[Article] A general model for detecting genetic determinants underlying longitudinal traits with unequally


248 91 229KB

English Pages 14 Year 2005

Report DMCA / Copyright

DOWNLOAD PDF FILE

Recommend Papers

[Article] A general model for detecting genetic determinants underlying longitudinal traits with unequally

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

Biostatistics (2005), 6, 3, pp. 420–433 doi:10.1093/biostatistics/kxi019 Advance Access publication on April 14, 2005

A general model for detecting genetic determinants underlying longitudinal traits with unequally spaced measurements and nonstationary covariance structure WEI HOU, CYNTHIA W. GARVAN, WEI ZHAO Department of Statistics, University of Florida, Gainesville, FL 32611, USA MARYLOU BEHNKE, FONDA DAVIS EYLER Department of Pediatrics, University of Florida, Gainesville, FL 32611, USA RONGLING WU∗ Department of Statistics, University of Florida, Gainesville, FL 32611, USA [email protected]

S UMMARY A mixture model for determining quantitative trait loci (QTL) affecting growth trajectories has been proposed in the literature. In this article, we extend this model to a more general situation in which longitudinal traits for each subject are measured at unequally spaced time intervals, different subjects have different measurement patterns, and the residual correlation within subjects is nonstationary. We derive an EM–simplex hybrid algorithm to estimate the allele frequencies, Hardy–Weinberg disequilibrium, and linkage disequilibrium between QTL in the original population and parameters contained in the growth equation and in the covariance structure. A worked example of head circumference growth in 145 children is used to validate our extended model. A simulation study is performed to examine the statistical properties of the parameter estimation obtained from this example. Finally, we discuss the implications and extensions of our model for detecting QTL that affect growth trajectories. Keywords: Longitudinal trait; EM algorithm; Genetic determinant; Mixture model.

1. I NTRODUCTION Many complex traits of interest to medical clinicians, such as head circumference growth, tumor progression, or drug response, display longitudinal or developmental features, changing their phenotypes with time or other independent variables. The developmental trajectories of these longitudinal traits have a genetic component (Atchley and Zhu, 1997), but the identification of the underlying genetic determinants or quantitative trait loci (QTL) is a difficult statistical issue. We have developed a statistical method for detecting genetic determinants that affect complex phenotypes undergoing developmental changes (Wu et al., 2002a, 2003a,b, 2004a,b; Ma et al., 2002). This method integrates mathematical models of growth curves into quantitative and molecular genetics theory within a QTL mapping framework. It does not ∗ To whom correspondence should be addressed.

c The Author 2005. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: [email protected]. 

A general model for the genetic study of longitudinal traits

421

directly estimate the expected phenotypic means of different QTL genotypes at different time points, but rather fits these means using growth curves as functions of time. Thus, the estimation of QTL effects on growth is equivalent to the estimation of model parameters describing the shape of growth curves. Although it was originally conceived to map the dynamic effects of QTL on growth trajectories, our statistical method has been extended to study the genetic mechanisms for many other biological or biomedical processes including allometric scaling (Wu et al., 2002b), thermal norm reaction, HIV-1 dynamics (Wang and Wu, 2004), tumor progression, circadian rhythms, and drug responses (Gong et al., 2004). Simulation studies suggest that this method can be used to detect the genetic machinery of each of these processes. We anticipate that the application of this method to practical genetic mapping data sets could lead to the discovery of genes that are biologically or clinically important. Our statistical models for detecting longitudinal QTL were previously constructed as finite mixture models (Wu et al., 2002a; Ma et al., 2002) with stationary first-order autoregressive residual covariance structures (Diggle et al., 2002). The autoregressive assumption is computationally convenient but the stationarity assumption may be unrealistic. Although variances can often be stabilized by transforming the data (Box and Cox, 1964; Carroll and Ruppert, 1984), it is more difficult to deal with nonstationary correlation structure. In Section 2, we incorporate nonstationary correlation structure into our statistical framework for detecting genetic determinants of growth trajectories. In Section 3, the model is further extended to accommodate unequally spaced measurement times and different measurement patterns. Section 4 describes an application to head circumference growth in children for different subjects from birth to age 7 years. Section 5 describes the wider implications and extensions of our approach. 2. A FRAMEWORK FOR HUNTING LONGITUDINAL QTL The statistical foundation of QTL mapping is a mixture model in which each observation y is assumed to have arisen from one of a known or unknown number of components. Assuming that there are L QTL genotypes contributing to the variation of a quantitative trait, this mixture model is expressed as y ∼ p(y|̟, ϕ, η) = ̟1 f (y; ϕ1 , η) + · · · + ̟ L f (y; ϕ L , η),

(2.1)

where ̟ = (̟1 , . . . , ̟ L )′ are the mixture proportions (i.e. QTL genotype frequencies) which are constrained to be nonnegative and sum to unity; ϕ = (ϕ1 , . . . , ϕ L )′ are the component (or QTL genotype) specific parameters, with ϕl being specific to component l; and η are parameters which are common to all components. To describe a longitudinal trait, for example growth, we need to measure trait values, y, at a series of hallmark time points (say τ ). The likelihood function of longitudinal data measured for a random sample of size n is constructed as L(̟, ϕ, η|y) =

n  i=1

[̟1 f (yi ; ϕ1 , η) + · · · + ̟ L f (yi ; ϕ L , η)] ,

(2.2)

where mixture proportions, ̟1 , . . . , ̟ L , can be viewed as the prior probabilities of QTL genotypes. Traditional QTL-detecting methods (Lander and Botstein, 1989) can be readily extended to accommodate time-dependent traits. The multivariate normal distribution of individual i measured at τ time points is expressed as   1 1 −1 ′  (yi − ml ) , f (yi ; ϕl , η) = exp − (yi − ml ) (2.3)  |1/2 2 (2π)τ/2 |

