Assurance in Intervention Research: A Bayesian Perspective on Statistical Power

Objective: This article introduces Bayesian assurance as an alternative to traditional power analysis in intervention research. Bayesian assurance is defined as the unconditional probability of identifying an intervention effect. Method: Assurance can be calculated as the expected statistical power based on a prior distribution of the unknown parameters related to the effect size. Using Monte Carlo simulation methods, we demonstrate Bayesian assurance in 2 small-scale randomized trials: a trial of motivational interviewing for patients with behavioral health disorders and a trial of a specialty mental health probation. Results: The findings suggest that traditional statistical power is highly sensitive to misspecification. Because assurance can be calculated across all possible effect sizes, it controls the uncertainty associated with the selection of a point effect size in traditional power estimation. Assurance usually produces larger sample-size estimates, and thus cutoff values for assurance may be lower than those typically used in classical power estimation. Conclusions: Compared to traditional power estimation, assurance appears to be more robust against inaccurate prior information. Assurance may be a preferred method for estimating sample sizes when prior information is poor and the costs of underpowering a study are great.