Clinically useful measures of effect in binary analyses of randomized trials

Abstract

The results of a randomized clinical trial can be reported using relative and/or absolute estimators of treatment effect. These various measures convey different information, and the choice can influence the physician's appreciation of the size of treatment effect and, subsequently, treatment decisions. We compare the estimators with respect to the clinically relevant information conveyed to physicians, and identify which clinical questions can and cannot be answered directly by each. We also identify opportunities for misinterpretation when one estimator is substituted for another, or when an estimator is mislabeled. Clinically important questions are addressed most directly by reporting both relative and absolute effects using relative risk and its complement, relative risk reduction, and risk difference and its reciprocal, number needed to treat. This is true of estimates of treatment effect derived from a single trial and also from meta-analysis of a group of trials. Because the control group's risk affects the numerical value of the odds ratio, the odds ratio cannot substitute for the risk ratio in conveying clinically important information to physicians. This is especially important when large treatment effects are shown in trials carried out in populations at high baseline risk.