Education and Mobility in Heterogeneous Labor Markets: Sonderausgabe von Heft 1/Bd. 226 Jahrbücher für Nationalökonomie und Statistik 9783110511307, 9783828203549

Citation preview

Education and Mobility in Heterogeneous Labor Markets Edited by Wolfgang Franz

With Contributions by Anton Flossmann and Winfried Pohlmeier (Universität Konstanz) Anja Kuckulenz and Michael Maier (ZEW Mannheim) Nikolai Stähler (Universität Mainz) Alfred Garloff and Anja Kuckulenz (ZEW Mannheim) Martin Heineck, Mathias Kifmann and Normann Lorenz (Universität Konstanz)

Lucius &c Lucius Verlagsgesellschaft Stuttgart 2006

Anschrift des Herausgebers des Themenheftes Prof. Dr. Dr. h. c. mult. Wolfgang Franz, Präsident Zentrum für Europäische Wirtschaftsforschung GmbH (ZEW) L7, 1 D-68161 Mannheim E-mail: [email protected]

Bibliografische Information der Deutschen Bibliothek Die Deutsche Bibliothek verzeichnet diese Publikation in der Deutschen Nationalbibliografie; detaillierte bibliografische Daten sind im Internet über http://dnb.ddb.de abrufbar. ISBN 3 - 8 2 8 2 - 0 3 5 4 - X ISBN ab 2 0 0 7 : 978-3-8282-0354-9

© Lucius & Lucius Verlagsgesellschaft mbH · Stuttgart · 2 0 0 6 Gerokstraße 51, D-70184 Stuttgart Das Werk einschließlich aller seiner Teile ist urheberrechtlich geschützt. Jede Verwertung außerhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlags unzulässig und strafbar. Das gilt insbesondere für Vervielfältigungen, Ubersetzungen und Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen.

Satz: Mitterweger 8c Partner Kommunikationsgesellschaft mbH, Plankstadt Druck und Bindung: Neumann Druck, Heidelberg Printed in Germany

Jahrbücher f. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2006) Bd. (Vol.) 226/1

Inhalt / Contents Editorial by W. Franz

4

Abhandlungen/Original Papers Flossmann, Anton L., Winfried Pohlmeier, Causal Returns to Education: A Survey on Empirical Evidence for Germany Kuckulenz, Anja, Michael Maier, Heterogeneous Returns to Training . . . Stähler, Nikolai, Employment Protection: Its Effects on Different Skill Groups and on the Incentive to become Skilled Garloff, Alfred, Anja Kuckulenz, Training, Mobility, and Wages: Specific Versus General Human Capital Heineck, Martin, Mathias Kifmann, Normann Lorenz, A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany

6-23 24-40 41-54 55-81 82-109

Jahrbücher f. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2006) Bd. (Vol.) 226/1

Editorial On the background of international competition, topics such as education and mobility have taken center stage in academia as well as in public discussions. Moreover, it is by now recognized that labor markets are anything but homogeneous given the tremendous share of low skilled unemployed people. Hence, the prerequisite of an informed discussion and policy advice on educational matters is a sound theoretical and empirical analysis of relevant features in this field. This is what this special issue is devoted to. The papers were presented at a conference organized by the Research Group "Heterogeneous Labor" supported by the Deutsche Forschungsgemeinschaft (German Science Foundation) in September 2005. All papers have undergone a referee process. A few remarks on the Research Group may be in order. The group, which started its work in May 2002, consists of eight different research projects in economics, business economics and econometrics located at the Department of Economics at the University of Konstanz and the Center for European Economic Research (ZEW). The Research Group "Heterogeneous Labor" investigates theoretically and empirically the determinants of the skill structure of labor as well as the causes and consequences of a shift of the skill structure. The focus is on non-competitive theories of the labor market and the role of the institutional design of labor market and the educational system. Ultimately the Research Group aims at making a scientific contribution to the debate on reforms in the educational system and the organization of the labor market in the light of technical change and globalization. In what follows the conference contributions selected for this special issue are reviewed very briefly. Anton Flossmann and Winfried Pohlmeier (University of Konstanz) survey the empirical evidence on causal effects of education on earnings for Germany. Various studies are compared in light of their underlying assumptions which lead to rather different interpretations of the estimated causal effect. The authors focus on the question as to what extent causal return estimates are informative with respect to educational policy advice. Anja Kuckulenz and Michael Maier (ZEW Mannheim) pick up the topic of heterogeneity and present findings concerning the influence of continuing training on earnings. In their analysis with German data they elaborate on the decision of a person whether or not to engage in training measures. A crucial criterion of this decision is the expected return to training which in turn partly depends on unobservable characteristics. Using recent advances in estimating returns to schooling, Local Instrumental Variables, the authors are able to allow for selection on unobservables in their estimations. In his theoretical contribution Nikolai Stähler (University of Mainz) considers the effect of employment protection on the incentive to invest in skill formation. He uses a matching model with two educational levels to analyse the impact of employment protection laws on various skill groups and on the incentives to acquire further skills. In a theoretical framework where workers decide ex-ante on their skill formation, he shows that if workers obtain a sufficiently large fraction of the rent created by skill formation, employment protection can raise the share of skilled workers. Alfred Garloff and Anja Kuckulenz (ZEW Mannheim) use the interrelationships between continuing training, job mobility, and wages in order to test to what extent to which training-related knowledge is company specific in nature. Specifically, the correlation be-

Editorial · 5

tween training and mobility is examined and wage effects of mobility are estimated taking training participation into account. Results suggest that there is some specific human capital which is incorporated into training and lost when moving between jobs. Martin Heineck, Mathias Kifmann and Normann Lorenz (University of Konstanz) examine the effects of tuition fees on long-term students using duration analysis. Drawing on unique data from the University of Konstanz, they demonstrate empirically that longterm tuition fees lead to a reduction in the length of time students take to complete their studies - although the impact of fees differs between subjects.

Wolfgang Franz

Editor in Chief Z E W , Mannheim

Jahrbücher f. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2006) Bd. (Vol.) 226/1

Causal Returns to Education: A Survey on Empirical Evidence for Germany By Anton L. Flossmann and Winfried Pohlmeier, Konstanz* JEL C21, J24, J31 Returns to education, potential outcome approach, instrumental variables, unconfoundedness, control function approach.

Summary This paper surveys the empirical evidence on causal effects of education on earnings for Germany and compares alternative studies in the light of their underlying identifying assumptions. We work out the different assumptions taken by various studies, which lead to rather different interpretations of the estimated causal effect. In particular, we are interested in the question to what extend causal return estimates are informative regarding educational policy advice. Despite the substantial methodological differences, we have to conclude that the empirical findings for Germany are quite robust and do not deviate substantially from each other. This also holds for the few studies which rely on ignorability conditions, regardless of whether they use educational attainment as a continuous treatment variable or as a discrete treatment indicator. Own estimates based on the matching approach indicate that the selection into upper secondary schooling is suboptimal.

1.

Introduction

It is beyond dispute among economists as well as non-economists, that educational attainment and individual earnings are positively correlated. This observation is mostly explained by the general notion that education has a productivity enhancing effect which is valued by the labor market through higher wages. This is why the returns to education, here simply defined as the earnings differences related to difference in educational attainment, have become a popular monetary measure to assess the quality of a certain educational training in the light of the needs of the labor market. However, classical return estimates obtained from least squares estimation of earnings functions, which do not take into account the heterogeneity of individual returns and the endogeneity of the individual schooling decision, are less informative for educational policy devices. First, the fixed coefficient specification assumes that the individual returns are identical for all individuals with the same educational degree. Furthermore, they are identical across different subpopulations, e.g an employee with a lower secondary degree could (ceteris paribus) expect as much income as an employee with an upper secondary

* This paper is a completely revised and updated version of Pohlmeier's (2004) survey presented at the BMBF-Workshop "Investition in Humankapital", Bonn, June 7th, 2 0 0 4 . Financial support by the DFG through research group "Heterogenous Labor" at the University of Konstanz and the ZEW, Mannheim, is gratefully acknowledged.

Causal Returns to Education: A Survey on Empirical Evidence for Germany · 7

degree if he or she had graduated from upper secondary school. More important, conventional return estimates do not provide any information to what extent more education is causally responsible for more income. This is because they summarize an overall effect of education on earnings which depend on a self-selection process due to individual differences in benefits, preferences, and costs of education that drive the schooling decision. In order to learn anything regarding the impact of the quality and/or quantity of educational attainment on individual earnings, which solely can be attributed to the educational attainment and not to self-selectivity, causal returns to education have to be estimated. Based on the potential outcome approach going back to Roy (1951) and Rubin (1974), the research interest has been centered around estimating causal returns to education which compare earnings of a worker with a given educational attainment to a counterfactual situation, i.e. the worker's earnings in case he had received a different educational attainment. This paper provides a selective survey of recent studies on the causal effects of education on earnings for Germany. Besides obvious differences considering the data base used as well as the definition of the education variable, the studies differ in terms of the underlying assumptions that identify the causal effect, the estimation method, and the type of causal effect estimated. We work out the methodological differences between various studies and the consequences for the interpretation of their empirical findings in Section 2. In Section 3 we discuss studies in the tradition of the Becker-Mincer type of human capital earnings function which uses schooling as a continuous dependent variable. The case of the causal impact of educational career choices within the structured school system, where educational input (program choice) is taken as a discrete treatment variable is presented in Section 4. Section 5 draws attention to further aspects of the impact of education on earnings including the role of program and income risk, overeducation, and the impact of school quality on earnings. Section 6 concludes and gives an outlook on future research.

2.

Causal returns

Although estimates of the returns to education generally rely on estimating a relationship between an individual's log earnings, In Y, and a variable which proxies the individual's educational attainment, S, the underlying assumptions and the resulting interpretation of the return estimates are rather different. To clarify these differences, we will use the correlated random coefficient representation In Υ = a + β S,

(2.1)

which nests a number of different specifications. Here α and β represent random variables correlated with observable and unobservable individual attributes (i.e. experience, gender, cognitive and noncognitive skills, preferences, costs of education etc.). Although the earnings function is usually motivated by the reasoning of human capital theory, it can simply be taken as a hedonic price equation relating educational attainment to the individual's wage. If educational attainment S is conventionally measured in terms of years of schooling 1 , the returns to education are given by ^=/î*lnY|

1

s = s + 1

-lnY|

5 :

(2.2)

Studies on the German school system usually use the number of school years required to obtain the highest degree the individual holds.

8 · A.L. Flossmann and W. Pohlmeier

Thus, the return rate is heterogeneous and varies across individuals. The focus of interest is no longer a single coefficient, but different moments of β. The mean causal return rate Ε [β] (mean partial derivative) reflects the effect of education on earnings in the sense of a laboratory experiment, because it denotes the expected income increase that arises from an one year increase of educational training of an individual that is randomly drawn from the population. In terms of the nomenclature of the econometric evaluation literature, Ε [β] is nothing else but the average treatment effect for the case of a continuous treatment variable. Due to the self-selection, the expected return rates for individuals with different schooling levels are different; E [ / 3 | S = so] E [ / 3 | S = s j ] . Even without imposing nonlinearities between log earnings and schooling, e.g. through polynomials in S, the return rates vary across education levels. The classical Mincer type of earnings function arises as a special case of (2.1 ) for a = α ο + ε and β = β0·. In Y = In Y(S) = a 0 + ßo

s

+ ε>

the difference turns out to be insignificant.

Causal Returns to Education: A Survey on Empirical Evidence for Germany · 17

5. Some neglected issues 5.1.

Overqualification

Overeducation can be defined as the part of a worker's overall educational attainment that is not required to perform the current job. In this sense, the returns to overeducation shed light on the efficient use of education and skilled labor. Obviously, overeducation, as a fuzzy concept, is difficult to measure. In the literature, a range of subjective and objective measures for overeducation have been proposed. Subjective measures are more or less based on the worker's self-assessment of the skills required for his current position compared to his own educational attainment. Objective measures usually relate the legally required education for a worker's current position to the actual level of education. If available, combinations of the two extreme concepts are conceivable and require that both a subjective and an objective criterion have to be satisfied to define a worker as being overqualified. Based on 25 econometric studies, Groot & van den Brink (2000) show in their meta-analysis that the estimates of (noncausal) returns to overeducation is rather sensitive to the measurement concept. However, in general they find broad international evidence that the returns to overeducation are considerably lower than the returns to required schooling. To our knowledge, the studies by Jochmann & Pohlmeier (2003) and Maier, Pfeiffer & Pohlmeier (2003) are the only studies explicitly focussing on the causal effects of overeducation. While the former study treats overeducation as a binary event within a Bayesian evaluation framework, the latter study is based on a random coefficient earnings function for specific skill groups where schooling additional to the required schooling serves as the continuous treatment variable. Table 4 below contains the estimation results for male fulltime employed German workers (BIBB-IAB, 1998/99 sample) based on the unconfoundedness approach proposed by Wooldridge (2002, 2004). Table 4: Causal Returns to Overeducation Skill Group

Ε[(β]

Unskilled Vocational training Foreman, sen. craftsman University graduates

.049 .082 .061 .237

(1.18) (23.25) (12.58) (7.66)

Source: Maier et al. (2003, p. 144), t-values in parenthesis, trimmed observations

For the group of skilled and high skilled workers, they find no evidence that the average causal returns to overeducation are lower than the average causal returns to required education. This somewhat stands in contrast to the evidence found for the majority of traditional studies. Therefore they conclude that overeducation should be viewed as a rational investment strategy, especially by a large part of the group of skilled workers. Only for the group of unskilled workers do the average returns to overeducation turn out to be very low. However, Maier et al. (2003) find considerable heterogeneity in the returns to overeducation, as well as in the returns to required education.

18 · A.L. Flossmann and W. Pohlmeier 5.2.

Income risk and program risk

Heterogeneity and self-selectivity also have serious implications for the assessment of the quality of educational policies beyond the econometric problem of identifying causal mean returns. Focussing only on mean returns ignores that the quality of educational policies also depends on the riskiness of the educational programs supplied, i.e. on V [β], the variance of the causal return effect. If policy makers are risk averse, they should also be concerned about the risk of an educational program. In order to estimate the program risk, additional identifying assumptions are required. For the case of a comparison of two educational programs (binary treatment model) the information on the covariance between a and β is equivalent to information on the correlation between the potential outcomes Yo and Yj. V [In Ya - In Y 0 ] = V [In Yj] - V [In Y 0 ] - 2 Cov [α, β].

(5.6)

Pohlmeier & Flossmann (2005) point out that the term V [In YJ] — V [In YQ] is a causal measure of the residual earnings inequality which can be estimated without additional identifying assumptions other than those used for the estimation of the mean causal returns. Hence the residual earnings inequality V [In Yt - In Y 0 ] > V [In Yj] - V [In Y 0 ]

(5.7)

can serve as an informative lower bound for the program risk if V [In YJ]—V [In YQ] > 0. However, it is important to note, that causal mean and variance effects are ex-post measures based of the random variable earnings (i.e. more precisely on the realization of this random variable). Thus the variance effect captures both individual income variation resulting from individual heterogeneity unknown to the econometrician, but known to the individual, income risk due to risk concerning the educational program (failure risk, matching risk of choosing the appropriate program etc.), and income risk due to the labor market conditions (e.g. macro shocks). Surprisingly, the aspect that educational choice is also based on risk evaluations and the trade-off between risk and return has been tackled only in a few studies. Notable exceptions are the papers by Belzil & Hansen (2002) and Hogan &c Walker (2001), who model the individual schooling decision as a dynamic programming decision under uncertainty. Hartog & Serano (2002) estimate the impact of income risk on the education decision on the basis of a static expected utility maximization calculus. They find out that income risk is a non-negligible determinant of the schooling decision. Heckman, Lochner & Todd (2003) compute the internal return rate of education taking into account income risk about the future income stream. Using US Census data (1940-1990), they find that the internal rate decreases substantially if one accounts for income risk. However, none of the studies mentioned above tries to combine the idea of the potential outcome approach with the aspect of educational choice under uncertainty.

5.3.

School quality and returns to education

Causal return estimates are helpful to discriminate between successful and less successful programs. In particular, a comparison of the mean causal returns with the mean causal returns for the treated and the untreated may yield valuable information on how well

Causal Returns to Education: A Survey on Empirical Evidence for Germany · 19

students are selected into different educational programs. From an educational economics point of view, however, causal return estimates are only of limited value, because they do not provide any specific information on whether educational programs are to be (re-) designed to improve the students' labor market prospects. For the improvement of the design of educational programs, it is therefore of greater importance to investigate how certain aspects of educational programs (e.g. expenditures per student, class size, studentteacher ratio, enumeration of teachers, assessment test for applicants) have an effect on the labor market performance in later years. For the US, Card and Krueger (1992, 1996) provide supporting empirical evidence for a positive link between school quality and subsequent earnings. 4 For various methodological reasons, their results were questioned by a number of studies including Betts (1995), Heckman, Layne-Farrar &c Todd (1996) as well as Hanushek, Rivkin & Taylor (1996). Using data for West Germany, Baumgartner (2004) investigates the impact of class size on early career earnings and finds no significant empirical evidence for this relationship. Similar to the case of the return estimates, the problem of heterogeneity and self-selectivity arises here, too, such that good scholars are more likely to attend good schools. Hence, it is impossible to infer that the above average performance/earnings of a worker educated at a 'good' school is causally related to the higher quality of the educational input. It may rather be the result of a self-selection process. Whether the selectivity problem is severe for the case of Germany is at least doubtful, because the dominating public school system guarantees fairly equal standards in terms of the schools' endowments. However, the low variation across observational units in Germany is likely to aggravate the problem of finding significant empirical support for a positive link between quality and earnings. Nevertheless, an analysis on the causal link between the quality of educational attainment and individual labor market success on large scale micro-data would nicely complement studies on the link between school quality and the outcomes of learning (e.g. based on the PISA or TIMSS data). 6.

Conclusion

This paper surveys the empirical evidence on causal effects of education on earnings for Germany and compares alternative studies in the light of their underlying identifying assumptions. We point out, that given these assumptions, the various studies estimate rather different causal effects. But despite the substantial methodological differences, we have to conclude that the findings are quite robust and do not deviate substantially from each other. This also holds for the few studies which rely on ignorability conditions, regardless whether they use educational attainment as a continuous treatment variable or as a discrete treatment indicator. The classical IV estimates are somewhat less robust compared to the results obtained by the control function and ignorability approaches, and are sensitive to the instrument chosen. Studies which additionally estimate the causal returns for the treated and the untreated indicate that the selection gain for graduates of the upper secondary school level is not significantly different from zero, if compared to the earnings of workers with intermedi4

See also Brewer/Eide/Ehrenberg by Brewer/Ehrenberg (1996).

(1999), Dale/Kmeger (2002), Light/Strayer (2000) and the survey

20 · A.L. Flossmann and W . Pohlmeier

ate secondary schooling. This at least raises some doubts about the effectiveness of the selection process at the early stage of the educational career. In this survey we restricted ourselves to the estimation of the mean causal returns to education that have been estimated for Germany. However, in order to use the potential outcome approach for educational policy analysis, information beyond the mean effects of a treatment is often required. For example, the policy maker could be interested in the extent of variation of a certain educational treatment effect, or in the proportion of individuals who gain from schooling. Methods which try to estimate evaluation parameters beyond the mean treatment effects are discussed, for example, by Heckman, Smith & Clements ( 1 9 9 7 ) , Chib and Hamilton ( 2 0 0 0 , 2 0 0 2 ) , and Carneiro, Hansen Sc Heckman (2003). Although the econometric evaluation literature offers a variety of sophisticated methods for the estimation of causal effects on the basis of non-experimental data, at the end, the choice and applicability of these estimators depend on the available data. In Germany, the information in the data is limited, if not marginal, for the purpose of estimating causal returns to education. In the future, a lot more effort should be taken in collecting appropriate and rich data sets, especially regarding the measurement of cognitive and non-cognitive skills, in order to give valid answers for questions in educational policy analysis.

