Moment-based Estimation of Latent Class Models of Event Counts

This paper develops and implements a GMM estimator for latent class models suitable for count data. The estimator uses conditional moment restrictions derived from standard count models. Both the efficient and consistent variants are considered. The implementation of optimal GMM based on semiparametric estimates of the weighting matrix appears to be problematic as the matrix is not guaranteed to be positive definite. A suboptimal variant which ensures positive definiteness is found to work well in computer simulations. The paper compares maximum likelihood and GMM estimators for Poisson based mixtures in two applications to U.S. health utilization data for the elderly from the National Medical Expenditure Survey.