499) 0.134 (0.139) 2.662*** (0.926) 0.106 (0.249) 24 0.560 0.405

2.228*** (0.578)

GMM (4c) Unempl

Table 7.6 Estimation results of the long-term impact of remittances on GDP per capita and unemployment in Romania (1991–2015, detrended series)

170 C. Incaltarau et al.

51.72

50.97 0.629 0.300 0.222 0.650 0.139 0.601 0.001 1

102.3

105.0 0.954 0.195 0.143 0.660 0.083 0.615 0.000 1

Notes: Robust standard errors are given in parentheses. Significance levels: * p < 0.1, ** p < 0.05, *** p < 0.01. In GMM estimations, remittances variable was treated as endogenous. First and second lags of both remittances and migration were used as instruments. Using second lags as instruments explains the lower number of observations. Instruments validity tests results are also displayed in the last part of the table (weakiv command in Stata). AIC and BIC stand for the Akaike Information Criterion and Bayesian Information Criterion. Source: authors’ estimations

BIC Hansen test CLR AR K J Rank Wald Endogenous regressors

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . . 171

Infrastructure quality

Public expenditure

Openness

Investments

Human capital

Age dependency

T2004–2008

T2002–2008

Migration*T2004–2008

Migration*T2002–2008

Migration

lnYt-1

0.0220* (0.0113) 0.0174 (0.0172) 0.0194*** (0.00485) 0.00627* (0.00348) 0.0186*** (0.00622) 0.0320*** (0.00947) 0.0198 (0.0118) 0.0275 (0.0275) 0.0234** (0.0103) 0.00710* (0.00367) 0.0170** (0.00695) 0.0295*** (0.00921)

0.116 (0.0989)

Prais–Winsten OLS (1a) (1b) lnY lnY 0.525*** 0.489*** (0.112) (0.120) 0.00427* 0.0145 (0.00223) (0.0116) 0.0105 (0.0108)

0.236** (0.101) 0.0128 (0.0127) 0.00987 (0.0209) 0.0176** (0.00652) 0.00584 (0.00361) 0.0116 (0.00731) 0.0343*** (0.00963)

0.0120* (0.00578)

(1c) lnY 0.469*** (0.112) 0.0108* (0.00510)

0.0333** (0.0148) 0.0380 (0.0269) 0.0244*** (0.00591) 0.00852*** (0.00315) 0.0203*** (0.00697) 0.0271*** (0.01000)

GMM (1d) lnY 0.573*** (0.127) 0.00551*** (0.00161)

0.353 (0.421) 0.666 (0.497) 0.284** (0.104) 0.147 (0.177) 0.467 (0.320) 0.321** (0.131)

0.143** (0.0557)

OLS (2a) Unempl

0.364 (0.438) 0.518 (0.447) 0.234 (0.191) 0.137 (0.165) 0.494 (0.365) 0.379 (0.305)

1.316 (2.896)

0.0368 (0.295) 0.115 (0.312)

(2b) Unempl

Table 7.7 Estimation results of the long-term impact of migration on GDP per capita and unemployment in Romania (1991–2015)

1.979 (3.713) 0.412 (0.519) 0.622 (0.480) 0.269* (0.145) 0.147 (0.186) 0.526 (0.409) 0.358* (0.185)

0.121 (0.234)

0.0606 (0.167)

(2c) Unempl

2.118 (1.447) 3.354 (2.661) 1.307 (1.159) 0.380 (0.290) 0.771 (0.650) 0.504* (0.294)

0.364* (0.216)

GMM (2d) Unempl

172 C. Incaltarau et al.

5.922*** (1.536) 25 0.989 0.983 71.99 61.02

6.130*** (1.618) 25 0.994 0.989 69.91 56.50

5.671*** (1.434) 25 0.991 0.985 72.71 59.30 0.409 0.282 0.244 0.421 0.186 0.470 0.000 2

6.567*** (0.686) 24 0.988 0.981

13.90 (41.46) 25 0.616 0.458 106.1 115.8

13.83 (41.15) 25 0.620 0.391 109.8 122.0

18.20 (49.01) 25 0.621 0.393 109.8 121.9 0.772 0.186 0.282 0.172 0.583 0.082 0.234 2

203.2 (124.9) 25

Notes: Robust standard errors are given in parentheses. Significance levels: * p < 0.1, ** p < 0.05, *** p < 0.01. In GMM estimations, migration and human capital variables were treated as endogenous. While both first and second lags of migration, human capital and net wage were used as instruments in the GDP per capita models, first lags were included in the unemployment models. This explains the lower number of observations is the migration models. Instruments validity tests results are also displayed in the last part of the table (weakiv command in Stata). AIC and BIC stand for the Akaike Information Criterion and Bayesian Information Criterion. Source: authors’ estimations

Observations R2 Adj. R2 AIC BIC Hansen test CLR AR K J Rank Wald Endogenous regressors

Constant

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . . 173

Observations R2 Adj. R2 AIC

Constant

Infrastructure quality

Public expenditure

Openness

Human capital

Age dependency

T2002–2008

Remittances*T2002–2008

Remittances

lnYt-1

0.0112 (0.00921) 0.0327** (0.0118) 0.00157 (0.00318) 0.00835 (0.00591) 0.0435*** (0.0109) 3.257** (1.194) 25 0.987 0.981 66.82

Prais–Winsten OLS (3a) lnY 0.481*** (0.116) 0.0445*** (0.0105) (3b) lnY 0.528*** (0.133) 0.0437 (0.0333) 0.0736* (0.0371) 0.184 (0.141) 0.00879 (0.0121) 0.0295** (0.0122) 0.00107 (0.00270) 0.00543 (0.00604) 0.0379*** (0.0113) 4.759** (1.711) 25 0.990 0.983 66.49 0.0169** (0.00721) 0.0375*** (0.0106) 0.00148 (0.00221) 0.00306 (0.00467) 0.0444*** (0.00844) 2.988*** (0.790) 24 0.987 0.981

GMM (3c) lnY 0.443*** (0.0839) 0.0483*** (0.00936)

0.208 (0.374) 0.172 (0.328) 0.0599 (0.174) 0.310 (0.325) 0.455*** (0.156) 38.49 (36.96) 25 0.556 0.407 107.7

0.896*** (0.301)

OLS (4a) Unempl

3.321 (2.147) 4.045* (2.148) 12.22* (6.570) 0.726 (0.782) 0.231 (0.250) 0.0178 (0.112) 0.281 (0.245) 0.318*** (0.0993) 49.46 (74.88) 25 0.667 0.500 104.5

(4b) Unempl

Table 7.8 Estimation results of the long-term impact of remittances on GDP per capita and unemployment in Romania (1991–2015)

0.640*** (0.0903) 0.540*** (0.122) 0.169*** (0.0393) 0.135* (0.0734) 0.270*** (0.0687) 81.98*** (8.946) 24 0.905 0.871

0.754*** (0.196)

GMM (4c) Unempl

174 C. Incaltarau et al.

57.06

54.30 0.248 0.111 0.098 0.183 0.109 0.235 0.000 1

116.2

115.5 0.142 0.203 0.081 0.973 0.040 0.276 0.000 1

Notes: Robust standard errors are given in parentheses. Significance levels: * p < 0.1, ** p < 0.05, *** p < 0.01. In GMM estimations, remittances variable was treated as endogenous. First and second lags of remittances and log of net wage were used as instruments. Using second lags as instruments explains the lower number of observations. Instruments validity tests results are also displayed in the last part of the table (weakiv command in Stata). AIC and BIC stand for the Akaike Information Criterion and Bayesian Information Criterion. Source: authors’ estimations

BIC Hansen test CLR AR K J Rank Wald Endogenous regressors

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . . 175

176

C. Incaltarau et al.

References Abdih Y, Chami R, Dagher J, Montiel P (2012) Remittances and institutions: are remittances a curse? World Dev 40(4):657–666. https://doi.org/10.1016/j.worlddev.2011.09.014 Acosta P (2007) Entrepreneurship, labour markets and international remittances: evidence from El Salvador. In: Özden Ç, Schiff M (eds) International migration, economic development & policy. Basingstoke; World Bank, Palgrave Macmillan, Washington, D.C., New York, pp 141–160 Acosta PA, Lartey EKK, Mandelman FS (2009) Remittances and the Dutch disease. J Int Econ 79 (1):102–116. https://doi.org/10.1016/j.jinteco.2009.06.007 Adams RH, Page J (2005) Do international migration and remittances reduce poverty in developing countries? World Dev 33(10):1645–1669. https://doi.org/10.1016/j.worlddev.2005.05.004 AGS (2008) Angajatorul roman (public si privat) si piata fortei de munca. Raport comparativ privind opiniile angajatorilor publici si privati din Romania privind evolutia pietei muncii (Romanian employer (public and private) and the labor market. Comparative report on private and public employers views on the evolution of the labor market in Romania). Romanian Government, Bucharest Albu LL, Nicolae M (2003) Use of households survey data to estimate the size of the informal economy in Romania. The informal economy in the EU accession countries. Size, scope, trends and challenges to the process of EU enlargement. Bertelsmann Foundation, World Bank, Center for the Study of Democracy, Vienna Institute for International Economic Studies, Sofia, Vienna Ambrosini JW, Mayr K, Peri G, Radu D (2015) The selection of migrants and returnees in Romania. Econ Transit 23(4):753–793. https://doi.org/10.1111/ecot.12077 Amuedo-Dorantes C, Pozo S (2004) Workers’ remittances and the real exchange rate: a paradox of gifts. World Dev 32(8):1407–1417. https://doi.org/10.1016/j.worlddev.2004.02.004 Andren D, Roman M (2016) Should I stay or should I go? Romanian migrants during transition and enlargements. In: Kahanec M, Zimmermann KF (eds) Labor migration, EU enlargement, and the great recession. Springer, Berlin Heidelberg, pp 247–270 Atoyan R, Christiansen LE, Dizioli A, Ebeke C, Ilahi N, Ilyina A, Mehrez G, Qu H, Raei F, Rhee A, Zakharova D, International Monetary Fund (2016) Emigration and its economic impact on Eastern Europe. IMF staff discussion notes 7. International Monetary Fund, Washington, DC Azam M (2015) The role of migrant workers remittances in fostering economic growth. Int J Soc Econ 42(8):690–705. https://doi.org/10.1108/ijse-11-2013-0255 Baldwin-Edwards M (2008) Migration policies for a Romania within the European Union: navigating between Scylla and Charybdis. In: Silasi G, Simina OL (eds) Migration mobility and human rights at the Eastern border of the European Union. Editura Universitatii de Vest, Timisoara, pp 258–288 Barajas A, Chami R, Fullenkamp C, Gapen M, Montiel P (2009) Do workers’ remittances promote economic growth? IMF working paper 153. International Monetary Fund, Washington, DC Barro RJ, Sala-i-Martin X (2004) Economic growth. MIT Press, Cambridge Bettin G, Presbitero AF, Spatafora NL (2017) Remittances and vulnerability in developing countries. World Bank Econ Rev 31:1–23. https://doi.org/10.1093/wber/lhv053 Blouchoutzi A, Nikas C (2010) The macroeconomic implications of emigrants’ remittances in Romania, Bulgaria and Albania. Post-Communist Econ 22(4):547–558. https://doi.org/10.1080/ 14631377.2010.518479 Blue SA (2004) State policy, economic crisis, gender, and family ties: determinants of family remittances to Cuba. Econ Geogr 80(1):63–82. https://doi.org/10.1111/j.1944-8287.2004. tb00229.x Borozan D (2017) Internal migration, regional economic convergence, and growth in Croatia. Int Reg Sci Rev 40(2):141–163. https://doi.org/10.1177/0160017615572889 Botezat A, Pfeiffer F (2020) The impact of parental labour migration on left-behind children's educational and psychosocial outcomes: evidence from Romania. Popul Space Place 26(2): e2277. https://doi.org/10.1002/psp.2277

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

177

Careja R (2013) Emigration for development? An exploration of the state’s role in the developmentmigration nexus: the case of Romania. Int Migr 51(5):76–90. https://doi.org/10.1111/imig. 12115 Carothers T (1998) Romania: the political background. In: Tindemans L (ed) Democracy in Romania. International Institute for Democracy and Electoral Assistance, Stockholm Catrinescu N, Leon-Ledesma M, Piracha M, Quillin B (2009) Remittances, institutions, and economic growth. World Dev 37(1):81–92. https://doi.org/10.1016/j.worlddev.2008.02.004 Chami R, Fullenkamp C, Jajah S (2003) Are immigrant remittance flows a source of Capital for Development? IMF working paper 189. International Monetary Fund, Washington, DC Chirila V, Chirila C (2017) The analysis of Romania’s external migration and of the causality between remittances and Romania’s economic growth. Amfiteatru Econ 19(46):696–710 Chiswick B, Miller PW (eds) (2015) Economics of international migration. North-Holland, Amsterdam Cindrea I (2007) The crisis on the labor market in Romania. Theor Appl Econ 4:25–28 Cosci S, Mirra L (2017) A spatial analysis of growth and convergence in Italian provinces: the role of road infrastructure. Reg Stud 52:516–527. https://doi.org/10.1080/00343404.2017.1334117 Cucuruzan RE (2009) Migraţia şi mobilitatea forţei de muncă din România în contextul integrării europene. Editura Fundaţiei pentru Studii Europene, Ideea Europeană, Cluj-Napoca de Haas H (2012) The migration and development pendulum: a critical view on research and policy. Int Migr 50(3):8–25. https://doi.org/10.1111/j.1468-2435.2012.00755.x Devarajan S, Swaroop V, Zou H-f (1996) The composition of public expenditure and economic growth. J Monet Econ 37(2):313–344. https://doi.org/10.1016/S0304-3932(96)90039-2 Dincu A, Orban M, Bageac I (2015) Romania din afara Romaniei. Dermatol Sin 20 Dumont JP, Garcon JP (2010) The crisis and its impact on migration and migrant employment. In: IEMed mediterranean yearbook. European Institute of the Mediterranean, Barcelona, pp 271–276 Duval L, Wolff F-C (2010) Remittances matter: longitudinal evidence from Albania. PostCommunist Econ 22(1):73–97. https://doi.org/10.1080/14631370903525611 Duval L, Wolff F-C (2016) Emigration intentions of Roma: evidence from central and south-East Europe. Post-Communist Econ 28(1):87–107. https://doi.org/10.1080/14631377.2015.1124558 Earle JS, Pauna C (1996) Incidence and duration of unemployment in Romania. Eur Econ Rev 40 (3):829–837. https://doi.org/10.1016/0014-2921(95)00093-3 Earle JS, Pauna C (1998) Long-term unemployment, social assistance and labor market policies in Romania. Empirical Econ 23(1):203–235. https://doi.org/10.1007/bf01205685 Edwards AC, Ureta M (2003) International migration, remittances, and schooling: evidence from El Salvador. J Devel Econ 72(2):429–461. https://doi.org/10.1016/S0304-3878(03)00115-9 Elrick TIM, Ciobanu O (2009) Migration networks and policy impacts: insights from Romanian– Spanish migrations. Global Netw 9(1):100–116. https://doi.org/10.1111/j.1471-0374.2009. 00244.x Elsner B (2013) Does emigration benefit the stayers? Evidence from EU enlargement. J Population Econ 26(2):531–553. https://doi.org/10.1007/s00148-012-0452-6 European Commission (2004) Regular report on Romania’s progress towards accession. European Commission, Brussels European Commission (2020) 2019 annual report on intra-EU labour mobility. Publications Office of the European Union, Luxembourg Finlay K, Magnusson L, E Schaffer M (2013) WEAKIV: Stata module to perform weakinstrument-robust tests and confidence intervals for instrumental-variable (IV) estimation of linear, probit and tobit models. Boston College Department of Economics, Boston Fratesi U, Riggi MR (2007) Does migration reduce regional disparities? The role of skill-selective flows. Revf Urban Regl Dev Stud 19(1):78–102. https://doi.org/10.1111/j.1467-940X.2007. 00125.x Gallagher T (2004) Furtul unei naţiuni: România de la comunism încoace (Theft of a Nation: Romania Since Communism). Humanitas, Bucureşti

