Using re-sampling methods in mortality studies

Traditional methods of computing standardized mortality ratios (SMR) in mortality studies rely upon a number of conventional statistical propositions to estimate confidence intervals for obtained values. Those propositions include a common but arbitrary choice of the confidence level and the assumption that observed number of deaths in the test sample is a purely random quantity. The latter assumption may not be fully justified for a series of periodic "overlapping" studies. We propose a new approach to evaluating the SMR, along with its confidence interval, based on a simple re-sampling technique. The proposed method is most straightforward and requires neither the use of above assumptions nor any rigorous technique, employed by modern re-sampling theory, for selection of a sample set. Instead, we include all possible samples that correspond to the specified time window of the study in the re-sampling analysis. As a result, directly obtained confidence intervals for repeated overlapping studies may be tighter than those yielded by conventional methods. The proposed method is illustrated by evaluating mortality due to a hypothetical risk factor in a life insurance cohort. With this method used, the SMR values can be forecast more precisely than when using the traditional approach. As a result, the appropriate risk assessment would have smaller uncertainties.