A Bayesian approach to epidemiological studies with misclassification in a binary risk factor
Bayesian statistical methods permit greater flexibility than most frequentist method by allowing misclassification rates to differ between the validation study and the remainder of the study. Bayesian approaches were developed using the freeware package WinBUGS. Simple methods may be applied to tables of summary data from unmatched or individually matched case-control studies to correct for misclassification in a single risk factor. A Bayesian model for prospective case-control studies has been developed which permits greater flexibility in the types of relationships between covariates, and also between the probability of misclassification and other covariates, than is allowed by other methods. The literature has been concerned about the application of prospective Bayesian methods to retrospectively collected data, but typically assumes that this is unimportant for maximum likelihood estimates without misclassification. Empirical work shoed very little difference between estimates obtained from prospective and retrospective approaches in the frequentist and Bayesian frameworks. Therefore it is justified to use prospective fixed effects Bayesian methods for retrospective data with or without misclassification. The methods developed quantify and correct for misclassification in the most common study designs encountered in epidemiology. The only additional cost would be the data collection for the internal validation study, but without this there is no way to evaluate or correct for potential bias.