422

W. H OU ET AL .

where yi = [yi (t)]1×τ = [yi (1), . . . , yi (τ )] is a vector of observation measured at τ time points and ml = [µl (t)]1×τ = [µl (1), . . . , µl (τ )] is a vector of expected values for QTL genotype l at different points. At a particular time t, the relationship between the observation and expected mean can be described by a linear regression model, yi (t) = xi µl (t) + ei (t), where xi is the indicator variable denoted as 1 if a QTL genotype l is considered for individual i and 0 otherwise; ei (t) as the residual error sequences assumed i.i.d. normal with mean 0 and variance σ 2 (t); and covariance structure σ (t j1 , t j2 )={ei (t1 ), ei (t2 )}. The use of a traditional multivariate approach to map QTL for growth processes is limited in two aspects: individual expected means of different QTL genotypes at all points and all elements in the matrix  need to be estimated, resulting in substantial computational difficulties when the vector and matrix dimension is large; failure to incorporate the underlying biological principle of growth limits its practical applicability and interpretation. We have developed a new statistical framework for detecting QTL that affect growth trajectories (Wu et al., 2002a; Ma et al., 2002). This framework includes two tasks. Firstly, we need to model the time-dependent expected means of QTL genotype l using a growth equation, µl (t) = g(t;  m l ),

(2.4)

which is specified by a set of curve parameters  m l . Several mathematical models, such as logistic or S-shaped curves, have been proposed to describe nondescreasing growth trajectories. The fundamental biological aspects of mathematical modeling of growth curves have been founded by earlier mathematical biologists, e.g. Von Bertalanffy (1957), and recently revisited by West et al. (2001). The overall form of the growth curve of QTL genotype l is determined by the set of curve parameters contained in  m l . If different genotypes at a putative QTL have different combinations of these parameters, this implies that this QTL plays a role in governing the differentiation of growth trajectories. Thus, by testing for differences in  m l among different genotypes, we can determine whether there exists a specific QTL that confers an effect on growth curves. Secondly, we need to model the structure of the within-subject covariance matrix using the stationary AR(1) model, with covariance structure cov(t1 , t2 ) = σ 2 ρ |t2 −t1 | . The parameters that determine the covariance structure are denoted by  v = (ρ, σ 2 ). We have implemented the EM algorithm, originally proposed by Dempster et al. (1977), to estimate three groups of unknown parameters in a QTL-hunting model, that is, the population genetic parameters  p ) that specify QTL genotype frequencies in the population, curve parameters ( m l ), and covari( v ). These unknowns correspond to mixture proportions (denoted by πl ), componentance parameters ( specific parameters (denoted by ϕl ), and common parameters (denoted by η), respectively, in the mixture model described by (2.1). A detailed description of the EM algorithm was given in Wu et al. (2002a) and Ma et al. (2002). To deal with heteroscedastic residual variance, we embedded the transform-both-sides (TBS) approach (Carroll and Ruppert, 1984) into the growth-incorporated finite mixture model (Wu et al., 2004b). Both empirical analyses with real examples and computer simulations suggest that this TBS-based model can increase the precision of parameter estimation and computational efficiency. Furthermore, this model preserves original biological interpretation of the curve parameters although statistical analyses are based on transformed data.

A general model for the genetic study of longitudinal traits

423

3. M ODEL GENERALIZATION 3.1

Measurement schedule

We now generalize our QTL-hunting model to consider typical characteristics of actual longitudinal data: irregularly spaced measurement times not common to all subjects and nonstationary correlation structure. Let yi = [y(ti1 ), . . . , y(tiτi )] be the vector of τi measurements on subject i and let ti = (ti1 , . . . , tiτi ) be the corresponding vector of measurement times. The longitudinal measurements for subject i at different time points can be approximated by a multivariate normal distribution written as   1 1 −1 ′  i (yi − mli ) , exp − (yi − mli ) (3.1) f (yi ) =  i |1/2 2 (2π)τi /2 |

where the observation vector at different time points, yi = [y(ti1 ), . . . , y(tiτi )], the mean vector for QTL genotype l, mli = [µl (ti1 ), . . . , µl (tiτi )], and the (τi × τi ) within-subject covariance matrix,  i , are all subject-specific. 3.2

