Uncertainty analysis of multiple epidemiological studies using frequency distributions of relative risks

Abstract

A new format for presenting uncertainty in the results of multiple epidemiologic studies of the same outcome is suggested. A set of 95% confidence intervals for relative risk, RR, is transformed to a frequency distribution of the normalized deviations, ln(RR)/SE(ln(RR)), from the null value In(RR)=0 (RR=1). It is assumed that deviations from RR=1 are due to unaccounted residual biases and we compare the distribution of these deviations with the distributions of the actual errors in physical measurements where the true values have subsequently become known, and the incidence of large errors can be estimated. Comparison of these distributions can, by analogy, help to understand how convincing is the evidence of elevated risk in observational studies.