Outcome switching in cohort studies of interventions: meta-epidemiological study.

Abstract

To study the prevalence and characteristics of outcome switching, the completeness of outcome prespecification, and factors associated with outcome switching in observational cohort studies of interventions. Longitudinal meta-epidemiological study. Registry records and journal publications. Controlled cohort studies investigating the effects of interventions. Eligible studies were registered on ClinicalTrials.gov within one month of their start date (2014-16) and had published results in peer reviewed journals by 2024. Firstly, proportion of studies with outcome switching identified by comparing the prespecified outcomes in the registry and those reported in the journal publication of results. Discrepancies were categorised as omission (prespecified primary outcomes not reported), downgrading (prespecified primary outcomes reported as non-primary), upgrading (prespecified non-primary outcomes reported as primary), and introduction of new primary outcomes (not registered as an outcome). Secondly, proportion of studies with completely prespecified primary outcomes, defined as registry entries that include the measurement variable, analysis metric, method of aggregation (the statistic summarising the outcome within each study group), and time point. Of 9965 registration records screened, 124 eligible studies with results published between 2015 and 2024 were included. Only 30 studies (24%) completely prespecified their primary outcomes. Outcome switching occurred in 60 (48%) studies, but only two provided an explanation. The most common types of switching were omission (n=32, 26%) and downgrading (n=32, 26%), followed by the introduction of new primary outcomes (n=25, 20%), and upgrading (n=2, 2%). Among 57 studies with outcome switching other than omission (ie, outcome results were reported), statistically significant results were favoured in 77% (44/57) by introducing or upgrading a new significant primary outcome or downgrading a non-significant one. No study characteristics were significantly associated with outcome switching in multivariable logistic regression. Outcome switching and inadequate outcome prespecification were common in cohort studies of interventions. Most changes were unexplained and favoured statistically significant results, raising concerns about potential selective reporting and highlighting the need for improved transparency in outcome reporting. Open Science Framework (https://osf.io/xn5zt/).