In their randomized clinical trial of misoprostol in patients taking nonsteroidal anti-inflammatory drugs (NSAIDs), Graham and colleagues [1] state that "treatment (misoprostol compared with placebo) was the only statistically significant (P = 0.008) term retained in the final model" and that "the number of duodenal erosions at entry. was not statistically significant (P = 0.10)". These statements raise three concerns. First, the total number of "outcomes" in the duodenal ulcer group was only 17 (5 receiving misoprostol and 12 receiving placebo). Second, as emphasized by Concato and colleagues [2], this small number would only allow for the modeling of a maximum of two explanatory variables (10 outcomes per variable). Adjusting for any more variables, as the authors seem to have attempted, would lead to "overfitting" of their model and would reduce the precision of their regression coefficient estimates, as confirmed by the wide confidence interval (CI) associated with the odds ratio for treatment effect (95% CI, 2.0 to 43.2). The small number of outcomes could also lead to "underfitting" of the data and undermine their statement that "age, sex, smoking, alcohol consumption, and type of arthritis were not statistically associated with duodenal ulceration". Third, the criteria used to select the independent variables entered in the stepwise forward logistic regression are unclear.

We are concerned that readers may interpret this study as proof that misoprostol has a large protective effect on duodenal ulceration among patients receiving NSAIDs independent of other "important ulcerogenic" risk factors. Emphasis on the discussion of the diminished statistical power resulting from so few end points (for which the authors are not responsible) when doing multivariable analysis would have avoided the danger of seriously overstating their conclusions.