Internal pilot sample size re-estimation for the B-free trial: A case study of a cluster randomized crossover trial

The intracluster correlation coefficient (ICC) is a critical parameter for determining sample size in cluster-randomized trials. However, accurate estimation of the ICC values are often challenging due to limited availability from participating sites. Even small deviations from the assumed ICC can result in underpowered studies.

We used simulations to evaluate the performance of blinded internal pilot sample size re-estimation (SSRE) on the B-Free trial, a cluster crossover trial with 20 hospitals and 12 planned 4-weeks periods with postoperative delirium as primary outcome. The SSRE process included: (1) calculating the initial sample size, (2) collecting interim data, (3) estimating the ICC, (4) re-estimating the sample size, and (5) increasing the number of cluster-periods if needed.

Across 1,000 simulations, interim ICC estimates closely matched true values when 25%, 50%, and 75% of the data were available, although variability was higher at 25%. Statistical power was restored to 80% and the Type I error was maintained at 5% when uncertainty in the ICC was modest (e.g., assumed 0.02 vs. observed 0.05). In the B-Free trial, the observed ICC at approximately 70% data availability closely approximated the final value. To achieve 81% power, the trial was extended by 5-6 periods at half of the participating sites.

With at least 50% of interim data, ICC estimates are generally reliable, and SSRE preserves Type I error rate. Incorporating an internal pilot into the design of cluster crossover trials is an effective strategy to maintain statistical power when ICCs are modestly misspecified.