|
Abstract:
|
For matched case-control studies, conditional logistic regression is the 'industry standard' method, producing well-behaved, attractively robust inference. Given the implicit notions of exchangeability in such highly stratified data, Bayesian approaches are also attractive and offer substantial flexibility for applications over the conditional framework. However, the Bayesian approach requires one to assume priors for the many nuisance parameters associated with stratification; in general, this choice actively influences the analysis and requires some care. We present a class of nuisance-parameter priors that, marginally, produce the conditional likelihood exactly. This allows advocates of the standard approach access to Bayesian flexibility, while Bayesians gain the benefits of conditioning. Further attractive but nonstandard features of the Bayesian model also are explored.
|
