Bayesian Estimation of Case-Crossover Designs with Repeated Events

• Activity Number: 496 - Outcome-Dependent Sampling in the Survival Setting
• Type: Topic Contributed
• Date/Time: Wednesday, August 2, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Lifetime Data Analysis Interest Group
• Author(s): Daniel L. Gillen* and Sevan Gulesserian
• Companies: University of California, Irvine and University of California, Irvine
• Keywords: Bayesian ; case-crossover ; repeated events ; correlated data ; partial likelihood

Abstract

The case-crossover design is commonly implemented. The power of this design stems from the inherent matching of within-subject invariant covariates, allowing for control of these potential confounding factors by design. In practice, conditional logistic regression parameter estimates are computed by noting an equivalence between the conditional logistic likelihood and the partial likelihood arising from Cox's proportional hazards model. However, in many studies a subject may experience the event of interest repeatedly over the course of follow, resulting in multiple matched case and control observations per subject or cluster. In this talk we propose the direct estimation of cluster-specific covariate effects via a novel semi-parametric hierarchical Bayesian approach. Our approach allows for estimation and comparison of marginal covariate effects as well as cluster-specific random effects via posterior estimates. The research is motivated by, and applied to, data obtained from a case-crossover study of N=7751 children sampled from Orange County, CA seeking to quantify the effect of air pollution exposure on the risk of asthma-related emergency room admissions.