References Angrist, }., G. Imbens, D. Rubin (1996), Identification of Causal Effects Using Instrumental Variables. Journal of the American Statistival Association, Vol. 91, pp. 444^472. Baumgartner, H. (2004), Are There Any Class Size Effects on Early Career Earnings in West Germany. DIW Discussion Paper, No. 4 1 7 . Becker, S.O., F. Siebern-Tbomas (2001), Returns to Education in Germany: A Variable Treatment Intensity Approach. Economics Working Papers from European University Institute. Belzil, C., J. Hansen (2002), Earnings Risk and the Demand for Higher Education. IZA Discussion Paper 513. Betts, J.R. (1995), Does School Quality Matter? Evidence from the National Longitudinal Survey of Youth. Review of Economics and Statistics, Vol. 77, pp. 2 3 1 - 2 5 0 . Blundell, R., L. Dearden, B. Sianesi (2005), Evaluating the Impact of Education on Earnings in the UK: Models, Methods and Results from the NCDS. Journal of the Royal Statistical Society, Series A, Vol. 168, pp. 4 7 3 - 5 1 2 , IFS Working Papers W03/20, Institute for Fiscal Studies. Bound, J., D. Jaeger, R. Baker (1995), Problems with Instrumental Variables Estimation When the Correlation between Instruments and Endogenous Explanatory Variables is Weak. Journal of the American Statistical Association, Vol. 90, pp. 4 4 3 - 4 5 0 . Brewer, D., R. Ehrenberg (1996), Does It Pay to Attend an Elite Private College? Evidence from the Senior Class of 1980. Research in Labor Economics, Vol. 15, pp. 2 3 9 - 2 7 1 . Brewer, D., E. Eide, R. Ehrenberg (1999), Does It Pay to Attend an Elite Private College? Cross Cohort Evidence on the Effects of College Type on Earnings. Journal of Human Resources, Vol. 34, pp. 1 0 4 - 1 2 3 . Card, D. (1995), Earnings, Schooling and Ability Revisited. Research in Labor Economics, Vol. 14, pp. 2 3 - 4 8 . Card, D. (2001), Estimating the Returns to Schooling: Progress on Some Persistent Econometric Problems. Econometrica, Vol. 69, pp. 1 1 2 - 1 1 6 0 . Card, D., A. Krueger (1992), Does School Quality Matter: Returns to Education and the Characteristics of Public Schools in the United States. Journal of Political Economy, Vol. 109, pp. 1-40.

Causal Returns to Education: A Survey on Empirical Evidence for Germany • 21

Card, D., A. Krueger (1996), Labor Market Effects of School Quality: Theory and Evidence. Pp. 97-140 in: G. Burtless (ed.), Does Money Matter? The Link Between Schools, Student Achievement and Adult Success, Brookings Institution, Washington D.C. Carneiro, P., K. Hansen, ]. Heckman (2003), Estimating Distributions of Treatment Effects with and Application to the Returns to Schooling and Measurement of the Effects of Uncertainty on Schooling Choice. International Economic Review. Cbib, S., B.H. Hamilton (2000), Bayesian Analysis of Cross-section and Clustered Treatment Models. Journal of Econometrics, Vol. 97, pp. 25-50. Chib, S.,B.H. Hamilton (2002), Semiparametric Bayes Analysis of Longitudinal Data Treatment Models. Journal of Econometrics, Vol. 110, pp. 67-89. Dale, S., A. Krueger (2002), Estimating the Payoff to Attending a More Selective College: An Application of Selection on Observables and Unobservables. Quarterly Journal of Economics, Vol. 117, pp. 1491-1528. Franz, W. (2002), ArbeitsmarktÄokonomik. 5. Auflage, Springer, Berlin. Groot, W., H.M. van den Brink (2000), Overeducation in the labor market: a metaanalysis. Economics of Education Review, Vol. 19(2), pp. 149-158. Hanushek, E., S. Rivktn, L. Taylor (1996), Aggregation and the Estimated Effects of School Resources. The Economic and Social Review, Vol. 78, pp. 611-627. Hartog, J., L. Serano (2002), Earnings Risk and Demand for Higher Education: A Cross-Section Test for Spain. IZA Discussion Paper, Vol. 641. Heckman, ]. (1979), Sample Selection as a Specification Error, Econometrica, Vol. 47, pp. 153-161. Heckman, ]. (1997), Instrumental Variables: A Study of Implicit Behavioral Assumptions Used in Making Program Evaluations. Journal of Human Resources, Vol. 32, pp. 441—462. Heckman, ]., A. Layne-Farrar, P. Todd (1996), Human Capital Pricing Equations with an Application to Estimating the Effect of Schooling Quality on Earnings. Review of Economics and Statistics, Vol. 78, pp. 562-610. Heckman, ]., X. Li (2003), Selection Bias, Comparative Advantage and Heterogeneous Returns to Education: Evidence for China in 2000. Discussion paper Nr. 829, IZA. Heckman, J., L. Lochner, P. Todd (2003), Fifty Years of Mincer Earnings Regressions. NBER Working Paper, Vol. 9732. Heckman, J., R. Robb (1986), Alternative Methods for Evaluating the Impact of Intervention, Longitudinal Analysis of Labor Market Data. Heckman, J., J. Smith, N. Clements (1997), Making the Most Out of Programme Evaluations and Social Experiments: Accounting for Heterogeneity in Programme Impacts. Review of Economic Studies, Vol. 64, pp. 487-535. Heckman, }., E. Vytlacil (1998), Instrumental Variables Methods for the Correlated Random Coefficient Model: Estimating the Rate of Return to Schooling when the return is Corelated with Schooling. Journal of Human Resources, Vol. 23, pp. 974-987. Heckman, ]., E. Vytlacil (2005a), Econometric Evaluation of Social Programs, in: J. Heckman, E. Learner (eds.), Handbook of Econometrics, Vol. 6, Elsevier Science, Amsterdam, forthcoming. Heckman, J., E. Vytlacil (2005b), Structural Equations, Treatment Effects, and Econometric Policy Evaluation. Econometrica, Vol. 73, pp. 669-738. Hogan, V., I. Walker (2001), Education Choice under Uncertainty. Working Paper, University College Dublin / University of Warwick. Ichino, Α., R. Winter-Ebmer (1999), Lower and Upper Bounds of Returns to Schooling: An Exercise in IV Estimation with Different Instruments. European Economic Review, Vol. 43, pp. 889-901. Ichino, Α., R. Winter-Ebmer (2004), The Long Run Educational Costs of World War II. Journal of Labour Economics, Vol. 22, pp. 57-86. Imbens, G. (2004), Nonparametric Estimation of Average Treatments Effects Under Exogeneity. Review of Economics and Statistics, Vol. 86, pp. 4-29. Imbens, G., J. Angrist (1994), Identification and Estimation of Local Average Treatment Effects. Econometrica, Vol. 62, pp. 467-476.

22 · A.L. Flossmann and W. Pohlmeier

Jochmann, M., W. Pohlmeier (2003), The Causal Effect of Overeducation on Earnings: Evidence from a Bayesian Approach. Pp. 93-108 in: F. Büchel, A. de Grip, A. Mertens (eds.), Overeducation in Europe: Current Issues in Theory and Policy. Edward Elgar, Cheltenham. Jochmann, M., W. Pohlmeier (2004), Der Kausaleffekt von Bildungsinvestitionen: Empirische Evidenz für Deutschland. Pp. 1-24 in: W. Franz, H.J. Ramser, M. Stadler (eds.), Bildung, Mohr Siebeck, Tübingen. Kling, J. (2001), Interpreting Instrumental Variables Estimates of the Returns to Schooling. Journal of Business and Economic Statistics, Vol. 19, pp. 358-364. Lauer, C., V. Steiner (2000), Returns to Education in West Germany, ZEW Discussion Paper 00-04. Lechner, M. (2001), Identification and Estimation of Causal Effects of Multiple Treatments under the Conditional Indepencence Assumption. Pp. 43-58 in: M. Lechner, F. Pfeiffer (eds.), Econometric Evaluation of Labour Market Policies, Physica/Springer, Heidelberg. Light, Α., W. Strayer (2000), Determinants of College Completion: School Quality or Student Ability?. Journal of Human Resources, Vol. 35, Pp. 299-332. Maier, M. (2004), Causal Effects of Education in Germany. Unpublished Discussion Paper, ZEW, Mannheim. Maier, M., F. Pfeiffer, W. Pohlmeier (2003), Overeducation and Individual Heterogeneity. Pp. 133-152 in: F. Büchel, A. de Grip, A. Mertens (eds.), Overeducation in Europe: Current Issues in Theory and Policy. Edward Elgar, Cheltenham. Maier, M., F. Pfeiffer, W. Pohlmeier (2004), Returns to Education and Individual Heterogeneity. ZEW Discussion Paper, (04-34). Pfeiffer, F. (1994), Selbständige und abhängige Erwerbstätigkeit. Campus Verlag, Frankfurt a.M. Pischke, ]., J. DiNardo (1997), The Return to Computer Revisited: Have Pencils Changed the Wage Structure Too?. Quarterly Journal of Economics, Vol. 112, pp. 291-303. Pischke, J., T. von Wächter (2005), Zero Returns to Compulsory Schooling in Germany: Evidence and Interpretation. IZA DP No. 1645. Pischke, J.S., A. Krueger (1995), A Comparison of East and West German Labor Markets Before and After Unification. Pp. 405-445 in: R. Freeman, L. Katz (eds.), Differences and Changes in Wage Structures. Chicago University Press. Pohlmeier, W. (2004), Bildungsrenditen. Pp. 17-20 in: Investitionsgut Bildung, Workshop "Investition in Humankapital", Bonn, June 7th, 2004. Bundesministerium für Bildung und Wissenschaft. Pohlmeier, W., A. Flossmann (2005), The Potential Outcome of Schooling: Individual Heterogeneity, Program Risk and ResidualWage Inequality. Paper presented at the conference "Heterogenous Labor and Education", DFG Research Group 454, Mannheim, September 12-13, 2005. Powell, J.L. (1994), Estimation of Semiparametric Models. Pp. 2443-2521 in: R.F. Engle, D.L. McFadden (eds.), Handbook of Econometrics, chap. 41, Elsevier Science. Roy, A. (1951), Some Thoughts on the Distribution of Earnings. Oxford Economic Papers, Vol. 3, pp. 135-146. Rubin, D.B. (1974), Estimating Causal Effects of Treatments in Randomized and NonRandomized Studies. Journal of Educational Psychology, Vol. 66, pp. 688-701. Schnabel, I., R. Schnabel (2002), Family and Gender Still Matter: The Heterogeneity of Returns to Education in Germany. University of Mannheim, unpublished working paper. Skarupke, R. (2005), Renditen von Bildungsinvestitionen. Paneldaten-Schätzungen für die Bundesrepublik Deutschland, Peter Lang, Frankfurt a.M. Staiger, D., J.H. Stock (1997), Instrumental Variables with Weak Instruments, Econometrica, Vol. 65, pp. 557-586. Steiner, V., C. Lauer (2000), Private Erträge von Bildungsinvestitionen in Deutschland. ZEW Discussion Paper, Vol. 00-18.

Causal Returns to Education: A Survey on Empirical Evidence for Germany · 23

Vytlacil, E. (2002), Independence, Monotonicity, and Latent Variable Models: An Equivalence Result. Econometrica, Vol. 70, pp. 331-341. Wooldridge, J.M. (2000), Introductory Econometrics: A Modern Approach. South-Western College Publishing. Wooldridge, J.M. (2002), Econometric Analysis of Cross Section and Panel Data. MIT Press, Cambridge, MA. Wooldridge, J.M. (2004), Estimating Average Partial Effects under Conditional Independence Assumptions. Unpublished working paper, Department of Economics, Michigan State University. Prof. Dr. Winfried Pohlmeier, Department of Economics, Box D124, University of Konstanz, D-78457 Konstanz, Germany. Phone ++49/+7531/88-2660. E-mail: [email protected]

Jahrbücher f. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2006) Bd. (Vol.) 226/1

Heterogeneous Returns to Training An Analysis with German Data Using Local Instrumental Variables By Anja Kuckulenz and Michael Maier, Mannheim* JEL J31, C14, C21 Continuing training, treatment effects, semiparametric estimation, local instrumental variables.

Summary Empirical work on the wage impact of training has noted that unobserved heterogeneity of training participants should play a role. The expected return to training, which partly depends on unobservable characteristics, is likely to be a crucial criterion in the decision to take part in training or not. We try to account for this fact by using recent advances in estimating returns to schooling, which allow for selection on unobservables, and apply it to estimating the impact of training on earnings. Allowing heterogeneity to be unobserved by the econometrician, but assuming that individuals may act upon this heterogeneity, completely changes the interpretation and properties of commonly used estimators. Our results based on local instrumental variables suggest that traditional estimates of the wage impact of training overestimate this effect.

1.

Introduction

Investment in continuing vocational training constitutes a m a j o r part o f post school human capital increment. A survey conducted by the I W Cologne (Institut der deutschen Wirtschaft Köln) in cooperation with the chambers o f commerce and crafts shows that German companies consider training important to enhance their competitiveness. In 2 0 0 1 , firms invested nearly 1 7 billion Euro in continuing vocational training (Weiß 2 0 0 3 ) . Almost all companies offer training and about half o f the employees participated in seminars, courses or on-the j o b training. Since 1 9 9 8 , when the last IW-survey on continuing vocational training was conducted, companies have cut the duration o f training spells and expenditure per employee to improve the efficiency o f training. Since the total number of training participants increased at the same time, total investment in continuing vocational training remained at the same spending level as in 1 9 9 8 (Weiß 2 0 0 0 ) . Participation depends on workplace and employee characteristics. In Germany, around 8 0 percent o f high qualified workers take part in training at least once in 1 9 9 7 - 1 9 9 9 ,

* W e thank Alexandra Spitz, T h o m a s Zwick, and seminar participants in Mannheim for helpful comments. W e also thank Iliyan Stankov for his research assistance. This work was supported by the Deutsche Forschungsgemeinschaft (DFG) through the research group Heterogeneous Labor. Neither the Bundesinstitut für berufliche Bildung (BIBB), the Institut für Arbeitsmarkt- und Berufsforschung (IAB) nor the Zentralarchiv (ZA) take any responsibility for the analysis or the interpretation of the data presented here.

Heterogeneous Returns to Training • 25

but less than 30 percent of those workers which are less qualified participate (Kuckulenz/Zwick 2003). Employees in small firms participate less in training than those in large firms and also women, foreigners, and workers above forty years of age receive less training (Pischke 2001). Obviously, heterogeneity of employees plays a role not only in obtaining skills, but also in economic consequences of education and training. As noted by the OECD (2004), only little is known about how the training impact on wages varies across heterogeneous training participants. The small empirical literature with German data has shown that training type as well as worker, job, and firm characteristics determine the wage impact of training. Recently, Pannenberg (1998), Jiirges and Schneider (2005) and Kuckulenz and Zwick (2003) have compared the wage effects of subgroups of employees with German data. Other work concentrates on certain aspects of heterogeneity, e.g. differences in training returns between employees with different educational backgrounds (Lynch 1992/Blundell/Dearden/Meghir 1996, OECD 1999, OECD 2004), age (OECD 2004), men and women (Pischke 2001) or tenure (Pannenberg 1998). Some of these papers come up with rather high estimates for the impact of training on wages which could be explained by unobserved factors: e.g. whether an employee is on a promotion path, climbing a steep ladder upwards or how able and motivated someone is. Pfeiffer and Reize (2001) interpret their results to show that training and career paths are intertwined and that higher wages may not actually be the consequence of training, but result from excellent career management. Likewise, Pischke (2001) finds that selection in training seems not to be based on wage levels but rather on earnings growth. For an overview of recent international estimates of wage returns to continuing vocational training, compare Leuven (2005). Recent literature on the returns to schooling provides methods which allow that returns may vary across schooling types and participants. Carneiro, Heckman and Vytlacil (2003) apply methods which allow for the likely fact that the expected return of the investment in human capital plays a role when deciding about the investment. Among others, also Blundell, Dearden and Sianesi (2005) have noted the importance of allowing for (observable) heterogeneity in returns to education. Selection into training may depend much more on individuals' ability and motivation than does selection into schooling, where family background characteristics are the main determinants (Ammermüller 2004, Lauer 2002). Also, training costs and maybe even training returns are more obvious and hence may play a more crucial role for the decision to take part in training than for the schooling decision. Our study uses recent econometric methods, which allow for selection on unobservables, and apply it to estimating the impact of training on earnings. With German survey data from 1998/1999, we explore the heterogeneity of training returns and how this may affect participation in training. By using the Local Instrumental Variables (LIV) method, we account for the heterogeneity of training returns in our analysis and allow for observed as well as unobserved heterogeneity. Therefore, the assumptions are much less stringent than those of ordinary least squares or conventional IV regressions. In fact, if unobserved heterogeneity is relevant and individuals act upon it, OLS and linear IV estimates can be seriously misleading (see Heckman/Vytlacil 2005). The following section provides a brief discussion of the theoretical background and previous empirical work. In the following, first the econometric method used is introduced, second the data set is described, and third the implementation and the estimation results are presented. The last section concludes and gives an outlook.

2 6 · A. Kuckulenz and M . Maier

2.

Background discussion

It has been noted in the literature that bargaining and rent-sharing between employer and employee should have an impact on the share of the rent generated by training which is granted to the training participant (e.g. Dear den/Reed/Van Reenen 2000 and Arulampalam/Booth/Bryan 2004). Therefore, heterogeneity in training returns cannot only be explained by differences in productivity effects of training, but also by differences in individual, firm, and job characteristics which relate to the bargaining power of employer and employee. Lazear (1979) notes that wages and productivity at a given point in the career do not have to correspond. Employees may first receive wages that are lower than their productivity and at a later stage of their professional career, they can profit from early investments in their human capital. In contrast, training demand should be highest for firm entrants and also the productivity effect of training should then be highest for this group. The return to training for workers with low qualifications should be higher if individuals with low qualifications are constrained in their choice of education. On the other hand, it may be that employer provided training is complementary to education (Blundell/Dearden/Meghir 1999) and therefore favors higher skilled employees. Kuckulenz and Zwick (2003) find with German data that the effect of training on earnings differs between agents with a broad spectrum of different characteristics and between firms with different characteristics. High-skilled employees profit more from training than low-skilled workers, the training earnings mark-up increases with professional experience but decreases with company tenure. Employees with previous unemployment spells and employees with temporary contracts profit less from training. Smaller firms, firms in a good economic situation, and firms that share profits with their employees pay a higher training earnings mark-up. The authors interpret these findings as evidence that the training wage effect not only depends on the productivity increase induced by training, but also on the bargaining position between employer and employee. Hence, the increase in productivity caused by training must not directly correspond to the wage effect of training. Nevertheless, the wage impact of training is frequently taken as (the lower bound of) the productivity impact of training (Blundell/Dearden/Meghir 1999). The decision to take part in continuous training is likely to be influenced by the expected returns to training; i.e. those workers for whom the expected return is higher will obtain more training than other workers for whom the expected return is lower. Hence, participants and nonparticipants in training are unlikely to have the same observed and hypothetical returns. Severe econometric problems are therefore posed by the endogeneity of training decisions. While former empirical work with German data has extensively analyzed the wage effect of training, none of them has accounted for the likely possibility that worker selection into training is based on the expected heterogeneous return to continuous training. Previous work has solved the endogeneity problem by using a Heckman-type selection correction term from a training participation equation (e.g. Lynch 1992). Also Blundell, Dearden and Meghir (1996) argue that continuous training might be correlated with transitory shocks to productivity and therefore include a Heckman correction term into their wage growth equation. Other authors tried to tackle the endogeneity problem by using instrumental variable estimation (Leuven/Oosterbeek 2001 or Kuckulenz/Zwick 2003) or nonparametric matching methods (Gerfin 2004). Also, fixed effects estimators have been used (Booth/Bryan 2005, Pischke 2001 or Barron/Berger/Black 1999), which produce unbiased estimates whenever unobserved individual effects are permanent. Leu-

Heterogeneous Returns to Training · 27

ven/Oosterbeek (2002) use a different approach to estimate the causal effect of training on wages by using information about workers who planned to participate in training but did not due to some exogenous event. They use this group of workers as the comparison group and assume that within their sample, participation in training is random. The estimated least-squares coefficient of the individual's choice parameter is only then to be interpreted as the causal effect of training on wages if workers are randomly assigned to take part in training. We have argued above that employees are either chosen by the employer providing training or that they select themselves into training and this implies that standard estimations using least squares produce biased results. Therefore, we rely on recent advances in estimating the returns to schooling using evaluation methods. In the literature, which was mainly spurred by Heckman and co-authors, schooling is treated as an endogenous variable in the standard wage function. While former work on training in Germany relied on the unconfoundedness or selection on observables assumption, we want to explicitly allow for heterogeneity of wage effects of training and for selection on unoberservables. With detailed information about the qualification profile and professional history of workers, the organizational and technological condition of workplaces, as well as some employer characteristics, we are able to explain a large part of the variation in wages. Nevertheless, some characteristics which are crucial for the selection into training are missing; above all ability, motivation and the information whether individuals are on a promotion path. Former work has also shown that workers with higher wage growth participate more often in training (Pischke 2001). With our cross section data, we cannot account for this directly since we observe wages only once. The advantage of the econometric model we are using is that it allows the effect of training to vary both in terms of observed and unobserved factors. Firms may offer training to those workers who are expected to be more productive after training or those workers who expect wage gains from training participation may select themselves into training courses. Since the probability of treatment increases with the gains from treatment, we allow the impact of training on earnings to differ across individuals and for selection on gains. Hence, we assume that individuals are forward looking agents who have expectations on the impact of training participation on their wage. Adequate instrumental variables have to be found that explain the selection into training participation in order to correct for treatment selection. We should stress again that under the heterogeneity assumptions stated above, conventional IV methods will not yield unbiased results. To get reliable results, much stronger assumptions on effect heterogeneity or individual choice behavior have to be imposed, which might be implausible in our case. Therefore, if no stronger assumptions can be made, evaluation methods like the local IV model are necessary to estimate the impact of training on earnings. 3.

