In previous columns [1,2], I highlighted how the determination of the association between an exposure (e.g., vitamin E supplementation) and a disease (e.g., coronary heart disease [CHD]) is not quite so straightforward as it might first appear. Before any definitive conclusions about the association can be drawn, we must consider the effect of potential confounding variables. Confounders can obscure the association of real scientific interest and easily lead the unwary astray. For example, an apparent association between the use of vitamin E supplements and the occurrence of CHD may be confounded, or distorted, by the effects of a third variable (e.g., smoking history) that is associated with the exposure and disease. Recognizing that confounding is a ubiquitous problem, what can be done about it? Confounding that cannot be controlled in the design of a study (e.g., by matching on known confounders) must be adjusted for in the analysis. In a previous column [2], I discussed a tried and trusted method of adjusting for confounding in the analysis, namely stratification. With stratification, confounding is controlled by assessing the association of interest within distinct groups of individuals who are relatively homogeneous with respect to the confounding variable (or variables). For example, when we stratify the analysis by smoking history, comparison of the association between vitamin E supplements and risk of CHD within any strata cannot be confounded by smoking because the strata are, by definition, constituted by individuals with the same smoking history. The principle underlying stratification is simple and intuitive: stratification removes the variability of the confounding variables (within any strata), thereby ensuring that these variables cannot influence the association between exposure and disease. Although stratification is a very effective and robust technique for adjusting for confounding, it cannot always be applied in practice. For example, stratification becomes far less appealing when there are many potential confounding variables, resulting in strata with too few individuals to make meaningful comparisons with any reasonable degree of precision. That is, as the number of confounders increases, the stratification necessarily becomes finer and finer, resulting in strata that contain too little information to reliably estimate the association of interest. In this column, I discuss an alternative approach that is widely used in practice, regression adjustment. © 2006 Elsevier Inc. All rights reserved.
