DOI: 10.12688/f1000research.189494.1 ISSN: 2046-1402

Composite endpoints sharing a name do not always share a definition — a registry-based audit of five systematic reviews, and why heterogeneity statistics do not detect the problem

Mahmood Ahmad, Awab Ahmad, Rida Aleem
Show PDF Cite
Background Systematic reviews are assembled from included studies identified by citation, and then match outcomes using their reported names. However, both steps assume a name can identify a quantity. We decided to test that assumption against trial registrations — this is where outcome definitions are recorded before results. Methods For five separate reviews across three drug classes, we searched ClinicalTrials.gov and PubMed. We recorded each query verbatim with its date. We exhausted pagination and reported recall against the review’s own included set. Each study was then resolved by registration identifier rather than citation string. For each included trial, we read every registered outcome at every rank and then decomposed each composite into its individual components. Pools were formed only where the components matched. In cases where they did not match, the declined pool was computed and published beside the refusal. Results Trials described as reporting the same composite often counted different events. For example, eleven apixaban trials in one indication carried five distinct primary component sets. Two separate trials from a single development programme, with names differing by a hyphenated suffix, counted VTE-related death and all-cause death respectively. Four thromboprophylaxis trials showed the primary event rate ranging from 1.39% to 8.99%. Heterogeneity statistics did not detect these mismatches, and incoherent pools returned I 2 of 0.0%, 3.9% and 83.6%. In one review, the coherent pool was the more heterogeneous of the two. Reading the outcomes below the primary moved the poolable set from three to eight on one review, and from eight to three on another. In a third, it left it unchanged. Conclusions Composite endpoints which share a name frequently do not share a definition, and no heterogeneity statistic reveals it. Coherence must then be established by reading definitions and comparing components.

More from our Archive