178

C. Incaltarau et al.

Gamlen A (2014) The new migration-and-development pessimism. Prog Hum Geogr 38 (4):581–597. https://doi.org/10.1177/0309132513512544 Giuliano P, Ruiz-Arranz M (2009) Remittances, financial development, and growth. J Devel Econ 90(1):144–152. https://doi.org/10.1016/j.jdeveco.2008.10.005 Goschin Z (2017) Exploring regional economic convergence in Romania. A spatial modeling approach. Eastern J Eur Stud 8(2):127–146 Grigoras V (2006) Consequences and plans. Incomes and investments from migration Bucharest. Living abroad on a temporary basis: the Romanians and the economic migration: 1990–2006. Open Society Foundation, Bucharest Hansen L (1982) Large sample properties of generalized method of moments estimators. Econometrica 50(4):1029–1054 Holland D, Fic T, Rincon-Aznar A, Stokes L, Paluchowski P (2011) Labour mobility within the EU - the impact of enlargement and the functioning of the transitional arrangements. National Institute of Economic and Social Research, London Horvath I (2012) Migratia internationala a cetatenilor romani dupa 1989. In: Rotariu T, Voineagu V (eds) Inertie si schimbare: dimensiuni sociale ale tranzitiei in Romania. Polirom, Iasi, pp 199–222 Incaltarau C, Maha LG (2012) The impact of remittances on consumption and investment in Romania. Eastern J Eur Stud 3(4):31–86 Incaltarau C, Maha LG (2014) The effects induced by the recent economic crisis on the labour market policies in Romania. Revista de cercetare si interventie sociala 47:44–66 Incaltarau C, Pascariu G (2019) Migration and resilience in the eastern European Neighbourhood: remittances as a mechanism for boosting recovery after shocks. In: Rouet G, Pascariu G (eds) Resilience and the EU's eastern Neighbourhood countries. From theoretical concepts to a normative agenda. Palgrave Macmillan, Cham, pp 475–514. https://doi.org/10.1007/978-3030-25606-7_16 Jackman M (2014) A note on the labor market effects of remittances in Latin American and Caribbean countries: do thresholds exist? Dev Econ 52(1):52–67. https://doi.org/10.1111/ deve.12034 Jawaid ST, Raza SA (2016) Effects of Workers’ remittances and its volatility on economic growth in South Asia. Int Migr 54(2):50–68. https://doi.org/10.1111/imig.12151 Jongwanich J (2007) Workers' remittances, economic growth and poverty in developing Asia and the Pacific countries. United Nations, Economic and Social Commission for Asia and the Pacific, Bangkok Kaczmarczyk P, Okólski M (2008) Demographic and labour-market impacts of migration on Poland. Oxf Rev Econ Policy 24(3):599–624 Katseli LT, Lucas R, Xenogiani T (2006) Effects of migration on sending countries: what do we know? OECD Publishing, Paris Kelly T (1997) Public expenditures and growth. J Dev Stud 34(1):60–84. https://doi.org/10.1080/ 00220389708422503 Kireyev A (2006) The macroeconomics of remittances the case of Tajikistan. IMF working paper no. 2. International Monetary Fund, Washington, DC Kourtit K, Newbold B, Nijkamp P, Partridge M (eds) (2020) Handbook of the economic geography of cross-border migration. Springer, Berlin Kraay AC (2015) Weak instruments in growth regressions: implications for recent cross-country evidence on inequality and growth. Policy Research working paper 7494. World Bank, Washington, DC Kubis A, Schneider L (2016) Regional migration, growth and convergence—a spatial dynamic panel model of Germany. Reg Stud 50(11):1789–1803. https://doi.org/10.1080/00343404. 2015.1059932 Kumar RR, Stauvermann PJ, Patel A, Prasad S (2018) The effect of remittances on economic growth in Kyrgyzstan and Macedonia: accounting for financial development. Int Migr 56 (1):95–126. https://doi.org/10.1111/imig.12372

7 Outmigration and Remittances as Facilitating Conditions for Economic Transition. . .

179

Lartey EKK (2013) Remittances, investment and growth in sub-Saharan Africa. J Int Trade Econ Dev 22(7):1038–1058. https://doi.org/10.1080/09638199.2011.632692 Lazaroiu S, Alexandru M (2008) Permanent or circular migration? Policy choices to address demographic decline and labour shortages in Europe. In: Hönekopp E, Mattila HR (eds) Permanent or circular migration?: policy choices to address demographic decline and labour shortages in Europe. International Organization for Migration, Regional Mission for Central and South Eastern Europe, Budapest, pp 215–258 Le T (2009) Trade, remittances, institutions, and economic growth. Int Econ J 23(3):391–408. https://doi.org/10.1080/10168730903119443 Litan C (2009) On the macroeconomic impact of remittances in Romania. Studia Universitatis Babes Bolyai Oeconomica 54(2):21–36 Lucas RE (1988) On the mechanics of economic development. J Monet Econ 22(1):3–42. https:// doi.org/10.1016/0304-3932(88)90168-7 Mankiw NG, Romer D, Weil DN (1992) A contribution to the empirics of economic growth. Q J Econ 107(2):407–437. https://doi.org/10.2307/2118477 Mara I (2012) Surveying Romanian migrants in Italy before and after the EU accession: migration plans, labour market features and social inclusion. The Vienna Institute for International Economic Studies, Vienna Massey DS, Arango J, Hugo G, Kouaouci A, Pellegrino A, Taylor JE (1993) Theories of international migration: a review and appraisal. Population Devel Rev 19(3):431–466. https://doi.org/ 10.2307/2938462 Mastrorillo M, Fagiolo G (2015) International migration and school enrollment of the left-behinds in Albania: a note. Eastern Europ Econ 53(3):242–254. https://doi.org/10.1080/00128775.2015. 1033244 McCormick B, Wahba J (2003) Return international migration and geographical inequality: the case of Egypt. J Afr Econ 12(4):500–532. https://doi.org/10.1093/jae/12.4.500 Mishra P (2015) Emigration and wages in source countries: a survey of the empirical literature. In: Lucas REB (ed) International handbook on migration and economic development. Edward Elgar Publishing, Cheltenham, pp 241–266 Naiditch C, Vranceanu R (2009) Migrant wages, remittances and recipient labour supply in a moral hazard model. Econ Syst 33(1):60–82. https://doi.org/10.1016/j.ecosys.2008.07.003 Neef R (2002) Aspects of the informal economy in a transforming country: the case of Romania. Int J Urban Reg Res 26(2):299–322. https://doi.org/10.1111/1468-2427.00381 Nijkamp P, Poot J, Sahin M (eds) (2012) Migration impact assessment; new horizons. Edward Elgar, Cheltenham Nijkamp P, Poot J, Bakens J (eds) (2015) The economics of cultural diversity. Edward Elgar, Cheltenham OECD (2013) Coping with emigration in Baltic and east European countries. OECD Publishing, Paris. https://doi.org/10.1787/9789264204928-en Prais SJ, Winsten CB (1954) Trend estimators and serial correlation. Working paper 383. Cowles Commission Ruiz I, Shukralla E, Vargas-Silva C (2009) Remittances, institutions and growth: a semiparametric study. Int Econ J 23(1):111–119. https://doi.org/10.1080/10168730802696715 Sandu D (2006a) Consequences and plans. Mentalities. Living abroad on a temporary basis: the Romanians and the economic migration: 1990–2006. Open Society Foundation, Bucharest Sandu D (2006b) Exploring Europe by work migration: 1990–2006. Living abroad on a temporary basis: the Romanians and the economic migration: 1990–2006. Open Society Foundation, Bucharest Sandu D (2010) Modernising Romanian society through temporary work abroad. In: Black R, Engbersen G, Okólski M, Pantîru C (eds) A continent moving west? : EU enlargement and labour migration from central and Eastern Europe. Amsterdam University Press, Amsterdam, pp 270–287

180

C. Incaltarau et al.

Sandu D, Alexandru M (2009) Migratia si consecintele sale. In: Preda M (ed) Riscuri si inechitati sociale in Romania. Polirom, Iasi, pp 287–304 Sandu D, Radu C, Constantinescu M, Ciobanu O (2004) A country report on Romanian migration abroad: stocks and flows after 1989. Multicultural Center, Prague Scrieciu SS, Winker P (2002) The Romanian economy in transition: developments and future prospects. Working paper. School of Business Administration at the International University in Germany, Bruchsal Senbeta A (2013) Remittances and the sources of growth. Appl Econ Lett 20(6):572–580. https:// doi.org/10.1080/13504851.2012.718057 Shioji E (2001) Composition effect of migration and regional growth in Japan. J Japanese Int Econ 15(1):29–49. https://doi.org/10.1006/jjie.2000.0461 Sikder MJU, Higgins V (2016) Remittances and social resilience of migrant households in rural Bangladesh. Migr Develop 6(2):253–275. https://doi.org/10.1080/21632324.2016.1142752 Solow RM (1956) A contribution to the theory of economic growth. Q J Econ 70(1):65–94. https:// doi.org/10.2307/1884513 Stanculescu MS, Marin M, Crai E (2010) Rapid assessment of the social and poverty impacts of the economic crisis in Romania. UNICEF, Bucharest Statistics RNIo (2014) Migratia internationala a Romaniei (Romanian International Migration), Bucharest Stoiciu V, Stănculescu M, Alexe I, Motoc L (2011) Impactul crizei economice asupra migratiei fortei de munca romanesti (The impact of the economic crisis on Romanian labor migration). Friedrich-Ebert-Stiftung, Bucharest Stough RR, Kourtit K, Nijkamp P, Blien U (eds) (2018) Modelling aging and migration effects on spatial labour markets. Springer, Berlin Tehseen Jawaid S, Raza SA (2012) Workers' remittances and economic growth in China and Korea: an empirical analysis. J Chin Econ Foreign Trade Stud 5(3):185–193. https://doi.org/10.1108/ 17544401211263946 Toth A, Toth G (2006) Consequences and plans. Entrepreneurial orientation. Living abroad on a temporary basis: the Romanians and the economic migration: 1990–2006. Open Society Foundation, Bucharest Vakulenko E (2016) Does migration lead to regional convergence in Russia? Int J Econ Policy Emerg Econ 9(1):1–25. https://doi.org/10.1504/ijepee.2016.074943 Voicu O, Toth G, Guga S (2008) Immigrant in Romania: perspectives and risks. Soros Foundation Romania, Bucharest Woodruff C, Zenteno R (2007) Migration networks and microenterprises in Mexico. J Devel Econ 82(2):509–528. https://doi.org/10.1016/j.jdeveco.2006.03.006 World Bank (WB) (2016) Migration and remittances factbook 2016, 3rd edn. World Bank Group, Washington, DC Wu S-Y, Tang J-H, Lin ES (2010) The impact of government expenditure on economic growth: how sensitive to the level of development? J Pol Model 32(6):804–817. https://doi.org/10.1016/ j.jpolmod.2010.05.011 Zamfir C, Stanescu I, Ilie S (2010) Dupa 20 de ani: optiuni pentru Romania (After 20 years: options for Romania). Romanian Academy, Romanian National Institute for Economic Research, Research Institute for quality of life, Bucharest

Chapter 8

Retirement, Relocation, and Residential Choices Ayoung Kim and Brigitte S. Waldorf

8.1

Introduction

In today’s aging societies, a good deal of the older population is faced with the decision of when to exit the labor force. The decision is often made jointly with a locational choice and a choice about housing consumption (Banks et al. 2012). Prior to retirement, the workplace location strongly influences where people live. Upon retirement, however, other factors such as closeness to children and grandchildren, climate, and amenities take on a pivotal role (Whisler et al. 2008; Wiseman and Roseman 1979; Serow 2001). Moreover—upon retirement—households typically receive less income and change their consumption patterns, spending relatively more on housing, food, and healthcare than on clothing, transportation, and household furnishing (Lee et al. 2014). As a result, many households downsize their home (Bian 2016) and move to places that are more affordable and easily accessible in anticipation that health and strength may become compromised (Abramsson and Andersson 2016). As people make these decisions based on their preferences and income constraints, we observe a multitude of “types” of retirees, ranging from wealthy golfplaying retirees in upscale resorts to retirees on small pensions below the poverty threshold. To better understand the wide range of possible manifestations of these linked decisions, this research focuses on the interplay of retirement, migration, locational choice, and housing consumption.

A. Kim (*) Department of Agricultural Economics, Mississippi State University, Starkville, MS, USA e-mail: [email protected] B. S. Waldorf Department of Agricultural Economics, Purdue University, West Lafayette, IN, USA e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 W. Cochrane et al. (eds.), Labor Markets, Migration, and Mobility, New Frontiers in Regional Science: Asian Perspectives 45, https://doi.org/10.1007/978-981-15-9275-1_8

181

182

A. Kim and B. S. Waldorf

Using data from the 2006 to 2017 American Community Surveys (ACS) of the United States Census Bureau, we select information on recent or new retirees, i.e., those who transitioned into retirement during the previous 12 months. As such, our study links mobility and housing consumption decisions not solely to age and aging, but more directly to the decision to retire and actually exiting the labor force. In general, new retirees are significantly more likely to move than comparable elderly who are still in the labor force or who have been retired for more than a year. However, we know little about differential migration propensities among recent retirees, what factors influence their destination choice, and how moving and retirement affect their housing consumption. Kim and Waldorf (2019) find that retired baby boomers who relocate have above-average incomes, leaving behind those with less economic power. In this paper, we focus on demographic attributes and show that they are actually more decisive than income when it comes to factors influencing retirees’ choices regarding whether to move upon exiting the labor forces, where to move and how much housing to consume. We address three specific research questions surrounding recent retirees’ mobility behavior. First, we ask how demographic and socio-economic attributes influence new retirees’ decisions to move. While the vast majority of new retirees do not move, those who move can choose from a large array of destinations. Our second research question thus investigates how demographic and socio-economic attributes affect retirees’ decisions for different types of destinations whereby we distinguish between short-distance moves and long-distance moves. We hypothesize that while the propensity to make a long-(short-) distance move increases (decreases) with income, the influential factors are age, marital status, and disability status. The third question focuses on differences in housing consumption between movers and stayers. We investigate the hypothesis that retirees who move also reduce their housing consumption, and that the type of move influences by how much they reduce their consumption. Ultimately, a synthesis of the various models suggests that the “dream” of moving and downsizing following retirement is more likely to be realized by healthy people who transition into retirement at a young age and whose spouse—if married—also exited the labor force. The chapter is organized as follows. The next section provides background information on later-life migration and housing consumption. This is followed by the empirical section, including the presentation of methods, data, and results. The chapter ends with a summary and conclusion.

8.2

Retirement and Migration in Aging Societies

The causes and consequences of aging populations are thoroughly discussed at global, national, and regional scales (for example, Alimi et al. 2018; Waldorf 2018; Lutz 2019). At the regional level, selective migration plays a pivotal role. Often the emphasis is on the migration behavior of young people, especially young educated people moving to growing urban agglomerations (Poot et al. 2008;

8 Retirement, Relocation, and Residential Choices

183

Waldorf 2009). Yet, migration of the older population1 also affects how aging societies take form at the local and regional level (Serow 2003). Since Wiseman and Roseman (1979) presented their typology of elderly migration, we have seen a remarkable growth of the literature on senior migration (for example, Serow 1987; Clark et al. 1996; Walters 2002; Plane and Heins 2003; Dorfman and Mandich 2016; Kim and Waldorf 2019). The studies have focused on the reasons why older people move, where they move, the personal and locational factors influencing their choices, and what the aggregate spatial consequences are. Age is the defining characteristic in these studies, directing our attention to the age-dependency of migration propensities, sometimes making rather vague distinctions between younger and older seniors, or delineating special age groups like the oldest old. The focus on age is understandable given that age is an easily identifiable and recorded attribute. Moreover, it is often used as a proxy for age-related attributes like, for example, frailty or being retired. However, age is neither sufficient to capture the full diversity of the older population nor can age be portrayed as a trigger for people’s actions or inactions. The older population is, as Wiseman and Roseman (1979) already proclaimed decades ago, a heterogeneous population. It is becoming even more diverse as, on average, people live longer, are healthier, and often choose to work well beyond the traditional retirement age of 65 years. Moreover, pension systems in aging societies often require that full retirement age must be pushed upwards. As a consequence, stages in the life cycle of older people can no longer be easily matched with particular age ranges and the relevance of a migration age schedule (Rogers 1988) becomes questionable. Instead, understanding older persons’ migration behavior requires a focus on life events such as transitioning into retirement. It is the major event in older persons’ lives, and their plans and dreams are anchored around that transition. Migration upon transitioning into retirement also implies choosing a new residence. The options are manifold, ranging from residing in a recreational vehicle to residing in a single-family home. Banks et al. (2012) allude to downsizing and housing costs as factors influencing the residential choices of retirees. Judd et al. (2012) emphasize that downsizing potentially has two components: (1) choosing a smaller house/less land and (2) choosing a less expensive residence. Choosing a smaller residence will reduce maintenance and free retirees of the burden of housework. Choosing a more affordable home will match the often lower incomes after people retire from their job. The empirical analysis presented in the next section contributes to the literature for three reasons. First, the analysis links migration directly to the transition into retirement; age is not neglected but it is associated with the age at which a person chooses to exit the labor force. Second, it also differentiates by moving distance which—in the U.S. context—can be proxied by whether a household moves within

1 Interestingly, migration at older age is more common in the USA than in Europe where the elderly tend to age in place (Banks et al. 2012; Kramer and Pfaffenbach 2016).