Modeling the covariance structure

In our previous work (Wu et al., 2004b), we implemented the TBS model to meet the variance stationarity assumption used in the AR(1) model (Wu et al., 2002a; Ma et al., 2002). N´u nez-Ant´on and Woodworth (1994) and N´u nez-Ant´on and Zimmerman (2000) proposed a transformation model for a nonstationary  i ) for subject i with τi observations at times correlation structure. In this model, the covariance matrix ( 0 < ti1 < ti2 < · · · < tiτi is expressed as ⎧ ⎨σ 2 ρ (tiλj2 −tiλj1 )/λ , λ = 0, σ (ti j1 , ti j2 ) (3.2) ⎩σ 2 ρ ln(ti j2 /ti j1 ) , λ = 0, where 1  j1  j2  τi , 0 < ρ < 1, and σ 2 , ρ, and λ, collectively denoted by  v , are unknown parameters to be fitted to the data. The parameter λ determines the behavior of the correlation between different measurements on the same subject (Zimmerman and N´u nez-Ant´on, 2001). The transformation approach for modeling the covariance structure in continuous time described by (3.2) corresponds to the AR(1) structure, when λ = 1. As with the Box–Cox transformation, the log transformation is obtained when λ is near zero.  i−1 ) of N´u nez-Ant´on and Woodworth (1994) used results in Barret (1979) to show that the inverse ( the covariance matrix constructed from (3.2) is tridiagonal, with ⎧

−1 λ λ )/λ −1 2(t λ −t λ ⎪  i−1 )11 = 1 − ρ 2(ti2 −ti1 )/λ , σ 2 (  i−1 )τi τi = 1 − ρ iτi i(τi −1) , σ 2 ( ⎪ ⎪ ⎪ ⎪

λ λ λ λ ⎪ −1 ⎪ ⎪  i−1 ) j ( j+1) = − 1 − ρ 2(ti( j+1) −ti j )/λ ρ 2(ti( j+1) −ti j )/λ , 1  j  τi − 1, σ 2 ( ⎪ ⎪ ⎨



−1 λ λ λ λ (3.3)  i−1 ) j j = 1 − ρ 2(ti j −ti( j−1) )/λ 1 − ρ 2(ti( j+1) −ti j) )/λ σ 2 ( ⎪ ⎪ ⎪ ⎪ ⎪ 2(t λ −t λ )/λ

⎪ 1 − ρ i( j+1) i( j−1) , 1 < j < τi , ⎪ ⎪ ⎪ ⎪ ⎩ 2 −1  i ) jl = 0, | j − l| > 1. σ (

A general form of the determinant can be derived by applying the Cholesky factorization to the covariance matrix (N´u nez-Ant´on and Woodworth, 1994), which is expressed as  i | = σ 2τi |

τi  j=2

1−ρ

2(tiλj −ti(λ j−1) )/λ



,

0 < ti( j−1) < t j , j = 2, 3, . . . , τi .

(3.4)

424

W. H OU ET AL .

Other options for modeling nonstationarity include the unstructured antedependence model, originally introduced by Gabriel (1962) and analyzed by Macchiavelli and Arnold (1994), Zimmerman and N´u nezAnt´on (1997), and Zimmerman et al. (1998). 3.3

The likelihood and estimation

Analogous to the treatment for (2.2), likelihood analysis based on irregularly spaced longitudinal observations combines the growth curve model (2.4) with the model (3.2) of the correlations among irregularly  p ,  m l ,  v ), the likelihood function, based again on the spaced repeated measurements. Letting  = ( assumption of multivariate normality for residual error sequences, is |y) = L(

n 

̟1 f 1 (yi ;  m 1 ,  v ) + · · · + ̟ L f L (yi ;  m L ,  v ) .

(3.5)

i=1

Suppose there is a QTL of major effect on growth curves which is segregating with two alternative alleles A and a in a natural population. The population frequencies of these two alleles are denoted by q and 1 − q, respectively, with the coefficient of Hardy–Weinberg disequilibrium (HWD) denoted as d (Lynch and Walsh, 1998). The population frequencies of the three QTL genotypes can be described in terms of the allele frequencies and HWD as ⎧ 2 ⎪ for QTL genotype A A, ⎨̟2 = q + d, ̟1 = 2q(1 − q) − 2d, for QTL genotype Aa, ⎪ ⎩ ̟0 = (1 − q)2 + d, for QTL genotype aa.

If two QTL are fitted to the data, we can estimate the allele frequencies and linkage disequilibrium (LD) between the QTL. Let q1 and 1 − q1 be the allele frequencies of A and a at QTL A and q2 and 1 − q2 be the allele frequencies of B and b at QTL B. Under the assumption of Hardy–Weinberg equilibrium, these two QTL, with the coefficient of LD denoted by D, form four gametes whose frequencies are (Lynch and Walsh, 1998) ⎧ p11 = q1 q2 + D, for gamete AB, ⎪ ⎪ ⎪ ⎨ p = q (1 − q ) − D, for gamete Ab, 10 1 2 (3.6) ⎪ p01 = (1 − q1 )q2 − D, for gamete a B, ⎪ ⎪ ⎩ p00 = (1 − q1 )(1 − q2 ) + D, for gamete ab.

