this study want to examine the effect of the AIC, BIC, AICc, CAIC, and HQIC information criteria in selecting the best regression model. the simulation study generated the longitudinal data with missing response. it manipulated many variables as mean and covariance structure (CP1~CP4), distribution shapes (Normal and non-Normal), sample size, and estimate method (ML and REML). The result discussed that AIC and HQIC had the best overall performance, but BIC, CAIC and HQIC performed better when the covariance patterns were more simple. No matter used which estimate, it was better that criteria based on total number of subjects than based on total number of observations.
comment
It is hard to understand how this generating method can represent the longitudinal design.
the simulation set the two models as true model with different distribution. I expect the AIC easily selected the complex model but there is not the pattern in reduce model as true model.