Under usual practice in epidemiologic research, covariate adjustment would be used to control for confounding effects. Exclusions, on the other hand, are justified as a part of an analytic strategy when there is evidence of an interaction altering the shape or direction of the exposure--outcome relationship across strata. This distinction applies with particular relevance in the analysis of the BMI--mortality relationship. Both covariate adjustment and exclusions have been used, with the need to control for interaction usually justified by concerns of reverse causality. The concern of reverse causality arises when conditions prevalent at baseline are thought to be associated with both lower average body weight and higher mortality risk, as would be the case with smokers. In the analysis, the choice must be made of whether to control for smoking status or exclude smokers altogether. Unfortunately, reverse causality is difficult to test for directly in observational data as deletions of very large subsets of the data can also lead to a different result, by chance alone. Findings that result after large-scale exclusions must therefore be tested in the standard statistical framework that can distinguish any new result observed after exclusions from one that could have occurred solely by chance. In statistical tests focused specifically on this question, the weight of the evidence suggests that interactions are not present in this context and that exclusion of subgroups (e.g., by smoking status, age, sex, use of alcohol) leaves the shape of the mortality curve unchanged, although it may alter the absolute level of risk. Unless some special question is being asked about subgroups, the author advises against large-scale exclusions for a common trait in analyses of the BMI--mortality relationship and emphasizes using representative study samples with measured exposure variables. A large-scale social experiment is cited that gives some evidence on the effect of a population-wide downward shift in BMI.