Confounding: Regression adjustment

Abstract

In previous columns [1,2], I highlighted how the deterination of the association between an exposure (e.g., viamin E supplementation) and a disease (e.g., coronary heart isease [CHD]) is not quite so straightforward as it might rst appear. Before any definitive conclusions about the ssociation can be drawn, we must consider the effect of otential confounding variables. Confounders can obscure he association of real scientific interest and easily lead the nwary astray. For example, an apparent association beween the use of vitamin E supplements and the occurrence f CHD may be confounded, or distorted, by the effects of third variable (e.g., smoking history) that is associated ith the exposure and disease. Recognizing that confounding is a ubiquitous problem, hat can be done about it? Confounding that cannot be ontrolled in the design of a study (e.g., by matching on nown confounders) must be adjusted for in the analysis. In previous column [2], I discussed a tried and trusted ethod of adjusting for confounding in the analysis, namely tratification. With stratification, confounding is controlled y assessing the association of interest within distinct roups of individuals who are relatively homogeneous with espect to the confounding variable (or variables). For exmple, when we stratify the analysis by smoking history, omparison of the association between vitamin E suppleents and risk of CHD within any strata cannot be conounded by smoking because the strata are, by definition, onstituted by individuals with the same smoking history. he principle underlying stratification is simple and intuiive: stratification removes the variability of the confoundng variables (within any strata), thereby ensuring that these ariables cannot influence the association between exposure nd disease. Although stratification is a very effective and robust