Invited commentary: shifting the burden of proof regarding biases and low-magnitude associations.

The issue of what exactly are the limits of epidemiology is now on the radar screen of the entire field. Shapiro (1) has added his thoughts to the public discussion by describing a sensitivity analysis of reported associations between oral contraceptive use and breast cancer. He addresses diagnostic selection bias and information bias as possible explanations for small relative risks, demonstrating quite clearly that the overall association of 1.07 for ever use of oral contraceptives reported in a "collaborative reanalysis" of 54 studies could easily be due to a rather low magnitude of each of these biases. His points are reasonable, but limited; his overall conclusions, however, are a mixed bag. On the one hand, he correctly cites the need to separate policy decisions from the strength of evidence regarding causality. I will return to this question at the end. On the other hand, he essentially advises that if bias can explain small associations, we should give up our scientific pursuit of the health issue in question ("...it is unlikely that refinements of epidemiologic or statistical methods, sensitivity analyses, validations studies, etc., can get around the difficulty" (1, p. 944)). This conclusion is not scientifically grounded for reasons I will discuss. 1) If bias is to be put forward as the hypothesized explanation for a set of findings, it should be subject to the same scrutiny as the hypothesis of causality: Shapiro asserts that, compared with women who do not take oral contraceptives, those who do are more likely to be diagnosed with breast cancer or are likely to be diagnosed at an earlier stage. Undoubtedly, there are diseases for which this diagnostic bias is likely. However, what evidence is there that gynecologists are more likely to instruct patients on breast serf-examination or to perform such examinations themselves when prescribing