The Importance of Disease Prevalence in Transporting Clinical Prediction Rules

Abstract

Because clinical prediction rules often are applied in new settings to calculate the probability of a disease, we evaluated the accuracy of three rules for predicting streptococcal pharyngitis in 310 patients. Use of the rules led to overestimations of disease probability in 47%, 82%, and 93% of the patients. When we used receiver-operating characteristic curve analysis, no rule lost power to discriminate streptococcal from nonstreptococcal causes of pharyngitis. The overestimations in disease probability likely were caused by differences in disease prevalence between our setting (5%) and the settings in which they were developed (15% to 17%). All rules led to accurate predictions when they were adjusted for the disease prevalence found in our setting using a likelihood ratio formulation of Bayes' theorem. The value of prediction rules, like that of other diagnostic tests, is affected by differences in disease prevalence in different settings. Failure to recognize and adjust for these differences may cause poor decision making or the premature dismissal of valid rules.