Penalized estimation of finite mixture models

Abstract

Economists often model unobserved heterogeneity using finite mixtures. In practice, the number of mixture components is rarely known. Model parameters lack point-identification if the estimation includes too many components, thus invalidating the classic properties of maximum likelihood estimation. I propose a penalized likelihood method to estimate finite mixtures with an unknown number of components. The resulting Order-Selection-Consistent Estimator (OSCE) consistently estimates the true number of components and achieves oracle efficiency. This paper extends penalized estimation to models without point-identification and to mixtures with growing number of components. I apply the OSCE to estimate players’ rationality levels in a coordination game.