Econometric model

The causal effects of training on earnings are analyzed within the framework of econometric evaluation methods. These methods take into account heterogeneous effects of training for each individual, which may depend on observable or unobservable factors. Allowing heterogeneity to be unobserved by the econometrician, but assuming that individuals act upon this unobserved heterogeneity completely changes the interpretation and properties of common estimators taking (observed) heterogeneity into account. Cameiro, Heckman and Vytlacil (2003) and Heckman and Vytlacil (2005) show that conventional

28 · A. Kuckulenz and M . Maier

IV estimators substantially misestimate the average marginal return and policy relevant effects. Two main streams of non-experimental methods taking unobserved heterogeneity into account can be divided. First, there are methods which control for the correlation between individual factors and program participation by using an adequate instrument. The second approach is to measure all individual factors that may be the cause of the correlation between individual factors and program participation and then, for example, match on these observed variables (Blundell/Dearden/Sianesi 2005). For a review of different approaches see Heckman, LaLonde and Smith (1999) or Caliendo and Hujer (2005). Imbens (2004) describes methods for selection on observables, Angrist (2004) focuses on models where selection is influenced by unobservable heterogeneity. While selection models try to model the complete selection process, the IV method, which is used here, focuses on searching a source of independent variation affecting the decision to participate but not the outcome (in our case, earnings). Other estimation strategies are based on difference in difference estimation which erase only time-invariant selection. In the following, a formal description of the basic framework of evaluation econometrics is given. Let D indicate the choice of treatment, that is β _ i 1 if the individual receives treatment, — 1 0 otherwise. Concerning the outcome variable, it is assumed that latent values exist for every possible value of the treatment variable. These latent outcome variables are denoted by Yj and Yo for D = 1 and D = 0, respectively. Only one of the two latent outcomes can be observed, as every individual can solely choose one treatment status. Therefore, the observed outcome is given by Y = D Y j + (1 — D)YQ. In the binary treatment case at hand, it means that every individual would receive an income in the treated as well as in the untreated case. Yj is observed for participants and YQ for nonparticipants. The causal effect of treatment D on the outcome variable Y is defined to be A = YI -

Y0.

(1)

This difference is unobservable for every individual, as either YJ or YQ cannot be observed. Different methods have been developed in the literature to overcome this problem. In general, averages of (1) for various subgroups are considered. The average treatment effect Δ α τ ε is the effect on an average individual of the population, whereas the average treatment effect on the treated Δ π and the average treatment effect on the untreated state the effects for the subpopulations of treated and untreated individuals, respectively. Formally, the effects are defined by Δ

ΑΤΕ

: =

£ [ Y I

_

YO]

Δ Π : = £ [ Y 1 - y 0 | D = l] A ™

:=

E[YJ -

Y0|D = 0]

( 2 )

(3) (4)

All effects can be defined conditional on X , for example Δ Α Τ Ε ( χ ) = £[Δ|Χ = * ] . In the empirical analysis of this paper, we use the Local Instrumental Variable (LIV) method of Heckman and Vytlacil (1999, 2000, 2001, 2005). First, the framework and underlying assumptions are described. Then we line out another causal effect, the marginal

Heterogeneous Returns to Training · 29

treatment effect, which was defined by Heckman and Vytlacil (1999) and relationships with the various types of causal effects are shown. Finally, the estimation strategy is outlined. The treatment indicator D is modelled by a latent index model: D = 1 ( μ ο ( Ζ ) - U D > 0).

(5)

1(A) is the indicator function, that is 1(A) = 1 if A is true and 1(A) = 0 otherwise. μ£)(Ζ) is a function of some instrumental variables Z. The latent outcomes are functions of some observable variables X and unobservable factors Ug and U¡, i.e. Y, = g(X, U¡), for i = l, 2. Participation in training corresponds to D = 1, nonparticipants are identified by D = 0. Heckman and Vytlacil (1999, 2000, 2001, 2005) state the following assumptions: • Given Χ, μβ(Ζ) depends in a nontrivial way on Z. This corresponds to the usual assumption of instrument relevance in linear IV models, i.e., the instruments have to influence the training decicion after controlling for other covariates X. • Ud is independent from X and all error terms in the model are independent from Ζ given X. This is the usual exclusion restriction of IV models which states that Ζ has no influence on the dependent variable after the covariates X are accounted for. A detailed discussion of these assumptions in the context of evaluation models is given by Vytlacil (2002). • The error term Un of the latent index model (5) is assumed to be absolutely continuous with respect to Lebesgue measure. • Furthermore, £|Yi | and £ | Yq| are assumed to be finite, which guarantees the existence of E[Y]. • For every individual, the probablility of participation P(D = 1) lies strictly between zero and one, given the observable characteristics X. With this setup, Heckman and Vytlacil (1999, 2000, 2001, 2005) define the marginal treatment effect, which is the causal effect of D given X and Ud : Δ μ τ ε ( χ , U) = E[Yj - Y 0 |X = χ,υΒ

= U]

(6)

The marginal treatment effect provides a framework to obtain expressions for various average treatment effects. Heckman and Vytlacil (1999,2000, 2001) derive the following relationships: (7) Δπ(χ) =

r1

Jo

A^ix) = f1

JO

du du.

(8)

(9)

Here, P(Z) is shorthand for P(D = 1|Z). Therefore, integration of the suitable weighted marginal treatment effects over the [0,1]-interval yields estimates of treatment effects for

30 · A. Kuckulenz and M. Maier

different subpopulations. The basic ingredient of this procedure is the marginal treatment effect. To get an estimate of it, the Local Instrumental Variables (LIV) estimator was proposed by Heckman and Vytlacil (1999, 2000, 2001, 2005): à ^

M

-

"

*

™

"

·

™

dP(z)

-

™

M

The LIV method estimates the marginal treatment effect for u = P(z). This can be seen by forming the derivative of the expectation of Y given P(Z) and noting that Y = D Y j + (1 - D ) Y 0 . The definition of the local average treatment effect A l a t e of Imbens and Angrist (1994) can be used to motivate the marginal treatment effect. The LATE is defined by ALATE(x,

P(Z), P(zf))

=

E[Y\P(Z) = P(z), X = χ] - E[Y\P(Z) = P(zf), X = x] P(D = l\Z = z) - P(D = 1\Z = z')

'

(11)

This is the treatment effect for the subgroup of individuals who change their treatment status due to a change of the instrument Ζ from ζ to z'· This subgroup of so-called compilers cannot be identified in a given dataset. For comments and criticism of this concept see Heckman (1997) and Angrist, Imbens and Rubin (1996) and the accompanying discussions. Considering P(z) —• P(z'), the expression of the LATE tends to the derivative of the conditional expectation of Y: lim P(z)->P(z0

dP(z)

(12)

The LIV-estimator estimates some sort of marginal LATE. Therefore, the marginal treatment effect can be interpreted as the effect on an individual with observable characteristics X and unobservables Up which is indifferent about participation.

4. Data For our analysis, we use a rich and representative German data set with information on 0.1 percent of all individuals employed in 1998/1999 - the BIBB/IAB data set "Qualification and Career Survey" . The cross-section data allow an assessment of the impact of training measures in 1996-98 on wages in 1998/1999. Our sample contains more than 7,400 male (full-time) employees from West Germany. We include about 70 explanatory variables that capture the salient employer and employee characteristics for wage determination (see also Table A l in the appendix for the complete list with detailed descriptions). The outcome variable is log midpoints of earnings in 1998/1999 from 18 earnings categories in the data. This variable has the advantage that earnings of highly paid workers are not censored from above. The key explanatory variable is participation in training courses and seminars that serve professional development during the years 1996 to 1998. This dummy might stand for quite substantial amounts of training, because employees might participate in various courses during 24 months. In addition, only formal training courses are included in the

Heterogeneous Returns to Training · 31

data set and short or informal training spells are explicitly excluded. Note that apprenticeship training is also excluded. Unfortunately, we don't have information about the content and the length of courses and we do not know the costs of training or who bears them. 5 8 percent of the employees participated in further training according to this definition (Table A l in the appendix). Participation differs tremendously for low and high skilled employees: while only around 3 0 percent of the workers without professional degree participated in training, 5 0 percent of the employees with a vocational school degree or an apprenticeship training took part in some kind of training during the last two years, and about 8 0 percent of high skilled employees ( master craftsman, university of applied sciences, and university degree) participated. Training participation also varies with age: 30—45 year old employees receive most training, older worker participate less (see Figures A l and A2 in the appendix). In our estimations, we use three different instrumental variables. The first instrument we use is the training intensity by industry, estimated from an earlier wave ( 1 9 9 1 / 1 9 9 2 ) of the BIBB/IAB-survey. Theoretically, it is not plausible that training intensity per sector in 1 9 9 1 / 1 9 9 2 influences individual wages in 1 9 9 8 / 1 9 9 9 . In contrasts, it is very likely that individuals which are employed in a sector with a high training intensity in 1 9 9 1 / 1 9 9 2 have a higher chance to participate in training in 1 9 9 8 / 1 9 9 9 than individuals in an economic sector with a low training intensity in 1 9 9 1 / 1 9 9 2 . Empirically, the training intensity per economic sector in 1 9 9 1 / 1 9 9 2 is not correlated with earnings seven years later but influences the individual probability to take part in training. Also, we imputed data from the Continuing Vocational Training Survey (CVTS 2 0 0 0 ) about sectoral shares of firms and shares of firms by employment size that include continuous training in their collective bargaining agreement. 1 There is no obvious link of this variable with individual earnings, using additionally sector dummies as covariates, but it influences the individual probability to take part in training. As third instrumental variable, we use a dummy variable indicating whether workers are employed in a modern job (in contrast to a traditional job) because the demand for training is higher in these jobs. After controlling for economic sector and occupation, we consider it as an valid instrument (compare results of the propensity score, Table A2). Further determinants of earnings are those found in the Mincer equation, i.e. actual work experience, 2 job tenure, former unemployment, and dummies for the highest educational achievement. 3 These variables are related to the situation in 1 9 9 8 / 1 9 9 9 . Together with these standard variables, we also include 11 dummies capturing the professional status, such as blue-collar or white-collar worker, civil servant or different sophistication levels of tasks for 1 9 9 8 / 1 9 9 9 . In addition, we use the following current job characteristics: computer use, profit-sharing, bonus payments, overtime work, whether a job is temporary, and 13 dummies for main job contents. These variables allow us to control a part of the individual heterogeneity between the employees. 4 Some of these variables (for example working overtime) can

1 2

3

4

The CVTS data is from 1999 and therefore fits well to the BIBB/IAB data set. We know when the individual started his or her first job and we include dummies for discontinuation such as unemployment. In Germany, the highest schooling degree is more informative for the level of education than years of schooling (Georegellis/Lange 1997). Some of these variables may also be endogenous in the earnings equation. We do not control for this, because those variables mainly serve as control variables for employee heterogeneity.

32 · A. Kuckulenz and M. Maier be interpreted as indicators of intrinsic motivation. Additional control variables explaining earnings are personal attributes. We include a dummy for children and non-German nationality. Finally, we add some employer characteristics: seven dummies for firm size, 5 dummies indicating the economic sector of the employer, 11 dummies for the federal state the firm is located in, and a dummy indicating whether the firm is in a good economic situation in 1 9 9 8 / 1 9 9 9 . Only full-time 5 employees (without self-employed) in West Germany are included, because in 1 9 9 8 there were still large differences in the labor market structures of the two parts of the country. The analysis is restricted to male employees, because the data do not allow us to model participation in the labor market simultaneously, which would be important for examining earnings effects for women.^ This reduces the sample to around 7 , 5 0 0 individuals. The descriptive characteristics of the variables used can be found in Table A l in the appendix. In order to obtain clean evidence on the earnings effects of employer-provided training, we exclude those training participants where we cannot be sure whether they were employed or unemployed while being trained (about 4 5 0 cases). The reason for this restriction is that we want to exclude training provided by government aimed at unemployed. Wage effects of training should differ for those employees which stay with a firm and those which move (Booth/Bryan 2 0 0 5 , Gerfin 2 0 0 4 , Loewenstein/Spletzer 1 9 9 8 , Lynch 1 9 9 2 , O E C D 2 0 0 4 , Garloff/Kuckulenz 2 0 0 6 ) . In our data set we can only identify very few individuals which change their employer after attending continuing training. We cannot show any significant difference between job stayers and movers and hence, we restrict our sample to those which stay with their employer.

5.

Implementation and results

The basic building blocks of the empirical analysis are estimates of Δ Μ Τ Ε ( χ , ü). For this purpose, estimates of the derivative of the conditional expectation of Y given X and P(Z) are needed. The latent outcome equations are specified as: In Υχ = αχ + Χθχ + ϋχ

(13)

In Y 0 = α 0 + Χ θ 0 + υ 0

(14)

The observable outcome is therefore given by In Y = D In Y j + (1 - D ) In Y 0 = a0 + Χθ0 + D(ax +DUx

5

6

+ ( 1 -D)U0.

- a0)

+ ΌΧ(θχ

-

θ0) (15)

We include only employees working 30 hours and above per week. Only 2.6 percent of the male employees work less than 30 hours. We also use a dummy for working overtime in order to take hours worked into account. The results do not change qualitatively, however, if we use log hourly wages instead of log earnings as the dependent variable. In order to include women, we would need to correct for sample selection in the earnings equation. This is impossible since only those women who participate in the labor market are included in the data.

Heterogeneous Returns to Training · 33

From this, the conditional expectation of In Y given X and P(Z) follows as E[ln Y|X, P(Z)] = «ο + Χθο + p ( z ) ( « 1 - «θ) + Ρ(.Ζ)Χ(Θι - θ0) + P(Z)E[U11Ρ(Ζ)] + (1 - P(Z))E[U 0 |P(Z)]

(16)

The derivative of (16) with respect to Ρ (Ζ) is given by 9£[ln Y|X, P(Z)] = ( a 1 - a o ) + X ( 0 1 - f l b ) + K(P(Z)) 3P(Z)

(17)

where K(P(Z)) =

8(Ρ(Ζ)Ε[[71|Ρ(Ζ)]^1-Ρ(Ζ))Ε[[70|Ρ(Ζ)])-To estimate pointwise estimates for all X and U (within the [0,1] interval) are needed. To reduce the dimension of the problem, the expectation is modelled as a partial linear model (see Carneiro/Heckman/Vytlacil 2003). The constant term and the term depending on X enter the conditional expectation linearly, whereas K(P(Z)) is modelled nonparametrically. To estimate these characteristics of the equation, the "double residual regression" of Heckman, Ichimura, Smith and Todd (1998) is used. This slight variation of the partial linear model of Robinson (1988) is tailored for the evaluation of binary treatment effects. First, equation (15) is rewritten in the following form: In Υ = «ο + X6>o + D(«i - ao) + DX(6>! - 0O) + D ^ i + (1 +P(Z)E[U1\P(Z)] + (1 - P(Z))£[U 0 |P(Z)] -P(Z)E[U1\P(Z)]-(1-P(Z))E[U0\P(Z)).

D)U0 (18)

The term D U j + (1 - D)U0 - P(Z)E[U1\P(Z)] - (1 - P(Z))£[U 0 |P(Z)] is gathered in an error term ε, which has mean zero given P(Z) by construction: In Υ = α 0 + Χθ0 + D(a 1 - a0) + DX(Û1 - θ0) +P(Z)E[U1IP(Z)]

+ (l-P(Z))E[U0¡P(Z))

+s

(19)

In parlance of partial linear models, the term P(Z) E[U1 |P(Z)] +(1 - P ( Z ) ) E[U0\P(Z)\ is the nonparametric component. From this, the conditional expectation of In Y given P(Z) follows: E(ln Y|P(Z)) = α 0 + £(Χ|Ρ(Ζ))Θ 0 + Ρ ( Ζ ) ( α ι - α 0 ) +P(Z)£(X|P(Z))(Ö! - θ0) + P(Z)E[Ui IP(Z)] +(l-P(Z))£[Uo|P(Z)]

(20)

Subtracting (20) from (19) yields: In Y - E [In Y|P(Z)] = (Χ - £[Χ|Ρ(Ζ)])Θ 0 + (D - Ρ(Ζ))χ («! - α 0 ) + (DX - Ρ(Ζ)£[Χ|Ρ(Ζ)])(0! - θ0) + ε

(21)

The conditional expectations £[Χ|Ρ(Ζ)] are estimated pointwise by local-linear regressions. After forming the differences, (21) is estimated by OLS. Using the estimated residuals from this regression, the derivatives of P(Z) E[Ui\P(Z)] +(1 -P(Z))£[Uo|P(Z)] can be estimated by the appropriate coefficients of local polynomial regressions. Using the

34 · A. Kuckulenz and M. Maier

empirical distributions of F(P(Z)\X), the weights for the integration of Δ Μ Τ Ε ( χ , u) over [0, 1] can be computed. Using the empirical distribution of X , unconditional treatment effects can be obtained. To judge the significance of the estimated effects, confidence intervals based on 50 bootstrap samples are computed. The propensity score is specified as a probit model. The estimated coefficients are contained in Table A2 in the appendix. All instruments are significant. The estimated treatment effects and the bootstrap confidence intervals are contained in Table 1. The point estimates of the treatment effects are negative. However, the confidence intervals show that the effects are not statistically significant. Therefore, no statements about the sign of the effect can be made. The wide confidence intervals show a considerable uncertainty about the causal effect of training. Table 1 : Estimates of the treatment effects Original sample

Bootstrap samples Confidence intervals 90% 95%

TUT ATE TT

-.077 -.073 -.063

(-.123, .015) (-.119, .015) (-.112, .014)

(-.145, .039) (-.142, .038) (-.134, .032)

The LIV estimates are lower than the relevant OLS estimate, which is 0.03 (t-value: 3.62) and considerably lower than the standard IV estimator, using the same instruments, which is 0.21 (t-value: 2.37). The latter results suggest a significant positive effect of training on wages. The insignificant LIV-estimates cannot support this conclusion. That is, using a method with considerably weaker assumptions on individual behavior than OLS or linear IV, the qualitative results of the latter cannot be confirmed. Therefore, the positive relationship claimed by conventional estimates has to be questioned. This is in line with the supposition stated in the literature, that estimates which do not account for (unobserved) heterogeneity and the selection in this regard are upward biased. We interpret our result that training does not have an impact on earnings itself but only in combination with unobserved factors. It is likely that training is part of a promotion path and that not a certain training, but a career track as a whole leads to earnings growth. Firms provide training to individuals only when the expected return of this investment is positive. Hence, training participants might be more able and motivated and therefore also be on such a track with higher earnings growth. When this unobserved heterogeneity is taken into account in the selection into training, the positive training impact estimated by conventional OLS or IV estimates vanishes.