Denote a two-QTL genotype by l1l2 , where l1 , l2 = 2, 1, 0 are the numbers of the capital allele at QTL 2 for A AB B, ̟ = 2 p p A and B, respectively. We express the genotype frequencies by ̟22 = p11 21 11 10 2 for A ABb, ̟20 = p10 for A Abb, ̟12 = 2 p11 p01 for Aa B B, ̟11 = 2( p11 p00 + p10 p01 ) for Aa Bb, 2 for aa B B, ̟ 2 ̟10 = 2 p10 p00 for Aabb, ̟02 = p01 01 = 2 p01 p00 for aa Bb, and ̟00 = p00 for aabb. Thus, we substitute QTL genotype frequencies in the mixture model (3.5) by gamete frequencies. The population genetic parameters are denoted by  p = (q, d) or (q1 , q2 , D) for the one- and twoQTL model, respectively. We develop an EM–simplex hybrid algorithm to obtain the maximum likelihood estimates (MLEs) of the population genetic parameters  p with a closed form of the EM algorithm (see Appendix) and the MLEs of the quantitative genetic parameters  m l and  v with the Matlab function fminsearch based on the modified simplex algorithm (Lagarius et al., 1998). The simplex algorithm, originally proposed by Nelder and Mead (1965), has been shown to be efficient for estimating the curve and matrix-structuring parameters (Zhao et al., 2004). In practical implementation, we integrate the simplex algorithm for quantitative genetic parameter estimates as an expanded M step to the EM algorithm designed to estimate population genetic parameters in each iterative loop.

A general model for the genetic study of longitudinal traits

425

The EM–simplex hybrid algorithm provides the point estimates of population genetic parameters, curve parameters, and matrix-structuring parameters. We derive the asymptotic variance–covariance matrix and evaluate the sampling errors of the estimates. The techniques for doing so involve calculation of the incomplete-data information matrix which is the negative second-order derivative of the incompletedata log-likelihood. The incomplete-data information can be calculated by extracting the information for the missing data from the information for the complete data (Louis, 1982). A different so-called supplemented EM algorithm or SEM algorithm was proposed by Meng and Rubin (1991) to estimate the asymptotic variance–covariance matrices, which can also be used for the calculations of the sampling  p ,  m l ,  v ). errors for the MLEs of the parameters ( We implemented an alternative algorithm for fitting the normal mixture model (3.5), the R function optim( ) (Venables and Ripley, 2002). This function uses several different algorithms, SANN, L-BFGS, CG, BFG, and Nelder–Mead, as default. The advantage of this function is that it directly provides estimates of standard errors for all parameter estimates. However, the EM–simplex hybrid algorithm is computationally more efficient. 3.4

Hypothesis tests

QTL-effect hypotheses. In contrast to traditional mapping approaches, our QTL-hunting model based on growth laws allows for the tests of a number of biologically relevant hypotheses (Wu et al., 2004a). These tests include a global test for the existence of significant QTL, a local test for the genetic effect on growth at a particular time point, a regional test for the overall effect of QTL on a particular period of growth, and an interaction test for the change of QTL expression across ages (Wu et al., 2004a). As an example, we formulate the hypothesis about the existence of a QTL affecting an overall growth curve as H0 : m 2 = m 1 = m 0 = m ,

(3.7)

i.e. the data can be fitted by a single logistic curve, whereas under the alternative, there are two different logistic curves. Under H0 , the log-likelihood ratio statistic, LR say, is asymptotically χ 2 -distributed with six degrees of freedom. We use an empirical approach for determining the critical threshold based on permutation tests, as advocated by Churchill and Doerge (1994). By repeatedly shuffling the relationships between marker genotypes and phenotypes, a series of the maximum LRs are calculated, from the distribution of which the critical threshold is determined. Population structure hypotheses. The occurrence of HWD and/or LD is related to population history and structure under the operation of various evolutionary forces, such as selection, mutation, and admixture (Lynch and Walsh, 1998). By testing these two parameters, our model can infer population structure and organization. For the one-QTL model, a test of whether the sampled population deviates from a Hardy–Weinberg equilibrium is formulated by H0 : d = 0 vs. H1 : d = 0. Similarly, if the two-QTL model is used, LD between two putative QTL can be tested by comparing the likelihood values under H0 : D = 0 and H1 : D = 0. When d = 0, the QTL genotype frequencies are determined solely by allele frequencies. When D = 0, the MLEs of gamete frequencies can be obtained with the constraint p11 p00 = p10 p01 . 4. A

WORKED EXAMPLE

4.1

Subjects

As part of our ongoing prospective longitudinal project on child development (Eyler et al., 1998a,b), we sampled a group of children (n = 145) from a human population in Florida who were born to normal, non-drug-using mothers. The children sampled were followed from birth to age 7 years. Interviewers and

426

W. H OU ET AL .

Fig. 1. Plots of head circumference vs. ages for each of the 145 children sampled from a natural population. The relationships between growth and age are displayed for (A) untransformed and (B) log-transformed data. Note that a few subjects have only a limited number of measurement points.

clinicians conducted a battery of psychological, neurological, and medical tests at nine age points: birth, 1 month, 6 months, 1 year, 18 months, 2 years, 3 years, 5 years, and 7 years. Growth data obtained at all nine age points included head circumference, weight, and length. However, the number and age points of measurements actually taken varied between subjects. As a case study for validating the model presented in this article, we used head circumference growth to test whether there exist specific QTL that trigger effects on growth trajectories. In order to adjust for gestations that were less than the typical 40 weeks, age in this study is defined as time since estimated conception, i.e. chronologic age plus actual weeks of gestation. To accurately reflect the time course of measures, we used the age in decimal years when the head circumference was measured. In 14 of the 1126 observations physically impossible changes in recorded head circumference from one time to the next (>5 cm) were considered to be in error and were eliminated.

