Simulated Maximum Likelihood Estimation of Discrete Models with Group Data

Abstract

This article has compared the performance of two methods of simulated maximum likelihood for the estimation of discrete choice models with group data. One method of simulated likelihood uses simulators which are sstatistically independent across individuals in the sample. The alternative method allows simulators to be correlated across individuals. the comparisons are based on the criteria of statistical efficiency and computation time cost. As the simulated maximum likelihood method with dependent simulators can take into account the presence of sufficient statistics in group data, it can have advantages over the simulated likelihood method with independent simulators in term of computation cost saving and statistical efficiency. The computation time cost of the simulated likelihood method with dependent simulators can be inexpensive as the method of simulated moments of McFadden (1989). This is so especially for group data with either large sample sizes or small number of groups. Besides theoretical analysis, Monte Carlo results provide some evidence.