Bayesian adaptive designs for the optimisation of implementation strategies: a virtual re-execution of the Physically Active Children in Education (PACE) trials.

backgroundCurrent designs used to optimise implementation strategies (e.g., sequential cluster randomised controlled trials (cRCTs) and multi-arm cRCTs) are inefficient and resource intensive. Bayesian adaptive designs may offer a more efficient alternative.

methodsWe conducted a virtual trial re-execution to assess the impact of using a Bayesian adaptive design for optimising an existing implementation strategy (Physically Active Children in education (PACE)) used to support the delivery of school-based physical activity. The two previous, sequential, two-arm cRCTs used to optimise PACE were combined into a single three-arm cRCT. We assessed the performance of a fixed version of this trial design compared to an adaptive version incorporating one, two, or three interim analyses in a virtual re-execution. Adaptions included arm dropping and early stopping for futility, noninferiority, or efficacy.

resultsAll adaptive designs stopped early for noninferiority, declaring the lower cost treatment as the optimal arm at interim analysis one. This was the same conclusion obtained in the fixed version of the three-arm cRCT and the original sequential two-arm cRCTs. Using adaptive designs, the same conclusion was reached using 50-75% fewer clusters. The adaptive trials would have taken approximately 40% less time than the sequential two-arm cRCTs, not accounting for the time to analyse the interim analysis. However, the treatment effect was biased towards the lower cost treatment, and this bias (away from the full-sample posterior estimate) increased when fewer clusters were randomised. The first 18 schools randomised were Catholic schools, and they responded better to the lower cost treatment compared to government schools. When the distribution of school type was balanced at each interim (i.e. matched the final sample proportions) the bias towards the lower cost treatment was reduced.

conclusionsBayesian adaptive designs offer improved efficiency for trials aiming to optimise implementation strategies, reducing the time and sample size needed to find the optimal strategy. However, care is required to ensure confounding demographics are balanced at each interim analysis to reduce the risk of making a type 1 or 2 error or an incorrect adaptive design decision (e.g. dropping an effective arm, incorrectly stopping early). CLINICAL TRIAL NUMBER: Not applicable.