Comparison of Mixed Models and Paired T-Test for Analyzing Crossover Clinical Trials in the Presence of Missing Data

Abstract

AB/BA crossover clinical trials are popular designs that can achieve high power with a lower number of subjects than other randomized control trial designs. They are often analyzed using paired t-test or mixed models, and like many clinical trials, are often impacted by missing data. Mixed models have been shown to produced more powerful and unbiased results in the presence of missing data than t-tests for other designs, but these two approaches have not been compared in crossover trials. We conducted a simulation study to compare the bias and power of paired t-tests and mixed models when analyzing an AB/BA crossover clinical trial in the presence of missing data. Several different missing structures were simulated under two within-subject correlations, ρ =0.3 and ρ =0.7. Both methods performed similarly when analyzing complete data, but the mixed model produced both equal or less bias estimates and higher power than the paired t-test under all simulation scenarios. In the worst-case scenario we considered, the t-tests resulted in percent bias up to -105% and power as low as 5% compared the mixed model’s percent bias of 1% and 57% power. In less severe cases, both methods had 0% bias, but mixed models still achieved an absolute power gain of 2%-6%. In the presence of missing data, the mixed model achieved higher power than the paired t-test under all simulated scenarios. The mixed model also achieved equal or less bias under all simulated scenarios. Therefore, mixed models should be used over paired t-test when analyzing AB/BA crossover clinical trial in the face of missing data.