6.

Conclusion

With German survey data from 1998/1999, we examine the heterogeneity of training returns and whether these may have an effect on training participation. Using the local IV method, which allows for the likely fact that the expected return of the investment in human capital plays a role when deciding about the investment, we are able to account for heterogeneity of training returns in earnings equations. The LIV estimator allows for

Heterogeneous Returns to Training · 35

observed as well as unobserved heterogeneity and selection into training may depend on both. Former work on the wage impact of training has suggested that selection on unobservables might be important and hence, traditional estimators used might incorporate an upward bias. Our LIV estimate is much lower than the relevant OLS and IV estimate (and furthermore, insignificant). We cannot find any causal effect of training on wages when taking into account that more able and motivated individuals participate in training, or those which are on a promotion path where training courses are part of the way. For future work it would be useful to use comprehensive information on career tracks and promotion in order to distinguish the impact of certain personnel measures. A A1.

Appendix Tables

Table A1 : List of Variables Used Variable Earnings Less than 600 DM Between 600 and 1000 DM Between 1000 and 1500 DM Between 1500 and 2000 DM Between 2000 and 2500 DM Between 2500 and 3000 DM Between 3000 and 3500 DM Between 3500 and 4000 DM Between 4000 and 4500 DM Between 4500 and 5000 DM Between 5000 and 5500 DM Between 5500 and 6000 DM Between 6000 and 7000 DM Between 7000 and 8000 DM Between 8000 and 9000 DM Between 9000 and 10000 DM Between 10000 and 15000 DM 15000 DM and more School Attainment Without School Leaving Certificate Lower Secondary School Intermediate Secondary School Entrance Examination for University of Applied Sciences High School Diploma

Share/Average

Notes

0.09% 0.12% 0.32% 1.20% 4.24% 7.54% 11.98% 14.75% 14.13% 12.19% 8.14% 7.15% 7.15% 4.04% 2.70% 1.51 % 2.22% 0.53% 2.28% 51.33% 25.20% 7.93% 13.25%

Vocational Training Without Professional Degree Full-Time Vocational School

12.08% 2.37%

Dual Apprenticeship

61.30%

Several years of professional training in school Several years of professional training in school and on-the-job

36 · A. Kuckulenz and M. Maier

Table A1: List of Variables Used (continued) Variable

Share/Average

Master Craftsman University of Applied Sciences University

12.64% 4.92% 6.35%

Training Training

58.08%

Professional Career Professional Experience Company Tenure

21.87 years 12.91 years

Unemployment

29.85%

Workplace Characteristics Computer Work Station

49.37%

Temporary Work Good Economic Situation Working Hours Overtime

49.36% 63.25% 177.21 hours 79.95%

Paid Overtime Overqualified Profit-Sharing Incentive Wage Good Economic Situation

35.93% 36.50% 9.20% 24.11% 63.25%

Modern Job

12.06%

Individual Children

Characteristics

48.51 %

Foreigner

5.43%

Not Married Handicapped

7.33% 4.85%

Other Variables Size of Firm Professional Status Federal State Economic Sector

Notes

Years from first job until today Years from starting to work for a company until today Dummy = 1 if a person was ever employed, otherwise 0 Work routine includes using the computer Working hours per month Dummy = 1 if a person works overtime, otherwise 0

Dummy = 1 if the company is in a good economic situation, otherwise 0

Dummy = 1 if a person has at least one child, otherwise 0 Dummy = 1 if a person does not have a German Nationality, otherwise 0

7 Categories 12 Categories 11 Categories: all Federal States of West Germany 5 Categories

Heterogeneous Returns to Training · 37

Table A2: Estimation results for the propensity score Parameters Training Intensity 91 Bargaining Agreement Modern Job Lower Secondary School Intermediate Secondary School Entrance Examination for University of Applied Sciences High School Diploma Full-Time Vocational School Dual Apprenticeship Master Craftsman University of Applied Sciences University Professional Experience Company Tenure Unemployment Computer Work Station Temporary Work Paid Overtime Working Hours Overqualified Profit-Sharing Incentive Wage Good Economic Situation Children Not Married Handicapped Constant Number of Observations LR-Test (χ 2 (72)) P-value of LR-Test

Estimates

Ζ - Values

1.26 1.84 0.18 0.29 0.29 0.30

5.29 3.03 3.42 1.91 1.92 1.91

0.23 0.01 0.19 0.23 0.21 0.10 0.05 0.35 -0.00 0.36 -0.24 -0.03 0.34 -0.03 0.04 0.13 -0.02 0.08 0.01 0.01 -3.37

1.43 0.07 2.58 2.61 2.00 0.90 0.72 5.02 -0.11 7.59 -2.52 -0.76 4.34 -0.69 0.59 3.04 -0.43 2.00 0.12 0.10 -11.52

7417 1987.10 0.00

Dummy variables are included for size of firm, professional status, federal state, and economic sector. Instruments included are: technical restructuring, organizational restructuring, three measures of personnel restructuring (hiring of additional workers, downsizing, and hiring of temporary workers), a dummy variable indicating whether workers are employed in a modern job, and sectoral shares of firms by employment size that include continuous training in their collective bargaining agreement.

38 · A. Kuckulenz and M. Maier

70%

60% 50% 40% 30% 20% 10%

0% 20-24

25-29

30-34

35-39

40-44

45-49

50-54

55-

Figure A1 : Training Participation per Age Groups

100% 90%

80% 70%

60% 50% 40% 30% 20% 10%

0% 1

2

3

1 - Without Professional Degree

4

5

6

4 - Master Craftsman

2 - Full-Time Vocational School

5 - University of Applied Sciences

3 - Dual Apprenticeship

6 - University

Figure A2: Training Participation per Qualification Groups

Heterogeneous Returns to Training · 39

References Ammermüller, A. (2004), PISA: W h a t M a k e s the Difference? Explaining the G a p in PISA Test Scores Between Finland and Germany. Z E W Discussion Paper N o . 04-04, M a n n h e i m . Angrist,J. (2004), Treatment Effect Heterogeneity in Theory and Practice. Economic Journal, 114, pp. C 5 2 - C 8 3 . Angrist, J., G. Imbens, D. Rubin (1996), Identification of Causal Effects Using Instrumental Variables. Journal of the American Statistical Association, 91, pp. 4 4 4 - 4 7 2 . Arulampalam, W., A.L. Booth, M.L. Bryan (2004), Are There Asymmetries in the Effects of Training on the Conditional Male Wage Distribution?. Discussion Paper 984, IZA, Bonn. Barron, J.M., M.C. Berger, D.A. Black (1999), D o Workers Pay for O n - t h e - j o b Training?. Journal of H u m a n Resources, 34(2), pp. 2 3 6 - 2 5 2 . Blundell, R., L. Dearden, C. Meghir (1996), The Determinants and Effects of Work-Related Training in Britain. Institute for Fiscal Studies, London. Blundell, R., L. Dearden, C. Meghir (1999), Work-Related Training and Earnings. Discussion paper, Institute for Fiscal Studies, London. Blundell, R., L. Dearden, B. Sianesi (2005), Evaluating the effect of education on earnings: Models, methods and results f r o m the National Child Development Survey. Journal of the Royal Statistical Society: Series A, 168, pp. 4 7 3 - 5 1 2 . Booth, A.L., M.L. Bryan (2005), Testing Some Predictions of H u m a n Capital Theory: N e w Training Evidence f r o m Britain. The Review of Economics and Statistics, 87(2), pp. 3 9 1 - 3 9 4 . Caliendo, M., R. Hujer (2005), The Microeconometric Estimation of Treatment Effects - An Overview. IZA Discussion Paper, 1653. Carneiro, P., J. Heckman, E. Vyctlacil (2003), Estimating the Return to Education W h e n It Varies Among Individuals, mimeo, University of Chicago. Dearden, L., H. Reed, J.V. Reenen (2000), W h o Gains W h e n Workers Train? Training and C o r p o r a t e Productivity in a Panel of British Industries. Garloff, Α., A. Kuckulenz (2006), Training, mobility and wages: specific versus general h u m a n capital. Jahrbücher für Nationalökonomie u n d Statistik, forthcoming. Georgellis, Y., T. Lange (1997), The Effect of Further Training on Wage G r o w t h in West Germany, 1 9 8 4 - 1 9 9 2 . Scottish Journal of Political Economy, 44(2), pp. 1 6 5 - 1 8 1 . Gerfin, M. (2004), Work-Related Training and Wages: An Empirical Analysis for Male Workers in Switzerland. Discussion Paper 1078, IZA, Bonn. Heckman, J.J. (1997), Instrumental variables: a study of implicit behavioral assumptions used in making p r o g r a m evaluations. Journal of H u m a n Resources, 32, pp. 4 4 1 - 4 6 2 . Heckman, J.J., H. Ichimura, J. Smith, P. Todd (1998), Characterizing selection bias using experimental data. Econometrica, 66, pp. 1 0 1 7 - 1 0 9 8 . Heckman, J.J., R.J. LaLonde, J.A. Smith (1999), The economics and econometrics of active labor markets programs. Pp. 1 8 6 4 - 2 0 9 7 in: A. Ashenfelter, D. Card (eds.), H a n d b o o k of Labor Economics, vol. 3, N o r t h Holland, Amsterdam. Heckman, J.J., E.J. Vytlacil (1999), Local instrumental variables and latent variable models for identifying and bounding treatment effects. Proceedings of the National Academy of Sciences, 96, pp. 4 7 3 0 ^ 4 7 3 4 . Heckman, J.J., E.J. Vytlacil (2000), The relationship between treatment parameters within a latent variable f r a m e w o r k . Economic Letters, 66, pp. 3 3 - 3 9 . Heckman, J.J., E.J. Vytlacil (2001), Local instrumental variables. Pp. 1 - 4 6 in: C. Hsiao, K. M o r i m u n e , J.J. Powell (eds.), Nonlinear statistical modelling, Cambridge University Press, Cambridge. Heckman, J.J., E.J. Vytlacil (2005), Structural Equations, Treatment Effects and Econometric Policy Evaluation. Econometrica, 73, pp. 6 6 9 - 7 3 8 . Imbens, G. (2004), N o n p a r a m e t r i c Estimation of Average Treatment Effects under Exogeneity: A Review. Review of Economics and Statistics, 86, pp. 4 - 2 9 . Imbens, G.W., J.D. Angrist (1994), Identification and Estimation of Local Average Treatment Effects. Econometrica, 62, pp. 467—475.

40 · A. Kuckulenz and M. Maier

Jiirges, H., Κ. Schneider (2005), Dynamische Lohneffekte berufliche Weiterbildung - Eine Längsschnittanalyse mit den Daten des SOEP. Discussion Paper No. 92, MEA, Mannheim. Kuckulenz, Α., T. Zwick (2003), The Impact of Training on Earnings - Differences Between Participant Groups and Training Forms. ZEW Discussion Paper No. 0 3 - 5 7 , Mannheim. Lauer, C. (2002), A Model of Educational Attainment Application to the German Case. ZEW Discussion Paper, (02-06). Lazear, E.P. (1979), Why Is There Mandatory Retirement?. The Journal of Political Economy, 87(6), pp. 1261-1284. Leuven, E. (2005), The Economics of Private Sector Training: A Survey of the Literature. Journal of Economic Surveys, 19, pp. 91-111. Leuven, Ε., H. Oosterbeek (2001), Firm-Specific Human Capital as a Shared Investment: Comment. American Economic Review, 91, pp. 342-347. Leuven, Ε., H. Oosterbeek (2002), A New Approach to Estimate the Wage Returns to WorkRelated Training. IZA Discussion Paper, No. 526. Loewenstein, M.A., J.R. Spletzer (1998), Dividing the Costs and Returns to General Training. Journal of Labor Economics, 16(1), pp. 142-171. Lynch, L.M. (1992), Private-Sector Training and the Earnings of Young Workers. The American Economic Review, 82(1), pp. 299-312. OECD (1999), Employment Outlook. OECD, Paris. OECD (2004), Employment Outlook. OECD, Paris. Pannenberg, M. (1998), Weiterbildung, Betriebszugehörigkeit und Löhne: Ökonomische Effekte des „Timings" von Investitionen in die berufliche Weiterbildung. Pp. 2 5 7 - 2 7 9 in: F. Pfeiffer, W. Pohlmeier (eds.), Qualifikation, Weiterbildung und Arbeitsmarkterfolg, vol. 31 of ZEWWirtschaftsanalysen, Nomos, Baden-Baden. Pfeiffer, F., F. Reize (2001), Formelle und informelle berufliche Weiterbildung und Verdienst bei Arbeitnehmern und Selbstständigen. Pp. 2 1 5 - 2 7 4 in: R.K. Weizsäcker (ed.), Bildung und Beschäftigung, no. 284 in Schriften des Vereins für Socialpolitik,. Duncker und Humblot, Berlin. Pischke, J.-S. (2001), Continuous Training in Germany. Journal of Population Economics, 14, pp. 523-548. Robinson, P.M. (1988), Root-n-consistent semiparametric regression. Econometrica, 56, pp. 931-954. Vytlacil, E. (2002), Independence, Monotonicity, and Latent Index Models: An Equivalence Result. Econometrica, 70, pp. 331-341. Weiß, R. (2000), Wettbewerbsfaktor Weiterbildung, Ergebnisse der Weiterbildungserhebung der Wirtschaft. Discussion Paper 242, Beiträge zu Bildungs- und Gesellschaftspolitik des Instituts der deutschen Wirtschaft, Köln. Weiß, R. (2003), Betriebliche Weiterbildung 2001 - Ergebnisse einer IW Erhebung, in: iw-trends 1/2003. Anja Kuckulenz, Centre for European Economic Research (ZEW), Department Labour Markets, Human Resources and Social Policy, P.O. Box 10 3 4 4 3 , D-68034 Mannheim, Germany. Phone ++49/+621/1235-287. E-mail: [email protected] Michael Maier, Centre for European Economic Research (ZEW), Department Labour Markets, Human Resources and Social Policy, P.O. Box 1 0 3 4 4 3 , D-68034 Mannheim, Germany. Phone ++49/+621/1235-362. E-mail: [email protected]

Jahrbücher f. Nationalökonomie u. Statistik (Lucius & Lucius, Stuttgart 2006) Bd. (Vol.) 226/1

Employment Protection: Its Effects on Different Skill Groups and on the Incentive to become Skilled By Nikolai Stähler, Mainz* JEL J24, J41, J42, J64, J68 Education, employment protection, unemployment, search and matching models.

Summary Employment protection affects labour market outcomes and hence the incentive to acquire skills. Using a matching model with two education levels in which workers decide ex-ante on their skill formation, it is shown that employment protection can raise the fraction of skilled workers. This will be the case if workers obtain a sufficiently large fraction of the rent created by skill formation. Furthermore, it will be shown that high-skilled workers face shorter unemployment duration and lower dismissal probabilities.

1.

Introduction

By comparing the data on the population percentage with tertiary education in each country with country-specific indices of the level of employment protection, as found in the the recent OECD reports (OECD 2004a and 2004b), we find that countries with high levels of employment protection tend to have a higher percentage of population with tertiary education. This observation seems to hold particularly in the European context. In countries rated as having relatively high employment protection, such as Belgium, Finland, Germany, the Netherlands, and Norway, the percentage of the population with tertiary education exceeds the average level within all OECD countries. Furthermore, in the OECD employment report (OECD 2004a) it is empirically shown that shorter unemployment duration, better re-employment probabilities and fewer dismissals are an accompaniment of education. But it is claimed in the report itself that "still, little is known about the labour market impact of (. . . ) learning." (OECD 2004a, p. 183). Next to the fact that the present paper gives a theoretical explanation for shorter unemployment duration, better re-employment chances and fewer dismissals of highly educated workers, it especially focuses on the effect that employment protection has on the ex-ante decision of individuals to obtain education. The existing literature that incorporates employment protection with education solely concentrates on education obtained "on-the-job" (see below a brief description of the literature in more detail). However, a huge part of individual skill formation takes place before entering the labour market (for example, the decision whether to get a university degree or not) as the OECD report (2004b) shows.

* I would like to thank Florian Baumann, Tim Friehe, Laszlo Goerke, Magnus Hoffmann, Martin Kolmar, Jörg Lingens, Daniel Scott, and Georg Tillmann for many helpful comments. I especially thank the Deutsche Forschungsgemeinschaft (DFG) for financial support.

42 · Ν. Stähler

Earlier, the theoretical literature about employment protection analysed the effects of employment protection in models with homogeneous workers. Employment protection decreases dismissals but, at the same time, decreases job creation (see Mortensen/Pissarides 1999, 2001, or Bettola 1990). Heterogeneous workers have only lately been introduced. Those models show that employed high-skilled workers benefit more from the same level of employment protection than employed low-skilled workers (see Guelfi 2004 and Kohns 2000, for example). This will also be confirmed in the present paper. Effects of employment protection on education have been analysed in models where education is considered to be "on-the-job training". Fella (2005) shows in a model with incomplete contracting that termination restrictions increase the firm's and the worker's incentive to invest in training. In his model, termination restrictions do not only include employment protection for workers, but he also investigates penalties in cases where the worker leaves. The reason for more training is that the restrictions assure a higher probability of obtaining benefits from the training investment for both sides, the firm and the worker. Burda (2003) presents a matching model with endogenous education decision. He models education as human capital investment made by firms. He finds through calibration that more employment protection decreases the incentive for education. While employment protection and "on-the-job training" both induce some costs for the firm, stronger employment protection results in lower education because the firm attempts to compensate for the additional costs of more employment protection by reducing education. In addition, Wasmer (2003) points out that the education decision is not independent of the aggregated state of the labour market. In particular, he shows that through different employment protection rules in the US and Europe, different investments in (specific and general) human capital can partly be explained. The investment in general and human capital is obtained by the worker "on-the-job". If employment protection is low, workers invest less in specific and more in general human capital as their general human capital will be rewarded in the case of a job loss. For sufficiently high employment protection, though, their expected returns of specific human capital investments will compensate for the risk of loosing the job (which is now lower with stronger employment protection) and, therefore, induces them to invest more in specific human capital. The argument is in agreement with Bean (1997), who claims that more job security may increase the worker's contribution to the firm. An ex-ante decision on skill formation of workers has been introduced by Sato and Sugiura (2003). They analyse the effects of subsidies on human capital investments and unemployment benefits on skill formation. In the present paper, firstly, I contribute to the discussion of the effects of employment protection on general human capital investment. I assume that only workers will decide whether they get educated or not. Secondly, I make an additional remark on the discussion that skill formation does not (only) take place "on-the-job" but is done ex-ante by the workers themselves (see OECD 2004b). As employment protection affects "onthe-job" skill formation, it is reasonable to ask the question which effect employment protection has on the ex-ante education decision of workers. Using a matching model in line with Mortensen and Pissarides (1994,1999,2001) with two skill groups, I show that, in contrast to Burda (2003), more employment protection increases the incentive to skill formation as long as the worker obtains a sufficiently large fraction of this investment and the education decision is taken by the worker ex-ante. I will proceed as follows. Section 2 describes the basic structure of the model. Section 3 deals with the effect of a change in employment protection on the labour market, ed-

Employment Protection · 43

ucation and unemployment. Section 4 calibrates the model to provide some additional insight. In section 5, the main findings are summarised. 2.

