A Continuous Labour Supply Model in Microsimulation: A Life-Cycle Modelling Approach with Heterogeneity and Uncertainty Extension

This paper advances a structural inter-temporal model of labour supply that is able to simulate the dynamics of labour supply in a continuous setting and addresses two main drawbacks of most existing models. The first limitation is the inability to incorporate individual heterogeneity as every agent is sharing the same parameters of the utility function. The second one is the strong assumption that individuals make decisions in a world of perfect certainty. Essentially, this paper offers an extension of marginal-utility-of-wealth-constant labour supply functions known as “Frisch functions” under certainty and uncertainty with homogenous and heterogeneous preferences. The lifetime models based on the fixed effect vector decomposition yield the most stable simulation results, under both certain and uncertain future wage assumptions. Due to its improved accuracy and stability, this lifetime labour supply model is particularly suitable for enhancing the performance of the life cycle simulation models, thus providing a better reference for policymaking.


Introduction
The empirical literature on structural labour supply models has gained an increasing interest over the past decades, as these models became important tools for policy makers to examine potential reform options. The continuous approach in the tradition [1] has been complemented by an approach based on a discrete choice specification, mainly inspired by [2,3]. The arbitrary discretisation of the working hours, however, can lead to a loss of information in the estimation, which can produce misleading results with inappropriate groupings. Hybrid approaches also exist, for example [4], where the male decision is discrete while the female decision is continuous. This paper examines the uses of continuous labour supply models in the context of forward simulations and explores the potential of incorporating better controls for heterogeneities in a microsimulation context.
The original interest in life-cycle labour supply was motivated by the need to investigate various dimensions of labour supply, such as the determinants of the shape of the life-cycle hours profile, the labour supply response to the aggregate wage, the changes and the source of the idiosyncratic year-to-year changes in labour supply. The existing literature, however, manages to shed little light on the original questions. It focuses mainly on one aspect of the inter-temporal hours variation, namely the labour supply response to the wage growth along a known life-cycle trajectory, whilst ignoring other aspects. One ignored aspect is the labour supply response to wage changes under uncertainty, meaning wage changes that determine individuals to revise their expectations of their future wages [5].
The life-cycle framework is proposed as an explanation for all components of the individual labour supply [5]. The life-cycle model can be used to explain the aggregate year-to-year movements in labour supply ''time effects'' [6], the systematic age effects in hours of work (''age effects'') and the differences across people with respect to their hours of work over the life-cycle ''person-specific effects''. With the integration the labour market behaviours, the life-cycle framework can also shed light on behaviours that span multiple phases in a person's life, such as consumption smoothing and borrowing [7,8].
In the context of estimating and simulating life-cycle labour supply, the choice of the labour supply elasticity to be simulated depends on the goal. If the goal is to compare the impact of wage variations across consumers on labour supply, the variation in the entire wage profile must be examined. Because the variation of the wage profile affects the value of the marginal utility of wealth, the Frisch elasticity cannot be used. Estimating the full impact on wages requires estimating the effect of shifts of the wage profile on hours of work besides the estimation of the inter-temporal elasticity. The estimation of the full impact of wage changes, both evolutionary and parametric, are of core importance for policy evaluation. Assuming that tax and benefit reforms represent unanticipated shifts in net real wages today and in the future, the elasticity measuring the cumulated response to evolutionary changes and parametric shifts in the life-cycle wage profile is the most appropriate for describing the response to these reforms [9].
Given the complex nature of the human economic behaviours, it is unsurprising to observe a large degree of population heterogeneity. One method to incorporate individual heterogeneity in the model is to estimate individual specific effect, which can be estimated either as fixed or random effects. The choice between fixed and random effects relies mainly on whether the individual specific effects can be assumed independent of the explanatory variables included in the model. Assuming that the individual component is correlated with the explanatory variables triggers many problems in a simulation. The estimated coefficients cannot be used to generate a conditional prediction of individual earnings without specifying the joint process determining the individual specific effects and the explanatory variables [10,11]. The impracticality of this option, together with the fact that the fixed-effect specification cannot accommodate covariates that are constant over time, constrained most studies to maintain the assumption of a zero correlation between the individual specific effects and the other covariates, a rather strong and improbable assumption. The main drawback of error components models is they provide less stable simulations due to the stochastic components, which affect the reproducibility of the results.
Another way to incorporate the heterogeneity effects is to use random coefficient models. There are discussions on implementing the random coefficient models in discrete models, usually in the form of mixed logit, e.g., [12,13], but the discussions on the continuous model are limited. Provided that heterogeneity is present in empirical models of labour supply, the application of random coefficients models is necessary to avoid biased estimates. The main drawback of these models, however, is their high computational cost. Given this limitation, existing studies suggest that if heterogeneity is non-existent or the bias is insignificant, the standard fixed coefficient models represent the optimal choice [14]. This paper proposes a structural inter-temporal model of labour supply, which estimates and predicts the dynamics of labour supply in a continuous setting. It aims to capture the individual heterogeneity to a larger extent than the existing labour supply models while maintaining the consistency with the lifetime economic theory. The model is estimated using both a transformed fixed effect specification that circumvents the standard problems mentioned above, and a random coefficient specification. Additionally, the model incorporates uncertainties regarding future wages to further explore heterogeneity.