184

A. Kim and B. S. Waldorf

Table 8.1 Research design Research question 1. Factors influencing decision to move; relative importance of income 2. Differential impact of demographic and economic attributes on the type of move (long-distance or out-of-state versus short distances or in-state move)

Type of model Model 1. Logit

Dependent variable Moving vs. staying

Model 2. Multinomial logit

3. Differential reduction of housing consumption by type of move

Model 3. Negative binomial Model 4. Logit

Moving status M ¼ 1, 2, 3 corresponding to: 1 ¼ does not move 2 ¼ in-state move 3 ¼ interstate move # of rooms Homeownership

Explanatory variables Income Education Age, sex, race marital status Disability Metro

Fixed effects Year t State at t-1

Moving status M ¼ 1, 2, 3 Income Education Age, sex, race marital status Disability Metro

the state or moves across state boundaries. Relocating across state boundaries involves substantial changes, ranging from building a new social network, to creating new activity spaces, to obtaining a new driver’s license and voter registration. Relocating across state boundaries is also quite costly. Pecuniary moving costs are of course higher, but so are the non-pecuniary costs of leaving behind family, friends, and familiar service providers such as doctors and even hairdressers. We argue that the factors influencing the two types of moves differ substantially. Finally, most migration studies do not consider the residential choices that people make when settling in the destination. Our analysis looks at the residential choices of those retirement migrants and compares them to the housing consumption of stayers. Thus, it will shed light on the diverse housing preferences of the older population, in particular, assess to what extent retirement migration is associated with reduced housing consumption.

8.3 8.3.1

Empirical Analysis Research Design

We design the empirical analysis for retirees in the USA who have been out of the labor force for less than 1 year. Table 8.1 summarizes the various elements of the research design. The first research question looks at the factors influencing the

8 Retirement, Relocation, and Residential Choices

185

propensity to move after transitioning into retirement. We choose a logit model to capture the systematic variations in the binary dependent variable (moving versus staying). The set includes both economic variables, i.e., total family income and educational attainment, and demographic variables, i.e., age, race, sex, marital status, disability status, and metropolitan status. Education is used as a wealth indicator. Education proxies the permanent income over a person’s working life and is strongly correlated with wealth (Wolla and Sullivan 2017). In addition, we capture spatial and temporal variation using fixed effects. The second research question further differentiates the moving status by including the destination choice, operationalized as moving within the state of residence (short-distance move) versus moving to a different state (long-distance move). This yields a limited dependent variable with three possible outcomes and a multinomial logit model is chosen. The right-hand side variables are the same as for the first research question. The third research question turns to mover/stayer differences in housing consumption. The overall hypothesis is that moves following the transition into retirement are associated with reduced housing consumption. We first operationalize housing consumption as the household’s number of rooms and model it using a negative binomial model. We then operationalize housing consumption as binary variable owning versus renting and use a logit model. In both cases, moving status is the key explanatory variable, and the set of controls includes the demographic and socio-economic variables.

8.3.2

Data

The empirical analysis is based on the American Community Surveys (ACS) conducted annually between 2006 and 2017. The ACS are cross-sectional household surveys administered by the U.S. Census Bureau2 and provide detailed demographic and socio-economic data for sampled households and their household members. Data and variable names are taken from ACS 2006 to 2017 via IPUMS USA (Ruggles et al. 2020). Our extracted sub-sample of retirees is based on carefully chosen selection criteria. An important feature of this study is that we are not interested in retirees per se, but only in those who exited the labor force in the previous year. Thus, the main sample selection criterion uses the combination of labor force status (labforce) at the time of the survey and whether the respondent worked 1 year prior to the survey (workedyr). As shown in Table 8.2, we select household heads who are not in the labor force at the time of the survey but were in the labor force a year earlier. To exclude outliers, we also require that the selected person is at least 60 years old. If married, we exclude those with very young spouses (under age 50). Moreover, we

2

We accessed the data from Minnesota University’s IPUMS USA (Ruggles et al. 2020).

186

A. Kim and B. S. Waldorf

Table 8.2 Sample selection Criterion Relate ¼ 1 labforce ¼ 1 workedyr ¼ 3 age  60 If marst ¼ 1, 2 ageSpouse  50 Race ¼ 1, 2 migrate1 ¼ 1, 2, 3

Explanation Household head Not in labor force Worked last year At least 60 years old Spouses of married household heads are at least 50 years old Respondent is black or white Did not move from abroad during the previous year

only select blacks and whites since other racial groups constitute quite small minorities among the retired population. This yields a total of n ¼ 221,942 observations in the sample, representing slightly more than 17 million household heads. Finally, we exclude those who moved from abroad during the previous year. Table 8.3 shows the sample summary statistics, broken down by moving status. The vast majority of households—93.1%—did not move during that first year of entering retirement. Among the movers, 65% chose a destination within their own state. With respect to demographic characteristics, recent retirees are, on average, almost 68 years of age, with intrastate movers being about a year and interstate movers even 2 years younger than stayers. Marital status also exhibits a noticeable, but not unexpected mover-stayer difference. Movers are substantially less likely to be married than stayers. That is most noticeable case for intrastate movers. Moreover, among those who are married, the spouse’s labor status is a decisive factor. Among married couples consisting of a recent retiree and a working spouse, moving is a rare event: 96% stay put, 2.8% move within the state, and only 1.4% move across state boundaries. Among married couples consisting of a recent retiree and a non-working spouse, the share of movers increases from 4.3% to 5.8%. Among married couples who move, the share of movers crossing state boundaries rises from 33% to 47%. Race and ethnicity also play important roles because—compared to the majority population (white and non-Hispanic)—blacks and Hispanics are more likely to move short distances (within own states) and are less likely to move across state boundaries. Since minority status in the USA is strongly associated with low economic power, we find the minority pattern replicated when matching migration status with income, educational attainment, and housing consumption. Average income is lowest among recent retires who move short distances and highest among longdistance movers. Stayers’ average income takes on a middle position. Highly educated retirees are overrepresented among the long-distance movers, those without a college education are overrepresented among short-distance movers. Homeownership is highest among stayers. Among those who moved, it is substantially higher for long-distance movers than for those moving within the state. The last set of variables includes several disability indicators, a very important consideration for the elderly population. Retirees’ health status influences whether to move and where to move. In our sample, we find that retirees with a disability are

8 Retirement, Relocation, and Residential Choices

187

Table 8.3 Summary statisticsa

Demographics Age: Mean (median) Female Family size Married, spouse in LFb Married, spouse not in LFb Single, widowed, separated/divorce White Black Hispanic Economics Total family incomec Educational attainment Less than high school High school Some college Bachelor Master degree or higher Housing consumption Home ownership Number of rooms Number of bedrooms Disability status Cognitive difficulty Ambulatory difficulty Indep. Living difficulty Self-care difficulty Vision / hearing difficulty

Stayers n ¼ 15,890,365

Movers within state n ¼ 766,819

across state boundaries n ¼ 405,531

68 (67) 40% 1.87 21% 38% 41% 92% 8% 3.7%

67 (66) 47% 1.63 13% 26% 61% 90% 10% 4.5%

66 (65) 42% 1.73 12% 43% 46% 95% 5% 2.2%

$75,079

$60,416

$85,054

8% 25% 29% 19% 20%

11% 25% 31% 18% 16%

4% 18% 32% 22% 25%

89% 6 4

52% 5 3

66% 6 4

3.2% 11.2% 3.7% 2.2% 10.8%

6.0% 17.9% 6.4% 3.8% 11.6%

2.4% 9.3% 2.4% 1.0% 8.4%

a

Weighted by perwt LF ¼ labor force c Inflation adjusted value in 2017 dollars b

under-represented among long-distance movers. Instead, retirees with a disability are more likely to stay put or move within the state.

188

A. Kim and B. S. Waldorf

8.3.3

Results

8.3.3.1

The Decision to Move upon Transitioning into Retirement

Table 8.4 summarizes the model estimates of recent retirees’ migration decisions. Model 1 is a logit model that estimates the probability of relocating, thus the choice options are staying put versus moving. All estimates are significantly different from zero. They suggest that the propensity to move increases with increasing income, diminishes with increasing age, is greater for retirees with a college degree than without a college degree, is smaller for women than for men, smaller for blacks than for whites, smaller for retirees without a disability than for retirees with a disability, and smaller for married than for unmarried retirees. Note that the marital status variable distinguishes not only between married and unmarried retirees, but also whether the spouse is still in the labor force. Model 1, for example, estimates that recent retirees’ odds of staying versus moving is 1.6 times higher if a non-working spouse is present, but 2.5 times higher if the spouse is still in the labor force. To demonstrate the importance of the variables, the standard approach is to present marginal effects or the relative odds. We are taking a different route and discuss the probabilities of moving for two “types” of retirees that represent low and high moving propensities. The first type, Type-A, has characteristics that are associated with a low propensity to move. That is, the Type-A retiree is a married black

Table 8.4 Estimation results of migration models Model 1

Intercept Fam income BA or more Age Age squared Female White M spouse in LF M spouse not in LF Disability Metrot-1 Model type Fixed effect on n Weighted n -2LogL

Movers vs. stayers b SEb 7.077 0.095 0.0003 0.000 0.006 0.002 0.224 0.003 0.001 0.000 0.013 0.002 0.024 0.004 0.916 0.003 0.465 0.002 0.209 0.002 0.034 0.002 Logit Yeart and Statest-1 221,942 17,062,715 8,321,903

Model 2 Short-distance movers vs. stayers b SEb 6.616 0.111 0.001 0.000 0.104 0.003 0.239 0.003 0.001 0.000 0.025 0.003 0.146 0.004 0.879 0.004 0.709 0.003 0.336 0.003 0.119 0.003 Multinomial Logit Yeart and Statest-1 221,942 17,062,715 9,701,575

Long-distance movers vs. stayers b SEb 1.334 0.192 0.001 0.000 0.210 0.003 0.070 0.005 [0.00003] 0.000 0.002 0.004 0.469 0.007 0.915 0.005 0.037 0.004 0.097 0.004 0.150 0.004

Note: estimates in brackets are not significantly different from zero at α ¼ 5%

8 Retirement, Relocation, and Residential Choices

189

Fig. 8.1 Estimated Probabilities of Moving for Selected Scenarios. Note: Type-A retiree is a married black woman without a college degree. She does not have any disability and lived in a metro area. Her husband is working, and the family income is $60,000. Type-B is an unmarried white college-educated man who has a disability. His income is $60,000, and he did not live in a metro area at t-1. The scenarios are described in the text

woman without a college degree. She does not have any disability and lived in a metro area during the previous year. Her husband is still working, and the family income is $60,000.3 As shown in Fig. 8.1, there are three variables that make a substantial difference in her propensity to move. The first one is the age at which she enters retirement. Her probability of moving is about 6.74% if she enters retirement at age 60, but it drops significantly to only 3.98% if she waits to retire until age 70. The second variable that makes a big difference is her marital status. Her moving probabilities increase by slightly more than 50% if her husband is not working (scenario I). Her moving probabilities are more than double if she is not married (scenario II). The third variable that is not only significant but also has a sizable impact is the disability status. Her moving probabilities increase by about 22% if she were disabled (scenario III). Varying Type-A’s income barely affects the probability of moving. Statistically significant but of small numerical magnitude is also the effects of sex, race, and metropolitan status.

3

The fixed effects for the calculations are Year ¼ 2017 and State ¼ Wyoming.

190

A. Kim and B. S. Waldorf

To illustrate the range of the predicted probabilities, we also introduce Type-B that encompasses attributes associated with a high moving probability. Type-B is an unmarried white college-educated man who has a disability and did not live in a metropolitan area during the previous year. To make him comparable to Type-A, we assign him a family income of $60,000 and vary the age of entering into retirement from 60 to 70. If he enters retirement at the age of 60, he will move with a probability of almost 20%. Should he work 10 years longer, then his chance of moving upon entering retirement will drop to 12.1%. Type-B retirees represent the upper limit of the predicted moving probabilities, and changing marital status or disability status will substantially drop the chances of moving. Model 2 is a multinomial choice model and represents a more differentiated perspective that distinguishes not only between staying and moving, but also between moving short-distances (within the state) and long-distances (across state boundaries). That is, households are assumed to choose between staying in the current residence, moving to another residence in the same state, and moving to a residence in a different state. Staying in the current residence is the reference category of the estimated parameters. Almost all variables are estimated to have a non-zero effect on the probabilities of moving4 and, interestingly, the direction of the effects is not always identical for long- and short-distance moves. Income is negatively associated with the propensity to move within the state, but positively associated with the propensity to move longer distances across state boundaries. In both cases, however, the magnitudes of the estimated effects are negligible. For example, a Type-A retiree earning $30,000 has an estimated 2.93% probability of moving within the state. If her income triples to $90,000, then the estimated probability slightly diminishes to 2.81%. In contrast, her probabilities of migrating across state boundaries slightly increase from 2.96% when having an income of $30,000 to 3.17% with a $90,000 income. Educational attainment also has opposite effects on the two types of moves. Having a college degree lowers the probability of making an intrastate move. The magnitudes of the effect are small, with the odds ratio of being close to 1. On the other hand, having a college degree positively affects the propensity to move across state borders, with an estimated odds ratio of 1.2. Race and disability status have opposite effects on intra- and interstate moves. Being white increases (decreases) the probability of moving across state boundaries (within the state). Having a disability increases the probability of an intrastate move and lowers the probability of moving across state boundaries. The marital status variable is the most important variable shaping the choice probabilities of recent retirees. Figures 8.2 and 8.3 show estimated probabilities for short-distance and longdistance migration compared to the probability of staying. Turning first to the shortdistance (in-state moves), a 60-year old Type-A retiree moves within the state with a

4 Only the parameter estimates for the squared age variable and the variable female are not significantly different from zero for the migration across state borders.

Fig. 8.2 Estimated Probabilities for Short-Distance (left) and Long-Distance (right) Migration. Note: Type-A retiree is a married black woman without a college degree. She does not have any disability and lived in a metro area. Her husband is working, and the family income is $60,000. Type-B is an unmarried white college-educated man who has a disability. His income is $60,000, and he did not live in a metro area. The scenarios are described in the text

8 Retirement, Relocation, and Residential Choices 191

192

A. Kim and B. S. Waldorf

Fig. 8.3 Ratio of Long-Distance Migration Probabilities and Short-Distance Migration Probabilities. Note: Type-A retiree is a married black woman without a college degree. She does not have any disability and lived in a metro area. Her husband is working, and the family income is $60,000. Type-B is an unmarried white college-educated man who has a disability. His income is $60,000, and he did not live in a metro area. The scenarios are described in the text

probability of less than 3%. At age 70, the probability is cut by more than a percentage point to only 1.8%. If the husband of a Type-A retiree stops working, these probabilities increase slightly. Apparently, for short-distance moves, the spouse’s labor force status is not a decisive factor. If, however, the Type-A retiree is not married, then the in-state moving probabilities are more than double. If the Type-A retiree becomes disabled, then the probability of moving within the state increases by around 50 percent. For long-distance migration propensities, the demographic variables operate very differently than in the case of short-distance migration. If a Type-A retiree’s husband does not work, the household’s long-distance migration propensities increase substantially, and they are estimated to be as high as if the Type-A retiree was not married. Thus, there is no difference between the scenarios I and II. If Type-A retiree becomes disabled, then long-distance moves become slightly less likely. Thus, Type-A as originally defined does not constitute the lower boundary of the longdistance migration probabilities. Migration propensities typically follow a distance decay pattern whereby longer moves are less likely than shorter moves. However, what we see for recent retirees is a remarkable deviation from the distance decay pattern. For several scenarios, we find that the probabilities of making a long-distance move across state borders far exceed those of making shorter in-state moves.

8 Retirement, Relocation, and Residential Choices

193

Table 8.5 Estimation results: housing consumption

Intercept Short-distance mover Long-distance mover Total family income BA or higher Age Age squared Female White Married, spouse in LF Married, spouse not in LF Disability Metrot-1 Model type Fixed effects n Weighted n -2LogL LogL AIC

Model 3 Number of rooms b SEb 1.568 0.082 0.155 0.005 0.143 0.006 0.001 0.000 0.089 0.002 0.005 0.002 0.000 0.000 [0.001] 0.002 0.017 0.004 0.140 0.002 0.133 0.002 0.030 0.002 0.013 0.002 Negative binomial Yeart and Statest-1 221,942 17,062,715

Model 4 Ownership b SEb 5.211 0.070 1.943 0.003 1.654 0.004 0.008 0.000 0.305 0.002 0.156 0.002 0.001 0.000 0.221 0.002 0.862 0.002 1.187 0.003 1.335 0.002 0.420 0.002 0.315 0.002 Logit Yeart and Statest-1 221,942 17,062,715 10,939,941

1,262,976 952,897

Note: estimates in brackets are not significantly different from zero at α ¼ 5%

Figure 8.3 displays the ratios of estimated long versus short-distance migration probabilities for various scenarios. If the migration behavior follows the typical distance decay pattern, then the ratios will be substantially smaller than one. However, we see that at age 60, the ratios are greater than one for all groups except for retirees with a disability (scenario III). With increasing age, short-distance migration becomes more prevalent relative to long-distance migration and the ratio declines. For 70-year old Type-A retirees and 70-year old retirees in scenario II, the ratio even drops below one. Note that married retirees whose spouse is not in the labor force (scenario I) are the most likely to move across state borders. In fact, at age 60, they are more than twice as likely to make a long-distance move to another state than to move within the state.

