Bayesian methods for pilot studies.

Background/aims: The use of pilot studies to help inform the design of randomized controlled trials has increased significantly over the last couple of decades. A pilot study can provide estimates of feasibility parameters, such as the recruitment, compliance and follow-up probabilities. The use of frequentist confidence intervals of these estimates fails to provide a meaningful measure of the uncertainty as it pertains to the design of the associated randomized controlled trial. The objective of this article is to introduce Bayesian methods for the analysis of pilot studies for determining the feasibility of an associated randomized controlled trial. Methods: An example from the literature is used to illustrate the advantages of a Bayesian approach for accounting for the uncertainty in pilot study results when assessing the feasibility of an associated randomized controlled trial. Vague beta distribution priors for the feasibility parameters are used. Based on the results from a feasibility study, simulation methods are used to determine the expected power of specified recruitment strategies for an associated randomized controlled trial. Results: The vague priors used for the feasibility parameters are demonstrated to be considerably robust. Beta distribution posteriors for the feasibility parameters lead to beta-binomial predictive distributions for an associated randomized controlled trial regarding the number of patients randomized, the number of patients who are compliant and the number of patients who complete follow-up. Ignoring the uncertainty in pilot study results can lead to inadequate power for an associated randomized controlled trial. Conclusion: Applying Bayesian methods to pilot studies' results provides direct inference about the feasibility parameters and quantifies the uncertainty regarding the feasibility of an associated randomized controlled trial in an intuitive and meaningful way. Furthermore, Bayesian methods can identify recruitment strategies that yield the desired power for an associated randomized controlled trial.