The number needed to treat (NNT) index has been proposed as a clinically useful measure to assess the results of randomized trials and other clinical studies. In its usual form, NNT indicates the expected number of patients who must be treated with an experimental therapy in order to prevent one adverse event, compared to the expected event rates under the control therapy. It can be formulated as a function of the proportions of patients who respond to treatment by more than a certain amount, the clinically important difference. We may also wish to evaluate two group studies comparing treatment and control responses, and to consider net benefit from treatment (by also allowing for individuals who deteriorate as well as those who respond positively). In this paper, we investigate the effect on NNT caused by measurement errors in continuous outcome measures. Such errors can lead to bias in the estimated proportions of subjects with clinically important responses, and hence bias the associated values of NNT. General expressions for the bias are derived, and enumerated for typical scenarios. For many situations, reliability of 80 per cent or more in the observations is required to restrict the bias to tolerable levels. Copyright © 2001 John Wiley & Sons, Ltd.