4.2

Results

Figure 1 illustrates the plots of head circumference against ages from birth to age 7 years for all subjects. Compared to growth trajectories based on raw data (Figure 1A), growth trajectories based on the log transformation display smaller and more uniform variation throughout head development (Figure 1B). This difference suggests that the TBS-based model is preferable to a model that uses untransformed data. Statistical tests suggest that head circumference growth trajectories for all subjects, except for those with too few repeated measurements, can be well fitted by the Jenss and Bayley (1937) model, g(t) = α + βt − eγ −δt ,

(4.1)

where the growth parameters  m = (α, β, γ , δ) describe initial growth (α), velocity of growth (β), and rapid early childhood or acceleration (γ and δ). As evidence, we randomly chose four subjects from the study samples, in each of which the Jenss–Bayley model is a good fit to head circumference data (Figure 2). In this typical example, therefore, (4.1) is used to describe the mean value of QTL genotype l at time t as shown by (2.4).

A general model for the genetic study of longitudinal traits

427

Fig. 2. Four subjects randomly sampled from the set of subjects studied. The growth trajectories of head circumference can be fitted by the Jenss–Bayley model (4.1).

To maximize the likelihood, we used two different algorithms, the Matlab function fminsearch and the R function optim( ). We tested five different methods included in the R function optim( ). The SANN method could not converge. The L-BFGS-B, CG, and BFG methods converged but could not generate positive definite Hessian matrices. The Nelder–Mead method achieved the highest maximum likelihood among the converged methods and generated a positive definite Hessian matrix. Both the Matlab function fminsearch and the R function optim( ) have consistently detected a significant QTL with major effects on differentiation in growth trajectories of head circumference and also provided similar estimates of the parameters and of their standard errors (Table 1), but the Matlab function fminsearch used 20 min lesser than the R function optim( ) to complete parameter estimates for our example on a PC with 1.7 GHz CPU and 512 MB RAM. The LR test statistics calculated from the Matlab function fminsearch under the full model (there is a QTL) and the reduced model (there is no QTL, i.e.  m 2 =  m 1 =  m 0 =  m ) is 62.87, greater than the critical threshold value (27.6) at the 0.01 significance level. The critical value for claiming the existence of this QTL was determined on the basis of simulation studies. In this example, the empirical estimate of the critical value is obtained from 300 simulations. Let A and a denote the alleles that cause increased or decreased head circumference growth, respectively. The increasing allele is more common than the decreasing allele, with qˆ = 0.58 for allele A and 1 − qˆ = 0.42 for allele a. The coefficient of HWD between the two alleles was estimated as dˆ = 0.036. The hypothesis test indicated that the population from which our samples were drawn is at Hardy–Weinberg equilibrium because dˆ is not significantly different from zero based on the log-likelihood test. Given considerable differences among the estimates of the three QTL genotype frequencies, this QTL should display a fairly high heterozygosity in the population. We calculated the three genotype frequencies as 0.37 for the homozygote ( A A) comprising of the increased allele, 0.21 for the homozygote (aa) comprising of the decreased allele, and 0.42 for the heterozygote ( Aa). The parameters describing growth curves of different QTL genotypes were estimated with reasonable precision, with sampling errors of the MLEs less than one tenth of the estimated values (Table 1). However, the estimation precision varies between parameters. The Jenss–Bayley model contains two parts, one reflected by a linear regression described by α and β and the second part reflected by an exponential function described by γ and δ. In each part, the slopes, β or δ, have always larger sampling errors than the intercepts, α or γ . Of the three covariance parameters, the transformation parameter (λ) has poorer estimation precision than the correlation (ρ) and variance (σ 2 ).

428

W. H OU ET AL .

Table 1. MLEs of population and quantitative genetic parameters for a major QTL affecting growth trajectories of head circumference in 145 normal children based on two different algorithms, fminsearch and optim( ). The standard errors of the MLEs are given in parentheses Parameters Genotype A A The Matlab function fminserach Frequency 0.37 αˆ l 48.62 (0.33) βˆl 0.58 (0.04) γˆl 4.20 (0.07) δˆl 2.14 (0.10) σˆ 2 ρˆ λˆ LR The R function optim( ) αˆ l βˆl γˆl δˆl

σˆ 2 ρˆ λˆ

0.21 47.01 (0.31) 0.60 (0.04) 4.34 (0.08) 2.29 (0.11)

Genotype aa 0.42 46.26 (0.47) 0.52 (0.07) 3.90 (0.11) 1.74 (0.17)

0.0007 (0.000048) 0.0072 (0.0047) −1.12 (0.08) 62.87

48.58 (0.32) 0.59 (0.04) 4.22 (0.07) 2.16 (0.10)

σˆ 2 ρˆ λˆ LR Monte Carlo simulation αˆ l βˆl γˆl δˆl

Genotype Aa

46.88 (0.30) 0.62 (0.04) 4.38 (0.08) 2.36 (0.11)

46.43 (0.49) 0.50 (0.07) 3.89 (0.11) 1.71 (0.17)

0.0007 (0.00003) 0.0069 (0.0034) −1.11 (0.08) 62.87

49.02 (0.44) 0.60 (0.07) 4.20 (0.08) 2.10 (0.13)

47.02 (0.39) 0.60 (0.05) 4.29 (0.09) 2.29 (0.14)

46.00 (0.64) 0.50 (0.09) 3.91 (0.09) 1.71 (0.15)

