Accuracy of the Number Needed to Treat Compared With Diagnostic Testing

Abstract

Context.— Number needed to treat (NNT) seems simple and is widely used. But its derivation from the absolute risk reduction is difficult to visualize. Like diagnostic sensitivity, absolute risk reduction is a measure of treatment accuracy. Thus, NNT is a measure of accuracy. NNT is inversely proportional to the relative risk reduction and the baseline risk that may be torturous when accuracy is poor. In order to better visualize the accuracy and weaknesses of NNT, NNT is compared with variables in diagnostic science, such as diagnostic sensitivity, specificity, and positive predictive value, that are often better understood. Objective.— To better understand the accuracy of NNT. Data Sources.— Receiver operating characteristic curves are used to help visualize accuracy. It is shown that baseline risk and prevalence are highly correlated. Like positive predictive value, NNT is dependent on prevalence. Similar to diagnostic testing, symptoms and additional testing can increase prevalence and improve accuracy of the NNT. Examples are shown where changes in prevalence cause alterations in NNT. Moreover, data indicate that when accuracy of NNT is low, although the average NNT may be favorable, some subgroups may exhibit very poor response and even harm. It is shown that manipulations to increase prevalence can produce smaller, more selective groupings that can improve the accuracy and reduce the cost of expensive drugs. Conclusion.— When the power of prediction is poor, the value of NNT must be carefully deliberated because it may be misleading. Indeed, the upper confidence interval may be a better reflection of NNT than the average.