A Life-Cycle Model of Labour Supply
Our model follows the theoretical model introduced by [15,16] and [17], and the unifying labour supply framework introduced by [9]. The model aims to estimate the effect of evolutionary wage changes assuming no parametric shift in the wage profiles. The labour supply responses are estimated and simulated under two scenarios: (i) people make decisions in a world of perfect certainty; (ii) people are uncertain about their future wages.

A Life-Cycle Model of Labour Supply under Certainty
Under the assumption of certainty, the effect of evolutionary wage changes on hours of work represents the conventional ''marginal-utility-of-wealth constant'' inter-temporal elasticity of labour supply obtained from the Frisch labour supply equations and the Euler condition. The marginal-utility-of-wealth parameter serves as a sufficient statistic that captures the information from the other periods needed to solve the maximization process in the current period [9].
The inter-temporal elasticity estimated from the Frisch specifications is relevant for predicting the individual labour supply into the future assuming a steady state, mainly due to the presence of the marginal utility of wealth, which is individualized, constant over time, and accounts for the worker's future plans [18].
Following [16,19], the theory underlying our model of lifetime hours of work is an extension of permanent income theory [20] to a situation where the relative price of consumption and leisure varies over the life-cycle. The permanent income hypothesis can be extended to lifetime labour supply to assume that individuals/ households look into the future when deciding the current number of hours supplied on the labour market. This theory allows the distinction between a consumer's dynamic behaviour and the factors determining the differences in hours worked between consumers. This separation leads to a manageable empirical model that accommodates the differentiation between labour supply responses to evolutionary changes and those to parametric changes in the wage profiles.
An economic model of labour supply under certainty with homogenous preferences and heterogeneous individual effects. We start with an economic model of life-cycle labour supply decisions assuming that workers make decisions regarding future income in an environment of perfect certainty. Workers are assumed to choose consumption and hours of work at each age to maximize a lifetime preference function, strongly separable over time, subject to wealth constraints. The model described below is designed for single decision-makers, but an extension to joint decision makers is straightforward.
Men's labour supply behaviour is assumed independent, while women's labour supply is conditioned on other household incomes besides their own earnings. Assuming that consumer i at age t has a utility given by the concave function U t (C t ,L t ,X t ), where C t is consumption at age t, L t is the number of hours of leisure at age t and X t is a vector of ''taste shifters'' at age t, X t can include both observed and unobserved variables. Due to the assumption of separating utility, the lifetime preference function is the sum of discounted future utilities at moment t~a (the beginning of the active life), where t is age and a the age of entering the labour market: The lifetime (active life) is assumed to consist of T{az1 periods, where T is retirement age. r is rate of time preference used for discounting the value of future utility. Formally, the consumer has to choose C t and L t at each age to maximize their lifetime preference equation (1) subject to a lifetime wealth constraint.
A 0 represents the assets at the start of the active life, H t hours worked at age t, W t the exogenous wage rate at age t, and R t the discount rate. In period t, the consumer can borrow/lend at a constant interest rate r t . To create the Frisch labour supply functions, we assume the contemporaneous individual utility function at age t is: G is a monotonically increasing function of C it , s is a timeinvariant preference parameter common across consumers. Y it is an age-specific ''tastes'' parameter which depends on the characteristics which influence the utility at age t via where v Ã it is the contribution of unobserved characteristics and y Ã is a vector of preference parameters. The preference parameter includes the participation decision. Assuming an interior optimum, the implied Frisch labour supply function or marginal utility of wealth constant labour supply function is obtained from maximizing the utility in period t subject to the lifetime wealth constraint.
The first order conditions of the Lagrange function result in the Frisch or ''l constant'' consumption and labour supply functions. For hours worked at age t, this implies: a~1= 1{s ð Þ, y~ay Ã , v it~a v Ã it , and b~a r{r ð Þ are assumed constant across consumers and time. Equation (2) is the Frisch or marginal utility of wealth constant labour hours of work function if the individual chooses to work. It decomposes the labour supply decision regarding hours worked into time t, personal characteristics X it , the wage rate W it and the l component, which is the sufficient statistic summarizing the information for each consumer from the other periods. The optimal value of l is obtained by substituting the l constant consumption and labour supply functions into the budget constraint. l is a function of initial assets, lifetime wages, interest rates, rates of time preference and tastes. l is the correspondent statistic to the permanent income from the permanent income theory in [20], thus a permanent component, which together with current wage determines the consumer's current consumption and labour supply. l also incorporates historic and future information on lifetime wages and assets that are relevant for current consumption and labour supply choices. The conclusion is that, assuming certainty, the ''l constant'' consumption and labour supply functions fully characterize a consumer's dynamic behaviour [9,15,18].
