In 1990, my colleagues and I reported that the mortality outcomes in randomized, primary prevention trials of cholesterol-lowering measures, when aggregated with the results of a meta-analysis, showed an excess of suicide and traumatic deaths (that is, "non-illness mortality") in the treated men (odds ratio, 1.76; P = 0.004) [1]. Cummings and Psaty [2] examined this issue by analyzing non-illness mortality occurring in multiple risk factor intervention trials; however, these trials, by their design, preclude attributing any outcome to one particular intervention.

Cummings and Psaty also redid the meta-analysis of the primary prevention trials of non-illness mortality and found that this treatment effect is not statistically significant (odds ratio, 1.42; P > 0.05). The difference between our findings and theirs is due to Cummings and Psaty's use of recently reported "intention-to-treat" outcomes from the World Health Organization (WHO) trial, their exclusion of the Los Angeles Veterans Affairs trial because of incomplete data, and their inclusion of the Expanded Clinical Evaluation of Lovastatin (EXCEL) study. Reasonable arguments can be made against the latter two decisions.

Although complete data from the Veterans Affairs trial would be preferred, excluding this trial from analysis is neither satisfying nor free of potential bias. Inclusion of the EXCEL study is dubious because no traumatic deaths occurred, and therefore its consideration only dilutes a treatment effect apparent in trials in which at least one non-illness-related death occurred. Moreover, it was not designed as a primary prevention trial, and, of the deaths not caused by trauma, 70% occurred in persons known to have heart disease at baseline [3].

Beyond specific methodologic arguments, it is important to note that primary prevention trials have been repeatedly analyzed because one of the outcomes—increased mortality in the treatment groups—is unexpected and worrisome. Iterative analyses, however, may tempt us to believe the particular analysis that produces the desired finding. One way to safeguard against this is to present the results from all analytic permutations and to compare these results with a "favorable" outcome. For example, when non-illness mortality and cardiac mortality were directly compared in various groupings of the eight largest single-intervention, randomized trials, we noted that the increase in non-illness mortality with cholesterol interventions appears to be more robust than the cardiac mortality reduction [4] (Table 1).