Bias in the evaluation of low-magnitude associations: an empirical perspective.

In nonexperimental research, the possible existence of bias can never be entirely eliminated, and our ability to infer causation is correspondingly limited (1). If, however, in any reasonably well-conducted study the magnitude of an association is large relative to any plausible biases that may be present, we may judge that it would persist even if those biases could be eliminated (2). In cancer epidemiology, for example, it is no accident that one of the early associations to be identified was between smoking and lung cancer (3), the presence of bias notwithstanding. Both the disease and the exposure were common, and the association was exceedingly strong. That association was correctly identified despite clear evidence (in retrospect) of selection bias (the inclusion of controls with smokingrelated disorders such as chronic bronchitis) (4, 5). By contrast, if an association is of relatively low magnitude (defined here as a relative risk estimate of less than 2.0), it may not be possible to judge whether or not it can be entirely accounted for by bias. Yet, we are confronted by the dilemma that in the evaluation of the risk of common diseases in relation to common exposures, even small relative risk increments, well below 2.0, may have profound public health implications. If, for example, the current use of estrogens for 5 or more years increases the risk of breast cancer by some 1.35-fold, as has been suggested (6), that association would pose a major public health problem: The