Effect of unequal variance on statistical tests for mixed paired and two-sample designs

Observations from studies comparing two treatments can involve both paired and independent samples. While tests have been proposed to utilize both sample types, most of these either assume equal variances across groups or have not been tested under the effect of unequal variance. Two methods have been proposed to handle the unequal variance case, but their relative performance is unknown. In this paper, a new procedure for the unequal variance case is proposed and an extensive simulation study is conducted to compare its performance to previously proposed methods. The new method is shown to be the most powerful test in certain cases. The simulation study suggests that tests that do not assume equal variance should be recommended as the default procedures.