As shown by [9], in a world of perfect certainty, l can be captured as an individual specific effect which is constant over time, thus the changes in individual wages have no impact on l. Following [9,18], we approximate l by: D i is a vector of observed time-invariant or rarely changing characteristics, Q Ã , y Ã t , h Ã are parameters assumed constant across consumers, and a Ã i is an error term. This specification imposes strong restrictions, as it assumes each worker knows the number of years he will work T i , and total lifetime income. It also incorporates the effect of interest rates and rates of time preference in the intercepts and the other parameters. The time invariant characteristics D can therefore be decomposed into an observed part D 1 and an unobserved part, D 2 . The observed part can be approximated by a linear form of D i1 , whereas the unobserved part, assumed to follow a normal distribution, is included in the error term. Thus, the individual unit effect can be estimated by a linear function of time-invariant characteristics and a normally distributed error that captures the unexplained individual effect. Therefore, we can rewrite equation (2) as: The mechanism for predicting lifetime wage profiles is approximated by: L i is the individual-specific effect, P it is a vector of personal and professional career characteristics, p o,k is a vector of coefficients (time and individual invariant), k is the number of personal and professional career characteristics and e it the error term.
An expansion of the economic labour supply model under certainty with heterogeneous preferences. Next, the model is extended by allowing for increased individual heterogeneity in labour supply preferences. The preference parameter s*N 0,C ð Þ is individual-specific and time-invariant. a i , b i , a ln l i { ln s i ð Þare assumed to be individual-specific and follow normal distributions across individuals. The lifetime labour supply function becomes: where a ln l i { ln s i ð Þf 0 zz 0,i is the individual effect (timeinvariant), and b i~f1 zz 1,i and a i~f2 zz 2,i are the random coefficients of age and wage. f is the mean intercept of hours of work, b the mean slope of age, a the mean slope of ln(wage) and z k,i , k~0, 1, 2 the deviation of individual intercepts and slopes from the mean values of f 0 , f 1 , f 2 : y is a vector of fixed coefficients, individual and time invariant. We can also allow for increased heterogeneities in the wage estimations using the similar random coefficient approach.
An expansion of the economic labour supply model under uncertainty. To further expand the model, we assume that individuals act in a world of uncertainty, where they adjust their expectations every period based on current and past information. The uncertainty can be accounted for via a forecasting error which incorporates past and future differences between realised values and their expectations.
Agents are assumed to form an expectation of wage adaptively based on past information. Due to the imperfection of forecasting, the wealth constraint needs to be updated every period. The updated wealth constraint in period k, WC k , can be expressed as the sum between assets in period 0, the discounted sum of total earnings between period 0 and period k, and the sum of expected total earnings, with discount factor d t between period k and retirement (period T): Alternatively, it can be expressed as the sum between assets in period 0, the expected discounted total earnings over the active life formed before entering the labour market and an adjustment error. The adjustment error equals the sum between the difference between realised earnings and their expectations made before entering the labour market over the period 0 to k (part A), and the expectations adjustment of the total earnings made in period t+1 to T (part B).
WC 0 z k I t reflects the information available to an agent in period t (e.g., personal and labour characteristics), and k is the error adjustment term. Current expectations of lifetime wealth reflect past expectations and an error adjustment term that could either lower or raise the total constraint. To further simplify the equation, the interest rate is assumed certain. The error adjustment term is expressed as: Þ is a vector of all wage information until period t, v t H ð Þ is a vector of working hours information until period t, and v t X ð Þ is a vector of personal characteristics until period t. As working hours are modelled using the wage rate and personal characteristics, the new adjusted lifetime wealth constraint can be proxyed by: p is a scale parameter in the approximation, X it is a vector of personal characteristics which affect the form of the current expectation, and B is the vector of coefficients of personal characteristics. Combining the anticipated paths for wages and income with the approximated empirical Lagrange function modifies the lifetime constraint. Therefore, l becomes: Å is a vector of time-variant characteristics, m is an adjustment ratio, which shows the marginal change of working hours caused by the error term LH i =LA 0 k . B explains the role of personal characteristics in the expectation. It is not possible, however, to distinguish B from the personal characteristics coefficient vector y in equation (3).

