Confounding

Abstract

When published, a randomized parallel-group drug trial essentially includes a table listing all of the factors, otherwise called baseline characteristics, known possibly to influence outcome. E.g., in case of heart disease these will probably include apart from age and gender, the prevalence in each group of diabetes, hypertension, cholesterol levels, smoking history, other cardiovascular comorbidities, and concomitant medications. If the prevalence of such factors is similar in the two groups, then we can attribute any difference in outcome to the effect of test-treatment over reference-treatment. However, if this is not the case, we have a problem which can be illustrated by an example. Figure 28.1 shows the results of a study where the treatment effects are better in the males than they are in the females. This difference in efficacy does not influence the overall assessment as long as the numbers of males and females in the treatment comparison are equally distributed. If, however, many females received the new treatment, and many males received the control treatment, a peculiar effect on the overall data analysis is observed: the overall regression line is close to horizontal, giving rise to the erroneous conclusion that no difference in efficacy exists between treatment and control. This phenomenon is called confounding, and may have a profound effect on the outcome of a trial. In randomized controlled trials confounding is, traditionally, considered to play a minor role in the data. The randomization ensures that no covariate of the efficacy variable is associated with the randomized treatment (Cleophas et al. 2006a). However, the randomization may fail for one or more variables, making such variables confounders. Then, adjustment of the efficacy estimate should be attempted. Methods include subclassification (Cochran 1968), regression modeling (Cleophas et al. 2006a), and propensity scores (Rosenbaum and Rubin 1983; Rubin 1997). This chapter reviews these three methods and uses hypothesized and real data examples for that purpose.