Robustness of Optimal Designs to Misspecification of Intraclass Correlations for Three-Level Cluster-Randomized Trials in Education

• In Event: 35.034 - Some Sharp Statistics

Abstract

When designing a cluster-randomized trial (CRT), misspecification of an intraclass correlation (ICC) is common because the true value is unknown until after the experiment concludes. An efficiency analysis is conducted to understand the robustness of an optimal design for a three-level CRT in education to misspecifications of one or both ICCs, and changes in costs. Findings suggest an optimal design is reasonably robust to misspecification of one or both ICCs. High relative efficiency (≥0.90) is maintained for misspecification of the Level 3 ICC between 60-70% under-specification to 93-240% over-specification, depending on the true value. The ranges are wider, for the Level 2 ICC. Small misspecification of the other ICC, and changes in the cost of additional participants have little impact.

Author

• Carl Westine, The University of West Georgia