Empirical Implementation
This section introduces the econometric and the simulation techniques for estimating and simulating lifetime labour supply, both under certainty and uncertainty. The model is structured in three parts: a selection model, a model for lifetime wage profiles and a model for lifetime labour supply profiles.

Data
We use the German Socio Economic Panel (GSOEP) data from 1996 to 2003 for estimation and 2004-2005 data for the simulation validation exercise. GSOEP is a nationally representative dataset of individuals living in Germany [21]. We consider only adult respondents below the retirement age (65). Table 1 presents basic employment statistics.
The dependent variable is the logarithmic value of actual working hours per week. The wage rate in the estimation and the simulation is calculated using actual working hours. ''Participation'' is defined as employment in the formal labour market, except for vocational training, zero working hours, military service and community service, which are all modelled as ''nonparticipation'' in the selection model.

The Empirical Model of Lifetime Labour Supply under Certainty
This subsection develops the econometric model for the labour supply under certainty. The empirical model follows the structure of the economic model: the first model assumes homogenous preference parameters, and second one allows for a higher degree of heterogeneity amongst individuals.

An Econometric Model of Labour Supply under Certainty
with Homogenous Preferences. For estimating a lifetime labour supply models which incorporates a lower degree of heterogeneity, we proposes a model similar to [22,23]. The model extends the standard fixed effects specification to an estimation procedure which enables an efficient estimation of time-invariant and rarely changing variables by applying a ''fixed effects vector decomposition''.
The lifetime labour supply model follows the structure of the economic model in (2): y it is the natural logarithm of hours worked per week, f is the intercept of the base unit, X is a vector of time varying variables (natural logarithm of wage, age, age squared, children, health dummies, household type dummies, cumulated experience until last year (full time, part time, unemployment), other household income (only from women), sector dummies), D are time invariant variables (education, education squared, cohort dummies), u i are the unobserved individual specific effects, v it is an iid error term, and and % are parameters common to all individuals and constant over time.
The lifetime labour supply model is estimated using a threestage procedure. To start, the model (4) is estimated using the standard fixed effects estimator and it is possible to extract the individual fixed effects or unit effects as: The estimated individual specific effectû u i captures the unobserved individual-specific effect, the observed individualspecific effects D, the individual means of the residuals v v i and the individual means of the time-varying variables. In the second step, the estimated individual specific effect is regressed against the observed time-invariant characteristics and the rarely changing variables to obtain the unexplained part of the individual-specific effects. The individual specific effects are decomposed as follows: % c z ci is the explained part. h i is the residual from equation (5) and captures the unexplained part of the individual-specific effect: In the third stage, the individual effect from model (4) is substituted with the unexplained part of the decomposed individual fixed effect vector obtained in the previous stage, resulting in an error that is no longer correlated with the time varying covariates included in the model. Therefore, model (6) can be estimated consistently by pooled OLS: The Monte Carlo simulations conducted by [22] reveal the circumstances under which the FEVD is inferior to the pooled OLS, random effects (RE) and fixed effects (FE). OLS is more appropriate when there are no individual effects, RE when the individual effects are uncorrelated with the other explanatory variables, and FE when the assumptions of the RE are violated and the within-variance of the variables of interest is sufficiently large compared with the between-variance. If this condition does not occur, the vector decomposition technique has better finite sample properties in estimating models that have time-invariant or rarely changing variables correlated with individual-specific effects.
The lifetime wage process is estimated based on model (7) using the same procedure: w it is the natural logarithm of gross hourly wage is, X it is a vector of time-varying variables, and D i a vector of time invariant characteristics. j 0,i is the individual specific effect and e it is the error term. The time-varying covariates in the wage equations are age, age squared, children, health dummies, household type dummies, cumulated unemployment experience until last year, and sector dummies. The time-invariant variables are education, education squared and cohort dummies.
