Sensitivity of five information criteria to discriminate covariance structures with missing data in repeated measures designs.

Abstract

This study analyzes the effectiveness of different information criteria for the selection of covariance structures, extending it to different missing data mechanisms, the maintenance and adjustment of the mean structures, and matrices. The Monte Carlo method was used with 1,000 simulations, SAS 9.4 statistical software and a partially repeated measures design (p=2; q=5). The following variables were manipulated: a) the complexity of the model; b) sample size; c) matching of covariance matrices and sample size; d) dispersion matrices; e) the type of distribution of the variable; f) the non-response mechanism. The results show that all information criteria worked well in Scenario 1 for normal and non-normal distributions with heterogeneity of variance. However, in Scenarios 2 and 3, all were accurate with the ARH matrix, whereas AIC, AICCR and HQICR worked better with TOEP and UN. When the distribution was not normal, AIC and AICCR were only accurate in Scenario 3, more heterogeneous and unstructured matrices, with complete cases, MAR and MCAR. In order to correctly select the matrix it is advisible to analyze the heterogeneity, sample size and distribution of the data.