Working Papers
A Note on Identification in the Multinomial Probit Model in the Presence of Weak Correlation
Abstract: An identification problem with the Multinomial Probit Model arises when correlations among choice alternatives are weak. Then even if formal conditions for model identification are satisfied, practical estimation of the unrestricted variance parameters is tenuous. An estimable specification of the model, in which all variance parameters are fixed, resolves the problem. This paper demonstrates the identification problem and presents an MCMC algorithm to estimate this restricted specification of the Multinomial Probit model.
Analysis of School Absenteeism for Single and Full Parent Families: Finite Mixture Roy Approach
Abstract: This paper analyzes factors affecting school absenteeism due to an injury or illness among the US school student population between 6 and 15 years of age. The number of missed school days is an overdispersed count, modeled in a flexible semiparametric way, using the Finite Mixture Roy (FMR) model for count variables, developed by Munkin (2022). The full/single parent family status (treatment) is potentially endogenous to the dependent variable. The Roy modeling structure captures observed heterogeneity defined by mother’s marital status. The FMR model further controls for unobserved heterogeneity using finite mixtures. The objective is to identify components within the data in both states, where the assumption of homogeneity in marginal and treatment effects is more realistic. The considered application motivates two additional features of the model. First, to better understand the structures of the latent components their probabilities are modeled as functions of regressors. Secondly, mother’s income is allowed to enter the treatment equation nonparametrically. The FMR model is estimated with two components in each states, interpreted as healthy and unhealthy students. Marital status decreases annual missed school days by about 13 percent for a randomly selected individual, but increases it by 9 percent for those families who actually select to have a single parent, which is evidence of adverse selection.
Abstract: This paper estimates the effect of income on the decision to consume alcohol and tobacco products, and on the corresponding levels of expenditure for households in Turkey using a semiparametric Bayesian approach and data derived from the 2010 Turkish Household Expenditure Survey. We find that unlike alcohol, which remains a normal good, tobacco products become an inferior good at high income levels. However, for smokers and drinkers only, tobacco and alcohol products respectively are normal goods. The results support the claim that taxing tobacco products is likely to keep individuals from smoking for the lower income groups, which include young adults.
Biases in the Maximum Simulated Likelihood Estimation of the Mixed Logit Model
Abstract: Jumamyradov and Munkin (2021) showed that the maximum simulated likelihood (MSL) estimator produces significant biases when applied to the bivariate normal and bivariate Poisson-lognormal models. Their conclusion is that similar biases can be present in other models generated by correlated bivariate normal structures, which include most commonly used specifications of the mixed logit (MIXL) models. This paper conducts a simulation study analyzing MSL estimation of the error-components (EC) MIXL. We find that the MSL estimator produces significant biases in the estimated parameters, leading to up to 12% difference in the true and estimated marginal effects. Supplemental appendix
Abstract: Jumamyradov and Munkin (2021) showed that the maximum simulated likelihood (MSL) estimator produces significant biases when applied to the bivariate normal and bivariate Poisson-lognormal models. Their conclusion is that similar biases can be present in other models generated by correlated bivariate normal structures, which include most commonly used specifications of the mixed logit (MIXL) models. This paper conducts a simulation study analyzing MSL estimation of the error-components (EC) MIXL. We find that the MSL estimator produces significant biases in the estimated parameters, leading to up to 12% difference in the true and estimated marginal effects. Supplemental appendix