8.3.3.2

Housing Consumption

Table 8.5 reports the estimation results of the two housing consumption models. Model 3 summarizes the estimates of the negative binomial model predicting the number of rooms. The estimates are significant for all variables except for the gender variable. Income, age, having a college degree, being white, and being married are

194

A. Kim and B. S. Waldorf

positively related to the number of rooms. Having a disability lowers the expected number of rooms. The magnitude for most of these effects are, however, barely influential. What makes a difference is the moving status. Compared to stayers, movers have a smaller house or apartment. On average, it has one room less. A 65-year old TypeA retiree, for example, who does not move, has 6.93 rooms in her place. If she moved to another residence within the same state, she downsized to a residence with 5.94 rooms. If she moved to another residence in a different state, she downsized by a little bit less, choosing a residence with 6.01 rooms. Also influential is marital status. Being married is, on average, associated with almost one additional room. This additional housing consumption does not vary by whether the spouse is in the labor force. The logit model on homeownership paints a similar picture of reduced housing consumption upon retiring and relocating. Moving status is by far the most influential factor for predicting homeownership among recent retirees. Reduced homeownership rates among movers with the reduction slightly bigger for those who move short distances. For example, a 65-year old Type-A retiree who stays has an 88.7% chance that she owns her home. If she moved within the state, her homeownership probability drops substantially by 36 percentage points to 52.9 percent. If she made a long-distance move to a different state, the drop is a little bit smaller, amounting to 29% points. Marital status and disability status also have decisive impacts on homeownership. If the Type-A retiree is not married, her probability of homeownership drops by 18 percentage points if she does not move, and by about 28 percentage points if she moves. If she has a disability, the homeownership probabilities also decline but by a substantially smaller amount, 5 percentage points if she does not move, 10 percentage points if she moves.

8.4

Summary and Conclusions

This paper focuses on the subset of the older people that just exited the labor force and plan arrangements for retirement. No longer bound to living close to the place of work, many older people wish to move to a location where they can enjoy cultural amenities, a pleasant climate, or proximity to children and grandchildren. Such plans typically include a desire to adjust their housing consumption. Downsizing is often the preferred option after children left the parental house and parents desire to have a hassle-free lifestyle without having to put too much time and energy into housework and gardening. This paper analyzes the nexus between retirement, relocation, and housing consumption. Our empirical analysis—using a large sample of recent retirees in the USA—suggests four noteworthy results. First, we find that income is not the decisive factor influencing whether recent retirees are moving. Instead, age, marital status, and disability status strongly affect the chances that a retiree will relocate after

8 Retirement, Relocation, and Residential Choices

195

transitioning into retirement. Second, when differentiating the decision to move by migration destination, we find that several factors—including income, educational attainment, and disability status—have opposite effects on short-distance moves versus long-distance moves. Marital status plays by far the most influential role for both short- and long-distance moves. In particular, recent retirees who are unmarried or whose spouse is also not in the labor force have remarkably high propensities to make a long-distance move. Third, we find that the moving behavior upon entering retirement is not always compatible with the distance decay of typical migration patterns. That is, long-distance moves are more likely than short-distance moves. This is particularly the case for the able-bodied young retirees with a non-working spouse. Lastly, we find very consistent results regarding reduced housing consumption following retirement-induced migration. Compared to stayers, retirees who move have residences with fewer rooms and are more likely to rent rather than own their residences. The reduction in housing consumption is slightly more pronounced among short-distance movers than among long-distance movers. Synthesizing the results suggests that retirement in that perfect location is primarily a question of demographics and health. People who keep working well into their 70s are not likely to move to Florida or other warm places when they eventually exit the labor force. An interesting group is composed of married couples with a substantial age difference, often with the husband being older than the wife. Our findings suggest that a working spouse deters retirement migration and it is not clear whether, and under what circumstances, the working spouse is willing to retire early so as to realize the dream. Perceptions of differential health, life expectancy, economic power, and independence may all play into such considerations. Future research should focus on two extensions. First, measuring the influence of income and wealth on migration needs to be improved. The data only allow us to consider current income rather than life time income. Moreover, we cannot measure people’s wealth directly but only proxy wealth using an education variable. Second, the sample analyzed in this study links retirement, migration, and housing consumption more directly than samples delineated by age-thresholds. Nevertheless, future research should make these linkages even more explicitly by using panel data. Panel data will allow us to reconstruct the migration and housing consumption histories of people before and after they transitioned into retirement. For example, it may very well be that people already downsize before retiring, or that they first choose to rent before deciding on buying a home in their preferred retirement location. Without detailed migration and consumption histories, these issues will remain unresolved.

References Abramsson M, Andersson E (2016) Changing preferences with ageing–housing choices and housing plans of older people. Hous Theory Soc 33(2):217–241

196

A. Kim and B. S. Waldorf

Alimi OB, Maré DC, Poot J (2018) More pensioners, less income inequality? The impact of changing age composition on inequality in big cities and elsewhere. In: Modelling aging and migration effects on spatial labor markets. Springer, Cham, pp 133–159 Banks J, Blundell R, Oldfield Z, Smith JP (2012) Housing mobility and downsizing at older ages in Britain and the USA. Economica 79(313):1–26 Bian X (2016) Leverage and elderly homeowners’ decisions to downsize. Hous Stud 31(1):20–41 Clark DE, Knapp TA, White NE (1996) Personal and location-specific characteristics and elderly interstate migration. Growth Chang 27(3):327–351 Dorfman JH, Mandich AM (2016) Senior migration: spatial considerations of amenity and health access drivers. J Reg Sci 56(1):96–133 Judd B, Bridge C, Davy L, Adams T, Liu E (2012) Downsizing in later life: myths and realities concerning the movement of older people in the housing market. In: Workshop, vol 15 Kim A, Waldorf BS (2019) Chapter 12: baby boomers’ paths into retirement. In: Franklin R (ed) Population, place, and spatial interaction: essays in honor of David Plane. Springer Verlag, Singapore, pp 225–247 Kramer C, Pfaffenbach C (2016) Should I stay or should I go? Housing preferences upon retirement in Germany. J Housing Built Environ 31(2):239–256 Lee S, Sohn SH, Rhee E, Lee YG, Zan H (2014) Consumption patterns and economic status of older households in the United States. Mon Labor Rev 2014(9):1–19 Lutz W (2019) World population trends: global and regional interactions between population and environment. In: Population and environment. Routledge, Warszawa, pp 41–65 Plane DA, Heins F (2003) Age articulation of US inter-metropolitan migration flows. Ann Reg Sci 37(1):107–130 Poot J, Waldorf B, van Wissen L (2008) Migration in a globalised world: a new paradigm. In: Poot J, Waldorf B, van Wissen L (eds) Migration and human capital. Edward Elgar, Cheltenham Rogers A (1988) Age patterns of elderly migration: an international comparison. Demography 25 (3):355–370 Ruggles S, Flood S, Goeken R, Grover J, Meyer E, Pacas J Sobek M (2020) IPUMS USA: Version 10.0 [dataset]. Minneapolis, MN: IPUMS. https://doi.org/10.18128/D010.V10.0 Serow WJ (1987) Determinants of interstate migration: differences between elderly and nonelderly movers. J Gerontol 42(1):95–100 Serow WJ (2001) Retirement migration counties in the southeastern United States: geographic, demographic, and economic correlates. The Gerontologist 41(2):220–228 Serow WJ (2003) Economic consequences of retiree concentrations: a review of north American studies. The Gerontologist 43(6):897–903 Waldorf BS (2009) Is human capital accumulation a self-propelling process? Comparing educational attainment levels of movers and stayers. Ann Reg Sci 43(2):323–344 Waldorf BS (2018) Inhabitants of earth. In: How to feed the world. Island Press, Washington, pp 5–23 Walters WH (2002) Later-life migration in the United States: a review of recent research. J Plan Lit 17(1):37–66 Whisler RL, Waldorf BS, Mulligan GF, Plane DA (2008) Quality of life and the migration of the college-educated: a life-course approach. Growth Chang 39(1):58–94 Wiseman RF, Roseman CC (1979) A typology of elderly migration based on the decision making process. Econ Geogr 55(4):324–337 Wolla SA, Sullivan J (2017) Education, income, and wealth. Federal Reserve Bank of St. Louis Research, St. Louis

Chapter 9

The Development of Uncertainty in National and Subnational Population Projections: A New Zealand Perspective Michael P. Cameron, Kim Dunstan, and Len Cook

Abstract Projections or forecasts of future population, and the age and sex structure of the population, are key inputs to decision-making. However, these projections have an associated uncertainty that is often underappreciated by decision-makers. Moreover, a decision-maker faced with multiple population projections has no clear basis on which to decide between alternative projections. In this chapter, we outline the sources of uncertainty in population projections. We then describe the development of population projection methods in New Zealand, where the representation of uncertainty has become central to official projections produced by Statistics New Zealand, as well as alternative projections produced by the National Institute of Demographic and Economic Analysis. Finally, we outline a model-averaging approach that can be used by decision-makers to combine the information from multiple independent population projections. This may provide a more accurate approach to the use of population projections in decision-making, without requiring substantial modelling capability. This approach is illustrated with the example of subnational areas for the Waikato Region of New Zealand.

M. P. Cameron (*) School of Accounting, Finance and Economics, University of Waikato and National Institute of Demographic and Economic Analysis, University of Waikato, Hamilton, New Zealand e-mail: [email protected] K. Dunstan Statistics New Zealand, Wellington, New Zealand e-mail: [email protected] L. Cook National Institute of Demographic and Economic Analysis, University of Waikato, Hamilton, New Zealand © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 W. Cochrane et al. (eds.), Labor Markets, Migration, and Mobility, New Frontiers in Regional Science: Asian Perspectives 45, https://doi.org/10.1007/978-981-15-9275-1_9

197

198

9.1

M. P. Cameron et al.

Introduction

Projections (or forecasts)1 of the future population are a key input for national and local government decision-makers, planners, private sector developers, businesses, community and advocacy groups, and many others. Having an indication of the future trajectory of population is valuable for guiding decisions on planning infrastructure and service provision, among other things (e.g. Auerbach and Lee 2001). Population projections can be used to anticipate and plan for future demands for public services and infrastructure, with lead times that reflect the nature of the resource base that will be required. In most cases, population projections not only seek to establish the future level of the population, but also the age and sex structure of the population and periods of differential rates of growth. Health, education, and income support are key areas of public policy that involve long payback periods for investment, and where the investment varies across age groups. For example, in New Zealand, the ‘From Birth to Death’ reports of the Social Monitoring Group of the Planning Council in the 1980s spanned a comprehensive range of areas, providing analysis of future trends with relevance to social policy. This required not only projections of the population level, but also the age structure of the population. These reports were used extensively in the development of social policy throughout this period. However, as American baseball player Yogi Berra is reputed to have once quipped, ‘It’s tough to make predictions, especially about the future’. Projections of future population are no different in this respect and come with a high degree of ex ante uncertainty, particularly for small populations and over longer projection horizons (Cameron and Poot 2011). Ex post, when compared with actual populations, all projections are imperfect and result in an error, where the actual population differs from what was projected. The tolerance of projection error will differ with the use of the projection. There are different costs arising from underestimating demand compared with overestimating demand. These vary between service and infrastructure types depending on the economic, social, and environmental costs of building overcapacity relative to the costs of building insufficient capacity. For example, in the mid-1970s when population projections had yet to take into account the impact of

1

The distinction between a projection and a forecast may appear to be semantic, but the distinction is fundamental to how end-users should interpret and use these information tools. A projection is a realization of a model based on a particular (and known) combination of assumptions in terms of the model parameters. Thus, the projection is fully determined by the choice of parameters (which might be chosen empirically or stochastically, as we discuss later in the chapter). Moreover, a projection is based on current policy settings and does not typically try to anticipate future policy changes, some of which may be in response to projected population changes themselves. In contrast, a forecast is a prediction of a likely future. It is the preference for a particular set of parameter values that turns a projection into a forecast. For further discussion of this point, see Dion and Galbraith (2015). We prefer to use the term ‘projection’, which we maintain throughout the chapter, although we recognize that many end-users will interpret projections as forecasts.

9 The Development of Uncertainty in National and Subnational Population. . .

199

the then rapid decline in fertility, electricity planners were examining the need for nuclear generation. A Royal Commission on Nuclear Power Generation in New Zealand was set up in 1976, and by the time it had reported back to the Government in 1978, population projections had lowered, and the continued use of hydro-electric energy was judged as likely to meet foreseeable future need. As this example suggests, the many and varied uses of population projections require a consideration of both error and uncertainty, which has not always been well conveyed by conventional deterministic2 projections. In this chapter, we draw on a long literature considering the sources of error and uncertainty in population projections (Alho and Spencer 1985, 1997; Alho 1990; Alho et al. 2008). We use New Zealand as a case study, describing the historical development of population projections, from early efforts to more recent methodological advances. New Zealand is an appropriate case study, as Statistics New Zealand3 (Stats NZ) has long been at the forefront of developments in population projections methodology, most recently in the adoption of stochastic population projections at the national level.4 Importantly, the stochastic approach to projections helps to convey uncertainty in the projections in a much more meaningful way than the traditional deterministic approach. Alongside the official projections produced by Stats NZ, staff at the National Institute of Demographic and Economic Analysis (NIDEA; formerly the Population Studies Centre) at the University of Waikato, led by Emeritus Professor Jacques Poot (now at the Vrije Universitat Amsterdam), pioneered the development of stochastic subnational population projections. Most recently, NIDEA has incorporated a gravity model of internal migration (Poot et al. 2016) into the population projections framework (Cameron and Poot 2014a, b). The approaches adopted by Stats NZ and NIDEA are different but complementary, and we argue that taking a model-averaging approach across these two different methodological approaches may actually lead to less dependence on a single projection source and better decision-making. The remainder of the chapter proceeds as follows: First, we briefly discuss general methodological approaches to population projections, distinguishing time series econometric approaches from the cohort component model. We also describe the key differences between top-down and bottom-up approaches to subnational population projections. In Sect. 9.3, we identify the main sources of uncertainty in population projections, particularly model uncertainty and parameter uncertainty. Section 9.4 outlines the historical development of population projection methods in New Zealand, at both Stats NZ and NIDEA, and Sect. 9.5 briefly describes the ex post accuracy of projections. Finally, Sect. 9.6 describes a model-averaging 2

A deterministic projection typically combines a given set of input parameters, compared with a stochastic projection which varies randomly according to the probability distributions of those input parameters. 3 For simplicity, we refer to Stats NZ, although before 1995 the organization was known as the Department of Statistics. 4 New Zealand is currently one of only a few countries, along with the Netherlands, where the national statistics agency produces official population projections using a stochastic approach.

200

M. P. Cameron et al.

approach as a way forward in dealing with model uncertainty and illustrates this with particular examples of subnational projections from the Waikato region of New Zealand. A novel contribution of this approach is the averaging of stochastic and deterministic projections by weighting them based on measures of their relative uncertainty. Section 9.7 concludes this chapter.

9.2

Population Projections

A variety of methods are available to demographers and population modellers for use in preparing population projections, whether at the national level or subnationally.5 These methods range from simple naïve models (e.g. extrapolation of past time series trends), through the ‘traditional’ demographic cohort component model, to sophisticated approaches such as agent-based modelling or microsimulation. There has been renewed interest in recent times in exploring the relative merits of different projection methodologies (e.g. see Raymer et al. 2012). Naïve models are most appropriate where there exists a lack of data available for parameterizing more complex models. The simplest of these models involves an assumption of linear population growth based on past growth trends or exponential growth (i.e. a constant population growth rate) based on past growth trends. Somewhat surprisingly, at the subnational level these naïve models tend to perform reasonably well, in terms of forecast accuracy, when compared with more sophisticated models (van der Gaag et al. 2003). Moreover, age- and/or sex-specific projections models can be estimated, either separately or jointly, to derive a projected age-sex distribution for the population. However, such naïve models fail to incorporate age structural effects, cohort effects, or major structural changes such as population ageing. Moreover, they are difficult to defend in discussions with planners or other end-users, as they lack a sound theoretical basis, and they fail to incorporate known or suspected drivers, including demographic and policy drivers, of population change. For a planner or policy-maker who hopes that their plans or policies will influence future population growth or decline, these models offer little guidance.6 This can reduce the ‘buy-in’ from important end-users of projections, hampering their eventual use.7 In contrast, the traditional workhorse model for demographic projections—the cohort component model (CCM)—explicitly accounts for the demographic drivers of population change. The CCM is based on a demographic identity—the population 5

For simplicity, we set aside the case of small area projections, such as for suburbs or sub-county areas, and household projections, where in both cases an even wider variety of projections methods are arising (e.g. see Cameron and Cochrane 2017). 6 Conversely, in the hands of a savvy operator, a naïve model and a simple assumption about a change in the growth rate might be used to justify almost any policy or plan. 7 For example, Cameron et al. (2007) used additional economic development as a driver of migration, in their subnational population projections for the Hamilton sub-region of New Zealand.

