Estimating risks for matching factors in case-control studies

Matching for factors such as age and sex is a convenient method for minimizing confounding in case-control studies, but it does not allow inferences about the effects of the matching factors unless case ascertainment is virtually complete and the distribution of the matching factors in the source population is known. When this is so, the effect of a particular factor can be estimated by comparing the population distribution of that factor with what is observed in the case series. Such a comparison, however, may itself be confounded by other factors that are related to both the matching factors and the disease under investigation. This article proposes a method for evaluating matching factors as risk factors, which uses information on the distribution of potential confounders in the reference series and exposure relative risk estimates to adjust the person-time proportionality constant in a Poisson regression model. The method is particularly suited to data sets in which many of the elementary matching strata contain few or no cases and/or controls. It makes use of standard analytic procedures, but requires the estimation of an additional variance-covariance component for the estimated Poisson regression coefficients. Further factors that may confound the relationship between exposure and disease are easily accommodated. The method is demonstrated in two examples: a matched case-control study of drugs in relation to the rare blood dyscrasia, agranulocytosis, that was conducted in Europe and Israel, and a case-control study of ovarian cancer in Australia.