TRIUMPH-1 extension data: what happens after the trial stops

TRIUMPH-1 met its primary endpoint and did not meet one of its hierarchically tested secondary endpoints, a result which appears in the paper and in almost none of the reporting. Under a pre-specified testing hierarchy a failure at that point formally precludes claims on the endpoints below it, whatever the nominal p-values show.

The endpoint hierarchy was pre-specified in the statistical analysis plan and is reproduced in the supplementary appendix. Readers relying on the abstract will not see it. The Journal reads the analysis plan before the paper for exactly this reason.

Randomisation was stratified by the usual variables. Of those randomised to active treatment, roughly 7% discontinued before the end of the trial, and the reasons are reported: gastrointestinal adverse events predominate, followed by withdrawal of consent and loss to follow-up. How those participants are handled is the difference between the two estimands and therefore the difference between the two headline figures.

Cross-trial comparison is the commonest error in coverage of this field. Populations differ in baseline weight, diabetes status and lifestyle-intervention intensity; durations differ; estimands differ. Two programmes reporting 15% and 20% may or may not reflect a difference between the drugs, and only a head-to-head trial settles it.

"The mean is not the interesting part and it never has been," said Dr Éliane Casgrain, health economist, Université de Montréal. "Show me the decile plot. That is where the clinical question lives."

The full publication is in the Journal’s trial-data explorer with the estimand, the comparator, the duration and the withdrawal rate recorded alongside the headline figure.