9 The Development of Uncertainty in National and Subnational Population. . .

201

at time t + 1 is equal to the population at time t, plus any increases due to births or in-migration between times t and t + 1, and minus any decreases in population due to deaths or out-migration between times t and t + 1. Thus, the CCM requires the modeller to first project the three components of population change: (1) births, typically projected by means of age-specific fertility rates applied to women of childbearing ages; (2) deaths, typically projected by means of age-sex-specific mortality (or its complement, survivorship) rates applied to the population of each age and sex; and (3) migration, which may be projected in a number of different ways (van der Gaag et al. 2003). Moreover, other drivers of population change may be included in the model by first modelling their influences on fertility, mortality, and/or migration. Because these drivers can be included directly in the model, the CCM is more readily accepted as an appropriate model by end-users.8 A further appeal of the CCM is that it can be extended to other demographic projections, such as ethnic population or labour force projections, by including additional components. For ethnic population projections, this is typically the addition of a paternity component, to allow for ethnic births generated by males partnering with females of other ethnic groups, and an inter-ethnic mobility component, to allow for the net effect of people changing ethnic identification over time (Statistics NZ 2017a). For labour force projections this is typically the addition of labour force participation rate and hours worked components (Statistics NZ 2017b). More sophisticated models incorporating agent-based modelling or microsimulation are also available (e.g. Benenson 1998; Fontaine and Rounsevell 2009). These approaches involve using large unit-record datasets to estimate transition probabilities between ‘states’. These states might include spatial location, partnership or marriage, birth of children, mortality, and so on. Thus, these transition probabilities can be used to project forward the population in terms of the components of demographic change in the CCM. However, these models are extraordinarily data-intensive, since they require projecting individual-level or household-level data (in the case of agent-based modelling) or representative individuals or households (in the case of microsimulation). Often, the transition probabilities are not directly measured or measurable and have to be assumed, resulting in complex and non-transparent models. In addition to the model, a relevant consideration for a suite of national and subnational population projections is whether they will be developed using a top-down or bottom-up approach (Willekens 1983). A top-down approach projects the population at the national level first, using a national-level model, then projects each subnational area either separately or as part of a multiregional model. The subnational projections are constrained to sum to the previously determined national projection. The top-down approach has the advantage of ensuring a sufficiently bounded national population projection, driven by known drivers of population change at the national level. In contrast, a bottom-up approach projects each

8 Although end-users may disagree with the assumptions or the effect of drivers used in these models.

202

M. P. Cameron et al.

subnational area (again, either separately or as part of a multiregional model) in an unconstrained manner. A national population projection can then be derived by summing the subnational projections. The bottom-up approach has the advantage of adhering to subnational drivers of population change, without requiring ex post adjustments to the projections to ensure consistency with a national total. However, the bottom-up approach also has several limitations. First, subnational drivers of population change are likely to be subject to greater relative uncertainty than national drivers. A bottom-up approach may therefore lead to national projections with implausibly wide uncertainty, depending on how uncertainty is aggregated to larger geographies. Second, where the components of population change are defined by ethnicity, assumptions of commonality in basic structure across subnational areas are unlikely to apply. Third, national and regional projections also become less timely using the bottom-up approach, as they are dependent on the derivation of projections at smaller geographic levels. This is a disadvantage for producers wanting an agile national-level projection process. Fourth, small area projections need to be derived with full granularity (e.g. single-year of age, ethnicity) if this is desired for higher geographies. Regardless of the overall approach adopted, even deterministic processes need constant vigilance in assessing the relevance of the parameters that are included. For example, Stats NZ had to decide during the early 1970s how much of the decline in fertility that rapidly arose in New Zealand reflected a significant permanent reduction in births, rather than a deferment of births. Stats NZ decided not to participate in the World Fertility Survey by the United Nations Fund for Population Activities in 1973. The first official projections to recognize the massive fall in fertility were not prepared until 1978.

9.3

Uncertainty

Understanding the uncertainty in population projections is important for their accurate interpretation. When measures of uncertainty are provided with projections, these can be used to (1) distinguish within projections what trends are more certain than others; (2) make informed judgements about how uncertainty changes with different projection horizons; and (3) to convey the important notion that the future does not follow smooth and certain paths, but is likely to fluctuate within a range of possibilities depending on interventions by policy agencies. There are several sources of uncertainty in population projections. Starting with Alho’s (1990) sources of error in population projections as a base, we distinguish three types of uncertainty in population projections. The types of uncertainty are not necessarily mutually exclusive, and some examples of uncertainty could be classified as belonging to more than one type. First, model uncertainty arises from an incomplete understanding of all the interacting and overlapping processes, their drivers and determinants, which contribute to future population outcomes. Any model is necessarily a simplification of the underlying reality, and as a result,

9 The Development of Uncertainty in National and Subnational Population. . .

203

there is uncertainty in terms of which model might be ‘best’. The degree of covariance between model parameters and the persistence or otherwise of shocks to time series are further sources of model uncertainty. Moreover, model choice is an important source of uncertainty, and while it may be attractive to argue otherwise, all population projection models require some degree of expert judgement. This judgement may be exercised at a fine operational level, such as in determining the valid range of values that parameters may adopt or at a broader methodological level, such as determining which drivers of population change are included, and which are excluded, from the model. Second, parameter uncertainty arises due to sampling and measurement error in the model parameters. Sampling errors arise where parameters are estimated from sample data, including estimation from short or incomplete time series. For example, the introduction of the methodology for integrating administrative data records with the results of the 2018 Census enumeration in New Zealand has changed the share of each ethnic and age group counted in census statistics. In providing more accurate measures of how the population is distributed among these groups in 2018 compared to earlier census, the rate of change between 2013 and 2018 has a different error structure than any earlier measures of intercensal change. Measurement error arises due to random errors or systematic bias in the estimation of the parameters. Both sources of error will be present in any model, even if the correct underlying model of the population processes was known. Third, random variation is a source of uncertainty in any model. Population processes (migration, fertility, mortality) can be thought of as stochastic processes that, in addition to being measured with error, are subject to random fluctuations between different population groups, regions, and over time. As such, there will always be uncertainty in any projection of future population, arising from random variation that cannot be deterministically accounted for. In contrast, stochastic models (see below) account for the irregularity experienced in the past, so that it can be applied to later observations, usually of one or more of the key parameters in a deterministic model. Uncertainty in population projections can be modelled or expressed in different ways. To simplify, population projection models can be generally characterized in terms of whether they are deterministic or stochastic (probabilistic). A deterministic projection tracks a single path forward in time, based on a single parameterization of the model. The uncertainty in a deterministic projection may be determined analytically from the uncertainty in the underlying parameter distributions or in the components of population change (e.g. see Alho and Spencer 1985, 1997). In contrast, a stochastic projection attempts to directly model the uncertainty in the projected population (e.g. see Tuljapurkar 1992; Tuljapurkar et al. 2004; Cameron and Poot 2011). Often this is achieved by simulation, i.e. running the model many times with different parameter values (drawn from distinct distributions for each parameter) and summarizing the projection interval based on the number of simulated projections that fall within given bounds (Pflaumer 1988). However, in both cases—analytically or simulation-based—measures of uncertainty are somewhat limited to measures of the parameter uncertainty and random

204

M. P. Cameron et al.

Naonal populaon Customised

Naonal ethnic populaon

Subnaonal ethnic populaon

Subnaonal populaon

Small area ('suburb') populaon

Naonal labour force

Naonal family & household

Subnaonal family & household

Fig. 9.1 Demographic projections produced by Statistics NZ

variation in a given model. In practice, most researchers do not adopt and run many different model and covariance structures in order to obtain an assessment of model uncertainty (for a notable exception, albeit based on relatively naïve time series models of total population, see Abel et al. 2013). Instead, a single modelling approach is adopted, and small variations on the model are used to derive measures of uncertainty. Most often, some variants of the CCM are used.

9.4

Historical Development of Population Projections in New Zealand

Stats NZ has a long history of producing demographic projections. National population projections were regularly produced from the 1950s, while labour force and Māori population projections were produced less regularly. For instance, the Hunn report in 1961 included projections, which were probably the first special purpose population projections to specifically capture the distinct demographic characteristics of New Zealand Māori for some 50 years ahead. In the mid-1970s, Stats NZ began projecting the population of subnational areas, a role that was previously undertaken by the Ministry of Works. An internally consistent suite of projections developed through the 1990s, which reflected New Zealand’s changing population composition and demand for demographic projections at different geographies (Fig. 9.1). The advantages of an independent national statistical organization (NSO) producing projections for all local government areas go beyond the consistency with national-level projections. The motivations of local councils—typically in terms of economic growth—often encourage aspirational population projections. These serve a role but are invariably based on aspirational demographic trends rather than prevailing trends. The role of the NSO is to provide objective statistical information according to international best practice methodology, not to customize its statistics

9 The Development of Uncertainty in National and Subnational Population. . .

205

according to national or local marketing or political aspirations. An NSO should also have a rich understanding of the range of quality issues associated with its sources. While there is a role for local experts to inform local population assumptions, NSOs provide a holistic perspective on all local areas across the country.9 The elicitation of expert opinion on projection assumptions is a science in itself and is discussed in Garthwaite et al. (2005), O’Hagan et al. (2006), and Dias et al. (2018); see also Statistics Canada (2019) for a practical example of elicitation used in official projections. Pacific population projections were first published in the mid-1990s, to complement Māori ethnic population projections. By 2017, Stats NZ was publishing national population projections for five overlapping ‘level 1’ ethnic groupings— Māori; Pacific; Asian; European; or other (including New Zealander); and Middle Eastern/Latin American/African—as well as for the three largest ‘level 2’ ethnic groups (Chinese; Indian; and Samoan) (Statistics NZ 2017a). Methodologically, official projections have continued to evolve. The cohort component method and a top-down approach—which enable the progressive release of higher level (e.g. national) projections without dependence on the lowest level projections—have persisted as the backbone of official projections. One of the more significant methodological developments by Stats NZ has been the application of a stochastic approach to national-level projections. The adoption of a stochastic approach has been progressive, beginning with national population projection in 2012, national labour force projections in 2015, and national ethnic population projections in 2015. The advantages of a stochastic approach are discussed in Alho (1997, 2005), Booth (2006), Bryant (2003, 2005), Dunstan and Ball (2016), and Keilman (1991), among others. The driver for Stats NZ to adopt stochastic projections has been less about improving the accuracy of projections and more about improving their interpretability. This reflects that conventional deterministic projections are poor at conveying uncertainty, especially for key demographic characteristics (e.g. dependency ratios, death numbers). For most characteristics, the uncertainty indicated by deterministic scenarios is neither consistent between characteristics nor consistent across the projection period. Given the inherent uncertainty of the future, encouraging users of projections to think about uncertainty is important. However, it is only feasible for users to think about uncertainty if that uncertainty is conveyed to them appropriately. Rather than using scenario variants that convey uncertainty only qualitatively (e.g. high, medium, low), stochastic projections will typically be summarized by means of percentiles that convey the probability of certain outcomes occurring (within the constraints of existing policy settings). For example, stochastic 9

Local council input has been a feature of Stats NZ’s small area (‘suburb’) population projections since they began in the 1990s. Local planners are well placed to inform assumptions about the timing and place of local developments such as new subdivisions and major building projects (e.g. retirement homes, prisons), and about zoning changes that can impact housing and population density. Their input makes the resulting projections more attuned to local plans, if not attuned to local aspirations.

206

M. P. Cameron et al.

Table 9.1 Method of assumption formulation for New Zealand population projections (published 2016) Assumption Base population Fertility

Mortality Migration

Sex ratio at birth

Median (50th percentile) Variance/distribution of values Empirical model: official population Expectation (judgement): variance estimates based on census and postvaries by age-sex enumeration survey Expectation (judgement): long-term Empirical model: ARIMA (0,1,0) total fertility rate of 1.85 births per model fitted to total fertility rate for woman 1977–2016 June yearsa Empirical model: coherent functional demographic model fitted to age-specific death rates for 1977–2015 June years Expectation (judgement): long-term Empirical model: ARIMA (1,0,1) annual net migration of 15,000 model fitted to net migration for 1988– 2016 June yearsa Empirical model: median and variance from sex ratio at birth for 1900– 2015 December years

a

It is worth noting that ARIMA models assume that future projection errors will have the same pattern as past errors

projections can be used to indicate the probability that a certain population characteristic will reach a given level by a given year, or will be within a given range by a given year, or will reach a given level within a given timeframe. They can also provide a projection interval that represents, for a given probability, the range within which the population is projected to be at certain dates. Different percentiles need to be provided for users with different risk appetites. As percentiles are non-additive, flexibility is also needed to derive percentiles for customized purposes (e.g. for ad hoc age groupings). There are many different ways to model uncertainty and to produce stochastic projections. Both expectation (e.g. Billari et al. 2014; Lutz et al. 2014; Lutz 2009; Lutz and Scherbov 1998) and empirical approaches (e.g. Lee and Carter 1992; Woods and Dunstan 2014) are used internationally to model uncertainty, and both have their advantages and disadvantages. Empirical models have been most intensively applied where demographic trends have been largely monotonic and sustained, as in the case of death rates and life expectancy. Stats NZ has taken a pragmatic approach by using a mix of expectation and empirical methods to produce projections (Table 9.1). Stats NZ provides an independent and comprehensive set of public-good demographic projections, but of course it is not the sole producer of projections in New Zealand. We focus our attention on one such producer, which has also been an innovator in population projections methods. In the early 2000s, the Population Studies Centre (PSC) at the University of Waikato began to provide bespoke subnational population projections for local councils and central government agencies. This began with work for the Bay of Plenty region (Bedford and Dharmalingam 2002; Bedford 2005), in response to a view by local government, particularly in the fast-growing Tauranga City and Western Bay of Plenty District, that official projections did not reflect what they perceived as expansive population growth

9 The Development of Uncertainty in National and Subnational Population. . .

207

prospects. This was a view shared by many local council planners, and soon after, Jacques Poot at PSC was contracted to produce projections for the Hamilton sub-region (Cameron et al. 2007, 2008a) and for Thames-Coromandel District (Cameron et al. 2008b). One of the key innovations in that early PSC work was that the projections were ‘end-user informed’. This meant that end-users had input into the assumptions underlying the population projections model, i.e. that the models would ‘incorporate local information by experts and end-users with respect to the assumptions that drive the projections’ (Cameron et al. 2007). This allowed these bespoke projections to satisfy end-users (i.e. local council planners) that the projections would accurately reflect their beliefs about the growth trajectory of their local area.10 One additional innovation in these early projections was that migration was modelled on the basis of age-sex-specific net migration rates, as opposed to modelling a total number for net migration, which would then be distributed by age and sex (Cameron et al. 2007). The advantage of net migration rates is that they adjust dynamically to reflect the underlying size of the population, without further intervention on the part of the modeller. Subsequently, the PSC population model was included within the Waikato Regional Council’s integrated WISE (Waikato Integrated Scenario Explorer) model (Rutledge et al. 2008, 2010). Given end-users’ concerns about accuracy of the projections, understanding and representing uncertainty in the projections was an important consideration in the PSC work. This led to the projections model being extended to produce stochastic population projections (Cameron and Poot 2011). This model was then used in subsequent subnational projections for the Waikato and Bay of Plenty regions (Cameron and Cochrane 2014; Cameron et al. 2014). The methodology for population projections at the re-branded National Institute of Demographic and Economic Analysis (NIDEA; formerly the Population Studies Centre) has continued to evolve. Recent developments include revising the projections model so that internal migration within New Zealand is estimated using a gravity model (Poot et al. 2016; Cameron and Poot 2014a, b). Along with other origin–destination models, the gravity model has several advantages over timeseries-based approaches for projecting net migration or net migration rates. First, it explicitly models all of the underlying inward and outward flows of migrants to and from each subnational area. This allows the projections model to demonstrate not only the total migration flows, but also their origins and destinations, which has been a key request from end-users. Second, it more explicitly accounts for the underlying drivers of population movement, at least for internal migration (international migration, which is much more volatile, still requires more traditional methods for projection). Third, the uncertainty in the gravity model parameters are estimated, and importantly the covariance between the parameters is estimated, and these can be used in deriving stochastic population projections. Fourth, the role of end-users in

10 Although, as it turns out, those beliefs were often aspirational rather than realistic. See later in this section.

208

M. P. Cameron et al.

pre-determining the outcome of the projections is limited, although recent projections by NIDEA had adopted a Bayesian approach to avoid overly optimistic assumptions driving the projections (Cameron and Cochrane 2014; Cameron et al. 2014).

