This paper provides structural estimates of a dynamic model of schooling, work, and occupational choice decisions based on 11 years of observations on a sample of young men from the 1979 youth cohort of the National Longitudinal Surveys of Labor Market Experience (NLSY). The structural estimation framework that we adopt fully imposes the restrictions of the theory and permits an investigation of whether such a theoretically restricted model can succeed in quantitatively fitting the observed data patterns. We find that a suitably extended human capital investment model can in fact do an excellent job of fitting observed data on school attendance, work, occupational choices, and wages in the NLSY data on young men and also produces reasonable forecasts of future work decisions and wage patterns.

One of the long-standing issues in the literature on transfer programs for the U.S. low-income population concerns the high cumulative marginal tax rate on earnings induced by participation in the multiplicity of programs offered by the government. Empirical work on the issue has reached an impasse partly because the analytic solution to the choice problem is intractable and partly because the model requires the estimation of multiple sets of equations with limited dependent variables, an estimation problem which until recently has been computationally infeasible. In this paper we estimate a model of labor supply and multiple program participation using methods of simulation estimation that enable us to solve both problems. The results show asymmetric wage and tax rate effects, with fairly large wage elasticities of labor supply but very inelastic responses to moderate changes in cumulative marginal tax rates, implying that high welfare tax rates do not necessarily induce major reductions in work effort.

Over the past decade, a substantial literature on the estimation of discrete choice dynamic programming (DC-DP) models of behavior has developed. However, this literature now faces major computational barriers. Specifically, in order to solve the dynamic programming (DP) problems that generate agents' decision rules in DC-DP models, high dimensional integrations must be performed at each point in the state space of the DP problem. In this paper we explore the performance of approximate solutions to DP problems. Our approximation method consists of: 1) using Monte Carlo integration to simulate the required multiple integrals at a subset of the state points, and 2) interpolating the non-simulated values using a regression function. The overall performance of this approximation method appears to be excellent, both in terms of the degree to which it mimics the exact solution, and in terms of the parameter estimates it generates when embedded in an estimation algorithm.

Statistical inference in multinomial multiperiod probit models has been hindered in the past by the high dimensional numerical integrations necessary to form the likelihood functions, posterior distributions, or moment conditions in these models. We describe three alternative approaches to inference that circumvent the integration problem: Bayesian inference using Gibbs sampling and data augmentation to compute posterior moments, simulated maximum likelihood (SML) estimation using the GHK recursive probability simulator, and method of simulated moment (MSM) estimation using the GHK simulator. We perform a set of Monte-Carlo experiments to compare the performance of these approaches. Although all the methods perform reasonably well, some important differences emerge. The root mean square errors (RMSEs) of the SML parameter estimates around the data generating values exceed those of the MSM estimates by 21 percent on average, while the RMSEs of the MSM estimates exceed those of the posterior parameter means obtained via agreement via Gibbs sampling by 18 percent on average. While MSM produces a good agreement between empirical RMSEs and asymptotic standard errors, the RMSEs of the SML estimates exceed the asymptotic standard errors by 28 percent on average. Also, the SML estimates of serial correlation parameters exhibit significant downward bias.

This research compares several approaches to inference in the multinomial probit model, based on Monte-Carlo results for a seven choice model. The experiment compares the simulated maximum likelihood estimator using the GHK recursive probability simulator, the method of simulated moments estimator using the GHK recursive simulator and kernel-smoothed frequency simulators, and posterior means using a Gibbs sampling-data augmentation algorithm. Each estimator is applied in nine different models, which have from 1 to 40 free parameters. The performance of all estimators is found to be satisfactory. However, the results indicate that the method of simulated moments estimator with the kernel-smoothed frequency simulator does not perform quite as well as the other three methods. Among those three, the Gibbs sampling-data augmentation algorithm appears to have a slight overall edge, with the relative performance of MSM and SML based on the GHK simulator difficult to determine.

This paper generalizes the normal probit model of dichotomous choice by introducing mixtures of normals distributions for the disturbance term. By mixing on both the mean and variance parameters and by increasing the number of distributions in the mixture these models effectively remove the normality assumption and are much closer to semiparametric models. When a Bayesian approach is taken, there is an exact finite-sample distribution theory for the choice probability conditional on the covariates. The paper uses artificial data to show how posterior odds ratios can discriminate between normal and nonnormal distributions in probit models. The method is also applied to female labor force participation decisions in a sample with 1,555 observations from the PSID. In this application, Bayes factors strongly favor mixture of normals probit models over the conventional probit model, and the most favored models have mixtures of four normal distributions for the disturbance term.

This study uses data from the Panel Survey of Income Dynamics (PSID) to address a number of questions about life cycle earnings mobility. It develops a dynamic reduced form model of earnings and marital status that is nonstationary over the life cycle. The study reaches several firm conclusions about life cycle earnings mobility. Incorporating non-Gaussian shocks makes it possible to account for transitions between low and higher earnings states, a heretofore unresolved problem. The non-Gaussian distribution substantially increases the lifetime return to post-secondary education, and substantially reduces differences in lifetime wages attributable to race. In a given year, the majority of variance in earnings not accounted for by race, education and age is due to transitory shocks, but over a lifetime the majority is due to unobserved individual heterogeneity. Consequently, low earnings at early ages are strong predictors of low earnings later in life, even conditioning on observed individual characteristics.