0.0010 (0.0007) 0.00102 (0.0073) −1.01 (0.10)

The growth curves for the three genotypes at the detected QTL are shown in Figure 3. The different growth curves of the QTL genotypes show the dynamic change of QTL expression over time. It appears that the QTL displays an increased effect on head growth from birth to 1.5 years old and maintains a consistent large effect on head growth thereafter (Figure 3). The heterozygote comprising the increasing and decreasing QTL alleles show intermediate growth relative to the two homozygotes, suggesting that this QTL affects growth trajectories in an additive gene action mode (Figure 3). Approximate parallelism of the three curves of the QTL genotypes (Figure 3) implies that the major gene has little interaction with age. We performed a Monte Carlo simulation study to investigate the statistical properties of parameter estimates from our model. We mimicked the application described above by simulating three growth curves

A general model for the genetic study of longitudinal traits

429

Fig. 3. Three growth curves (foreground) each presenting one of the three groups of genotypes at the major QTL detected by our nonstationary model. The differentiation of growth curves shows the effect of the QTL on growth trajectories.

or QTL genotypes each following a different logistic equation. Our model was then used to estimate the simulated data and calculate the curve and matrix parameters. The same simulation scheme was repeated 400 times to estimate the true sampling errors of the MLEs of the parameters, which are given in Table 1. All the parameters including genotype frequencies, curve parameters, and matrix parameters are well estimated. The estimation precision based on this simulation study is broadly in agreement with that from the numerical analysis. 5. D ISCUSSION In our previous publications, a series of statistical models for characterizing specific QTL regulating growth processes have been developed by implementing growth laws derived from physiological principles (West et al., 2001) or built on the fitness of observational data (Von Bertalanffy, 1957). Although these models have been used successfully for gene detection from practical longitudinal data in forest trees (Wu et al., 2003a) and HIV-1 dynamics (Wu et al., 2004c), they are still limited unless some underlying assumptions are relaxed. These assumptions, including equally spaced time intervals, identical measurement schedule for all subjects, and within-subject stationary correlation structure, are often not met in practice, especially in human genetic research where subjects under study cannot be experimentally under full control. The goal of this article was to derive a more general model by relaxing each of these assumptions, so as to widen the applicability of QTL-detecting models. We used an example for child head circumference growth to validate the usefulness of our new model. A QTL that affects growth trajectories of child head circumference from birth to age 7 years has been successfully detected in a sample of normal children (n = 145) drawn from a population, representative of children delivered at our hospital. This QTL was detected to exert a consistent effect on head size growth after age 2 years. Samuelsen et al. (2004) also speculated genetic control over child head size growth in a Norwegian population. The standard errors of parameter estimates in our example were evaluated from the Fisher information matrix derived on the basis of the second derivative of the likelihood function. The results are broadly concordant with those from an alternative approach based on simulation studies that mimics the structure of the example used.

430

W. H OU ET AL .

Our new model is robust in that it can handle any pattern of longitudinal data, including missing values provided that these are missing at random. Its additional value lies in its ability to describe nonstationary covariance structures while avoiding the overparameterization of the unstructured multivariate normal model. In this study, our derivation takes advantage of stationary variance on the log-scale through implementing the TBS model. It is possible that nonstationarity in variance still exists after transformation, in which case a more general model is needed which allows for both the within-subject residual variance and correlation to be time dependent (N´u nez-Ant´on and Zimmerman, 2000; Zimmerman and N´u nez-Ant´on, 2001). In some earlier work, several authors have explored the feasibility of modeling nonstationary variance (Wolfinger, 1996; N´u nez-Ant´on, 1997; Zimmerman et al., 1998). Our model is developed to detect the QTL responsible for growth trajectories. The incorporation of molecular markers into this model can discover the position of a QTL on the genome. An additional task for such genetic mapping is to derive the conditional probabilities of unknown QTL genotypes given known marker genotypes. For experimental crosses or well structured pedigrees, these conditional probabilities that correspond to mixture proportions in the finite mixture model described by (2.1) are expressed as a function of the recombination fraction between the markers and putative QTL. If a mapping population is randomly drawn from a natural population, as used in this study, we can implement a closed-form solution, as derived in Lou et al. (2003), to estimate the allelic frequencies of QTL alleles and the LD between the markers and QTL. The estimation of the recombination fraction for a cross or pedigree and the LD for a natural population is a first step toward position cloning QTL affecting a complex trait. Given marked evidence for the existence of QTL for growth in child head circumference, it would be worthwhile genotyping molecular markers and, ultimately, identifying the underlying QTL hidden in this population. We intend to implement the model proposed in this article and its extensions in our web-based software, FunMap, for functional data analysis of genetic factors that determine growth trajectories (Ma et al., 2004).

ACKNOWLEDGMENTS We thank Professor Peter Diggle and two anonymous referees for their constructive comments on the manuscript. This work is partially supported by National Institutes of Health (No. DA05854) to Marylou Behnke and Fonda Davis Eyler and an Outstanding Young Investigators Award (No. 30128017) of the National Natural Science Foundation of China and the University of Florida Research Opportunity Fund (No. 02050259) to Rongling Wu. The publication of this paper is approved as journal series R-10068 by the Florida Agricultural Experiment Station.

