MMRM versus MI in dealing with missing data--a comparison based on 25 NDA data sets

Both multiple imputation (MI) and mixed-effects model repeated measures (MMRM) approaches appear to be better choices than the traditional last-observation-carried-forward (LOCF) approach in analyzing incomplete clinical trial data sets in drug development research. However, relative performances of these two approaches are unknown in controlling type I error rate and statistical power in the hypothesis testing of determining the efficacy of an investigational drug. Little research has been done in comparing robustness of the two approaches in analyzing ignorable missing data of clinical trials. In this research, a comparison between the MI and MMRM approaches is made in analyzing the simulated incomplete data sets and 25 New Drug Application (NDA) data sets of neuropsychiatric drug products. The MMRM approach appears to be a better choice in maintaining statistical properties of a test as compared to the MI approach in dealing with ignorable missing data of clinical trials.