Equivalence between conditional and random-effects likelihoods for pair-matched case-control studies

• 成果类型: Article
• 署名作者: Rice, Kenneth
• 署名单位: University of Washington; University of Washington Seattle
• 刊物名称: JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
• ISSN/ISSBN: 0162-1459
• DOI: 10.1198/016214507000001463
• 发表日期: 2008
• 页码: 385-396
• 关键词: investigating underlying risk; EFFECTS MODELS; ratio; misclassification; Heterogeneity; estimators; priors

摘要

Two approaches dominate the analysis of highly stratified data sets: use of conditional likelihood and random-effects models. Conditioning is more traditional and has attractive robustness properties. Contemporary random-effects approaches rely more on accurate assumptions but can be applied to a larger class of problems. For pair-matched studies with arbitrary numbers of categorical covariates, we reconcile these two approaches, showing that the conditional approach has an exact interpretation as a specific type of random-effects model. The random-effects models that provide equivalence with conditioning naturally inherit all the attractive properties of that approach, and we argue they are a pragmatic model choice. We also discuss desirable characteristics of the random-effects model that are entirely novel and that have foundational implications. For example, although our model is specified entirely within a full-likelihood framework, its assumed random-effects distribution does not need a parametric form; we make explicit the robustness that this provides. In addition, our model forces a priori exchangeability of case and control status, an attractive property for objective analyses. The equivalence of the two approaches justifies extensions of conditional likelihood methods to new situations. We give a motivating example from a real problem concerning misclassification error in a genotyping process, and discuss refinements specific to genetic applications.