The model

The economy is populated by a continuum of individuals which is normalised to one. The economy is continuous in time, and individuals die at rate S. They are replaced by new born individuals without education at the same rate. At the beginning of their lives and, therefore, before they enter the labour market, each individual has to decide whether to become educated or not. Firms are also measured in a continuum, while their number is determined by the condition of free market entry. Firms supply jobs, and workers offer their work. The productivity of jobs determines how many jobs are supplied and whether existing jobs are destroyed. These two features are captured in the job creation and the job destruction conditions. Unemployment exists due to market frictions and matching problems between employers and employees. There are two separate labour markets, one for the low-skilled and one for the high-skilled. Once educated, high-skilled workers do not move into the low-skilled sector. The number of high-skilled workers shall be denoted by A* = χ and the number of lowskilled workers by A ' = (1 — x), respectively. A Cobb-Douglas matching function, equal in both sectors, M(A'u', A V ) , with the typical properties and i = l, s for low- and highskilled is assumed to capture the market frictions and indicate that job-seeking takes time. 1 The sector-specific unemployment rate is denoted by « ' , while v1 is the corresponding vacancy rate. Unemployment and vacancies exist at the same time. I define the ratio Θ' = AV/A'M' = v'/u', which is called market tightness. Note that it could be reasonable to assume that the matching processes in the high- and low-skilled sector differ. The qualitative results of the following analysis are not influenced by such an assumption. Therefore, for simplicity, I do without the differentiation and assume the matching function to be equal in both sectors. 2 The rate at which vacancies are filled is the number of matches, given by the matching function, divided by the number of vacancies in the corresponding sector. In terms of the market tightness, this can be written as q(ß'), with q'{9') < 0. Analogously, the probability of an unemployed worker finding a job in each sector can be calculated as θ·η(θ·) with (fl'q(e'))' > 0. The productivity of a given job can be decomposed additively into a global component p + a', with as > 0 and J = 0, and an idiosyncratic component e'. The idiosyncratic component is a zero-mean shock with e' € [e¡; e„], e' is distributed according to the cumulative distribution function G(e'), with g(e') being the corresponding density function 1

2

Petrongolo/Pissarides (2001, p. 396) claim that "(t)he stylized fact that emerges from empirical literature is that there is a stable aggregate matching function of a few variables that satisfies the Cobb-Douglas restrictions with constant returns to scale in vacancies and unemployment". Therefore, the loss of generality assuming a Cobb-Douglas matching function can be justified by the consistence of such a function with empirical facts. Furthermore, it simplifies the analysis later on. For a Cobb-Douglas matching function, the differentiation can simply be done by assuming functions of the form M ! ( A ! « S , A V ) = (Aius)a(Asvs)n-a) and Ml(Alu\ Ali/) = (A'u' fiA'v1)"-"^, where 0 < α, σα < 1, for example. I will briefly refer to the resulting quantitative difference in Section 3.

4 4 · Ν. Stähler

and being equal in both sectors. Shocks occur to every job at a Poisson rate λ. The shock changes the idiosyncratic productivity, drawing a new é out of the distribution. If the idiosyncratic productivity falls below the endogenous reservation productivity e^, the job will be destroyed. There are vacancy costs c for each vacancy. Employment protection is represented by dismissal costs Τ which have to be paid by the firm when firing a worker. Τ is equal in both sectors since I assume the same level of employment protection to hold for each skill group. Τ can be interpreted as legal costs for dismissals. 3 Employers and employees bargain over wages according to a Nash-bargaining process. The wages depend on the job-specific productivity. It is assumed that after each shock the wages are renegotiated. Following Pissarides (2000), only unemployed workers look for a job, while the employed stay on their job until they are dismissed or die. I assume that individuals are low-skilled at the beginning of their lives, but heterogeneous in their ability to learn. Ability χ is uniformly distributed in an interval χ e [0,1], with zero representing the highest ability. The lower an individual's ability to learn, the more units of education χ an individual needs to become high-skilled. A unit of education costs k. Education takes place infinitely fast at the beginning of the period when the individual is born. To decide whether to get educated or not, each individual compares the utility value of being low-skilled with the utility value of being high-skilled minus the individual costs of education. I assume that in case of equality, the individual gets educated. As I assume a uniform skill distribution and because the population size is normalized to unity, the resulting threshold value of χ is equal to x. Each individual with χ > χ will stay low-skilled (and vice versa). The game structure is arranged according to the following time line. Birth I

Education I

Work (Labour market outcome) I

time

In the first step, after individuals have been born, they make their education decision and then, in the second step, enter the labour market. The labour market outcome is then determined as described above. The game is solved by backward induction. Because of the properties of a linear homogeneous matching function, the individual's decision about education does not influence the sector-specific labour market outcome and, therefore, the education decision of others. This is fairly important, and I will return to this point in more detail later on, when turning to the education decision. In the following analysis, I am going to describe the structure of the labour market and then explain the education decision.

3

It can be reasonable to a s s u m e that dismissal costs are higher in the high-skilled sector. T h i s can be justified by the a s s u m p t i o n that high-skilled w o r k e r s k n o w m o r e a b o u t their legal rights and can a f f o r d better lawyers to enforce their rights. Further, one c a n a s s u m e that workers, when being dismissed, receive a severance payment. For simplification, the basic m o d e l setup does without this a s s u m p t i o n . T h e implications will be discussed in Section 3.

Employment Protection · 45

2.1. The labour market Firms In both sectors, each firm has one job to offer which can be either filled or vacant. The number of firms is endogenously determined. The value of the firm J(e') depends on the productivity (overall and idiosyncratic), wages w(e') and option value. Therefore, the value of the firm satisfies the Bellman equation r](€')=p

+

i i a'+e -w(e )+X

β" J(x)dG(x)

+ G ( ^ ) ( - T ) - / ( e ' ) - « / ( Λ (1)

r is the interest rate of the economy. In the case of a dismissal, which occurs with probability XG(e^), firms have to pay the amount Τ to some third party. Following Pissarides (2000), newly created jobs are endowed with the highest possible idiosyncratic productivity and satisfy the Bellman equation r f - ^ p +

,0.«. a'+el-w^+X

J*"Kx)dG(x)

+ G(4)(-T)

- y0·'

δ]°'\

(2)

with w 0 ·' as initial wage payment. Analogously to the above formula, one can derive the Bellman equation for a vacancy V' as rV' = -c + 4(0') [/°·' - V'] .

(3)

The possibility of free market entry for firms implies that vacancies will be created as long as their present value is greater than zero. Accordingly, in equilibrium, V' = 0 holds. Therefore equation (3) implies

r

0,/

(4)

q{6>)

in equilibrium. Equation (4) says that the discounted value of a newly created job has to equal the search costs c per period multiplied with the average search duration l/q(9'). Workers Workers can either be employed (with utility W(e')) or unemployed (with utility I/'). If they are unemployed, they get the option value of finding a job, since, I assume for simplicity that there are no unemployment benefits. The Bellman equation for the unemployed satisfies rU' = e'q(ß') ( V a ' - 17') - SU',

(5)

where W 0 ·' is the utility of a newly founded job. The equation for an employed worker reads rW(e') = w(e') + λ β " W(x)dG(x)

+ G(e'd)U' -

W(e')

SW(e').

(6)

46 · Ν. Stähler

T h e employed workers obtain the utility of the wage (depending on the sector the corresponding worker is working in) and the option value of a change in the idiosyncratic component. If the idiosyncratic component falls below a certain value e'^, the j o b is destroyed and the corresponding worker becomes unemployed. If the job has just been created, the corresponding Bellman equation for a newly employed worker is given by rW0·'

= w0'1 + λ

Γ" -Jfd

W(x)dG(x)

+ G(e'd)U

-

w0·'

sw

0,1

(7)

J o b Destruction and J o b Creation Wage bargaining is modeled as a Nash bargaining between workers and firms, where the bargaining power of workers is β. The workers' fall back utility is the utility of unemployment, U', while the firms' fall back utility are the negative dismissal costs, — T . T h e bargaining can be represented as w(e') = arg max ^ W ( f ) — U'^j (/(e) + T ) 1 - ^ for the wage of continuing jobs and uP'' = arg max

— U'j^/0''1

^ for newly created

jobs, because when firms and workers bargain for the first time, firms do not face the threat of having to pay the dismissal costs in the case bargaining fails (see Mortensen and Pissarides, 2 0 0 1 ) . The resulting wages are w{e') = ß[p + a' + e' + (r + λ ) Τ + c6'] and ufi·' = w(eu) — ß(r + S + λ ) T . Using these wages, we can derive the J o b Destruction Condition (further described as J D ) , which describes under which circumstances a j o b is destroyed, as Ρ + ' + < , + ν τ τ τ s

Γ ( ^dG(z> ζ~

=

-(r+á)T·

(8)

and the J o b Creation Condition (further described as J C ) , which shows when jobs are created and therefore determines market tightness Θ'. It can be derived from using equation (4) and equalises the cost of searching with the productive value of a job for the employer. Hence, it can be stated as \

r + λ + δ

ä /

- ( ΐ - β ) Τ = — j - 4 q{Q>)

(9)

By solving equations (8) and (9) simultaneously, one obtains the reservation productivity ej and the market tightness Θ' in each sector. Because of the higher sector-specific productivity ρ + a s in the high-skilled sector, there is fewer job destruction in the high-skilled sector than in the low-skilled sector. J o b s remain productive longer and are, therefore, destroyed later than in the low-skilled sector. This yields ej
θΚ Empirical

research tends to confirm this hypothesis, as can be seen in Boeri et al. ( 2 0 0 4 ) . 2.2.

Education decision

Individuals enter the labour market unemployed. The utility value of an individual is therefore represented by equation (5). 5 Thus, for an individual to get educated, the util-

5

For this kind of modeling see also Kolm/Larsen (2003), Sato/Sugiura (2003) or Pissarides (2000), chapter 7.

Employment Protection · 47 ity of being unemployed for the high-skilled worker minus the individual costs of education has to equal or exceed the utility of being unemployed of the low-skilled worker, Us — kx > UK The threshold value of χ for an individual to still be willing to get educated can thereby be calculated as χ = u . In equilibrium, W(eu) — U' = V' = 0 in steady state and, therefore, ] 0 · ' = V =

{1

holds (see Pissarides 2000, p. 41, where ). Substituting this into (5) leads to

_ß ) \ r + s ) & which yields 1

ßc

(l-ß)(r

+ S)k

φ5-θ1).

(io)

As equation (10) shows, there is no education without bargaining power (β — 0). This is because without bargaining power, workers only receive their reservation wage. This is true across both sectors. The higher the bargaining power, the higher the worker's share of a firm-worker-pair. Therefore, the worker receives higher payoffs on his education investment. Equation (10), furthermore, states that the decision to get educated mainly depends on the differences of the market tightnesses and, therefore, on the probabilities of finding a job in each sector. The market tightness in each sector is determined in the second stage of the game, which is the labour market outcome. This outcome influences the education decision of individuals before they enter the labour market. But the education decision itself does not influence the labour market outcome in the second stage. This is due to the matching function, which is linear homogeneous, and the free market entry for firms. Independent of how many people in the economy are low- or high-skilled, in a steady state, the sector specific market tightness remains the same for given exogenous parameters.6 Therefore, the individual's decision which sector to go to does not affect the decision of others and vice versa. The outcome of the second stage of the game (the labour market outcome) enters into the first stage (education decision) as a given parameter, while the first stage does not influence the second stage of the game. If more people decide to become educated, more vacancies in the high-skilled sector are supplied. Therefore, the original sequential structure can be dropped, as there is no reaction function of individuals.

2.3.

Unemployment

Unemployment is determined by inflows (XG(e'j)( 1 — u')A' + SA') into and outflows (e'q(ß')u' + u'A'S) out of unemployment in each sector. In steady state the changes of

6

In the end, market tightness in each sector only depends on unemployment and vacancy rates, not on the number of people in each sector. Increasing the number of, for example, unemployed people in one sector leads to a decrease in market tightness. This increases the probability of a vacancy being filled, which increases the value of the vacancy (see equation (3)). Because of free market entry, a corresponding number of vacancies will be offered, lowering market tightness to its original equilibrium value again. This also works when the number of unemployed is decreased or vacancies are decreased and increased.

48 · Ν. Stähler unemployment will be zero. Thus, the steady state unemployment rate for each sector can be derived as àG(ó + s (11)

XG(e^) + e'q(0') + S

To calculate the total number of unemployed in the whole economy, I multiply the unemployment rate (11) with the corresponding number of workers in this sector A' and add the resulting numbers of unemployed workers of each sector (u* = «'(1 — x) + u$x). Since population is normalised to one, u* simultaneously gives the total number of the unemployed in the economy as well as the overall unemployment rate. Doing so after some rearranging yields u* =

3.

(1 - xWsq(0s)(G(€l,

) + δ) + (x9lq(el) + G(el.) + S)(G(€S.) S2 2 [XG(el, ) + elq(9l) + á][AG(e*) + 6sq(es) + S]

+8)

.

(12)

Analysis

It is straightforward to show that employed high-skilled workers benefit more from employment protection than low-skilled workers. For this to be the case, ^ < ^ γ has to hold, which has already been shown by Guelfi (2004) and Kohns (2000) in a slightly different model framework. It can be easily shown that it also applies in this model (for a proof, see Appendix A). As we already know from the above analysis that high-skilled workers face a higher re-employment probability and shorter unemployment duration than low-skilled workers (resulting from 0 s > #'), the question of how an increase in employment protection changes the ex-ante decision of individuals on skill formation arises. To show which effect a change in employment protection has on education, recall equation (10). Differentiating (10) with respect to firing costs Τ yields dx dT

ßc (des (1 -ß){r + X)k \dT

d6l\ dT

By totally differentiating the JD and JC, equations (8) and (9), we can derive dQi _

(1 - ß)XG(e'd)

\fl'«(0')

_ G{eld) (r + S + XG(€sd))\ G(6*)(r

where 0 < η = — ^ ¡ j θ^ =

+

i +

A.G(^))/

G(esd)(r + 8 + XG(e'dm l

s

(G(e d)-G(e d)W*q(0*)

_

A

g

(16)

~

< 1 for a Cobb-Douglas matching function. Equa-

tion (16) states that as long as the bargaining power of workers exceeds some threshold value (β >J>), education increases when employment protection is increased. The threshold value β consists of the relative duration of unemployment between low- and highskilled (first term in brackets) minus the relative dismissal probability between low- and high-skilled (second term in brackets). This is weighted by the relation between dismissal probabilities in the low- and high-skilled sector and the average unemployment duration in the high-skilled sector (

)•> as well as the relation of the differences of

changes in market tightness and dismissal probabilities ( ^

1

^ G(

s

)

The intuition is clear. The probability of finding a job in each sector decreases with each increase in employment protection. Obviously, it depends on the relative decrease of those probabilities. If the probability in the low-skilled sector decreases by more, it becomes relatively easier to find a job in the high-skilled sector. Therefore, more people get educated. The higher the bargaining power, the higher the worker's share of the surplus of a firm-worker-pair. If the bargaining power of workers is high enough (β > β), the worker obtains a sufficiently large fraction of the rent created by skill formation. In this case, therefore, more workers are willing to become educated. Different dismissal costs in both skill sectors do not change the qualitative results of the above analysis. The only difference is that if dismissal costs are higher in the low- (high) skilled sector, the original market tightness of the corresponding sector will be lower, changing the original education decision according to equation (10). One should note that there is one exception. Assume that employment protection in the high-skilled sector is stronger. Then, the higher expected dismissal costs can overcompensate the productivity advantage over the low-skilled sector, yielding a lower market tightness in the high-skilled sector. That means that there will be no education from the start, as it does not pay from the workers point of view. 9 In this case, only a decrease in employment protection can result in more education. It is also reasonable to assume that workers are entitled to receive a severance payment when being dismissed. Again, in principle, the argument does not change and education will still increase when increasing the severance payment, as long as workers obtain a

9

N o t e that according to equation (10) that would result in a negative x. As this is economically impossible, this option is excluded. All mathematical solutions for χ < 0 are set to 3c = 0 as there cannot be a negative education level.

50 · Ν. Stähler

sufficiently large fraction of the rent created by their education investment. The difference in having severance payments is that the wage claim of workers increases, since their fall back utility increases by the possibility to receive the severance payment in the case of a dismissal. An increase of the level of severance payment also decreases market tightness in both sectors, and the effect of education depends on the relation of the decrease of market tightnesses according to equation (13). It has to be noted, however, that the decrease of market tightness in each sector differs compared to the situation with pure firing costs as presented in the above model, when severance payments exist. From a theoretical point of view, it is not possible to say whether the decrease of market tightness in presence of severance payments exceeds or falls short of the decrease presented in the paper at hand due to the additional wage effect. Nevertheless, for education to increase along with an increase in severance payments, a computable threshold value for β will result, which resembles the one of equation (16). 1 0 It is straightforward to show that the fact that more workers get educated reduces unemployment. To prove this, I differentiate equation (12) with respect to k. du*

=

(G(esd) + 8Wlq(9l)-(G(Á)

+

8)9sq(es)

(XG(e'd) + θ'ς(θι) + í)(XG(e¿) + 6sq(6s) + δ)

'

(17)

where | | < 0. Since it has already been demonstrated that < and 6 s > θ', it is obvious that the first term on the rhs, is negative. Therefore equation (17) is positive. N o w decreasing education costs k increases the education level χ and therefore reduces unemployment. The economic intuition is as follows. As can be seen in equation (11) in combination with (17) the unemployment rate in the high-skilled sector is lower than in the low-skilled sector. If more people move from the low-skilled sector to the high-skilled sector, less people are unemployed and therefore the overall unemployment rate falls. Still, the effect of employment protection on overall unemployment is ambiguous since the effects on the unemployment rates in each sector are unclear, like in many existing publications (see Pissarides 2000). Employment protection leads to fewer dismissals, but at the same time to fewer vacancies, which implies fewer inflows into and fewer outflows out of unemployment. It is not clear which of the two effects is dominant. In the present model, an additional effect is introduced. As has been shown above, education can be positively influenced by employment protection. Education in itself negatively influences the unemployment rate. Therefore, the effect of employment protection on unemployment differs through the educational channel from standard results. In case unemployment increases (decreases), this effect is weakened (boosted) by the negative effect education has on unemployment when employment protection is increased and β>β. 4.

Calibration

To better assess the magnitude of the effects of employment protection on different skill groups and on the incentive to become skilled, I perform a simple calibration of the model. 10

A proof for the argumentation about different sector specific levels of employment protection and severance payments will be send upon request.

Employment Protection · 51

Table 1 : Parameter Values Parameter Ρ a$ r

λ c ß S k η

Scenario 1 (Sc.1)

Scenario 2 (Sc.2)

Scenario 3 (Sc.3)

1.75 0.5 0.05 0.7 0.3 0.5 0.05 7.5 0.5

1.75 0.5 0.05 0.3 0.3 0.5 0.05 7.5 0.5

1.75 0.3 0.03 0.3 0.3 0.5 0.1 3.5 0.5

Following Mortensen and Pissarides (1999), I assume that the distribution of idiosyncratic productivity e is uniform, here over [ - 1 , 1 ] and, therefore, G(e) = \ \ \ + e] being the cumulative distribution function. Furthermore, q{9') = (θ')~η, with 0 < η < 1. The values of the parameters are presented in Table 1 and intend to reflect the European case (Scenario 1) and the U.S. case (Scenario 2). Scenario 3 is added to show that it also holds for other parametric specifications. The values are chosen to obtain an unemployment rate of about 8.8% for Europe (Scenario 1) and 6.5% for the US (Scenario 2) according to the actual employment report of the OECD (2004a) at an employment protection level Τ = 0. The resulting unemployment rate for Scenario 3 is 5.7%. Furthermore, the bargaining power is chosen to be β = η, according to the condition of Hosios (1990). 11 Broersma and van Ours (1999) find empirically that it is reasonable to assume η = 0.5. In all three scenarios, β > β. The above theoretical analysis predicts an increase in the level of education if employment protection is increased. It is apparent from Figure 1 that the level of education χ increases with employment protection Τ in all three scenarios as all three graphs have a positive slope. 12 That is

Figure 1: Education Level

11

The qualitative results of the above calibrations do not change if this is not the case. But following a standard assumption when calibrating matching models, this is assumed to hold here as well.

