Determining the number of components in mixture regression models: an experimental design

Abstract

Despite the popularity of mixture regression models, the decision of how many components to retain remains an open issue. This study thus sought to compare the performance of 26 information and classification criteria. Each criterion was evaluated in terms of that component's success rate. The research's full experimental design included manipulating 9 factors and 22 levels. The best results were obtained for 5 criteria: Akaike information criteria 3 (AIC3), AIC4, Hannan-Quinn information criteria, integrated completed likelihood (ICL) Bayesian information criteria (BIC) and ICL with BIC approximation. Each criterion's performance varied according to the experimental conditions.