9.5

The Accuracy of Historical Population Projections

In a world increasingly bombarded by ‘forecasts’, it is also worth reflecting on the relative predictability of demographic events. Fundamentally, the population can only change through births, deaths and migration—although ethnic populations are susceptible to changing patterns of identification over time (O’Donnell and Raymer 2015), and local populations are susceptible to volatility in migration flows that are often the most significant component of population change. But this fundamental simplicity does make population size and age-sex composition more predictable than other phenomena. Somewhat trivially, even if population projections are not used by a decision-maker or are used but not understood, then an assumption about the future population structure and level will be implicit in the action taken. The Population Monitoring Group of the New Zealand Planning Council played a pivotal role in bringing to the attention of the public the policy implications of population change, and since this was disestablished in 1991, no other similar ongoing forum of population experts has existed in New Zealand. In policy areas such as housing, justice, health, urban infrastructure, and immigration, policy is less likely to respond in good time to the certainties we already are aware of that will drive the future population, and its structure, in each place. A Governmental Committee to Review Power Requirements met each year up to the late 1970s to advise Parliament of the implications of population change on future electricity needs and inform planning of major energy projects by the State. The most recent expert review of population change was prepared by the Royal Society of New Zealand in 2014, but it has not had the attention it merited (see Royal Society of New Zealand 2014). As Patrice Dion, chief demographer at Statistics Canada observed, ‘A good population projection is not defined by whether or not it matches reality. It is defined by whether it was plausible and useful to policy-makers at the time the projection was created’.11 This is not to say that projection accuracy is unimportant, but that policy-makers and other users should question assumptions at the time the projections are produced, not some time later with the benefit of hindsight. Nonetheless, an assessment of projection accuracy can be useful for both producers and users, especially as an indicator of potential uncertainty in current projections. Surprisingly, routinely published assessments of accuracy are rare among national statistical agencies. This is in spite of errors in past projections providing

11

https://www.statcan.gc.ca/eng/blog-blogue/cs-sc/ptp.

9 The Development of Uncertainty in National and Subnational Population. . .

209

an indication of likely uncertainty in more recent projections (Keilman 1990, 1997). Stats NZ has published two reports, the first in 2008, and an update (which also assessed the accuracy of population estimates) in 2016 (Statistics NZ 2016). The general findings about accuracy are not surprising: • Projection accuracy tends to diminish the further out from the starting point and • Projection accuracy tends to diminish as populations become smaller, such as for smaller geographic areas or age groups. Each of the key assumptions—fertility, mortality, migration—has contributed to projection inaccuracy at different times. Despite New Zealand fertility levels around replacement level for the last four decades, small changes in the timing of childbearing have caused ups and downs in birth numbers, periodically placing strains on local maternity services (Sceats 1999). Mortality assumptions have a relatively small impact on total population projections, but a significant impact on projections of the population in older ages. An important development in the accuracy of mortality forecasting are models that draw on the latest age-and-sex-specific shifts in mortality patterns, such as those that use functional demographic models (Woods and Dunstan 2014). These are preferable to mortality models that focus on extrapolating life expectancy at birth, which usually assume age-specific mortality patterns are changing at the same proportionate rate. Migration assumptions have usually been the most inaccurate, for several reasons. First, it reflects the volatility of New Zealand’s international migration, with large swings in arrivals, departures, and net migration experienced over short periods of time. Second, it reflects that internal migration becomes an increasingly significant driver of population change as geography becomes smaller. Third, it reflects that migration is a much more complex process to model than fertility or mortality, given the multitude of factors that drive migration both into and out of an area, as well as its temporal and spatial variability. Quantifying the accuracy of past projections is useful as it reinforces that accuracy and uncertainty are intrinsically linked. Sub-populations that have high projection uncertainty are generally those where projections are less accurate. This does not mean that projections are worthless. It does, however, underscore the importance of regularly updating projections to reflect new demographic data and policy settings. It also reinforces the value of including quantified measures of uncertainty in projections to enable end-users to differentiate between more certain and less certain demographic outcomes. Indeed, this differential uncertainty is conveyed by stochastic projections. The 2016-base New Zealand population projections (Statistics NZ 2016) indicate a 65+ population of 1.62–2.06 million in 2068 (90% projection interval) or relative uncertainty of 12% around the median. In contrast, the 15–39 population has a projection of 1.32–2.31 million in 2068 or relative uncertainty of 28% around the median. Both future births and migration, and uncertainty thereof, play a role in the greater relative uncertainty for the younger population.

210

9.6

M. P. Cameron et al.

A Way Forward

Having recognized that population projections are subject to uncertainty, that there are several alternative population projections available from different providers, and that the degree of uncertainty (as conveyed by the accuracy of historical population projections) can be quite high, what is a decision-maker to do? How should a ‘preferred’ population projection for decision-making purposes be determined? In our experience, there is no one correct answer to this question. However, in this section we propose one way forward for the decision-maker, which does not involve deciding on a single preferred population projection, but instead makes use of multiple projections and their implicit or explicit uncertainty. Returning to the concepts of uncertainty we outlined in Sect. 9.3, there are essentially three sources of uncertainty: (1) model uncertainty; (2) parameter uncertainty; and (3) random variation. The latter two sources of uncertainty are explicitly accounted for in a stochastic population projection, and as Alho et al. (2008) note, in the case of deterministic scenario projections (high/medium/low) this uncertainty may be inferred from the range between the high and low projections. However, stochastic population projections do not typically account for model uncertainty.12 When faced with model uncertainty, and a multitude of possible (deterministic and/or stochastic) population projections, one fruitful solution may be to employ ‘model averaging’ (Claeskens and Hjort 2008; Bijak 2010). The overall approach is very simple—each projection is weighted using some schema of weights, and an overall projection, with or without an associated measure of uncertainty, is derived from the weighted average of the projections. This approach is increasingly common in forecasting applications—for instance, it is used in combining climate projections from different climate models to derive an overall ensemble projection for future climate (Tebaldi and Knutti 2007; Knutti et al. 2010). There are many model-averaging approaches that could be applied. One challenge in model averaging is how two stochastic projections can be averaged, or how stochastic projections may be averaged with deterministic projections. One solution is to use the median projection to represent a stochastic projection and then to weight the median projection based on a measure of its uncertainty. Deterministic projections can similarly be weighted on the basis of their uncertainty, with their uncertainty determined by the accuracy of past projections or based on expert judgement. We illustrate the model-averaging approach with a simple case study based on Stats NZ and NIDEA subnational population projections for 10 territorial

12

More accurately, while stochastic population projections could be structured to incorporate some forms of model uncertainty, to date we are unaware of any systematic attempt to incorporate model uncertainty into stochastic population projections. For a notable exception that employs Bayesian model averaging, albeit based on relatively naïve time series models of total population, see Abel et al. (2013).

9 The Development of Uncertainty in National and Subnational Population. . .

211

authorities13 in the Waikato Region of New Zealand. Specifically, we use 2013-base population projections (February 2015 release) from all the three variant deterministic projections (high, medium, and low) from Stats NZ and the median stochastic projection and associated measures of uncertainty from NIDEA. We take a simple weighted average of the medium Stats NZ projection and the median NIDEA projection, where the weights are the inverse of the projection interval. This approach ensures that the projection with the least uncertainty (in terms of parameter uncertainty and random variation, as discussed above) receives a higher weighting. Following Stoto (1983), we assume the range from high to low projection is a 67% projection interval for the Stats NZ projections (see also Alho et al. 2008). To ensure comparability, we use the 67% projection interval from the NIDEA stochastic projections. We then compare the two projections (Stats NZ and NIDEA) and the modelaverage projection with the realized population, based on the 2018 subnational population estimates, for each area. The currently available subnational population estimates for 2018, despite being only partly based on the 2018 Census, are still the best available estimate of subnational populations at the time of writing. The results of the model-average projection, along with the contributing NIDEA and Stats NZ projections, are summarized in Table 9.2. The Stats NZ medium projections are somewhat higher than the NIDEA median projections for all territorial authorities. However, the range from high to low projection for the Stats NZ projections is similar in percentage terms to the 67% projection interval for the NIDEA projections. This provides some confidence that interpreting the range from high to low variant as a 67% projection interval is not untenable. For most territorial authorities, the NIDEA projection interval is slightly smaller than the range for the Stats NZ projections, so the model-average projection is slightly closer to the NIDEA median projections than the Stats NZ medium projections. The last four columns of Table 9.2 present a median and 67% projection interval for the combined projections. This is derived by taking 100 draws from the NIDEA stochastic projections, and 100 draws from a normal distribution for the Stats NZ projections,14 and weighting each draw equally. The median projection by this method is similar to the model-average projection, but the range is instructive. When the two models (NIDEA and Stats NZ) agree, the range of the averaged projection is similar to those of each of the contributing projections, whereas when the two models disagree the

13

Territorial authorities are the smallest administrative unit in New Zealand. They do not necessarily share contiguous boundaries with regions. The Waikato region is comprised of all, or part, of 11 territorial authorities. However, the proportion of Rotorua District that lies within the Waikato Region is small, so it is excluded from consideration here. Taupo and Waitomo Districts also contain parts that lie outside the Waikato Region, but the populations of those areas are small. However, it should be noted that the NIDEA projections only project the parts of each territorial authority that lie within the Waikato Region and so will be underestimates relative to the total population for Taupo and Waitomo Districts. 14 This normal distribution assumes a mean equal to the medium projection and a standard deviation equal to half the difference between the low and high projections.

Territorial authority ThamesCoromandel Hauraki Waikato Matamata-Piako Hamilton Waipa Otorohanga South Waikato Waitomo Taupo 4.0% 2.8% 2.3% 2.6% 3.0% 4.0% 3.1% 3.6% 2.5%

18,800 71,900 34,100 163,800 51,800 9810 23,200 9220 35,700

19,150 73,300 34,800 166,700 52,700 10,050 23,700 9460 36,400

3.7% 3.9% 4.4% 3.6% 3.7% 4.9% 4.3% 5.2% 3.9%

18,450 70,500 33,300 160,800 50,800 9570 22,700 8980 35,000

18,795 70,275 33,101 160,178 49,916 9462 22,251 9083 35,336

18,412 68,162 32,205 155,487 47,635 8993 21,224 8825 34,676

18,789 69,121 32,580 157,550 48,364 9176 21,561 8987 35,108

Stats NZ 2013-base projection for 2018 Low Medium High Range 27,000 27,600 28,200 4.3%

NIDEA 2013-base projection for 2018 P16.5 Median P83.5 Range 26,833 27,410 27,504 4.2%

Table 9.2 Projections and model-averaging projections for territorial authorities in the Waikato region

18,795 70,275 33,101 160,178 49,916 9462 22,251 9083 35,336

18,440 68,470 32,327 156,262 47,935 9069 21,367 8865 34,774

18,771 70,235 33,082 160,149 49,620 9435 22,139 9090 35,374

Model-average projection Average P16.5 Median 27,504 26,898 27,459

19,173 72,452 34,395 164,962 52,174 9905 23,397 9315 35,976

P83.5 28,088

3.9% 5.7% 6.3% 5.4% 8.5% 8.9% 9.2% 4.9% 3.4%

Range 4.3%

212 M. P. Cameron et al.

9 The Development of Uncertainty in National and Subnational Population. . .

213

Table 9.3 Projection errors for 2018 for territorial authorities in the Waikato region Territorial authority Thames-Coromandel District Hauraki District Waikato District MatamataPiako District Hamilton City Waipa District Otorohanga District South Waikato District Waitomo District Taupo District WMAPE

Projection NIDEA Stats NZ median medium 27,410 27,600

Model average 27,504

SNPE 30,800

Error NIDEA median 11.0%

Stats NZ medium 10.4%

Model average 10.7%

18,789 69,121

18,800 71,900

18,795 70,275

20,600 77,800

8.8% 11.2%

8.7% 7.6%

8.8% 9.7%

32,580

34,100

33,101

35,500

8.2%

3.9%

6.8%

157,550 48,364 9176

163,800 51,800 9810

160,178 49,916 9462

165,900 54,800 10,400

5.0% 11.7% 11.8%

1.3% 5.5% 5.7%

3.4% 8.9% 9.0%

21,561

23,200

22,251

24,800

13.1%

6.5%

10.3%

8987

9220

9083

9570

6.1%

3.7%

5.1%

35,108

35,700

35,336

38,300

8.3% 8.5%

6.8% 4.8%

7.7% 7.0%

range of the averaged projection is much wider. This demonstrates that a failure to account for model uncertainty likely leads to measures of projection uncertainty that are overly conservative. Table 9.3 compares the three projections (NIDEA median; Stats NZ medium; model average) with the subnational population estimate for 2018. The error in the last three columns of Table 9.3 is expressed as the percentage error (with the population estimate as the denominator). The final row shows the weighted mean absolute percentage error (WMAPE) for each of the three projections. WMAPE is a suitable summary measure of the relative size of errors in population projections (Siegel 2002). It is clear from Table 9.3 that both the NIDEA and Stats NZ projections systematically underprojected the 2018 population. This is due to the historically high and unexpectedly sustained net international migration gain for New Zealand over this period. The projection error ranges from 10.4% to 1.3% for the Stats NZ medium projection and from 13.1% to 5.0% for the NIDEA median projection. Being a weighted combination of the two other projections, the error for the model-average projection is in between, ranging from 3.4% to 10.7%. The WMAPE is lowest for the Stats NZ medium projection (4.8%) and highest for the NIDEA median projection (8.5%), with the model-average projection in-between (7.0%). The results in Table 9.3 have implications for decision-makers. A decision-maker is faced with a choice of population projection on which to base their decision. However, they will not know in advance the error in any of the projections they could choose. Choosing one of the available projections could result in a projection

214

M. P. Cameron et al.

with a higher, or lower, relative error. The decision-maker faces a risk that they choose the projection that turns out to have the highest relative error. A risk-averse decision-maker could avoid the risk of choosing the ex post worst-performing projection by using a weighted average of the available projections. In our illustrative case, that would reduce the worst outcome (in terms of relative error) from a WMAPE of 8.5% to a WMAPE of 7.0%. Obviously, this is worse than the best case WMAPE of 4.8%; however, the decision-maker would not know ex ante which of the available projections will result in the smallest relative error.

9.7

Conclusion

Projections or forecasts of future population, and the age and sex structure of the population, are key inputs to decision-making. However, projections are not perfect predictions of a certain future; they must contain uncertainty. This uncertainty arises from model uncertainty, parameter uncertainty, and random variation. Decisionmakers must therefore be cognizant of this uncertainty and how it will affect their decisions. Moreover, decision-makers are often faced with multiple alternative population projections for the same area and/or group of interest. How should a decision-maker decide between these alternative projections? One consideration is to attempt to choose the projection that has the lowest degree of error. However, there is no way for the decision-maker to determine this ex ante and therefore no clear way to decide on a single preferred projection. We argue that alternative projections should be seen as complementary, rather than competing, inputs into decision-making. A decision-maker can make use of all of the available projections, rather than limiting themselves to one. This limits the risk of the decision-maker choosing as their preferred projection the projection that ex post results in the highest relative error. We have illustrated a model-averaging approach that makes use of all of the available projections and creates a single projection by weighting the available based on estimated uncertainty. This approach limits the risk of choosing the projection with the highest relative error. We acknowledge that our illustration is relatively simple, although the illustration does demonstrate how deterministic and stochastic projections may be combined. Nevertheless, we believe that illustrating the approach with a simple case study is useful in opening the conversation about model averaging as a useful tool for applied demography. We leave as an exercise for future researchers to determine an optimal model-averaging approach. In the meantime, the simple approach should provide decision-makers with a useful heuristic to assist in using a range of population projections for improved decision-making. Acknowledgements This research is supported by the New Zealand Ministry of Business Innovation and Employment (MBIE)-funded projects UOWX1404 Capturing the Diversity Dividend of Aotearoa New Zealand (CADDANZ) and CONT-29661-HASTR-MAU Nga Tangata Oho

9 The Development of Uncertainty in National and Subnational Population. . .

215

Mairangi. We thank Jacques Poot for his leadership and vision in the development of subnational stochastic population projections models in New Zealand and Ian Pool, John Bryant, Niko Keilman, for comments and suggestions on the development of these methods at the National Institute of Demographic and Economic Analysis (NIDEA) at the University of Waikato. The opinions, findings, recommendations, and conclusions expressed in this paper are those of the authors. They do not necessarily represent those of Stats NZ or NIDEA, who take no responsibility for any omissions or errors in the information contained here.

