PERFORMANCE OF INFORMATION CRITERIA FOR MODEL SELECTION IN A LATENT GROWTH CURVE MIXTURE MODEL

Abstract

Novel simulation studies are performed to investigate the performance of likelihood-based and entropy-based information criteria for estimating the number of classes in latent growth curve mixture models, considering influences of true model complexity and model misspecification. Simulation results can be summarized as (1) Increased model complexity worsens the performance of all criteria, and this is salient in Bayesian Information Criteria (BIC) and Consistent Akaike Information Criteria (CAIC). (2) The classification likelihood information criterion (CLC) and integrated completed likelihood criterion with BIC approximation (ICL.BIC) frequently underestimate the number of classes. (3) Entropy-based criteria correctly estimate the number of classes more frequently. (4) When a normal mixture is incorrectly fit to non-normal data including outliers, although this seriously worsens the performance of many criteria, BIC, CAIC, and ICL.BIC are relatively robust. Additionally, overextracted classes with trivially small mixture proportions can be detected when the sample size is large. (5) When there is an upper bound of measurement, although this worsens the performance of almost all criteria, entropy-based criteria are robust. (6) Although no single criterion is always best, ICL.BIC shows better performance on average.