52 · Ν. Stähler

because workers obtain a sufficiently large fraction of the investment in skill formation, since with these parameter constellations β > β (as can be seen in Figure 2) for all Τ that yield an economically reasonable solution for the unemployment rate. 1 3 Furthermore, using these parameter constellations, it can be shown that the levels of unemployment in both sectors and therefore in the entire economy fall when employment protection is increased. The simulation suggests that more employment protection can increase the level of education and decrease unemployment as long as the bargaining power of workers is great enough. 5.

Conclusions

The main result of this paper is that more employment protection increases the level of education acquired by workers ex-ante as long as workers obtain a sufficiently large fraction of the rent resulting from skill formation. This is due to the fact that increasing employment protection relatively increases the (re-)employment chances of high-skilled workers through market tightness. In addition to the better re-employment chances of high-skilled workers, it can be shown that high-skilled workers get dismissed later than low-skilled workers and benefit more from an increase in employment protection. Furthermore, a higher level of education unambiguously decreases unemployment. This paper complements Burda (2003) and Wasmer (2003). In their works, they suggest that education is achieved "on-the-job". While Burda (2003) finds that more employment protection decreases education, Wasmer (2003) finds the opposite. This is due to the fact that in the model presented by Burda (2003), the firm bears the costs of education, while in the model by Wasmer (2003), the worker faces those costs. Next to the fact that it is important who pays for education, I introduce a novel argument to the discussion. It is not only important who pays for education, but also when the education decision takes place. As a huge part of workers' skill formation happens before they enter the labour market (OECD 2004b), this is an interesting topic to think about. Despite the fact that it might seem odd to use employment protection as an instrument to increase education, 12

13

The calibration allows for the speculation that a lower level of education costs, k, amplifies the increase of education as the slope of scenario 3 is much steeper than the one of scenarios 1 and 2. Economically reasonable solutions are those solutions where the unemployment rates do not fall below zero, which can be a mathematical solution in case Τ > 0.85.

Employment Protection · 53

we have seen that education can be increased by more employment protection under the above mentioned conditions. Especially for developed countries, this argument might be interesting, as it is often claimed that in times of globalisation comparative advantages can only be maintained through greater innovation and well educated workers. We can additionally conclude that high levels of bargaining power amongst workers have positive effects on their education decision. Therefore, relatively high levels of workers' bargaining power and employment protection might be one of the influencing factors for education. Nevertheless, the results should not be misinterpreted as a pleading for more employment protection. As we have also seen, the ambiguous effect that an increase of employment protection has on overall unemployment cannot be resolved. We can only say that, as long as workers obtain a sufficiently large fraction of the rent created by ex-ante skill formation, the decrease (increase) of unemployment is boosted (reduced) through the educational channel when employment protection is increased.

Mathematical Appendix A Who benefits more from employment protection? Employed high-skilled workers obtain more benefits from employment protection if reservation productivity of high-skilled workers falls by more than the one for low-skilled workers when employment protection is increased ( ^ the JD and JC, equations (8) and (9), yields de

'd _

dT~

(r + S + λ)(Γ + S)cq'(e') {r

+ s + XGie'jVafm

The above condition ( ^


f )

= 1.

(1)

t

The hazard rate or transition probability hj(t) captures the probability that student i being still enrolled in the t^ 1 semester - ends his studies after t semesters. It is given by: 7 hi{t)

= Pr(Tj

= t\Tj

> t)

with

0 < h¡{t)

< 1.

(2)

Another important concept is the survival function G,(i) which measures the probability that student i has not ended his studies within t semesters but is still enrolled in semester t + 1. The survival function can be derived from the hazard rate in the following way: t G , ( í ) = Pr(T¡

> t) =

Y[

( 1 - hi(k)).

(3)

k=l

Note that the following relationship also holds:

olirti'1-'• !

1 η

(

V— -^ ^ y ^, t ) ) )1 + Σ h hί > y

i { t ) )

> ' ( * ι * ί ( * ) ) ]· \- ΣU J

1

1

( « )

In contrast to the continuous-time model this log-likelihood function cannot be split up into m different log-likelihood functions, one for each possibility /. Therefore the model has to be estimated simultaneously. 8 4.2.

Explanatory variables

For each possibility of terminating one's studies, the unconditional hazard rates for each semester deviate substantially from the average hazard rate for that possibility. For example, the probability of switching majors is quite high after the first semester, even higher after the second, but then fairly low for subsequent semesters. After the second and fourth semester, the probability of transferring to another university is quite high but relatively low and almost constant for all other semesters. The probability of obtaining a degree is basically zero for the first eight semesters but rises steadily beginning in the ninth semester. These patterns are difficult to capture by polynomials, so we used dummy variables sts χ which assume value one when a student is enrolled in semester x. For an observation of a student who is enrolled in his twelfth semester, sis 12 equals one and all the other dummy variables are equal to zero. We aggregated some of the dummy variables into one dummy variable. For example, for transferring to another university sts9,..., stslO s t s x )· Since the are aggregated into one dummy variable sis9-20 (with sfs9-20 = hazard rates also differed considerably for the different majors, we estimated the model for each major separately. Tuition fees enter our analysis in two different ways. On the one hand, we estimated the effect of having to pay in the next semester. On the other hand we considered the consequences of knowing that at some point later on tuition fees become due. We therefore employed two different variables: (i) An indicator for having to pay in the next semester Tuition fees probably have a direct effect when they become due in the next semester. We therefore created the variable Ind which is equal to one when our data indicate that the student has to pay tuition fees in the following semester. T o determine the value of Indt we only used information up to semester t — 1. This included all information available which indicated that the student qualified for an exemption like having been on leave, having studied abroad or having raised a child. Since we actually observe whether a student paid in the following semester in most cases, we were able to check the validity of our indicator. Table 4 shows how well our indicator performs. Considering all students who might have to pay in the next semester (i.e. those who have not finished their studies within four semesters beyond the standard period of study), we find that in 9 9 . 2 3 % of all cases our indicator predicts correctly.

8

We take advantage of the fact that this log-likelihood function is equal to the log-likelihood function of a multinomial logit model, as has been shown by Allison (1982). Estimates were performed with the software package Stata.

90 · M. Heineck, M. Kifmann, and Ν. Lorenz

Table 4: Relationship between indicator and tuition fees paid Student pays Yes No Has to pay according to indicator

Yes No

530 (11.63%) 9 (0.20%)

26 (0.57%) 3.992 (87.60%)

(ii) Indicators for anticipating that tuition fees can be due later on Unless students are completely myopic, there is also an indirect effect of tuition fees. In this case not only having to pay for the next semester but also anticipating that tuition fees can become due later on will influence students' behavior. We try to capture this by the dummy variable A which depends on whether the student is enrolled at a time when tuition fees had already been implemented. This variable is equal to one when tuition fees had already been implemented or when it was well-known that they will be implemented. Since the law was enacted in May 1997 we assume that students anticipated tuition fees by the fall semester 1997. We also created variables which depend on how long a student already anticipated tuition fees: • The variable R measures the share of a student's enrolled semesters in which he knew about tuition fees. For example R = 0.25 for a student in his eighth semester, when he was in his seventh semester in fall 1997. • The variable Rw = */R assumes that the impact of knowing about tuition fees grows at a decreasing rate with each additional semester in which students anticipate tuition fees. There were only minor differences between using R and Rw. We therefore restrict our analysis to R and A. Finally, we also included the following variables: • • • • •

the age of the student when begins his studies (age), gender (male), date of enrollment (enroll), size of class in the first semester (class-size), county of residence before enrollment (local).

The date of enrollment measures the exact date when students were registered at the University of Konstanz. It can serve as a proxy for ability as students with good school grades are frequently given a first choice. Furthermore, students who enroll early may Table 5: Explanatory variables time-independent

time-dependent

male age enroll local class-size

sts X Ind A R

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany · 91

be more motivated. With the variable local we analyze whether local students differed in their behavior from other students. Local students had their residence before enrollment either in the county Konstanz or in one of the adjacent counties Tuttlingen, Bodensee (Friedrichshafen) or Ravensburg. Table 5 gives an overview of all explanatory variables and shows which are time-dependent. 4.3.

The identification strategy

T o analyze the impact of tuition fees on the length of study it would be ideal to observe two different groups who begin their studies at the same time and under the same circumstances and for which the only difference is whether they have to pay tuition fees or not. Unfortunately, this is not feasible with our data set. Therefore, our identification strategy is to compare students who began their studies at different dates. These students differ in their exposure to tuition fees. An obvious setback of this approach is that our variables A and R may not only measure the impact of tuition fees but also the impact of other unobservable variables which changed at the same time when tuition fees were implemented. However, it is the only feasible approach given the data available. We return to this problem in section 6.2. when we discuss our results. In the following section we present the estimation results for each of the six majors. It was not always feasible to estimate all coefficients. In some cases, we could not determine the impact of the indicator variable Ind because there were not enough observations. For transferring to another university this was not even possible for any of the majors. 5. 5.1.

Regression results The effects of tuition fees

Table 6 shows our hypotheses for the effects of tuition fees on the different hazard rates. They apply to the indicators for anticipating that tuition fees might be due later on as well as to the indicator for having to pay in the next semester. In particular, we suppose the following impact on students' behavior: Table 6: Hypotheses on the effects of tuition fees on the hazard rates /'

Hazard rate for

1 2 3 4 5

Obtaining a degree Transfer Switching majors Dropping out Failing

Hypothesis + + -

+ +

1. Obtaining a degree Tuition fees can be avoided by studying faster. The hazard rate for obtaining a degree should therefore increase. 2. Transfer to another university Switching to a university outside of Baden-Württemberg which does not charge tuition fees is a possibility to avoid fees. The hazard rate for transferring to another university can therefore be expected to rise.

92 · M. Heineck, M. Kifmann, and Ν. Lorenz 3. Switching majors If a student switches majors, then the semesters he studied so far are relevant for whether tuition fees have to be paid. The expected costs of switching majors has therefore increased which should lead to a lower hazard rate for switching majors. 4. Dropping out Tuition fees have increased the costs of continuing one's studies for those students who expect that they need a long period of time to obtain a degree. The hazard rate for dropping out can therefore be expected to increase. In particular, the hazard rate should be larger if tuition fees have to be paid in the following semester. 5. Failing Tuition fees for long-term students create the incentive to take exams earlier and less well prepared. The hazard rate for failing can therefore be expected to increase. Table 7 shows our regression results for the different possibilities of terminating one's studies. We present the results for two regressions. In the first regression, we capture the anticipatory effects of tuition fees by the dummy variable A which is one when tuition fees had already been implemented or when it was known that they will be implemented. The first column of Table 7 shows how the hazard rate changes with A. In the third column, the effect of the indicator of having to pay tuition fees in the following semester is presented for the regression with A. Analogously, the second and fourth column display the results on the various hazard rates when R is used to capture the effects of anticipating tuition fees. This variable measures the share of a student's enrolled semesters in which he knew about tuition fees. The table shows the qualitative effects of the variables measuring the impact of tuition fees and states the significance levels. For the regression with A, the regression coefficients are presented in Appendix A.l. The results with R as the explanatory variable can be obtained from the authors. All significant effects are in line with our hypotheses. With the exception of the hazard rate for switching majors, the hazard rates always change significantly in at least two majors. Particularly strong is the effect of the variables measuring the anticipation of tuition fees. In most majors, there is a significant sign according to our hypotheses. The indicator which measures the impact of having to pay tuition fees in the following semester has a pronounced effect on the hazard rate for dropping out. In four of six majors, there is a significant positive sign. Table 7 also shows that our results do not depend strongly on whether we use A or R to measure the anticipatory effects of tuition fees. The signs differ only in three cases. 9 If the coefficient is significant for at least one of the variables, then the signs are always identical. 5.2.

Further results

In Table 8 we present the results for additional explanatory variables on the hazard rates. Since the results between the regressions using A and R hardly differ, we only present the results for the model based on A (see Appendix A. 1 for the exact values of the coefficients). 9

In Physics the signs of A and R differ with respect to the hazard rate for failing. The indicator for having to pay tuition fees in the following semester has a different sign for the hazard rate for obtaining a degree in Psychology and for the hazard rate for transferring to another university in Economics.

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany · 93

Table 7: Regression results for the effects of tuition fees Anticipation A R Obtaining a degree Biology Chemistry Economics Physics Psychology Public Adm. Transfer Biology Chemistry Economics Physics Psychology Public Adm. Switching majors Biology Chemistry Economics Physics Psychology Public Adm. Dropping out Biology Chemistry Economics Physics Psychology Public Adm. Failing Biology Chemistry Economics Physics Psychology Public Adm.

+***

+ + +

+ +*

Indicator fees in t + 1 Regression A Regression R + + +*

+ -

+

+ +**

+ +

+***

+ +

+

n.a. +*

+ n.a.

+ + n.a.

n.a. n.a. n.a. n.a. n.a. n.a.

n.a. n.a. n.a. n.a. n.a. n.a.

-

+**

+

+

n.a.

+*

-

-

+

+

-

-

_**

+ +

+ +

+

+*

+**

+*

+

-

-

+***

+ +**

+

-

+

+***

+** +***

+

+*

-

-

+

+

+***

n.a. n.a. + n.a.

n.a. n.a. + n.a.

+

+***

+

+ : positive effect; - : negative effect *: significance level 10%; **: significance level 5 % ; ***: significance level 1% n.a.: not available

Gender: A minus indicates that men have a lower conditional probability to terminate their studies in a particular way. We find that apart from Economics there are no significant differences with respect to obtaining a degree. The hazard rate for transferring to another university or switching majors, however, is significantly lower for men in Chemistry, Physics and Economics. Men therefore seem to be less "mobile" in these majors

94 · M. Heineck, M. Kifmann, and Ν. Lorenz

Table 8: Further regression results Degree Gender (male) Biology Chemistry Economics Physics Psychology Public Adm. Age Biology Chemistry Economics Physics Psychology Public Adm. Date of enrollment Biology Chemistry Economics Physics Psychology Public Adm. Local students Biology Chemistry Economics Physics Psychology Public Adm. Size of class Biology Chemistry Economics Physics Psychology Public Adm.

+

Transfer _*

Switching m.

Drop, out

+

-

-

_***

-

_***

-

-

+

-

+ -

+ + +*

+

-

+*

-

+

+

-

+

-

+*

+ +

-

+ +

-

-

-

+

-

—**

+**

+

-

+*

+

+

+ +

+*

-

+

-

+

-

+

-

-

+*

+ -

-

+***

+ + + +

+**

+* +*

-

-

_ *

-

+ _ * * *

+** -

+ + +**

-

+**

-

+*

+

+ +***

-

+ + +

+**

+*

-

+ + +

+ n.a.

+ +

-

Failing

-

-

+*

+

-

+

+ : positive effect; - : negative effect *: significance level 1 0 % ; **: significance level 5 % ; ***: significance level 1 % n.a.: not available

than women. Furthermore, men have a significantly lower hazard rate for dropping out in Biology and Chemistry. Age: The hazard rate for obtaining a degree tends to be lower for older students, i.e. students who started their studies at a higher age. Furthermore, the hazard rate for transferring to another university is significantly higher in all majors but Public Administration. Finally, the conditional probability for obtaining a degree is significantly lower in three

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany · 95

out of six majors. Older students therefore seem to be less successful. One explanation for this phenomenon is that older students may have a larger time gap between the end of secondary schooling and the take-up of their university studies. A reason may also be that older students already needed more time for secondary schooling which could indicate lower ability. Date of enrollment: Date of enrollment measures the exact date within the registration period when students registered at the University of Konstanz. We found that students who registered late had a significantly higher hazard rate for dropping out in all majors. Furthermore, the hazard rate for transferring to another university (Chemistry) and failing (Psychology, Public Administration) is significantly higher. For Economics and Public Administration we find a significantly lower hazard rate for obtaining a degree. These results are in accordance with a study by Hackl and Sedlacek (2002). They performed a duration analysis for students at the Vienna University of Economics and Business Administration. Students who enrolled at a later date were less successful in the sense that they had a significantly longer duration of their studies. How can this finding be explained? On the one hand, the practice that students with good school grades are given a first choice seems to be responsible for the importance of the date of enrollment. This is the case in Biology, Psychology and Public Administration which employed a Numerus Clausus throughout the observation period. On the other hand, motivated students will usually tend to enroll early even if there are no grade restrictions. In contrast, students who are not sure about which major to choose will enroll rather late. These students are probably less determined in their studies. Local students: Students who come from Konstanz or nearby show a significantly higher hazard rate for transferring to another university. It seems that these students are interested in getting to know something different after all. Otherwise, the results depend highly on the chosen major. In Public Administration, students have a significantly higher hazard rate for dropping out. One reason may be that this major is chosen as a way to stay in the Konstanz area. For other majors, this effect does not hold. Size of class: A larger class size significantly raises the hazard rate for transferring to another university or to drop out. The decline in the studying conditions due to more fellow students is probably responsible for this result. 6.

Progress of studies with and without tuition fees

6.1.

Cumulative incidence functions

The regression results presented in the previous section illuminate only partially how tuition fees change the behavior of students. In particular, it is not a priori clear whether more or less students obtain a degree or drop out as tuition fees tend to increase both hazard rates. Even if one considers the coefficients, it cannot be seen which effect is dominant. 1 0 Furthermore, the quantitative impact of tuition fees remains unclear. For this reason, we calculate cumulative incidence functions for the different ways of terminating one's studies. 11 For given values of the explanatory variables these functions 10

11

This problem is well-known in the context of a multinomial logit model which is formally identical to our competing risk model. Even if a variable has a positive coefficient with respect to one event, the probability for this event can decline if the probability for other events increases by more. See Kalbfleisch/Prentice (2002, p. 252).

96 · M . Heineck, M . Kifmann, and Ν. Lorenz

state for each possibility j of terminating one's studies the probability that a student has finished his studies within semester t. T h e functions are calculated in the following way. First, the probability of terminated one's studies in a certain way ; in period t is given by the product of the survival function G(í|x¿(í)) and the respective hazard rate

hi(t[i¿t(t)).

Summing up these probabilities until semester t yields the cumulative incidence function: t l'(t) = 7'(ί|χ,·(ί)) = Σ

G(k\xi(mh'(k\^(t)).

(14)

k=l Based on our regressions which use the variable A to capture the anticipatory effects of tuition fees, we calculate the cumulative incidence functions for two sample students « = 0 , 1 . For these students, all explanatory variables are equal to the average values for each major with the exception of the variables A and Ind which capture the impact of tuition fees: • For student 0, both A und Ind are equal to zero, i.e. this student is not affected by tuition fees. • For student 1, we set A and Ind equal to one. This student always knows about tuition fees and must pay the fees if he studies too long. Table 9 shows the estimated average length of study to obtain a degree for the six majors. In all majors it is lower for student l . 1 2 According to this figure, which is in the center of the public debate, the introduction of tuition fees is a success story. However, the cumulative incidence functions we present in the following show that the reasons why student 1 seems to study faster are quite different for the six majors. Table 9: Average length of studies for obtaining a degree

Biology Chemistry Economics Physics Psychology Public Adm.

student 0

student 1

difference

11.81 11.73 11.08 13.18 13.64 12.97

11.36 11.50 10.69 12.87 12.44 12.27

-0.45 -0.23 -0.39 -0.31 -1.20 -0.70

As an illustration, Figure 1(a) displays the cumulative incidence functions for obtaining a degree in Psychology. It shows that students start obtaining their degree after semester 8. The probability to graduate within 12 semesters is 1 7 . 7 % for student 0 and 3 0 . 2 % for student 1. Until semester 17, the cumulative incidence function has a larger value for student 1. Then, student 0 has a higher probability of obtaining a degree than student 1. The probability to graduate within 2 0 semester is 5 6 . 6 % and 5 5 . 5 % respectively.