References Abel GJ, Bijak J, Forster JJ et al (2013) Integrating uncertainty in time series population forecasts: an illustration using a simple projection model. Demogr Res 29:1187–1226 Alho JM (1990) Stochastic methods in population forecasting. Int J Forecast 6:521–530 Alho J (1997) Scenarios, uncertainty and conditional forecasts of the world population. J R Stat Soc Ser A Stat Soc 160:71–85 Alho J (2005) Remarks on the use of probabilities in demography and forecasting. In: Keilman N (ed) Perspectives on mortality forecasting: probabilistic models. Swedish Social Insurance Agency, Stockholm, pp 27–38 Alho JM, Spencer BD (1985) Uncertain population forecasting. J Am Stat Assoc 80:306–314 Alho JM, Spencer BD (1997) The practical specification of the expected error of population forecasts. J Off Stat 13:203–225 Alho J, Cruijsen H, Keilman N (2008) Empirically based specification of forecast uncertainty. In: Alho J, Hougaard Jensen S, Lassila J (eds) Uncertain demographics and fiscal sustainability. Cambridge University Press, Cambridge, pp 34–54 Auerbach AJ, Lee RD (eds) (2001) Demographic change and fiscal policy. Cambridge University Press, Cambridge Bedford R (2005) Demographic forecast, 2051: Movement and change in population and households in the Bay of Plenty. Research report commissioned by Bay of Plenty Regional Council. Population Studies Centre, University of Waikato, Hamilton Bedford RD, Dharmalingam A (2002) Migration and population forecasts for the Western Bay of Plenty, 2001-2051: Implications for the Labour Force. Paper presented at the 16th conference on labour, employment and work, Wellington, 21–22 Nov 2002 Benenson I (1998) Multi-agent simulations of residential dynamics in the city. Comput Environ Urban Syst 22:25–42 Bijak J (2010) Forecasting international migration in Europe: a Bayesian view. Springer, Dordrecht Billari F, Graziani R, Melilli E (2014) Stochastic population forecasting based on combinations of expert evaluations within the Bayesian paradigm. Demography 51:1933–1954 Booth H (2006) Demographic forecasting: 1980 to 2005 in review. Int J Forecast 22:547–581 Bryant J (2003) Can population projections be used for sensitivity tests on policy models?. New Zealand Treasury working paper 03/07. New Zealand Treasury, Wellington Bryant J (2005) What can stochastic population projections contribute to policy analysis? N Z Popul Rev 31:1–11 Cameron MP, Cochrane W (2014) Population, household, and labour force projections for the Waikato region, 2013–2063. Research report commissioned by Waikato Regional Council, University of Waikato, Hamilton Cameron MP, Cochrane W (2017) Using land-use modelling to statistically downscale population projections to small areas. Australas J Reg Stud 23:195–216 Cameron MP, Poot J (2011) Lessons from stochastic small-area population projections: The case of Waikato subregions in New Zealand. J Popul Res (Canberra) 28:245–265

216

M. P. Cameron et al.

Cameron MP, Poot J (2014a) Developing a systems-based multi-region stochastic population projections model for New Zealand. Paper presented at the 61st Annual North American Meetings of the Regional Science Association International, Washington, D.C., 12–15 Nov 2014 Cameron MP, Poot J (2014b) Projecting future inter-regional migration using age-gender-specific gravity models—application to New Zealand. Paper presented at the 16th conference on labour, employment and work, Wellington, 27–28 Nov 2014 Cameron MP, Cochrane W, Poot J (2007) End-user informed demographic projections for Hamilton up to 2041. Research report commissioned by Hamilton City Council, Population Studies Centre Discussion Paper No. 66, Population Studies Centre, University of Waikato, Hamilton Cameron MP, Cochrane W, Poot J (2008a) Population Projections until 2061 for Futureproof—the Hamilton sub-regional growth strategy. Research report commissioned by Hamilton City Council, Waikato District Council, and Waipa District Council, Population Studies Centre, University of Waikato, Hamilton Cameron MP, Cochrane W, Poot J (2008b) Thames-Coromandel District Council: population, household, and dwelling projections. Research report commissioned by Thames-Coromandel District Council, Population Studies Centre, University of Waikato, Hamilton Cameron MP, Jackson NO, Cochrane W (2014) Baseline and stochastic population projections for the territorial authorities of the Waikato Region for the period 2013–2063. Research report commissioned by Waikato Shared Services. National Institute for Demographic and Economic Analysis, University of Waikato, Hamilton Claeskens G, Hjort NL (2008) Model selection and model averaging. Cambridge University Press, Cambridge Dias L, Morton A, Quigley J (eds) (2018) Elicitation: the science and art of structuring judgment. Springer, New York Dion P, Galbraith N (2015) Back to the future: a review of forty years of population projections at Statistics Canada. Can Stud Popul 42:102–116 Dunstan K, Ball C (2016) Demographic projections: user and producer experiences of adopting a stochastic approach. J Off Stat 32:947–962 Fontaine CM, Rounsevell MDA (2009) An agent-based approach to model future residential pressure on a regional landscape. Landsc Ecol 24:1237–1254 Garthwaite PH, Kadane JB, O’Hagan A (2005) Statistical methods for eliciting probability distributions. Technical paper 1–2005. Carnegie Mellon University, Pittsburgh Keilman N (1990) Uncertainty in national population forecasting: issues, backgrounds, analyses, recommendations. Swets and Zeitlinger, Amsterdam Keilman N (1991) Uncertainty in national population forecasting: issues, backgrounds, analyses, recommendations. Int J Forecast 7:392–393 Keilman N (1997) Ex-post errors in official population forecasts in industrialized countries. J Off Stat 13:245–277 Knutti R, Furrer R, Tebaldi C et al (2010) Challenges in combining projections from multiple climate models. J Climate 23:2739–2758 Lee R, Carter L (1992) Modeling and forecasting US mortality. J Am Stat Assoc 87:659–671 Lutz W (2009) Toward a systematic, argument-based approach to defining assumptions for population projections. Interim report IR-09-037, International Institute for Applied Systems Analysis, Vienna Lutz W, Scherbov S (1998) An expert-based framework for probabilistic national population projections: The example of Austria. Eur J Popul 14:1–17 Lutz W, Butz W, Samir KC (eds) (2014) World J Am Stat Assoc in the 21st century. Oxford University Press, Oxford O'Donnell J, Raymer J (2015) Identification change and its effect on projections of the Aboriginal and Torres Strait Islander population in Australia. J Popul Res (Canberra) 32:297–319 O'Hagan A, Buck CE, Daneshkhah A et al (2006) Uncertain judgements: eliciting experts’ probabilities. Wiley, Chichester

9 The Development of Uncertainty in National and Subnational Population. . .

217

Pflaumer P (1988) Confidence intervals for population projections based on Monte Carlo methods. Int J Forecast 4:135–142 Poot J, Alimi O, Cameron MP et al (2016) The gravity model of migration: the successful comeback of an ageing superstar in regional science. Investig Reg J Reg Res 36:63–86 Raymer J, Abel GJ, Rogers A (2012) Does specification matter? Experiments with simple multiregional probabilistic population projections. Environ Plan A 44:2664–2686 Royal Society of New Zealand (2014) Our futures: te pae tawhiti. Royal Society of New Zealand, Wellington Rutledge DT, Cameron M, Elliott S et al (2008) Choosing regional futures: challenges and choices in building integrated models to support long-term regional planning in New Zealand. Reg Sci Policy Pract 1:85–108 Rutledge D, Cameron M, Elliott S et al (2010) WISE—waikato integrated scenario explorer, technical specifications version 1.1. Research report commissioned by Environment Waikato, Landcare Research, Hamilton Sceats J (1999) Cohort patterns and trends in the timing and spacing of births. In: Pool I, Johnstone K (eds) The lifecourses of New Zealand women: fertility, family formation and structure, fertility regulation, education, work and economic wellbeing. University of Waikato, Hamilton, Population Studies Centre, pp 41–44 Siegel JS (2002) Applied demography: applications to business, government, law and public policy. Academic Press, San Diego Statistics Canada (2019) Using expert elicitation to build long-term projection assumptions. Paper presented at the Eurostat-UNECE work session on demographic projections, Belgrade, 25–27 Nov 2019 Statistics NZ (2016) How accurate are population estimates and projections? An evaluation of Statistics New Zealand population estimates and projections, 1996–2013. Statistics NZ. http:// archive.stats.govt.nz/browse_for_stats/population/estimates_and_projections/how-accuratepop-estimates-projns-1996-2013.aspx. Accessed 21 Jan 2020 Statistics NZ (2017a) National ethnic population projections: 2013(base)–2038 (update). Statistics NZ. https://www.stats.govt.nz/information-releases/national-ethnic-population-projections2013base2038-update. Accessed 21 Jan 2020 Statistics NZ (2017b) National labour force projections: 2017(base)–2068. Statistics NZ. https:// www.stats.govt.nz/information-releases/national-labour-force-projections-2017base2068. Accessed 21 Jan 2020 Stoto M (1983) Accuracy of population projections. J Am Stat Assoc 78:13–20 Tebaldi C, Knutti R (2007) The use of the multi-model ensemble in probabilistic climate projections. Philos Trans R Soc A 365:2053–2075 Tuljapurkar S (1992) Stochastic population forecasts and their uses. Int J Forecast 8:385–391 Tuljapurkar S, Lee RD, Li Q (2004) Random scenario forecasts versus stochastic forecasts. Int Stat Rev 72:185–199 van der Gaag N, van Wissen L, Rees P et al (2003) Study of past and future interregional migration trends and patterns within European Union countries: in search for a generally applicable explanatory model. Netherlands Interdisciplinary Demographic Institute, The Hague Willekens FJ (1983) Multiregional population analysis for urban and regional planning. In: Batty M, Hutchinson M (eds) System analysis in urban policy making and planning. Plenum Press, New York Woods C, Dunstan K (2014) Forecasting mortality in New Zealand: A new approach for population projections using a coherent functional demographic model. Statistics New Zealand working paper no. 14–01, Statistics NZ, Wellington

Chapter 10

Applications of Machine Learning Models in Regional and Demographic Economic Analysis: A Literature Survey Mehmet Güney Celbiş

I am completely operational, and all my circuits are functioning perfectly. - Quote by HAL 9000, in 2001 A Space Odyssey.

10.1

Introduction

Big data is becoming increasingly available in regional and demographic economic analysis. Old challenges related to data unavailability are being replaced with difficulties dealing with more complex data sets that are often updated almost in real-time. Model-based, causality-oriented approaches are being substituted with black-box prediction-oriented techniques—primarily in nonacademic fields for expediting policy and management decisions. Big data sets allow for highly flexible relationships between variables, and machine learning (ML) techniques can enable researchers to discover highly complex relationships (Varian 2014). Complexity in this regard is related to “the unpredictable nature of non-linear and dynamic systems” (Nijkamp et al. 2001). Alpaydin (2016) defines ML as a way to achieve artificial intelligence (AI) and states that ML, which is grounded in statistical theory, is the driving force and a requirement for AI. The latter, in turn, can be defined as “The automation of activities that we associate with human thinking, activities such as decision-making, problem solving, and learning” (Bellman 1978). Regarding the ability of ML to discover complex relationships, Harding and Hersh (2018) state: Very often, machine learning tools enhance the existing econometric methodology by grounding modeling decisions in data as opposed to unreliable human intuitions, which manifest themselves as modeling choices.

M. G. Celbiş (*) Yeditepe University, Department of Economics, and United Nations University - Maastricht Economic and Social Research Institute on Innovation and Technology, Maastricht, The Netherlands e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 W. Cochrane et al. (eds.), Labor Markets, Migration, and Mobility, New Frontiers in Regional Science: Asian Perspectives 45, https://doi.org/10.1007/978-981-15-9275-1_10

219

220

M. G. Celbiş

These modeling choices in traditional economic research, as Athey (2018) points out, correspond to the data-driven “tuning” of the algorithms in ML models. While Athey (2018) also highlights the benefits of this approach, the above statement by Harding and Hersh (2018) may seem potentially provocative to many theoretical economists. In fact, it can be argued that in social sciences the automated and iterative processing of big data is not sufficient to solve problems related to causality by itself, and that careful research design is necessary (Grimmer 2015). Despite the fact that causality concerns are beginning to be addressed more commonly in ML models—particularly through the use of treatment effect approaches (Athey and Imbens 2016) —ML methods have been relatively uncommon in economic research (Athey and Imbens 2019). As we shall see in this literature survey, studies that use ML techniques in the area of regional and demographic economics are few—the majority of them being very recent—with very few studies dating earlier than 2017. Economists often aim to measure causal effects in the form of elasticities. Many common supervised ML techniques, however, focus on prediction and allow for all possible nonlinearities and interactions. This enables ML algorithms to uncover generalizable patterns and discover complex structures in the absence of prior modeling and specific functional forms (Mullainathan and Spiess 2017). Particularly when the data analyzed is “big,” ML techniques perform well in out-of-sample prediction (Harding and Hersh 2018; Mullainathan and Spiess 2017; Athey and Imbens 2019). While ML methods can identify influential predictors, they do not aim to present specific effect sizes, their directions, and significances. This being said, there are numerous similarities between econometrics and ML methods, as highlighted in a series of highly influential studies by Susan Athey.1 The rest of this chapter is structured as follows: Sect. 10.2 presents a survey of studies using ML methods in regional and demographic economic analysis. Section 10.3 provides a brief overview of ML applications in other policy-related economic research. Finally, in 10.4 I briefly discuss the prospects related to ML methods and how they can complement econometric approaches and list some R routines for commonly used ML models.

10.2

A Survey of Studies that Use ML Methods in Areas Related to Regional and Demographic Policy

Interestingly, an early application of ML techniques focusing on demographic attributes such as population and the level of education was a doctoral thesis by Fazenbaker (2009) where the aim was predicting intellectual productivity measured through patent data. The author never published his work in a peer-reviewed journal and continued his career as a rock musician. Athey and Imbens (2019) present a detailed overview on the overlaps between the two fields. The textbooks by Friedman et al. (2001) and James et al. (2013) can also serve as valuable sources to observe these similarities. 1

10

Applications of Machine Learning Models in Regional and Demographic. . .

221

Prior to 2017, ML methods were quite rare in the literature under focus. Xu et al. (2013) and Zhang and Zhang (2014) used neural network approaches to analyze unemployment trends and travel intentions, respectively, where the former study made extensive use of search engine data, and the latter used individual level surveys. In a similar manner, through the use of web-based data, Arribas-Bel et al. (2015) introduced ML methods for urban spatial analysis in their study where they examined the spatial dimension of tweeter activity in Amsterdam. The literature started to expand by 2017. In parallel, the ML methods used by researchers began to diversify. Web-based and satellite data has become more usable and popular among social scientists due to advances in ML. Glaeser et al. (2018) predicted the income levels in New York City using a support vector machine approach on data obtained from Google Street View. In another urban-level research, Arribas-Bel et al. (2017) utilized satellite imagery for predicting the Living Environment Deprivation (LED) index for the city of Liverpool and examined the relationship between LED and certain important land cover and structure variables using random forest and gradient boosting approaches. Satellite imagery has been also used for the grid-based estimation of population in the USA in a study by Robinson et al. (2017) where the authors used neural network techniques. Further spatial research on urban areas has been carried out by Comber and Arribas-Bel (2019); Comber (2019); Arribas-Bel et al. (2019) who apply ML techniques on the delineation of urban areas and to address matching tasks through the use of conditional random fields, random forests, and text-based algorithms among other methods. In the application of ML models, labor markets and unemployment have been subjects that draw much attention—perhaps due to the availability of large individual level and historical data sets. The topic of employment has also served as a scene for advances in methodological contributions. For instance, for the purpose of introducing new methods that do not require the specification of particular functional forms, Cook and Hall (2017) show that using neural network analysis performs better than certain traditional benchmark methods for predicting unemployment. Neural networks are used in labor-related research also by Lopez-Yucra et al. (2018) who investigate the determinants of child labor in Peru. From an urban-oriented perspective, Kaiser (2018) employed unsupervised Bayesian ML techniques for investigating segmentation in urban labor markets in relation to voluntary selection of informal employment. A macroeconomic approach towards the topic of unemployment is taken by Kreiner and Duca (2019) in their study on the main predictors of the rate of unemployment US, where the authors use neural networks and least absolute shrinkage and selection operator (LASSO) techniques. Some researchers, however, aimed to introduce new ML approaches for classifying occupations (Colace et al. 2019), and job vacancies (Boselli et al. 2017) through the analysis of text-based information using methods like neural networks and support vector machines. In another study with a specific focus on occupations, Saavedra and Twinam (2019) developed a new occupation scoreboard—through the use of LASSO and tenfold cross-validation—with the goal of estimating earning gaps with greater accuracy.

222

M. G. Celbiş

Labor markets have been studied using ML techniques in relation to migration as well. Tarasyev et al. (2018) used decision trees among several other ML approaches for predicting migration behavior for job searching purposes. A similar research question is addressed by Liu (2019) who examines the specialization of migrants, labor market competition, and employee mobility through unsupervised ML models.2 Concentrating further on migrant behavior, Iman and Tarasyev (2018) employed decision trees and K-nearest neighbors (KNN) approaches for predicting the perceptions of migrants regarding the labor market in Moscow. Aside from being studied in conjunction with labor markets, migration by itself has also been examined using ML approaches. In their study on US counties and world countries, Robinson and Dilkina (2018) used neural networks and gradient boosting for predicting migration and identifying its main determinants. Nigam et al. (2019), however, employed a diverse set of methods: random forest, gradient boosting, support vector machine, multilayer perceptron, and neural networks, for predicting migrants at risk and identifying the sources that render migration dangerous. Further ML-based research with demographic and regional scopes have been conducted by Wu et al. (2020) who use gradient boosting, random forests, and support vector regression for investigating the main predictors of traffic accidents in Zhongshan, while Mittal et al. (2019) who employ decision trees, random forests, and neural networks for exploring the determinants of crime in India. A chronological and categorical overview of the ML-based studies in the research areas within the scope of this survey is presented in Table 10.1, where the studies are grouped under four general thematic categories: Employment and the Labor Market, Migration and Mobility, Urban Data Processing, and Socioeconomic Development.

