Pitfalls in epidemiological analysis.

Abstract

Epidemiologists rely heavily on the relative risk in their analyses and presentations. As an index it is intelligible and intuitively appealing but can give an exaggerated impression of the strength of the association and is unreliable for comparisons. This can be shown by deriving relative risks from a normal correlation surface with an unimpressive level of correlation. Relative risks ought to be handled with caution; the underlying population risk and the relative size of exposed and reference categories should be reported. Efforts to control for additional variables, confounders, by some kind of multi-variate technique, another standard procedure, could easily give a false sense of security. From time to time it has been made clear in the literature that errors of measurement in the third variable or in the additional variables could lead to the appearance of false independent effects, but these warnings do not seem to have been heeded nearly as much as they deserve. A simulation experiment is used to bring the lesson home, with realistic numerical assumptions. A moderate degree of error contamination will produce spurious effects. This has nothing to do with sampling errors, large samples rather aggravate this danger. In meta-studies this is a source of error and conflicting results to take account of.