Interim Analyses: The Repeated Confidence Interval Approach

SUMMARY Most medical trials are monitored for early evidence of treatment differences or harmful side-effects and many sequential methods have been proposed for this. Similarly, data collected in a prospective epidemiological study are likely to be reviewed periodically in the course of the study. The repeated confidence interval approach, which combines aspects of sequential estimation and testing, allows a full exploration of the data at each interim analysis and does not depend on a rigidly enforced statistical stopping rule. In this paper we present the general principles underlying the construction of repeated confidence intervals and describe how they can be used in reaching a decision to terminate a study early. We discuss design considerations, which depend on the form of early stopping anticipated, and explain how the basic method can be adapted to cope with the problems of unpredictable group sizes or, more generally, unequal increments in information between analyses. Extensions of the method to handle survival data, categorical data, normal responses with unknown variance and multivariate normal observations are also presented.