10.3

Machine Learning Applications in Other Policy-Related Areas

Aside from the above reviewed literature on regional and demographic analysis, there are studies that use ML methods in other areas of economics that would be beneficial to briefly cover in this review. In general, the literature covered in this section aim to understand regional and macroeconomic outcomes through ML methods. For instance, from a regional perspective, Chang et al. (2014) combine spatial econometrics with regression trees for the identification of convergence clubs in Europe. Results related to economic convergence are provided also by Bang et al. (2017), who identify the determinants country-level economic growth and elaborate on their partial effects through ML techniques such as regression trees, neural networks, bootstrap aggregation, boosting, and random forests. GDP prediction

2 This dissertation is not yet available and the presented information relies on the available abstract of the study at: https://ir.uiowa.edu/etd/6977/.

10

Applications of Machine Learning Models in Regional and Demographic. . .

223

Table 10.1 List of Studies Author(s) Short title Employment and the labor market Saavedra A machine learning and approach to improving Twinam occupational income (2019) scores Liu (2019) Predicting labor market competition and employee mobility

Kreiner and Duca (2019) Colace et al. (2019) Tarasyev et al. (2018)a Kaiser (2018)

Boselli et al. (2017)

Cook and Hall (2017) Xu et al. (2013)

Can machine learning on economic data better forecast the unemployment rate? Towards labor market intelligence through topic modeling Machine learning in labor migration prediction Segmentation in urban labor markets: a machine learning application and a contracting perspective Using machine learning for labor market intelligence Macroeconomic indicator forecasting with deep neural networks Forecasting the unemployment rate by neural networks using search engine query data

Migration and mobility Nigam et al. Migration through (2019) machine learning lens: Predicting sexual and reproductive health vulnerability of young migrants

Methods

Result

Least absolute shrinkage and selection operator regression with tenfold cross-validation Unsupervised machine learning models

A new occupation score is developed that allows a better estimation of earning gaps Uncovering information regarding the specialization of migrants versus native employees, prediction of future employee turnover outcomes Prediction of unemployment and identification of main predictors

Neural networks, least absolute shrinkage, and selection operator Text classification

Decision trees among other ML models Unsupervised Bayesian machine learning

Bag of word Word2vec, neural networks, support vector machine Neural networks

Introduction of a new method for occupation classification Presentation of a model for predicting migration for job searching Finds that a significant fraction of people voluntarily choose non-formal employment Classification of online job vacancies

Neural networks

Presentation of a new technique for predicting unemployment Identification of the advantages of analyzing unemployment trends through web-search behavior

Random forest, gradient boosting, support vector machine, multilayer perceptron, sequential neural network

Identification of critical sources that make migration dangerous and the prediction of migrants at risk (continued)

224

M. G. Celbiş

Table 10.1 (continued) Author(s) Robinson and Dilkina (2018)

Short title A machine learning approach to modeling human migration

Methods Artificial neural networks, gradient boosting

Iman and Tarasyev (2018)

Machine learning methods in individual migration behavior

Decision tree, K-nearest neighbors

LopezYucra et al. (2018) Zhang and Zhang (2014)

Could machine learning improve the prediction of child labor in Peru? Analyzing Chinese citizens’ intentions of outbound travel: a machine learning approach Urban data processing Comber Demonstrating the utility (2019) of machine learning innovations in address matching to spatial socioeconomic applications Arribas-Bel Building (s and) cities: et al. (2019) Delineating urban areas with a machine learning algorithm Glaeser et al. (2018)

Big data and big cities

Robinson et al. (2017)

A deep learning approach for population estimation from satellite imagery Remote sensing-based measurement of living environment deprivation: Improving classical approaches with machine learning

Arribas-Bel et al. (2017)

Arribas-Bel et al. (2015) Comber and Arribas-Bel (2019)

Cyber cities: Social media as a tool for understanding cities Machine learning innovations in address matching: A practical

Artificial neural networks

Neural network, twicelearning

Result Prediction of migration between counties in the USA and between countries across the world, identification of main predictors Predicting the migrant’s perceptions regarding the Moscow labor market Prediction of child labor and the identification of main predictors Identification of the travel intentions of Chinese citizens

Conditional random fields, random forest

A more efficient method for address matching is introduced

Approximate densitybased algorithm for discovering clusters in large spatial databases with noise Support vector regression

Delineating urban areas as a unit of analysis for urban research, introducing a new ML approach Prediction of income in New York City using Google Street View Grid-based prediction of population in the USA

Convolutional neural networks Random forest, gradient boosting

Clustering algorithm

Conditional random fields, word2vec

Prediction of living environment deprivation in areas of Liverpool and the identification of main predictors. Evaluation of the impacts of main predictors Analysis of tweeter patterns for the city of Amsterdam Introduction of new machine learning (continued)

10

Applications of Machine Learning Models in Regional and Demographic. . .

225

Table 10.1 (continued) Author(s)

Short title

comparison of word2vec and CRFs Socioeconomic development Wu et al. Economic development, (2020) demographic characteristics, road network, and traffic accidents in Zhongshan, China Fazenbaker Exploring impact of edu(2009) cational and economic factors on national intellectual productivity using machine learning methods Mittal et al. Monitoring the impact of (2019) economic crisis on crime in India using machine learning

Methods

Result techniques for linking address pairs

Gradient boosting, random forest, support vector regression

The main predictors of traffic accidents in Zhongshan are identified

M5 rules, conjunctive rule, decision table

Identification of federallevel determinants of patent applications and grants in the USA

Decision trees, random forest, neural networks

Identification of the main predictors of crime in India

This is a work in progress. Therefore, the methods and the findings are at a preliminary discussion stage

a

has also been extended to predict recessions using ML models such as random forests by Nyman and Ormerod (2017, 2020). Focusing explicitly on economic growth, Cogoljević et al. (2018) use artificial neural networks to predict GDP using macroeconomic measures of energy, resource, and waste and show that the approach provides advantages in predicting economic activity. Concentrating on the causality between institutions and per-capita GDP, Gründler and Krieger (2018) use an ML approach (support vector machine) as an intermediate step for creating an explanatory democracy index variable and find that democracies grow faster than autocratic regimes.3

10.4

Concluding Remarks

Econometrics has been the main engine for understanding causal relationships between economic variables. For ML methods, however, substantiating theoretical economic models seldom exist. Nevertheless, researchers with experience in econometric methods will easily notice many similarities with ML methods, as ML is often considered as a branch of statistics. This being said, ML focuses on predictions and does not provide output in the form of elasticities or effect sizes. It does, however, 3 Using ML methods as intermediary steps for creating variables is an approach recommended by Athey (2018).

226

M. G. Celbiş

tell the researcher which variables play relatively more effective roles in improving predictive accuracy. Partial dependencies can be examined to some extent, and useful side products like measures of intrinsic similarities among observations can be obtained (Breiman 2001). Nevertheless, ML techniques may hardly be considered as a substitute for econometric methods, particularly if the research question requires the estimation of causal effect sizes. ML and econometrics can be very effective jointly, when used in a complementary manner. As discussed in this literature survey, in light of the recommendations of pioneering researchers in the field, ML methods can also be very useful when they are used as intermediate steps. The comparison of ML techniques to traditional econometric approaches deserves further attention. ML techniques are already becoming the core methods in many practical sectors, including health, marketing, and finance. Econometrics, however, faces the risk of being confined to the social science departments of universities, despite its profound technical depth and accuracy. One reason for this could be that “regular” patterns in data—which econometric approaches explain with high—are rare, and the availability of big data now emphasizes this fact in a stronger way. Einav and Levin (2014) highlight how big data applications aim to achieve accurate out-of-sample predictions in the presence of large amounts of potential predictors, through methods that have not attracted much attention in economics, where the interest of the latter mostly lies in understanding causal relationships. Studies that compare the performances of ML and econometric techniques are scarce. Following their analysis on movie box office figures, Liu and Xie (2019) observe that ML methods perform better in the presence of irregularities and nonlinearities in the data, while econometric models perform better in explaining longrun trends when there is less heterogeneity in the data. In another study, Bajari et al. (2015) used six ML methods aside of a linear regression and a logit model for estimating demand for salty snacks. Based on their results, the authors claim that the ML methods are “considerably more accurate.” The ability of ML techniques to capture highly complex relationships can help researchers to approach difficult research questions initially with ML models and proceed to construct econometric estimations based on findings of the ML procedures. The reverse can also be useful as a confirmation or a meta-robustness check of economic relationships. I believe ML methods will be increasingly costly to ignore by academic researchers. They are very positively received by policymakers and the private and public sector researchers, thanks to their intuitiveness. ML techniques provide very simple and easy to understand output for laypeople, compared to say, a regression table. However, one should be very careful in appreciating these attributes of ML methods, as concerns regarding causality which is of essential importance for policymakers are clearly present. Attention should be given to the current efforts of prominent statisticians and econometricians who aim to provide solutions to such concerns by introducing new tools into ML approaches, often by importing them from the domain of econometrics.

10

Applications of Machine Learning Models in Regional and Demographic. . .

227

Table 10.2 R Packages for Commonly Used Methods Method Decision tree Random forest Gradient boosting Support vector machine Neural network

Package rpart randomForest gbm, xgboost svm neuralnet

Author(s) Atkinson and Therneau (2000) Liaw and Wiener (2002) Ridgeway and Ridgeway (2004), Chen et al. (2015) Karatzoglou et al. (2006) Günther and Fritsch (2010)

The R software provides many functions for conducting ML-based research. To conclude this chapter, I present Table 10.2 which lists some of the popular ML routines in R.

References Alpaydin E (2016) Machine learning: the new AI. MIT Press, Cambridge Arribas-Bel D, Garcia-L’opez M-A, Viladecans-Marsal E (2019) Building (s and) cities: delineating urban areas with a machine learning algorithm. J Urban Econ 2019:103217 Arribas-Bel D, Kourtit K, Nijkamp P, Steenbruggen J (2015) Cyber cities: social media as a tool for understanding cities. Appl Spat Anal Policy 8(3):231–247 Arribas-Bel D, Patino JE, Duque JC (2017) Remote sensing-based measurement of living environment deprivation: improving classical approaches with machine learning. PLoS One 12(5): e0176684 Athey S (2018) The impact of machine learning on economics. In: The economics of artificial intelligence: an agenda. University of Chicago Press, Chicago, pp 507–547 Athey S, Imbens G (2016) Recursive partitioning for heterogeneous causal effects. Proc Natl Acad Sci 113(27):7353–7360 Athey S, Imbens GW (2019) Machine learning methods that economists should know about. Ann Rev Econ 11:685–725 Atkinson EJ, Therneau TM (2000) An introduction to recursive partitioning using the rpart routines. Mayo Foundation, Rochester Bajari P, Nekipelov D, Ryan SP, Yang M (2015) Machine learning methods for demand estimation. Am Econ Rev 105(5):481–485 Bang J, Sen T, Basuchoudhary A (2017) New tools for predicting economic growth using machine learning: a guide for theory and policy. In: Southern Economic Association Conference Bellman R (1978) An introduction to artificial intelligence: can computers think? Boyd & Fraser, San Francisco Boselli R, Cesarini M, Mercorio F, Mezzanzanica M (2017) Using machine learning for labour market intelligence. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, Berlin, pp 330–342 Breiman L (2001) Random forests. Mach Learn 45(1):5–32 Chang H-J, Shih Y-S, Su T-J (2014) Split selection methods for regression tree on detecting regional economic convergence. J Chin Statist Assoc 52(2):180–208 Chen T, He T, Benesty M, Khotilovich V, Tang Y (2015) Xgboost: extreme gradient boosting. R Package Version 0.4-2. pp. 1–4 Cogoljević D, Alizamir M, Piljan I, Piljan T, Prljić K, Zimonjić S (2018) A machine learning approach for predicting the relationship between energy resources and economic development. Phys A Statistic Mech Appl 495:211–214

228

M. G. Celbiş

Colace F, De Santo M, Lombardi M, Mercorio F, Mezzanzanica M, Pascale F (2019) Towards labour market intelligence through topic modelling. In: Proceedings of the 52nd Hawaii international conference on system sciences Comber S (2019) Demonstrating the utility of machine learning innovations in address matching to spatial socio-economic applications. Region 6(3):17–37 Comber S, Arribas-Bel D (2019) Machine learning innovations in address matching: a practical comparison of word2vec and crfs. Trans GIS 23(2):334–348 Cook T, Hall AS (2017) Macroeconomic indicator forecasting with deep neural networks. In: Federal Reserve Bank of Kansas City, Research Working Paper, pp 17–11 Einav L, Levin J (2014) Economics in the age of big data. Science 346(6210):1243089 Fazenbaker C (2009) Exploring impact of educational and economic factors on national intellectual productivity using machine learning methods. Virginia Commonwealth University, Richmond Friedman J, Hastie T, Tibshirani R (2001) The elements of statistical learning, vol 1. Springer, New York Glaeser EL, Kominers SD, Luca M, Naik N (2018) Big data and big cities: the promises and limitations of improved measures of urban life. Econ Inq 56(1):114–137 Grimmer J (2015) We are all social scientists now: how big data, machine learning, and causal inference work together. Political Sci Politics 48(1):80–83 Gründler K, Krieger T (2018) Machine learning indices, political institutions, and economic development. KOPS Universität Konstanz, Konstanz Günther F, Fritsch S (2010) Neuralnet: training of neural networks. R J 2(1):30–38 Harding M, Hersh J (2018) Big data in economics. IZA World of Labor, Bonn, p 451 Iman HS, Tarasyev A (2018) Machine learning methods in individual migration behavior. In: Russian regions in the focus of changes: conference proceedings.—Ekaterinburg, 2018. LLC Publishing office EMC UPI, Saint Paul, pp 72–81 James G, Witten D, Hastie T, Tibshirani R (2013) An introduction to statistical learning, vol 112. Springer, New York Kaiser M (2018) Segmentation in urban labor markets: a machine learning application and a contracting perspective Karatzoglou A, Meyer D, Hornik K (2006) Support vector machines in r. J Stat Softw 15(9):1–28 Kreiner A, Duca JV (2019) Can machine learning on economic data better forecast the unemployment rate? Appl Econ Lett 27:1–4 Liaw A, Wiener M (2002) Classification and regression by randomforest. R News 2(3):18–22 Liu Y. (2019). Predicting labor market competition and employee mobility — a machine learning approach. PhD Dissertation, University of Iowa. https://doi.org/10.17077/etd.3x1x-s7p3 Liu Y, Xie T (2019) Machine learning versus econometrics: prediction of box office. Appl Econ Lett 26(2):124–130 Lopez-Yucra K, Libaque-Saenz C, Lazo J, Bravo E (2018) Could machine learning improve the prediction of child labor in Peru? In: Information Management and Big Data: 4th Annual International Symposium, SIMBig 2017, Lima, Peru, September 4–6, 2017, Revised Selected Papers, vol 795. Springer, Berlin, p 15 Mittal M, Goyal LM, Sethi JK, Hemanth DJ (2019) Monitoring the impact of economic crisis on crime in India using machine learning. Comput Econ 53(4):1467–1485 Mullainathan S, Spiess J (2017) Machine learning: an applied econometric approach. J Econ Perspect 31(2):87–106 Nigam A, Jaiswal P, Arora T, Girkar U (2019) Migration through machine learning lens– predicting sexual and reproductive health vulnerability of young migrants. arXiv preprint arXiv:1910.02390 Nijkamp P, Poot J, Vindigni G (2001) Spatial dynamics and government policy: an artificial intelligence approach to comparing complex systems. In: Fischer M, Frolich J (eds) Knowledge, complexity and innovation systems. Springer, Boston, pp 369–401 Nyman R, Ormerod P (2017) Predicting economic recessions using machine learning algorithms. arXiv preprint arXiv:1701.01428

10

Applications of Machine Learning Models in Regional and Demographic. . .

229

Nyman R, Ormerod P (2020) Understanding the great recession using machine learning algorithms. arXiv preprint arXiv:2001.02115 Ridgeway G, Ridgeway MG (2004) The gbm package, vol 5. R Foundation for Statistical Computing, Vienna, p 3 Robinson C, Dilkina B (2018) A machine learning approach to modeling human migration. In: Proceedings of the 1st ACM SIGCAS conference on computing and sustainable societies, pp 1–8 Robinson C, Hohman F, Dilkina B (2017) A deep learning approach for population estimation from satellite imagery. In: Proceedings of the 1st ACM SIGSPATIAL Workshop on Geospatial Humanities, pp 47–54 Saavedra M, Twinam T (2019) A machine learning approach to improving occupational income scores. Explor Econ Hist 75:101304 Tarasyev AA, Agarkov GA, Hosseini SI (2018) Machine learning in labor migration prediction. In: AIP Conference Proceedings, vol 1978. AIP Publishing LLC, Melville, p 440004 Varian HR (2014) Big data: new tricks for econometrics. J Econ Perspect 28(2):3–28 Wu W, Jiang S, Liu R, Jin W, Ma C (2020) Economic development, demographic characteristics, road network and traffic accidents in Zhongshan, China: gradient boosting decision tree model. Trans A: Trans Sci 16:1–33 Xu W, Li Z, Cheng C, Zheng T (2013) Data mining for unemployment rate prediction using search engine query data. SOCA 7:33–42 Zhang C, Zhang J (2014) Analysing Chinese citizens’ intentions of outbound travel: a machine learning approach. Curr Issue Tour 17(7):592–609