A practical application for the efficient use of historical data in randomized controlled design using systematic review data.

Abstract

Randomized controlled trials are the gold standard for evaluating interventions but are resource-intensive and impose considerable burden on participants. Historical control designs, which incorporate data from previous trials into new analyses, could reduce sample size requirements, but their practical application is limited. This study explored the feasibility of constructing a historical control using the Bayesian meta-analytic predictive approach based on data from a systematic review of metformin in type 2 diabetes mellitus. We extracted baseline and end point glycated hemoglobin values and participant characteristics from 25 randomized trials (36 comparisons). Using the RBesT package in the R software, meta-analytic predictive priors were constructed, and their contribution to new data was quantified using the effective sample size metric. Between-trial heterogeneity (τ) was initially set to the default conservative value (0.5) and subsequently estimated and replaced in the analysis using the output from the metafor package based on the reported study-level variances. Additional stratified analyses based on follow-up duration and publication year were conducted. Initial models yielded low effective sample size values (1-4), indicating minimal information borrowing from historical data. The presence of a high level of heterogeneity (τ = 0.55-0.85) appeared to limit the priors' influence. Further investigation into a stricter data set showed that four outlier studies reduced the estimated τ from the data to 0.0935, which increased effective sample size from 18 (conservative model) to 52 (data-derived model), while estimated treatment effects remained stable (-1.34% to -1.36% glycated hemoglobin). Between-trial heterogeneity (τ) estimation critically influences the informativeness of meta-analytic predictive-based historical control analysis. Despite having 25 trials with the same condition, the effective sample size was very low. Data-driven τ estimation substantially increased effective sample size without altering the treatment effect estimates and could represent a valid alternative to historical controls in clinical trial design compared with default conservative choices.