APPENDIX In what follows, we give a closed-form solution for the EM algorithm to obtain the MLEs of population genetic parameters  p . The QTL genotype frequencies (̟l ) in the likelihood function, (3.5), are viewed as the prior probabilities. In the E step, calculate the posterior probability (li ) of subject i to carry a particular QTL genotype l and the ratio (li ) of the posterior to prior probability for the same subject, which are respectively expressed as πl fl (yi ;  m l ,  v ) , li = 2 l=0 πl f l (yi ;  m l ,  v ) li = 2

fl (yi ; m l , v )

l=0 πl f l (yi ;  m l ,  v )

=

li . πl

(A.1) (A.2)

A general model for the genetic study of longitudinal traits

431

In the M step, obtain the estimates of  p by solving the log-likelihood equations. For the one-QTL model, we have n (1i − 0i ) , (A.3) qˆ = n i=1 i=1 (21i − 2i − 0i ) √ −b ± b2 − 4cn dˆ = , (A.4) 2n where 2

b = (2q − 2q + 1) c = q(1 − q)

n  i=1

n  i=1

n  1i + (1 − 2q) [(1 − q)2i − q0i ], i=1

[q(1 − q)1i − (1 − q)2 2i − q 2 0i ].

In each M step, the estimate of d with two possible solutions based on (A.4) should be taken from its space, max[−q 2 , −(1 − q)2 ]  d  2q(1 − q). For the two-QTL model, we obtain the estimates of the gametic frequencies using n (222i + 21i + 12i + φ11i ) , pˆ 11 = i=1 2n n (220i + 21i + 10i + (1 − φ)11i ) pˆ 10 = i=1 , 2n n (202i + 12i + 01i + (1 − φ)11i ) pˆ 01 = i=1 , 2n n (200i + 10i + 01i + φ11i ) pˆ 00 = i=1 . 2n

(A.5) (A.6) (A.7) (A.8)

Iterations are repeated between (A.1)–(A.4) for the one-QTL model and between (A.1), (A.2), and (A.5)–(A.8) for the two-QTL model until convergence. According to the invariance property of the maximum likelihood method, the MLEs of the allele frequencies of the two QTL and their LD can be obtained from the MLEs of gamete frequencies by solving (3.6). R EFERENCES ATCHLEY, W. R. AND Z HU , J. (1997). Developmental quantitative genetics, conditional epigenetic variability and growth in mice. Genetics 147, 765–776. BARRET, W. (1979). A theorem on inverse of tridiagonal matrices. Linear Algebra and Its Applications 27, 211–217. B OX , G. E. P. AND C OX , D. R. (1964). An analysis of transformations. Journal of the Royal Statistical Society: Series B 26, 211–252. C ARROLL , R. J. AND RUPPERT, D. (1984). Power-transformations when fitting theoretical models to data. Journal of the American Statistical Association 79, 321–328. C HURCHILL , G. A. AND D OERGE , R. W. (1994). Empirical threshold values for quantitative trait mapping. Genetics 138, 963–971.

432

W. H OU ET AL .

D EMPSTER , A. P., L AIRD , N. M. AND RUBIN , D. B. (1977). Maximum likelihood from incomplete data via EM algorithm. Journal of the Royal Statistical Society: Series B 39, 1–38. D IGGLE , P. J., H EAGERTY, P., L IANG , K. Y. UK: Oxford University Press.

AND

Z EGER , S. L. (2002). Analysis of Longitudinal Data. Oxford,

E YLER , F. D., B EHNKE , M., C ONLON , M., W OODS , N. S. AND W OBIE , K. (1998a). Birth outcome from a prospective matched study of prenatal crack/cocaine use. Part I: Interactive and dose effects on health and growth. Pediatrics 101, 229–237. E YLER , F. D., B EHNKE , M., C ONLON , M., W OODS , N. S. AND W OBIE , K. (1998b). Birth outcome from a prospective, matched study of prenatal crack/cocaine use. Part II: Interactive and dose effects on neurobehavioral assessment. Pediatrics 101, 237–241. G ABRIEL , K. R. (1962). Antedependence analysis of an ordered set of variables. Annals of Mathematical Statistics 33, 201–212. G ONG , Y., WANG , Z. H., Z HU , Y. AND W U , R. L. (2004). A statistical model for high-resolution mapping of quantitative trait loci affecting pharmacodynamic processes. The Pharmacogenomics Journal 4, 315–321. J ENSS , R. M. AND BAYLEY, N. (1937). A mathematical method for studying the growth of a child. Human Biology 9, 556–563. L AGARIUS , J. C., R EEDS , J. A., W RIGHT, M. H. AND W RIGHT, P. E. (1998). Convergence properties of the Neler–Mead simplex method in low dimensions. SIAM Journal of Optimization 9, 112–147. L ANDER , E. S. AND B OTSTEIN , D. (1989). Mapping Mendelian factors underlying quantitative traits using RFLP linkage maps. Genetics 121, 185–199. L OU , X.-Y., C ASELLA , G., L ITTELL , R. C., YANG , M. K. C. AND W U , R. L. (2003). A haplotype-based algorithm for multilocus linkage disequilibrium mapping of quantitative trait loci with epistasis in natural populations. Genetics 163, 1533–1548. L OUIS , T. A. (1982). Finding the observed information matrix when using the EM algorithm. Journal of the Royal Statistical Society, Series B 44, 226–233. LYNCH , M.