When analysing the wage and the labour supply processes using panel data, the natural question that arises is whether the nonresponse or the missing observations are endogenously determined. One source of sample selectivity is the unbalanced nature of the panel. If the panel attrition is endogenous to the wage and the labour supply processes, sample selection would be informative for wage and hours, and therefore the estimates from the wage and hours equations would be biased. The panel attrition bias is disregarded from the analysis because previous studies using the GSOEP have indicated that the sample selection bias is not significant [24,25]. The dataset is popular for analysing labour market issues. Another sample selectivity bias comes from the fact that wages and working hours are observed only for the individuals in the labour market and the selection bias is determined by the differences between workers and non-workers. A selection bias arises when some component of the participation decision is relevant to the wage and hours processes. Disregarding these relationships, the estimates of wages and hours of work from the subsample of working individuals will be biased.
If the relationships between the participation decision and the wage and hours processes occur through the observables, the selection bias can be controlled by introducing the appropriate conditioning variables in the wage and hours equations. If the relationships between the participation decision and the wage and hours processes occur through the unobservable, meaning that the unobserved characteristics affecting the participation decision are correlated with the unobservable from the wages and hours equations, simply controlling for the observables is not enough to obtain unbiased estimates. If the observables are correlated with the unobservable, in order to get unbiased estimates, the wage and hours equations should include an estimate for the unobservable [26].
The selection model is defined using the latent variable model: The selection indicator s it defines the observed employment status y it : & For fixed u i the probability of observing y it~1 is given by: ð Þ is the cumulative distribution function and. v it is assumed to be normally distributed. With i.i.d. error terms, the log likelihood function for a fixed effect probit model is given by: For a fixed T and N??, the maximum likelihood is inconsistent because the number of unknown parameters grows with the number of individuals in the sample. This is the so called ''incidental parameter problem'' which impedes the fixed effects probit model to be estimated consistently for a fixed number of periods. To circumvent this problem and to estimate the selection model accounting for the unobserved heterogeneity, the model decomposes the unobserved effect u i , following the approach introduced by [27]. The assumption is that the unobserved effect can be modelled as: This equation assumes that the correlation between u i and x it acts only through the time averages of the exogenous variables, whereas a i represents the remaining part of the unobserved effect which is independent of the time-varying variables. Equation (8) can be substituted into the selection model as follows: To summarize, the selection model includes time-invariant characteristics (education, education squared, cohort dummies, nationality dummies), time-varying variables (age, age squared, interaction between education and age, cumulated experience (full time, part-time, unemployment), health dummies, household type dummies, children and other household income (for women only)) and the means for the time-invariant variables (except for age).
To test for selection bias in both hours and wage equations for men and women, the study applied the fixed effect specification. For people participating in the labour market, the inverse Mills ratios were computed and then introduced into the initial models for hours and wages. The final step involved estimating the augmented models by applying the FE specification and testing the coefficient of the inverse Mills ratios. To correct for the sample selection in the wage and hours equation, the inverse Mills ratios for each year were included as time varying variables in the FEVD estimation of the wage and hours models.
For simulating both labour supply and wage processes, the main requirement imposed on the estimation method is to provide consistent and unbiased estimates. Allowing for unobserved heterogeneity and for a correlation between the unobserved individual effects and the other explanatory variables requires the use of FE estimation. This estimation technique, however, is not ideal in a simulation context given that FE cannot be simulated. An alternative is the FEVD technique, which circumvents the problems of a standard FE model (i.e. assumes a correlation structure between the unobserved individual effects and the other explanatory variables) by decomposing the individual effect and estimating the last stage as a pooled OLS.

