A Comparison of Latent Mean Decomposition Approaches for Three-Level Cluster Randomized Trial Data

• In Event: Advances in Multilevel Modeling for Causal Inferences

Abstract

Multilevel models have critical importance when investigating a treatment effect and level-specific treatment by covariate interactions in a cluster-randomized-trial design. Covariate by treatment interactions can help applied researchers to better understand the nature of the intervention effect. However, investigating these interactions requires effort for the appropriate decomposition of the covariate. When working with reflective aggregations, this decomposition should be completed using expected cluster means rather than observed cluster means. The objective of this study is to compare approaches that employ expected means to form level-specific interactions in a three-level CRT design. An illustration and a small-scale simulation study to compare the two-stage adjustment procedure and the Bayes approach were provided. The simulation study indicated promising results with the two-stage approach.

Author

• Burak Aydin, Ege University