AND

WALSH , B. (1998). Genetics and Analysis of Quantitative Traits. Sunderland, MA: Sinauer.

M A , C.-X., C ASELLA , G. AND W U , R. L. (2002). Functional mapping of quantitative trait loci underlying the character process: a theoretical framework. Genetics 161, 1751–1762. M A , C.-X., W U , R. L. 20, 1808–1811.

AND

C ASELLA , G. (2004). FunMap: functional mapping of complex traits. Bioinformatics

M ACCHIAVELLI , R. E. AND A RNOLD , S. F. (1994). Variable order antedependence models. Commuications in Statistics: Theory and Methods 23, 2683–2699. M ENG , X.-L. AND RUBIN , D. B. (1991). Using EM to obtain asymptotic variance–covariance matrices: the SEM algorithm. Journal of the American Statistical Association 86, 899–909. N ELDER , J. A. AND M EAD , R. (1965). A simplex method for function minimization. Computer Journal 7, 308–313. ´ , V. (1997). Longitudinal data analysis: non-stationary error structures and antedependence models. N EZ -A NT ON N U´  Applied Stochastic Models and Data Analysis 13, 279–287. ´ , V. AND W OODWORTH , G. G. (1994). Analysis of longitudinal data with unequally spaced obserN U´  N EZ -A NT ON vations and time-dependent correlated errors. Biometrics 50, 445–456. ´ , V. N EZ -A NT ON N U´  699–705.

AND

Z IMMERMAN , D. L. (2000). Modeling nonstationary longitudinal data. Biometrics 56,

S AMUELSEN , S. O., S TENE , L. C. AND BAKKETEIG , L. S. (2004). Association of head circumference at birth among sibling pairs. Paediatric and Perinatal Epidemiology 18, 26–34.

A general model for the genetic study of longitudinal traits V ENABLES , W. N.

AND

433

R IPLEY, B. D. (2002). Modern Applied Statistics with S, 4th edition. New York: Springer.

VON B ERTALANFFY, L. (1957). Quantitative laws in metabolism and growth. Quarterly Review of Biology 32, 217–231. WANG , Z. H. AND W U , R. L. (2004). A statistical model for high-resolution mapping of quantitative trait loci determining human HIV-1 dynamics. Statistics in Medicine 23, 3033–3051. W EST, G. B., B ROWN , J. H. 628–631.

AND

E NQUIST, B. J. (2001). A general model for ontogenetic growth. Nature 413,

W OLFINGER , R. D. (1996). Heterogeneous variance–covariance structures for repeated measures. Journal of Agricultural, Biological, and Environmental Statistics 1, 205–230. W U , R. L., M A , C.-X., C HANG , M., L ITTELL , R. C., W U , S. S., Y IN , T. M., H UANG , M. R., WANG , M. X. AND C ASELLA , G. (2002a). A logistic mixture model for characterizing genetic determinants causing differentiation in growth trajectories. Genetical Research 19, 235–245. W U , R. L., M A , C.-X., L IN , M. AND C ASELLA , G. (2004a). A general framework for analyzing the genetic architecture of developmental characteristics. Genetics 166, 1541–1551. W U , R. L., M A , C.-X., L IN , M., WANG , Z. H. AND C ASELLA , G. (2004b). Functional mapping of quantitative trait loci underlying growth trajectories using a transform-both-sides logistic model. Biometrics 60, 729–738. W U , R. L., M A , C.-X., L ITTELL , R. C. AND C ASELLA , G. (2002b). A statistical model for the genetic origin of allometric scaling laws in biology. Journal of Theoretical Biology 217, 275–287. W U , R. L., M A , C.-X., YANG , M. C. K., C HANG , M., S ANTRA , U., W U , S. S., H UANG , M., WANG , M. C ASELLA , G. (2003a). Quantitative trait loci for growth in Populus. Genetical Research 81, 51–64.

AND

W U , R. L., M A , C.-X., Z HAO , W. AND C ASELLA , G. (2003b). Functional mapping of quantitative trait loci underlying growth rates: a parametric model. Physiological Genomics 14, 241–249. Z HAO , W., W U , R. L., M A , C.-X. traits. Genetics 167, 2133–2137.

AND

C ASELLA , G. (2004). A fast algorithm for functional mapping of complex

´ , V. (1997). Structured antedependence models for longitudinal data. N EZ -A NT ON Z IMMERMAN , D. L. AND N U´  In Gregoire, G., Brillinger, D. R., Diggle, P. J., Russek-Cohen, E., Warren, W. G., and Wolfinger R. D. (eds), Modelling Longitudinal and Spatially Correlated Data: Methods, Applications, and Future Directions. Lecture Notes Series in Statistics No. 122. New York: Springer, pp. 63–76. ´ , V. (2001). Parametric modeling of growth curve data: an overview (with Z IMMERMAN , D. L. AND N U´  N EZ -A NT ON discussion). Test 10, 1–73. ´ , V. AND E L BARMI , H. (1998). Computational aspects of likelihood-based N EZ -A NT ON Z IMMERMAN , D. L., N U´  estimation of first-order antedependence models. Journal of Statistical Computation and Simulation 60, 67–84.

[Received March 8, 2004; first revision June 9, 2004; second revision October 14, 2004; third revision November 18, 2004; fourth revision December 20, 2004; accepted for publication January 19, 2005]