12

The average length of study is calculated conditional on the first 2 0 semesters. For student 1, the remaining semesters should hardly increase the average length of study as his probability of studying more than 2 0 semester is only 1 % . For student 0, however, we may somewhat underestimate of average length of study because the probability of studying more than 2 0 semester may reach up to 6.6%. This implies that the reduction in the average length of study between student 0 and student 1 is probably even larger.

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany • 97

student 0

student 1

(a) Cumulative incidence functions

(b) Difference of cumulative incidence functions with 95 % confidence interval Figure 1 : Obtaining a degree in Psychology

Semester

98 · M. Heineck, M. Kifmann, and Ν. Lorenz

To analyze how tuition fees affect the probability of terminating one's studies within a certain period of time, we calculate the difference between the cumulative incidence functions of student 1 and student 0: D'(t)=l[(t)-I'0(t).

(15)

Using the Delta-method (see Greene 2003, p. 70), we determine pointwise confidence intervals. Based on these confidence intervals, we can say whether tuition fees have significantly changed the probability of terminating one's studies within semester t for each possibility /'. Figure 1(b) shows the difference between the two functions presented in Figure 1(a) and the corresponding 95% confidence interval. From semesters 8 to 13 the lower bound of the confidence interval is above zero. This implies that tuition fees have increased the probability of successfully finishing one's studies in these semesters at the 5% significance level. In the long run, there are no significant changes. In the following, we present graphs of the functions D'(t) for each possibility of terminating one's studies. Appendix A.2 contains the numerical values and states whether the functions D'(f) are significantly different from zero. Obtaining a degree: Figure 2 shows the difference between the cumulative incidence functions for obtaining a degree. For earlier semester we observe the following results: • An increase in the probability of obtaining a degree In Biology, Psychology and Public Administration the probability of obtaining a degree increases. This change is significant in Biology (semesters 8 to 10), Psychology (semesters 8 to 14) and Public Administration (semester 12). 15% 10% 5%

0%

1 -5% -10%

- o - Biology

α—π2

3

4

5

6

7

- λ - Chemistry - o - Economics Physics -X- Psychology - o - Public Adm.

-15% -20%

-25% Figure 2: Obtaining a degree - difference between cumulative incidence functions

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany • 99 • A decline in the probability of obtaining a degree In Chemistry, Economics and Physics the probability of obtaining a degree falls. This is significant in Chemistry (starting semester 10) and Economics (starting semester 9). In the long run the probability of obtaining a degree falls in all majors. However, this decline is significant only in Chemistry, Economics and Public Administration. Overall, we can distinguish four groups of majors: 1. Earlier degrees: In Biology und Psychology we observe a significant short run increase in the probability of obtaining a degree without a significant long run change. In these majors, students therefore seem to study more determined and obtain their degree earlier. 2. Fewer degrees: In Chemistry and Economics we find a significant decline in the probability of obtaining a degree for most semesters. 3. Mixed evolution: In Public Administration the probability of obtaining a degree increases significantly in the short run but declines significantly in the long run. 4. No significant effects: For Physics we observe no significant effects. Transfer to another

university:

Figure 3 demonstrates that tuition fees increase the probability to transfer to another university in all majors. These changes are significant in Physics, Economics and Public Administration. Students therefore seem to evade tuition fees in some majors by switching to another university. Switching majors: Figure 4 shows the result for the changes in the probability to switch majors. In almost all majors, tuition fees decrease this probability. Only in Psychology we find a non-significant increase. The decline is significant in Economics and Physics. This is in line with the hypothesis that students tend to switch their major less frequently because the expected costs have increased due to tuition fees. Dropping out: As Figure 5 illustrates, the probability of dropping out increases for all majors. Only in Psychology a temporary decline until semester 11 can be observed which is not significant. The increase is significant in Chemistry, Physics, Economics and Public Adminstration. This result is probably due to the fact that tuition fees raise the costs of continuing one's studies. Failing: Figure 6 shows that the probability of Physics where no change is observable. and Public Administration. A possible exams earlier and less well prepared to

6.2.

failing one's studies increases in all majors but A significant increase can be found for Biology explanation is that students tend to take their avoid tuition fees.

Discussion

Table 10 summarizes the results of our analysis based on the cumulative incidence functions. The sign of the difference between student 0 and 1 after 20 semesters is shown. An F indicates that students study faster, i.e. there is a significant short run increase in

100 • M. Heineck, M. Kifmann, and Ν. Lorenz

12%

10%

8%

- o - Biology -û- Chemistry -o~ Economics - + - Physics - χ - Psychology -o-Public Adm.

6%

4%

2%

!

0% 1

1 2

1 3

! 4

1 5

! 6

! 7

! 8

ι r r > ·. ! ! 1 I 1 ! 9 10 11 12 13 14 15 16 17 18 19 20

Figure 3: Transfer - difference between cumulative incidence functions

-α-Biology -ώ-Chemistry - o - Economics —t— Physics - χ - Psychology -o-Public Adm.

-5%

-6%

Figure 4: Switching majors - difference between cumulative incidence functions

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany · 101

4%

3% - o Biology -ώ- Chemistry - o - Economics -+- Physics - χ - Psychology -o- Public Adnr 0%

-1%

Figure 6: Failing - difference between cumulative incidence functions

102 · M. Heineck, M. Kifmann, and Ν. Lorenz

Table 10: Efferts of tuition fees after 20 semesters Biology Chemistry Economics Physics Psychology Public Adm.

Degree

Transfer

F -*** -** F -***

+ + +*** +** + +**

Switching m. -** -** + -

Dropping out + +*** +** +** + +***

Failing +* + + 0 + +*'

the probability of obtaining a degree but there are no significant long run effects with respect to this probability. Based on the significant changes after 20 semester, we can state the following results: • The significant decline in the probability of obtaining a degree in Chemistry, Public Administration und Economics is due to an increase in the probability of - a transfer to another university (Economics, Public Adminstration), - dropping out (all three), - failing one's studies (Public Administration). • In Physics there are no significant long-run effects with respect to the probability of obtaining a degree. A significant increase in transfers to other universities is compensated by a significant decline in the switch of majors. • In Biology und Psychology, virtually no long-run effects of tuition fees can be observed. It is interesting to compare these results with our regression analysis for the hazard rates in Table 7. The sign of the variables capturing the effect of tuition fees is frequently significantly positive and never significantly negative for the hazard rate for obtaining a degree. In contrast, the probability of obtaining a degree is significantly lower in three cases and never increases. In particular, in Public Administration the signs are opposite and significant. This apparent contradiction can be explained if one considers the relationship between the probability of obtaining a degree and the respective hazard rate. The probability of obtaining a degree in a certain semester is given by the product of the probability of still being a student in this semester and the hazard rate for obtaining a degree (see equation (14)). In Public Administration, there is a positive effect of tuition fees on switching majors, dropping out and failing (see Table 7). As a consequence, the probability that students are still studying is lower with tuition fees. This explains why the probability of obtaining a degree falls even though the hazard rate for obtaining a degree increases. As pointed out in the introduction, the main purpose for implementing tuition fees is to encourage students to study faster. Our results show that students of Biology, Psychology and, to some extent, Public Administration indeed seem to obtain their degree after a shorter duration of their studies. However, other majors show a different picture. In Chemistry, Public Administration and Economics, the probability of obtaining a degree is significantly lower after the introduction of tuition fees. An interesting question is why the results differ between the majors. One possibility is that the majors vary in the flexibility students have in accelerating their studies. In some majors, students may be able to avoid tuition fees by intensifying their studies and by passing the necessary exams earlier. In

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany · 103

other majors, this may not be possible and students drop out if they are not able or willing to pay tuition fees. Another explanation is that the majors attract distinct types of students who react differently to the introduction of tuition fees. It is also interesting to contrast our results with the changes in the average length of studies which we presented in Table 9. Based on this table, the introduction of tuition fees for long-term students seems to be a complete success. In all majors, students appear to study faster. However, this interpretation ignores that this figure is conditional on students obtaining a degree. In particular, it can decline for two reasons: (i) students obtain their degree in a shorter period of time. (ii) less long-term students obtain a degree. As our analysis shows the first reason applies to Biology, Psychology and Public Administration. In Chemistry, Economics and Physics, in contrast, the lower average length of study is entirely due to less students obtaining a degree. In particular, the probability of dropping out has increased in these majors. This can hardly be called a success. We therefore do not think that the average length of study is a good measure to capture the effects of tuition fees. Unfortunately, this statistic is usually regarded as the most important policy target. Finally, we return to our identification strategy. We compared students who are differently affected by tuition fees because they took up their studies at different dates. We chose this approach because no control group is available who studied at the same time but was not affected by tuition fees. Clearly, a limitation of this strategy is that we may measure the impact of other factors which changed over time apart from the introduction of tuition fees. In particular, reforms of the major programmes can be responsible for changes in the hazard rates. We therefore sent a questionnaire to the administrators of the different majors and found that the requirements for the majors Chemistry and Public Administration were reformed in 1999, one year after the introduction of tuition fees for long-term students. We suspect that the reform in Chemistry may partially explain the quantitatively strong effects which we observe in this major (see Figure 2). We can also not rule out that the introduction of an "orientation examination" after two semesters had an influence on student behavior. Under this scheme, which was introduced in all majors but Economics in the fall term 2000, students are required to pass a certain number of exams within two semesters. 13 Finally, there was some insecurity about the exact future of the Economics programme in 1997. This may have contributed to the large increase in the probability of transferring to other universities (see Figure 3). We also considered whether changes in the entry requirements for the different majors could be responsible for some of the effects we observe after the introduction of tuition fees. In particular, it may be important whether the majors had a Numerus Clausus (NC), i.e. were only open to students with a minimum grade average in high school. This, however, does not seem to be an important factor. In Biology, Psychology and Public Administration there was a NC throughout the observation period, while in Economics there were no entry restrictions. Chemistry and Physics had a NC for a short period of time (Chemistry for 4 years in the mid-90s, Physics for some time in the 80s). The abolition of the NC in Chemistry may have contributed somewhat to the decline in the probability of obtaining a degree after the introduction of tuition fees.

13

In Economics, such a requirement was already implemented in 1993.

104 · M. Heineck, M. Kifmann, and Ν. Lorenz 7.

Conclusion

This study examined the impact of tuition fees for long-term students at the University of Konstanz. In a duration analysis we examined how tuition fees change when and how students finish their studies in the majors Biology, Chemistry, Economics, Physics, Public Administration and Psychology. The effect of tuition fees was measured by an indicator for having to pay tuition fees in the following semester as well as by variables which capture whether students knew that tuition fees are due if they study too long. In most majors we found evidence that the introduction of tuition fees influenced students' behavior. With respect to the hazard rate, i.e. the conditional probability of terminating one's studies in a particular way, we observed the following significant effects: 1. The indicator for having to pay tuition fees in the following semester often raises the hazard rate for dropping out. In some majors, there is also evidence that the hazard rate for obtaining a degree and for transferring to another university has increased. 2. Knowing that tuition fees have to be paid frequently increases the hazard rate for obtaining a degree, for transferring to another university, for dropping out and for failing. Based on our regression results we furthermore examined how the probability of terminating one's studies within a certain period of time changes. In two majors we found that students obtain a degree in a shorter period of time. In three other majors, however, we observed that the probability of obtaining a degree generally decreased. In addition, the probability of transferring to another university, of dropping out and of failing significantly increased in the long run. The probability of switching majors fell significantly in two majors. In the political debate, the average length of study to obtain a degree plays a prominent role. Considering only this measure, tuition fees for long-term students appear to be successful. In all majors, this measure fell with the introduction of tuition fees. However, if one considers the different possibilities of terminating one's studies, things look differently. In three of the six majors examined, the length of study to obtain a degree is reduced only because less long-term students finish with a degree. In particular, the number of drop outs increased. Nevertheless, in two majors we found that the average length of study to obtain a degree decreased because students were actually studying faster. Our study therefore does not allow us to draw a clear-cut conclusion. In future research, it would therefore be desirable to analyze data from other universities. Of particular interest would be a comparison with universities in states which have not introduced tuition fees for long-term students. Finally, we want to comment on the implications of our results for the debate on the introduction of general tuition fees which have to be paid for each semester studied. These fees are likely to have stronger effects than the current fees for long-term students. In particular, they will also influence students who expect to finish within four semesters beyond the standard period of study. According to our analysis, general tuition fees should have a considerable impact on how fast students study and whether they obtain a degree. However, an important aspect will be whether fees are accompanied by a student loan scheme. This is not the case for the current tuition fees for long-term students.

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany · 105 A

Appendix

A.1

Regression results Biology

Degree

Chemistiy

Economics

Physics

Psychology

Public Adm.

sts8 sts9 sts10 sts11 sts12 sts13 sts14 sts15 sts16 sts17 sts18 sts19 sts20 male age enroll local class-size A Ind const

1.974967*** 20.2287*** 3.397996*** 19.86454*** 5.662304*** 23.24294*** 7.378845*** 23.60037*** 7.44879*** 23.91013*** 7.597168*** 23.88942*** 7.930193*** 24.3195*** 5.861437*** 23.98152*** 7.35232*** 24.0679*** 5.692042*** 21.58534*** 6.054108*** 21.58534*** 7.549235*** 21.58534*** 5.064344*** 21.58534*** 0.1164097 -0.2172269 -0.6435049* -2.022609*** -0.0082655 -0.0129179 0.2437513** -0.2182385 -0.0022371 0.0055694* 0.343907*" 0.162498 0.1848649 0.0634963 -6.303439*** -20.48888

4.260039*** 6.218726*** 7.184894*** 7.601967*** 7.644835*** 7.389974*** 7.071717*** 6.869842*** 6.837578*** 5.718039*** 5.718039*** 5.718039*** 5.718039*** -0.3737093*** -0.1685597 -0.0137154** -0.0372965 -0.0036107** 0.1465043 0.5877511* -7.273345***

19.45348*** 19.45348*** 19.45348*** 22.42413*** 23.62787*** 23.93926*** 24.13418*** 23.50445*** 23.35253*** 22.56984*** 22.89418*** 22.17571*** 22.71232*** -0.2914152 -0.7726283 -0.0012705 -0.1363979 0.0002042 0.0718659 0.408061 -22.64185

n.a. 20.44832*** 20.44832*** 21.87805*** 22.82276*** 23.21273*** 23.71829*** 23.73998*** 23.37639*** 23.07087*** 23.07087*** 23.07087*** 23.07087*** -0.0039631 -0.4874159*** -0.0051815 0.0210963 -0.0049904 0.8572541*** -0.1534556 -22.91521

2.920238*** 3.814612*** 4.71912*** 5.922132*** 6.477966*** 7.321886*** 7.160747*** 6.839535*** 6.512411*** 6.405023*** 5.605654*** 4.644008*** 5.101521*** 0.0375017 -0.1318994 -0.0145248*** -0.1887497*** -0.0003815 0.3594634*** 0.4849847*** -7.659718***

Transfer sts2 sts3 sts4 sts5 sts6 sts7 sts8 sts9-20 male age enroll local class-size A Ind const

0.0611056 -0.1781256 -0.6485812*** -1.313005*** -1.253751*** -1.565894*** -1.499817*** -2.311762*** -1.560806*** -1.108872*** -1.582918*** -2.548449*** -3.503091*** -2.51121*** -3.67331*** -3.57254*** -0.2641158* -0.7249841*** 1.513525*** 1.547521*** 0.0063263 0.0196962* 0.13792 0.3271722* 0.0223436*** 0.0001706 0.1184073 0.4445865 n.a. 2.208698* -8.379917*** -5.365878***

0.4735028** -0.9934174*** -0.5634423* -0.5434851 -0.6364782 -1.870737*** -1.479822** -1.210622*** -0.697343*** 1.758308*** 0.0113911 0.2115051 0.0047906* 1.223186*** -0.1728365 -7.867756***

0.6542962** 0.56272* -0.0116743 -0.2588272 -0.5298933 0.0681669 -0.3732931 -1.657214** -0.2133594 0.2443782 -0.320059 -0.551057 -1.197199 -1.450943* -1.745857*** -2.282797*** -0.9368295*** 0.0951684 2.518581*** 0.5301219** 0.0120502 -0.0085142 0.1190214 0.2316609 0.0120946*** 0.0190861 0.6986277** 0.1741252 0.787424 n.a. -9.923418*** -7.046656***

0.139246 -0.8213257*** -0.5084243* -2.728362*** -0.6988106** -1.551012*** -2.234751*** -2.068758*** -0.1462606 0.5455654 0.0131276 0.6970058*** 0.004051*** 0.6964993*** 1.509301*** -6.230482***

106 · M. Heineck, M. Kifmann, and Ν. Lorenz Biology

Chemistry

Economics

Physics

Psychology

Public Adm.

Switching majors sts2 0.5794609*** sts3 -1.313513*** sts4 -0.3539095 sts5 -1.552763*** sts6 -1.48928*** sts7 -2.302278*** sts8 -2.265815*** sts9-20 -2.485305*** male -0.1027515 age 0.8585485* 0.00213 enroll local 0.0221473 class-size 0.0009991 A -0.0936973 Ind n.a. const -5.474606***

0.6715915*** -0.5695386* -0.0717464 -1.523288*** -0.3078543 -2.732351*** -1.575297*** -3.734553*** -0.6544738*** 0.310408 0.0000255 0.1076915 -0.0009691 0.0124755 n.a. -3.364237**

0.5320003*** -0.7765511*** -0.3639707* -0.6948294*** -0.4598338* -1.393332*** -1.879352*** -2.714367*** -0.4164134*** 1.166683*** -0.0006809 0.0082337 -0.0034395** -0.1404207 n.a. -4.805014***

0.7819458*** -0.6241763* -0.6241763* -0.6241763* 0.6065015* -1.393406* -0.8931372 -2.308758*** -1.305817*** 1.287214* -0.0035324 -0.2519128 -0.0079171** -0.7096649** n.a. -4.983414***

0.3395517 -0.9537078 -1.507402** -1.507402** -0.1654198 -1.657781 -1.631124 -1.817148*** 0.4637762 -0.1159352 -0.0079007 -0.0275291 -0.0117013 0.06867 n.a. -4.023478**

0.7688363*** -0.9937873*** -0.1493226 -0.8820427** -0.9612757** -1.887952*** -1.570922*** -1.216996*** 0.1448964 0.3059622 -0.0045018 0.6096544*** 0.0026074* 0.0197209 n.a. -6.004521***

Dropping out sts2 sts3 sts4 sts5 sts6 sts7 sts8 sts9 sts10 sts11 sts12 sts13-20 male age enroll local class-size A Ind const

0.0903734 -1.140102*** -0.3294629 -0.3220855 -0.3093046 -0.9277298*** -1.271706*** -2.030767*** -1.532291*** -1.628181*** -2.277854** -0.5835983 -0.3007809** 0.5776969 0.023144*** -0.2600392* 0.0060457** 0.8883495*** -0.7177593 -3.593123***

0.0874903 -1.082016*** 0.0859385 0.4210878*** 0.5104502*** -0.3905777* -1.393037*** -1.499522*** -1.334647*** -0.7148301* -0.3481676 -0.6326656* -0.0307982 0.5092352* 0.0164711*** -0.4127154*** 0.0049849*** 0.4721371*** 0.1287419 -3.825025***

0.2274827 -0.936578*** 0.4603892** 0.4792717** 0.6741985*** -0.5987* -0.3674402 -0.9228817** -3.133373*** -1.396396*** -1.396396*** -1.396396*** -0.2346281 0.6684465 0.017397*** -0.1318082 0.0037631** 0.2664241* 2.05273*** -4.290384***

-0.2560648 -0.4539695 0.7889643*** 0.8743837*** 0.6298195*** 0.2247153 -0.1787363 -0.9492164** -0.6234914 -1.778823** -0.7667034 -0.0889769 0.0028107 0.4621017*** 0.0118794** -0.3689604*** 0.0074347 -0.1111032 1.97766*** -4.843065***

