Safety evaluation and confidence intervals when the number of observed events is small or zero

Abstract

A common objective in many clinical studies is to determine the safety of a diagnostic test or therapeutic intervention. In these evaluations, serious adverse effects are either rare or not encountered. In this setting, the estimation of the confidence interval (CI) for the unknown proportion of adverse events has special importance. When no adverse events are encountered, commonly used approximate methods for calculating CIs cannot be applied, and such information is not commonly reported. Furthermore, when only a few adverse events are encountered, the approximate methods for calculation of CIs can be applied, but are neither appropriate nor accurate. In both situations, CIs should be computed with the use of the exact binomial distribution. We discuss the need for such estimation and provide correct methods and rules of thumb for quick computations of accurate approximations of the 95% and 99.9% CIs when the observed number of adverse events is zero.