Composite Endpoints for Clinical Trials

Composite endpoints are commonly used in randomised controlled clinical trials because they offer potential advantages, such as smaller sample sizes, shorter completion times and the summary measure for a treatment effect if more than one outcome is important. However, they are also associated with risks and pitfalls, particularly if basic clinical and statistical requirements are not adequately respected. This article focuses on time-to-event composite endpoints that combine fatal and nonfatal events and aim to demonstrate efficacy of a treatment in long-term trials. The most important issues with the selection of individual endpoint components — the trial conduct, data analysis, interpretation and reporting — are discussed. Composite endpoints should be clinically meaningful and the expected effects on each component should be similar, based on biological plausibility. Surrogates as endpoint components must have been validated, i.e. a correlation to a hard clinical outcome must have been confirmed. Patients need to be followed up on assigned treatment (clinical status permitting) until death or end of planned follow-up in the absence of events, and must not be regarded as ‘a trial completer’ after occurrence of the first component event. Difficulties in interpretation arise when the results on single components of the composite endpoint go in opposite directions, and when hard clinical outcomes are combined with soft endpoints, particularly if the latter occur much more frequently but are of inferior relevance. Accordingly, all individual components of the composite need to be analysed separately. In this review, several examples from recently published trials are used to illustrate both the principal needs and the inherent problems of using a composite endpoint. More specific regulatory guidelines and better reporting standards are needed. The ultimate goals of a trial with a composite endpoint should be to solve a medical problem and to support clinical decision making, rather than just to provide statistical convenience.