An Econometric Model of Labour Supply under Certainty
with Heterogeneous Preferences. For the estimation of the lifetime labour supply model which incorporates a high degree of individual heterogeneity, the study proposes a mixed fixed and random coefficient model similar to [28]. The fixed coefficients are considered constant across consumers and time, and the random coefficients are unique for each individual, but constant over time.
The assumption is that the labour supply model can be approximated by: where J it is 1| KzM ð Þ vector of explanatory variables, q is the corresponding vector of coefficients, z 0,i is the individual specific effect and v it is the error term. The composite error term z 0,i zv it À Á is correlated over time. Serially correlated errors imply then the OLS or maximum likelihood standard errors are no longer valid. This dependence can be taken into account either by using the sandwich estimator for the standard errors, which does not make any assumptions about the distribution of withindependence of the residuals, or by modelling the dependence explicitly [28]. One way to model the dependence is to decompose the error component into a time-constant or a permanent error component z 0,i , unique for each individual, and a transitory error component v it , which varies across individuals and time. The permanent error component represents the combined effect of the omitted timeconstant covariates and of the unobserved heterogeneity. z 0,i and v it are assumed normally distributed, independent of each other. v it is assumed independent across individuals and over time. Model (10) can be rewritten by moving z 0,i in the intercept: Model (11) represents a random-intercept model, which is a regression model with an individual specific intercept. z 0,i can be interpreted as a random parameter that is estimated together with the variance of the v it . The parameters of the random-intercept model are estimated by maximum likelihood. To conclude, the linear random-intercept model allows the overall level of hours to vary across individuals after controlling for covariates.
Additional heterogeneity is incorporated by including additional random coefficients besides the random intercept, meaning that the effects of some covariates are allowed to vary across individuals. The resulting model is a mixed fixed and random coefficient model. By introducing individual-specific slopes, the assumption of parallel individual-specific regression lines is relaxed and our model becomes: f 0 represents the mean intercept of hours of work, f k , k~1,2 the mean slopes of the covariates chosen to have random coefficients, and z k,i , k[ 0,2 ½ the deviation of individual intercept and slopes from the mean values of f k .y is a vector of fixed coefficients, constant across individuals and time. X 0 it represent 1|K vector of covariates with fixed coefficients, whereas R 0 it , is 1|M vector of covariates with random coefficients.
The wage model follows a similar specification; besides the random individual-specific intercept, the model is specified with two additional random coefficients for age and education: R it is the vector the covariates estimated with random coefficients. P it is the vector of covariates with fixed coefficients. p 0 is the mean intercept of wage, and p 1 is the vector of the mean slopes of the variables R it . j k,i , k~0,1 represents the deviation of individual intercept and slopes from the mean values of the point estimators p 0 , p 1 : Given the high computation costs of estimating the selection model using a random coefficient specification, only fixed coefficient specification is used in the study. The inverse Mills ratios from the estimated probit regressions are included in the random coefficients wage equation of both men and women, whereas for hours, the additional term is included only for women.

The Empirical Model of Lifetime Labour Supply under Uncertainty
The extension of the empirical model of lifetime labour supply to incorporate uncertainty is straightforward. The methodology is the same above, except for an additional regressor that captures the forecasting error. This term is introduced to incorporate that individuals adapt their expectations regarding their future wages each period.

An Econometric Model of Labour Supply under
Uncertainty with Homogenous Preferences. Under uncertainty, the labour supply model assuming homogenous preferences is expressed as: ln represents the forecasting error equal to the cumulated discounted difference between the actual wage and the expected wage multiplied by the probability of being employed, from the start of the active life (a) until the current year. r j is the interest rate at time j. The selection and the wage models are the same as under certainty, assuming homogenous preferences.

