Bayesian analysis of binary data from an audit of cervical smears

Abstract

A large data set of unsatisfactory smear rates from 629 different cervical smear takers is used to illustrate a Bayesian approach to the analysis of medical audit data with a binary outcome. It is shown that this gives more relevant results in terms of characterizing individual smear takers than frequentist methods, and that the full Bayesian analysis is preferable to an empirical Bayes approximation when the number of different smear takers is small and at least one individual smear taker has contributed only a few smears. Plots of Bayesian posterior distributions with 95 per cent confidence sets are suggested as a useful way of feeding back the results to individual clinicians.