The effect of dependence on the performance of Bayes' theorem: an evaluation using a computer simulation

Abstract

A computer simulation is described which generates 'cases' of vaginal discharge. This simulation was used to evaluate the potential effects of dependence between clinical features on the diagnostic performance of Bayes' theorem. The following observations were made: (1) dependence between some but not all pairs of features reduced the overall diagnostic efficiency (judged by the number of true positive diagnoses); (2) the overall reduction in efficiency was never substantial; (3) the largest effects on the diagnosis of individual conditions was observed with rarer diseases, and with combinations of features which were inherently unlikely to be dependent. It is concluded that the diagnostic efficiency of Bayes' theorem will not be greatly influenced by dependence if a reasonable amount of commonsense is applied to the selection of the knowledge base.