Management of Baseline Measurements in Statistical Analysis of 2×2 Crossover Trials

Crossover designs have applications in a wide range of sciences. The simplest and most common of such designs are the two-period, two-treatment (2×2) crossover. As a consequence, each subject provides a 4×1 vector of responses for data analysis in the following chronological order: baseline (period 1), post-baseline (period 1), baseline (period 2), and post-baseline (period 2). We considered three types of analytic approaches for handling the baselines:1) analysis of variance (ANOVA) method which ignores the first or both period baselines or use a change from baseline analysis 2) analysis of covariance (ANCOVA) method which uses an analysis of covariance where linear functions of one or both baselines are employed as either period-specific or period-invariant covariates 3) Joint modeling method that conducts joint modeling of a linear function of the baseline and post-baseline responses with certain mean constraints for the baseline responses. The crossover clinical trial data was analyzed, using the proposed models.          Based on the results on real data among all mentioned models, the first model (direct comparison of post-treatment values) and the second model (post-treatment measurement subtracts corresponding baseline) had the lowest and the highest standard errors, respectively. With respect to Akaike Information Criterion (AIC), the fifth model (comparison of post-treatment values adjusted by all available baseline data) and the eighth model (comparison of post-treatment values adjusted by difference and sum of all available baseline data) had the lowest magnitude, and the ninth model (modeling period baseline jointly with post-treatment values) had the highest AIC for both variables which the values of AIC were 518.1, 520.9 and 1137.8, respectively. To sum up, it is found that baseline data of crossover trial may be used to improve the efficiency of treatment effect estimation when applied appropriately.