ABSTRACT
Having discussed some key concepts of statistical design in previous chapters, we turn to the topic of methods for analyzing data from clinical studies including clinical trials. We begin with the principles of data analysis and stress the importance of purpose-driven analysis. Often we read data analyses that involve regression models, where the outcome measure of each subject is regressed on treatment plus other covariates. We should first distinguish the purpose of such regression analyses. There could be two kinds of purposes. One is to study the treatment and covariate association with respect to outcome. The purpose can be to study how some covariates can influence the treatment effect so that, when planning a clinical trial, we can properly choose a population based on the information gathered from the data. Here the attention is on the covariates. Another kind of purpose is for assessing the treatment effect “adjusted for” covariates. The focus is on the treatment effect. Our interest in this and next chapter is placed on the latter. As we have alluded to in Section 2.8, in a randomized trial, we can choose to analyze the treatment effect without any covariate and the conclusions will be valid, irrespective of the observed treatment–covariate association. As a common practice, we usually adjust a covariate—prognostic factor—that is unbalanced between the groups. In this chapter, we study the method of analysis of covariance (ANCOVA), where the endpoint is a continuous outcome, to discuss the appropriateness of this common practice. We show that the degree of adjustment and the variance reduction depend on two factors: the covariate’s correlation or association with the outcome variable and the degree of between-group imbalance. We also extend the discussion on stratified analysis and logistic regression for categorical data. We continue the topic of covariate adjustment in survival data analysis in the next chapter.