An Econometric Model of Labour Supply under
Uncertainty with Heterogeneous Preferences. The heterogeneous preference model under uncertainty follows the same procedure as the homogenous preference model under uncertainty, with the exception that the forecasting error in equation (12) has a random coefficient. This change implies that the effects of the variables vary across individuals, but the selection and wage models remain the same as under certainty. Table 2 summaries the differences among all proposed models.

Simulations
This section describes how the models are applied to simulate the labour supply responses for 2004 and 2005. The accuracy of the projections obtained from these models is compared with the simple extended Heckman model that is commonly used in simulating continuous labour supply. Since the GSOEP dataset is a panel, the Heckman model is extended to incorporate the unobserved heterogeneity. While the estimation of the probit model is identical with one used for the selection model, the second step of the Heckman procedure is estimated by a standard RE model.
The simulations are performed using the estimates from the empirical models presented in the previous sections. The simulation follows the basic structure of a dynamic microsimulation. To simplify the exercise, this simulation only consists of the demographic and labour market module. The demographic module updates some basic demographic variables (e.g., age) over time. It also updates the variables that may interact with the demographic variables. The labour market module is the core part of the simulation, which updates the employment status, the wage, and hours of work. Figure 1  Both the estimation and the simulation is implemented using Stata. Fixed and random effects models are estimated using the standard panel commands xtreg, xtmixed is used for random coefficient model, and the fixed effect vector decomposition is estimated using the codes from [22].

Results
This section presents the estimation and the simulation results. All models, including wage and labour supply models are estimated separately for men and women.

Estimation Results
The parameter estimates for the probit models are largely skipped as the estimates are in line with previous findings. Inverse Mill's ratios are included in the wage and labour supply models, with the exception of male labour supply equation where the inverse Mill's ratio is not significant using any estimation method.
For the random coefficients models, a test is performed to verify whether the random intercept model is sufficient to capture the heterogeneity in the wage and hours estimation. The likelihoodratio test suggests that the random coefficients model fits better than the simple random intercept model, both for wage and labour supply. Tables 3 and 4 show the parameter estimates for the wage equations for women and men. Both model specifications are standard. The age effect is as expected for both men and women in all three models: a positive impact with decreasing marginal effects, showing that hourly wage has the standard humpedshaped age pattern. When estimated using FEVD, the wage profile of women shows a stronger curvature than for men. The larger coefficients of the linear and quadratic age variables show that the growth rate of wages is higher for women at younger ages, but the growth rate reduces more rapidly at later ages than for men. After the introduction of heterogeneity, the impact of age is reduced for women, whereas for men it is increased. At younger ages, however, both men and women have a similar curvature of the wage profile. The estimated rate of return to education differs quite a lot between men and women and between the three models. The model estimated with FEVD illustrates a positive return to education for women, whereas for the other models the return appears insignificant. For men, the returns appear to be negative in the FEVD variant model and insignificant in the other models.
Unemployment experience has a significant negative impact on wages for both men and women. The model with FEVD method illustrates the highest absolute impact, while the random coefficient model and the RE model show similar effects. In all three models, having or not having children influences wages negatively, and the impact in absolute value is higher for women than for men. The estimates of the inverse Mills ratios imply small positive correlations between the individual-specific error components of the selection model and the wage equation for men in all three models. For women, these correlations are negative in the mixed and the extended Heckman model.
To conclude, the estimation results from the wage models are in line with our expectations. Among all the specifications, the FEVD estimator appears to fit the data the best. The estimation results for the labour supply models are extensively examined in this study. Tables 5 and 6 present the coefficients estimated. Five labour supply models for both men and women are compared: the FEVD under certainty and uncertainty, the mixed fixed and random coefficients model under certainty and uncertainty, and the extended Heckman under certainty. The estimated coefficients are in general significant and stable across models.
The estimated wage elasticity is highly significant and stable for both men and women across models. Men record on average, in absolute value, a higher elasticity than women do. The highest    wage elasticity is found under certainty when estimated using the FEVD. Assuming that the effect of wage varies across individuals dampens the magnitude of the effect. Assuming uncertainty, the effect increases by 2 percentage points. The Heckman model provides higher wage elasticity estimates than the random coefficient model. As expected, the wage expectation correction has a small but significant negative effect on labour supply in both the FEVD and the random coefficient models. Incorporating that people adapt their expectations in a heterogeneous manner dampens the average impact of the wage expectation correction.
With regard to age, the labour supply profile of men illustrates a higher growth rate of hours worked at younger ages than the profile for women. The rate of growth in hours worked, however, reduces more rapidly toward later ages for men than for women. When the age effect is allowed to vary across individuals, its effect reduces in magnitude compared to the FEVD variant. The Heckman model provides the highest estimates for the age effect.
The estimates for the return to experience (full time and parttime) are similar across models: all key variables are significant, have the same shape and similar magnitudes. For men, the cumulated work experience has a negative effect on the lifetime labour supply response, which can be explained by the age-labour supply profile. For women, the cumulated work experience has a small positive effect on the labour supply response. Unemployment experience has a negative effect for the labour supply of both men and women, and the effect is higher in absolute value for women.
The presence of children in the family has a negative effect on the labour supply of women, whereas for men the effect is not significant. The effects are stable across models. For women, the estimates of the inverse Mills ratios differ greatly across models. The estimates of the Heckman model are similar to the random coefficient model under uncertainty. To conclude, the estimates are stable across the different model specifications and show on average a high level of significance. The FEVD model fits the best.

