ABSTRACT
In the evaluation of clinical trial data, regulators generally prefer simple hypothesis tests over model-based tests to determine whether a drug is effective. For example, in the evaluation of a drug that reduces blood pressure, a Z-test based on the difference between the test treatment and control of the mean change in systolic blood pressure from baseline where the important baseline factors are balanced between treatment groups would be preferred over a test based on a linear model of the change in systolic blood pressure that includes a term for treatment along with terms for other baseline variables that are imbalanced between the arms of the study. The underlying reason for this is that the model-based test is dependent on more estimates of related parameters than is the simple difference in means and thus requires more assumptions to be correct. In what follows we will look at the variance of a model-based estimate of the treatment effect that accounts for baseline variables that are imbalanced between the arms of the study and compare it with the variance of a simple difference in means to help understand why there is a preference for simple hypothesis tests.
