Abstract by David Lowe
Measuring the Performance of Information Criteria in a Repeated Measure Setting
Much store is placed in information criterion’s ability to select the best model. The purpose of this study is to compare the abilities of Akaike Information Criteria (AIC) and Bayesian Information Criteria (BIC) in identifying the correct covariance structures of repeated measures data. Data are generated under four covariance structures: compound symmetry, first-order autoregressive, unstructured, and exponential. Each of these covariance structures are created under various calibrations of the number of subjects, measurements per subject, and with missing data or not. AIC and BIC are used to compare models built with the four covariance structures to see whether the criterion picks the correct model. On average, BIC outperforms AIC in identifying the correct covariance structure, except in the unstructured case. Fewer observations per subject, and missing data both diminish the ability of AIC and BIC.