Interval Estimation For Summary Measures Of Attributable Risk From Case-control Studies (epidemiology).

Abstract

From a public health perspective, measures of the strength of association between exposure to a suspected risk factor and a disease outcome (e.g. relative risk, odds ratio) provide incomplete information, because they do not reflect the prevalence of exposure to the risk factor, and thereby the actual impact of association. In order to assess this impact, measures of attributable risk have been proposed in the literature. Under certain assumptions, the population attributable risk can be interpreted as the proportionate reduction in the total number of cases that would be expected (eventually) by removing the risk factor of interest. The general problem considered in this dissertation is the interval estimation for summary measures of attributable risk from the analysis of stratified case contol data (both matched and unmatched). The estimators proposed in this dissertaton for these attributable risk measures are functions of the Mantel-Haenszel odds ratio and, in typical studies of chronic disease outcomes, do not require the assumption of a rare disease. A product-multinomial method was proposed in order to obtain a large-sample variance from unmatched data, and evaluations of these methods using simulated data revealed good performance with respect to bias and confidence interval coverage. Furthermore, these results were consistent in two alternate limiting models (few strata with large stratum sample sizes and many strata with small stratum sample sizes). The methods developed here were consistently better than alternate methods from the literature when applied to simulated data. A matched-set-multinomial method was proposed for matched-set data, and provided direct expression for large-sample variances. The simplifications to these expressions for the case of matched pairs was presented in terms of the usual notation for that setting. Evaluation of these methods using simulated data revealed that they, too, performed well with respect to bias and confidence interval coverage. The Mantel-Haenszel based summary estimator proposed in this dissertation is recommended for both matched and unmatched case-control data. The product-multinomial and matched-set-multinomial methods developed in this dissertation are recommended for obtaining large-sample variances from unmatched and matched-set data, respectively.