THE FINITE-SAMPLE PROPERTIES OF MAXIMUM LIKELIHOOD ESTIMATORS IN MULTINOMIAL PROBIT MODELS

Abstract

In this paper we discuss the finite-sample properties of maximum likelihood estimators of multinomial probit model. Sample sizes ranging from 20 to 200 at interval of 10 are tested. For each sample size, we estimated 9 utility specifications under different error covariance matrix assumptions. Robust variance matrix is used to get the asymptotic covariance of estimators for each model. We use Monte Carlo experiments to generate the data sets and replicate 500 times for each experiment. We found that when the size of standardized parameter of generic variable is greater than 0.5 and the sample size is greater than 60, the odds of getting significant estimator that is close to the true value are larger than 0.9. When the size of standardized parameter is small, e.g., less than 0.25, a sample size of at least 170 is needed to get good results.