In earlier columns [1–3], I highlighted how the determiation of the association between an exposure (e.g., vitamin supplementation) and a disease (e.g., coronary heart disase [CHD]) is not quite so straightforward as it might first ppear. In particular, confounding variables very often obcure the association of real scientific interest. For instance, n apparent association between the use of vitamin E suplements and the occurrence of CHD may be confounded, r distorted, by the effects of a third variable (e.g., age or moking history) that is associated with the exposure and he disease. When it is not possible to control for confounding via the esign of a study (e.g., by means of randomization or by atching on known confounders), it is imperative to make n appropriate adjustment for confounding in the analysis. n earlier columns [2,3], I discussed two widely used methds of adjustment: stratification and regression adjustment. ith stratification, confounding is controlled by assessing he association of interest within distinct groups of individals who are more or less homogeneous with respect to the onfounding variable (or variables). The principle underlyng stratification is simple and intuitive: stratification reoves the variability of the confounding variables (within ny strata), thereby ensuring that these variables cannot nfluence the association between exposure and disease. owever, stratification becomes problematic when there are any potential confounding variables, resulting in a very arge number of strata with too few individuals to make eaningful comparisons with any reasonable degree of preision. An alternative approach for adjusting for confoundng is to estimate the exposure effect of interest within a egression model for the dependence of the disease outcome n the exposure and any potential confounders. For examle, an adjusted estimate of the association between vitamin supplements and the risk of CHD might be obtained from