Finite Mixture Modeling of Censored Regression Models

• Activity Number: 423
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Social Statistics Section
• Author(s): Maria Karlsson*+ and Thomas Laitila
• Companies: Umeå University and Örebro University/Statistics Sweden
• Keywords: Finite Mixture Models ; Censoring ; Tobit model ; EM-algorithm

Abstract

This paper proposes the use of finite mixture of regression models for censored response variables, i.e. a finite mixture of Tobit models, as a mean for modeling an unknown distribution. Thus, mixture is not primarily motivated by data containing observations from different populations but constitute a flexible way to consider different kind of features in the data, e.g., asymmetrically distributed error terms, not allowed by other suggested estimators of censored regression models. Monte Carlo simulation results on the properties of the estimator of the finite mixture of Tobit models are very promising compared to earlier suggested estimators. Results also indicate that applying a finite mixture model, in studies where the aim is to identify latent subpopulations, gives a risk to wrongly identify components if data is heteroscedastic.