Sensitivity analysis in Bayesian clinical trials was underused and poorly reported: a systematic survey.

backgroundThe increasing adoption of Bayesian methods in clinical trials necessitates robust sensitivity analyses to validate their assumptions, particularly the prior distribution. However, the design, implementation, and interpretation of these analyses are not well characterized. The purpose of this study is to examine the design, conduct and interpret characteristics of sensitivity analyses in Bayesian clinical trials.

methodsWe systematically searched PubMed, Embase, and the Cochrane Central Register of Controlled Trials to identify Bayesian clinical trials published from database inception through December 2024. Two reviewers independently performed study selection and full-text screening. Data regarding sensitivity analyses were extracted in duplicate using standardized, pilot-tested data collection forms. Multivariable logistic regression was employed to examine the association between five variables with more likely reporting sensitivity analyses for prior information.

resultsAmong the 171 included trials, 103 had an accessible study protocol or statistical analysis plan (SAP), of which 36 (35.0%) trials with available protocols or SAPs failed to pre-specify any sensitivity analyses. Sensitivity analyses were reported in 91 (53.2%) trials, with four reporting at least one sensitivity analysis that contradicted the primary analyses. 38 (22.2%) interpreted the results of sensitivity analysis in the discussion section. Post-hoc sensitivity analyses were explicitly described in 16 trials, only five provided a rationale for these analyses. 24 (23.3%) reported sensitivity analyses not pre-specified in the protocol, 47 (45.6%) omitted pre-specified sensitivity analyses from their results, none of these provided a justification. Of the 122 trials that reported using prior distributions, 33 (27.0%) conducted sensitivity analyses on the priors. Among the 58 trials that used informative priors, 22 (37.9%) performed sensitivity analyses; correspondingly, 11 out of 64 trials (17.2%) with non-informative priors did so. Multivariable analysis indicated that both the use of informative priors and the availability of a protocol or SAP were significantly associated with an increased likelihood of conducting sensitivity analyses for priors.

conclusionsThis study reveals gaps in the design, conduct, and reporting of sensitivity analyses for Bayesian trials, especially concerning priors. There is a pressing need to improve and standardize the use of sensitivity analyses in Bayesian clinical trials.