Simulation Results
The simulation exercise evaluates the model's out-of-sample prediction performance. Each model is estimated on the same dataset and is used to predict labour supply in years 2004 to 2005. The simulation uses a simple dynamic microsimulation framework, where only crucial demographic variables are updated. The performance of the simulation is judged according to the distance between the actual value and the predicted value. Table 7 gives a general overview of each model's performance using this indicator. Both simulated and actual values are coded in logarithm scale.
The errors in the simulation come from three sources. The first source is the selection model, common to all models. The second source is the continuous labour supply model, which covers five different model setups. The last source is the wage estimation, as in the models under uncertainty the residual of the wage equation is crucial. In general, the FEVD fixed coefficients under certainty and uncertainty perform the best in terms of mean value and standard deviation of the simulated residuals. In terms of unbiasedness, the two FEVD estimated models and the mixed coefficients model under certainty perform better than the Heckman extension model. The mixed coefficients estimation, however, can be significantly affected by outliers and has a large standard error for the random intercepts, which may lead to errors in the simulation. Figure 2 sorts the observations in terms of the absolute difference between simulated and actual values. Since the wage expectation adjustment (uncertainty) term does not play a major role numerically as suggested in the estimation result, the models with the uncertainty correction have a very similar curve compared with the ones without, provided the same estimation method is used. As a result, only 3 models are selected in Figure 2. The figure shows that the models estimated with the FEVD method performs the best in terms of the percentage of observations with the prediction error of less than 20 hours, while the Heckman extended model performs the worst judged by this criteria. This suggests that in a simulation study where the absolute error is a key factor, the proposed models should be estimated with FEVD. Figure 3 extends the previous graph by showing the results for different groups, whereas Figures 4 and 5 show the distribution of the simulation for different groups. The simulation shows that the Heckman model is biased for 2004 and it continues to get worse in 2005. This bias may come from the biased estimation of timeinvariant characteristics. The model estimated with mixed coefficients, although the mean predicted value is closer to the actual ones, it has a much larger standard error in the prediction. This finding is consistent with Figure 5, which shows that the mixed coefficients method has one of the most unbiased performances.
Overall, the FEVD variant of the model has the best simulation result according to the validation tests. The uncertainty extension, however, does not seem to have a large impact, this may partially be due to the error from the wage estimation. The mixed coefficient estimation seems to handle the heterogeneity better than the other models, but is less than ideal for simulations.

Conclusion
This paper develops a structural lifetime model for estimating and simulating continuous labour supply. The model is consistent with the lifetime economic theory and is able to capture the individual heterogeneity to a larger extent than existing labour supply models by using more refined estimation techniques, including fixed effect vector decomposition (FEVD) and the mixed coefficients estimation method. In addition, one variant of the model loosens the certainty assumption in life cycle modelling.   Instead, individuals are assumed to adjust their labour supply behaviour based on the differences between expected and actual earnings. The paper compares different combinations of the model specifications and estimation methods, as well as the standard random effects model (Heckman) for their simulation performances. In a simple simulation presented in this paper, models were estimated with different combinations of estimation techniques and uncertainty correction term. The results suggest that the