0.0229043 -0.9287206*** -0.2086328 -0.3479509* -0.2112073 -0.3923557* -1.103335*** -3.152148*** -2.427758*** -1.82287*** -1.459584*** -0.2992622 0.0307512 0.5986893*** 0.0199791*** 0.2544592** 0.0020393** 0.5432201*** 1.141365*** -5.171585***

-0.3277562 1.060357 0.0809847 -0.6658845 -0.0283409 0.765601 n.a. -6.513882

0.0291644 -0.7020984 0.001702 0.3399538 -0.0046668 0.8324706** 0.5165393 -3.734778

n.a. 2.200965 0.0045826 21.04945** -0.0332951 -0.154604 n.a. -30.72349

2.410151** 1.546876* 0.0544989* -1.664271 -0.0682431 22.42727*** 1.773447 -28.62266

0.3753721 1.336394*** 0.0259608* 0.0698376 0.0014566 1.438029*** 0.7164247 -9.644116***

Failing

0.0275029 -0.9369486*** -0.2469692 -0.2285155 -0.4700747** -0.9175168*** -1.50017*** -2.655011*** -1.807645*** -1.831407*** -1.232548** -1.138159* -0.2453875** 0.9873081*** 0.0256908*** -0.0460841 0.012736*** 0.1651401 1.544195 -5.682223***

male 0.6058492 age 1.620225 enroll 0.0144437 local -0.2837911 class-size 0.042646** A 1.135638** Ind n.a. const -14.51838***

Number of semesters Log-Likelihood

11,367 -4,491.30

5,943 -2,785.67

9,538

9,206

9,885

25,056

-5,182.13

-3,265.69

-2,956.45

-7,526.73

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany · 107

A.2

Differences of the cumulative incidence functions

1 2 3 4 5 6 7 8 9 10 11 12 13 15 16 17 18 19 20

0.02% 0.03% 0.04% 0.05% 0.06% 0.07% 0.08% 0.14%» 0.36%* 1.84%* 2.79% 1.99% -0.30% -2.55% -2.82% -3.62% -3.76% -3.94% -4.37% -4.41 %

0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% -0.23% -0.40% -4.42%** -8.77%** -12.63%*** -15.24%*** -17.53%*** -18.57%*** -19.26%*** -19.32%*** -19.37%*** -19.41%*** -19.46%***

0.00% 0.00% 0.00% 0.00% 0.00% -0.01 % -0.01 % -0.23% -1.74%** -4.90%*** -6.34%*** -8.55%*** -10.09%*** -11.02%*** -11.64%*** -12.13%*** -12.28%*** -12.41%*** -12.52%*** -12.62%***

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0.41 % 0.74% 0.87% 0.94% 0.97% 1.00% 1.03% 1.03% 1.03% 1.03% 1.03% 1.02% 1.02% 1.01% 1.01% 1.01% 1.01% 1.01% 1.01% 1.01%

2.04% 2.94% 3.16% 3.25% 3.27% 3.23% 3.22% 3.19% 3.19% 3.18% 3.17% 3.30% 3.37% 3.40% 3.42% 3.43% 3.43% 3.44% 3.45% 3.45%

3.16%*** 6.88%*** 7.65%*** 8.60%*** 9.33%*** 9.83%*** 9.96%*** 10.13%*** 10.30%*** 10.41%*** 10.42%*** 10.41%*** 10.40%*** 10.39%*** 10.38%*** 10.37%*** 10.36%*** 10.35%*** 10.35%*** 10.34%***

0.89%* 2.39%** 3.12%** 3.51%** 3.90%** 4.53%** 4.86%** 4.99%** 5.06%** 5.13%** 5.20%** 5.24%** 5.31%** 5.34%** 5.35%** 5.36%** 5.36%** 5.36%** 5.36%** 5.36%**

0.22% 0.57% 0.72% 0.92% 0.96% 1.09% 1.19% 1.23% 1.24% 1.25% 1.25% 1.24% 1.23% 1.21% 1.20% 1.19% 1.19% 1.18% 1.18% 1.17%

-0.24% -0.66% -0.73% -0.89% -0.94% -1.00% -1.02% -1.04% -1.06% -1.09%

-0.32% -1.37% -1.71% -2.33% -2.49% -3.06% -3.11% -3.27% -3.29% -3.31%

-0.89% -2.59%* -3.06%* -3.78%** -4.30%** -4.93%** -5.18%** -5.32%** -5.38%** -5.42%**

-0.80%** -2.40%*** -2.79%*** -3.15%*** -3.48%*** -4.52%*** -4.66%*** -4.88%*** -4.93%*** -4.98%***

0.05% 0.12% 0.14% 0.15% 0.16% 0.19% 0.20% 0.21 % 0.21 % 0.22%

g majors 1 2 3 4 5 6 7 8 9 10

0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% -0.02% -0.03% -0.05% -0.43% -1.61 % -0.75% -2.28% -3.41 % -4.41 % -4.88% -5.49% -5.78% -6.24%

0.00%* 0.00%* 0.00%* 0.00%* 0.00%* 0.00%* 0.00%* 0.00%* 1.77%*** 3.39%*** 8.48%*** 12.53%*** 12.84%*** 9.47%** 5.68% 3.35% 1.87% 0.65% -0.34% -1.11%

0.01%* 0.03%* 0.04%* 0.04%* 0.05%* 0.06%* 0.06%* 0.16%* 0.37% 0.78% 1.40% 5.46%** 3.17% -0.95% -4.01 % -6.03%** -7.64%** -8.31%*** -8.56%*** -8.94%*** 1.14%** 2.28%** 2.69%** 3.20%** 3.25%** 3.61%** 3.76%** 3.83%** 3.90%** 3.97%** 4.02%** 4.35%** 4.47%** 4.51%** 4.52%** 4.53%** 4.52%** 4.52%** 4.52%** 4.51%** -0.01% -0.12% -0.14% -0.21% -0.25% -0.29% -0.31% -0.33% -0.37% -0.42%

108 · M . Heineck, M . Kifmann, and Ν. Lorenz

Semester

Biology

Chemistry

Economics

Physics

Psychology

Public Adm.

-5.03%"* -5.06%*** -5.09%*** -5.10%*** -5.12%*** -5.13%*** -5.14%*** -5.14%*** -5.15%*** -5.16%***

0.21 % 0.20% 0.18% 0.16% 0.15% 0.14% 0.13% 0.13% 0.12% 0.12%

-0.46% -0.53% -0.58% -0.63% -0.67% -0.70% -0.73% -0.75% -0.77% -0.79%

3.25%*** 5.47%*** 6.05%*** 7.30%*** 8.24%*** 8.63%*** 8.70%** 8.70%** 8.68%" 8.64%** 8.49%** 8.27%" 8.13%** 8.02%** 7.92%** 7.85%** 7.79%** 7.73%** 7.68%** 7.63%**

1.16% 2.39%* 2.74%* 3.93%* 4.91 %* 5.85%* 6.09%* 6.36%* 6.50%* 6.51%* 6.58%* 6.62%* 7.98%" 8.58%** 8.91 % " 9.10%" 9.24%" 9.33%" 9.39%** 9.43%"

-0.30% -0.52% -0.70% -1.25% -1.78% -2.16% -2.40% -2.55% -2.64% -2.78% -2.84% -0.71 % 1.31% 2.02% 2.21 % 2.24% 2.21 % 2.16% 2.10% 2.05%

1.64%"* 3.07%"* 3.58%*" 4.53%*" 5.29%*" 6.08%*" 6.67%"* 6.95%"* 6.98%"* 7.04%*" 7.12%*" 7.73%*" 8.32%*" 8.46%*" 8.44%*" 8.38%*** 8.29%*" 8.21%*" 8.12%*** 8.02%*"

0.36% 0.61 % 0.84% 1.01 % 1.14% 1.22% 1.29% 1.36% 1.40% 1.43% 1.45% 1.46% 1.45% 1.45% 1.44% 1.44% 1.43% 1.43% 1.42% 1.42%

0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%

0.05% 0.10% 0.14% 0.18% 0.22% 0.26% 0.30% 0.33% 0.36% 0.39% 0.42% 0.51 % 0.57% 0.59% 0.60% 0.60% 0.61 % 0.61 % 0.61 % 0.61%

0.42%*" 0.80%*** 1.16%*" 1.49%*** 1.81 % * " 2.11%*" 2.39%*" 2.66%*" 2.92%*" 3.16%*** 3.35%*** 3.63%*** 3.73%*** 3.76%*** 3.77%*** 3.78%*** 3.77%*** 3.77%*** 3.77%*** 3.76%***

Switching majors (cont.)

11 12 13 14 15 16 17 18 19 20

-1.11% -1.12% -1.13% -1.14% -1.14% -1.15% -1.15% -1.15% -1.15% -1.15%

-3.32% -3.33% -3.33% -3.33% -3.34% -3.34% -3.34% -3.34% -3.34% -3.34%

-5.46%** -5.48%** -5.50%** -5.51%** -5.52%** -5.53%** -5.53%** -5.54%** -5.54%** -5.55%**

Dropping out

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0.89% 1.61% 1.86% 2.27% 2.63% 2.87% 3.01% 3.08% 3.10% 3.11% 3.08% 3.63% 3.81% 3.85% 3.86% 3.86% 3.86% 3.85% 3.85% 3.84%

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0.17%* 0.32%* 0.45%* 0.57%* 0.69%* 0.79%* 0.90%* 0.99%* 1.09%* 1.17%* 1.20%* 1.21%* 1.21%* 1.21%* 1.21%* 1.21%* 1.21%* 1.21%* 1.21%* 1.21%*

7.14%*** 12.36%*" 13.68%"* 15.99%"* 17.82%*" 19.18%"* 19.81%*" 20.19%*" 20.35%*" 20.51 % * " 20.58%"* 20.53%*" 20.35%*" 20.25%*" 20.18%*" 20.14%*" 20.11%"* 20.07%"* 20.04%*" 20.01 % * "

Failing

0.03% 0.05% 0.06% 0.08% 0.08% 0.09% 0.10% 0.10% 0.10% 0.11% 0.11% 0.11% 0.11% 0.11% 0.10% 0.10% 0.10% 0.10% 0.10% 0.10%

A Duration Analysis of the Effects of Tuition Fees for Long-Term Students in Germany • 109

References Allison, P. (1982), Discrete-Time Methods for the Analysis of Event Histories. Pp. 61-98 in: S. Leinhard (ed.), Sociological Methodology Jossey-Bass: San Francisco. Booth, Α., S. Satchell (1995), The Hazards of doing a PhD: An Analysis of Completion and Withdrawal Rates of British PD Students in the 1980s. Journal of the Royal Statistical Society, Ser. A, Vol. 158, pp. 297-318. Chtzmar, J. (2000), A Discrete-time Hazard Analysis of the Role of Gender in Persistence in the Economics Major. Journal of Economic Education, Vol. 31, pp. 107-118. Greene, W.H. (2003). Econometric Analysis. Prentice Hall: Upper Saddle River, New Jersey, 5 edn. Hackl, P., G. Sedlacek (2002a), Analyse der Studiendauer am Beispiel der Wirtschaftsuniversität Wien. Pp. 41-59 in: R. Dutter (ed.), Festschrift 50 Jahre Osterreichische Statistische Gesellschaft, Österreichische Statistische Gesellschaft: Wien. Hackl, P., G. Sedlacek (2002b), Forschungsbericht Studienverlaufsanalyse, Forschungsbericht, Wirtschaftsuniversität Wien, http://eeyore.wuwien. ac.at/stat4/forschungsbericht502.pdf (1 July 2005). Hamerle, Α., G. Tutz( 1989), Diskrete Modelle zur Analyse von Verweildauer und Lebenszeiten. Campus Forschung No. 568, Campus: Frankfurt. Kalbfleisch, ]., R. Prentice (2002), The Statistical Analysis of Failure Time Data. Wiley: Hoboken, New Jersey. Ministerium für Wissenschaft, Forschung und Kunst Baden-Württemberg (2003). Bericht zum Staatshaushaltsplan für 2004, Stuttgart. Sedlacek, G. (2003), Analyse der Studiendauer und des Studienabbruch-Risikos unter Verwendung der statistischen Methoden der Ereignisanalyse. Ph.D. thesis, Wirtschaftsuniversität Wien. Statistisches Landesamt Baden-Württemberg (2003), Langzeitstudierende: Anzahl sinkt auf insgesamt 19.600, Pressemitteilung 333/2003, http://www.statistik.baden-wuerttemberg.de/Pressemitt/2003333.asp (31 October 2003). Yamaguchi, K. (1991), Event History Analysis. Sage Publications: Newbury Park. Martin Heineck, Universität Konstanz, Fach D 136, D-78457 Konstanz. Prof. Dr. Mathias Kifmann (corresponding author), Universität Augsburg, Universitätsstr. 16, D-86159 Augsburg. Tel. ++49/+821/598-4206. E-mail: [email protected] Normann Lorenz, Universität Konstanz, Fach D 136, D-78457 Konstanz.

Helmut Kromrey

Empirische Sozialforschung Modelle und Methoden der standardisierten Datenerhebung und Datenauswertung 11., überarbeitete Auflage 2006. 565 S., kt. € 14,90/sFr 26,80 UTB 1040 (ISBN 3-8252-1040-5)

Die in elfter Auflage vorliegende Einführung in die standardisierte empirische Sozialforschung ist anwendungspraktisch orientiert. Ihr Aufbau orientiert sich am Ablauf eines realen Forschungsprozesses. Die grundlegenden wissenschaftstheoretischen und methodologischen Fragen werden nicht als Selbstzweck, sondern entsprechend ihrer Relevanz für die Forschungspraxis abgehandelt und durch sozialwissenschaftliche Beispiele veranschaulicht. Der Text setzt Vorkenntnisse nicht voraus und ist besonders an den Bedürfnissen der "Neueinsteiger" orientiert. Für "Wiederholer" und Praktiker werden in den Fußnoten fortlaufend Hinweise zur Vertiefung geboten sowie Querbezüge zu qualitativen Forschungsansätzen hergestellt.

Inhaltsübersicht Vorbemerkung: Wozu "Methoden empirischer Sozialforschung?" 1

Empirische Sozialforschung und empirische Theorie

2

Forschungsfragen, Forschungsdesign, Forschungsprozess

3

Die empirische "Übersetzung" des Forschungsproblems

4

Strategien der Operationalisierung und Indikatorenauswahl

5

Messung und Datenerhebung in den Sozialwissenschaften

6

Auswahlverfahren

7

Datenerhebungsverfahren und -instrumente der empirischen Sozialforschung

8

Methoden und Modelle der deskriptiven Statistik

9

Typen und Konzepte empirischer Sozialforschung: Eine Übersicht

ORDO

Jahrbuch für die Ordnung von Wirtschaft und Gesellschaft Begründet von WALTER EUCKEN und FRANZ BÖHM Herausgegeben von Hans Otto Lenel Helmut Gröner Walter Hamm Ernst Heuß Erich Hoppmann Wolfgang Kerber

Martin Leschke Ernst-Joachim Mestmäcker Wernhard Möschel Josef Molsberger Peter Oberender

Razeen Sally Alfred Schüller Viktor Vanberg Christian Watrin Hans Willgerodt

Band 56: 2005. 425 S„ geb. € 76,- / sFr 127,-. ISBN 3-8282-0327-2

Inhaltsübersicht Christian Watrin, Hayeks Theorie einer freiheitlichen politischen Ordnung Walter Hamm, Entartung des politischen Wettbewerbs Paul Kirchhof, Freiheitlicher Wettbewerb und staatliche Autonomie - Solidarität Viktor Vanberg, Auch Staaten tut Wettbewerb gut: Eine Replik auf Paul Kirchhof Paul Kirchhof, Der Staat tut dem Wettbewerb gut Alfred Schüller, Soziale Marktwirtschaft als ordnungspolitische Baustelle - Die Synthese zwischen "Freiburger Imperativ" und "Keynesianische Botschaft" ein nationalökonomischer Irrweg Ulrich Fehl, Warum Evolutorische Ökonomik? Carl Christian von Weizsäcker, Hayek und Keynes: Eine Synthese Manfred E. Streit, Die Soziale Marktwirtschaft - zur Erosion einer wirtschaftspolitischen Konzeption Drieu Godefridi, The Anarcho-Libertarian Utopia-A Critique

LUCIUS LUCIUS

Wolf Schäfer, Exit-Option, Staat und Steuern Katarina Röpke/Klaus Heine, Die 26. Ordnung - Vertikaler Regulierungswettbewerb, supranationale Rechtsangebote und europäischer Binnenmarkt Roland Vaubel, Das Papsttum und der politische Wettbewerb in Europa: Eine Übersicht Norbert Berthold/Sascha von Berchem, Lokale Solidarität - die Zukunft der Sozialhilfe? Andreas Freytag/Simon Renaud, Langfristorientierung in der Arbeitsmarktpolitik Dieter Schmidtchen/Roland Kirstein, Mehr Markt im Hochschulbereich: Zur Effizienz und Gerechtigkeit von Studiengebühren Jürgen Zerth, Ideal der flächendeckenden Versorgung im Gesundheitswesen: Idee oder Fiktion? Hartmut Berg/Stefan Schmitt, Zur Bestreitbarkeit von Märkten: Low-CostCarrier als neue Anbieter auf dem EU-Luftverkehrsmarkt

®

Stuttgart

Neue Theorien des Rechts Herausgegeben von Sonja Buckel, Ralph Christensen und Andreas Fischer-Lescano 2005. XVI11/444 S., kt. 24,90 /sFr 43,70. UTB 2744 (ISBN 3-8252-2744-8) Die Autoren geben einen Überblick über moderne rechtstheoretische Fragestellungen. Diese werden vor dem Hintergrund der aktuellen Herausforderungen für das Recht vorgestellt, um zum kritischen Mit- und Nachdenken der wichtigsten Richtungen anzuregen. Das Buch bietet ein breites Spektrum der neuen Theorien: rechtsphilosophische (Habermas, Brandom, Derrida, Lyotard etc.), rechtstheoretische (Alexy, Wiethölter), rechtssoziologische (Luhmann, Teubner, Koh), rechtsgeschichtliche (Amstutz, Fögen) und gesellschaftstheoretische (Foucault, Maus, Critical Legal Studies, Postmarxismus, Feminismus). Inhaltsübersicht Einleitung: Neue Theoriepraxis des Rechts

9. Gestaltung des Rechts: Agamben (Fabian Steinhauer)

A. Trennung und Verknüpfung von Recht und Politik

C. Fragmentierung des Rechts

1. Demokratischer Positivismus: Habermas/Maus (Peter Niesen/Oliver Eberl) 2. Dekonstruktion der Gerechtigkeit: Nietzsche/Derrida (Thomas-Michael Seibert) 3. Systemtheorie: Luhmann/Teubner (Gralf-Peter Calliess) B. Politik des Rechts 4.

5.

6. 7. 8.

Prozedurale Rechtstheorie: Wiethölter (Andreas Fischer-Lescano/Gunther Teubner) Partisanen der Rechtskritik: Critical Legal Studies etc. (Günter Frankenberg) Neo-Materialistische Rechtstheorie (Sonja Buckel) Macht und Recht: Foucault (Thomas Biebricher) Feministische Rechtstheorie (Sarah Elsuni)

LUCIUS "LUCIUS

10. Theorien der radikalen Fragmentierung: Ladeur/Lyotard/Weber (Matthias Kronenberger) 11. Neo-Pragmatismus: Brandom (Ralph Christensen/Michael Sokolowski) 12. Nachpositivistisches Rechtsdenken (Nikolaus Forgó/Alexander Somek) 13. Theorie der Interpretation: Davidson (Jochen Bung) 14. Psychoanalytische Rechtstheorien (Stefan Häußler) 15. ökonomische Theorie des Rechts (Felix Müller) D. Transnationaler Rechtspluralismus 16. Theorie transnationaler Rechtsprozesse (Felix Hanschmann) 17. Evolutorische Rechtstheorie (Andreas Abegg) 18. Deliberative Rechtstheorie (Timo Tohidipur)

®

Stuttgart