The Effects of Including Observed Means or Latent Means as Covariates in Multilevel Models for Cluster Randomized...

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Statistical Modeling

Background:

• Cluster randomized trials (CRTs) are efficient for interventions targeting groups.
• Increasing statistical power in CRTs is crucial for detecting treatment effects.
• Covariate adjustment in multilevel models can enhance CRT power.

Purpose of the Study:

• To compare methods of including covariates in two-level models for CRTs.
• To assess the impact of observed versus latent cluster means on statistical power.
• To evaluate the effect of Level 1 deviation scores on model performance.

Main Methods:

• Monte Carlo simulation study.
• Manipulation of key parameters: effect sizes, cluster sizes, number of clusters, intraclass correlation, missing data patterns, and covariate correlations.
• Comparison of multilevel models with observed cluster means, latent cluster means, and Level 1 deviation scores.

Main Results:

• No substantial differences between observed and latent mean models in convergence, Type I error, coverage, and bias.
• Potential coverage issues with latent means in small cluster sizes.
• Observed mean models performed comparably to latent mean models in power, but superiorly with small cluster sizes.

Conclusions:

• Both observed and latent cluster means are viable covariates in two-level models for CRTs.
• Observed cluster means offer a more robust approach, particularly with smaller cluster sizes.
• The findings inform optimal covariate inclusion strategies for maximizing power in CRTs.