Re: Fruit and Vegetable Intake and Overall Cancer Risk in the European Prospective Investigation Into Cancer and Nutrition

The excellent article in the Journal by Boffetta et al. (1) on fruit and vegetable intake and cancer risk offers a nice example on why inferences based on P values are problematic in large studies that show associations of very small magnitude. The reported associations between high fruit and vegetable intake and cancer risk reached nominal levels of statistical significance (P = .006 and P < .001, respectively), and the authors conclude that the study “supports the notion of a modest cancer preventive effect of high intake of fruits and vegetables” (1). The respective hazard ratios per 100 g/d intake of fruits and vegetables were 0.99 and 0.98, respectively, and 0.97 per 200 g/d intake of fruits and vegetables combined. The authors are commended for being cautious that residual confounding can affect such hazard ratios of very small magnitude (1). The main problem, however, is that for associations of very small magnitude, P values alone do not suffice to make inferences, even if there is absolutely no residual confounding. The inferences depend critically on how strong the associations are expected to be based on previous results, biological reasoning, or practical considerations, that is, the associations that would be important from a clinical or public health perspective.