Testing and estimation in labour supply and duration models

Abstract: Estimation and testing in labour supply and unemployment duration models are studied. In the first two papers estimation techniques designed to account for quantity constraints and perceived tax scales are developed, while the remaining three papers focus on specification tests.In Paper [1] we estimate a model of male labour supply incorporating direct information on individuals perceived hours restrictions. The effect of rationing for the response in the hours of work to changes in the economic environment is found to be small.The commonly used assumption that individuals know the tax scale in full detail when deciding on labour supply is relaxed in Paper [2]. It is shown how perceived tax scales can be estimated jointly with the structural parameters in a labour supply function. Based on Swedish data from 1980 differences between true and estimated perceived tax scales are found to be small.Paper [3] surveys different approaches to testing endogeneity in micro-econometric models. We focus on the linear model when the explanatory power is low. Extensions to a nonlinear labour supply model are briefly discussed. Monte Carlo experimens shown that the power of the tests are unacceptably low for low R2-values. An empirical labour supply model illustrates.In Paper [4] a Öagrange Multiplier (LM) test for normality in censored regressions is derived against the generalized log-gamma distribution. The test statistic coincides with the skewness part of previously suggested LM and conditional moment (CM) tests. The small sample performance of the test is examined and compares fabourably to an alternative test.A correctly specified distribution is a crucial assumption for consistent estimation in duration models. In Paper [5] LM tests for exponential and Weibull duration distributions are derived against the generalized gamma distribution. Small sample properties of the tests are examined and compared to a CM test. Using an empirical unemployment duration model estimated with Swedish data as an illustration we can not reject exponentially distributed durations.