Statistical Power and Optimal Design for Multisite and Cluster-Randomized Studies With Outcome Unreliability

• In Event: 31.038 - Power Analysis in Multilevel Modeling

Abstract

We derived power and optimal sample allocation formulas for two-level multisite and cluster-randomized studies when measurement error reduces the reliability of the outcome. We then conducted a series of analyses using the new formulas with varying degrees of measurement error and different cost structures to determine the effects, if any, on optimal sample allocation, power, and minimum detectable effect size. The results suggest that in both designs outcome unreliability introduces additional uncertainty regarding treatment effects, reduces the power to detect such effects, and consistently inflates the optimal individual sample size regardless of sampling cost ratio and variance component structures. However, the magnitude of these consequences was dependent on these parameters.

Authors

• Kyle T. Cox, University of Cincinnati

• Benjamin Kelcey, University of Cincinnati