Cluster-Level Analyses to Estimate a Risk Difference in a Cluster Randomized Trial With Confounding Individual-Level...

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Epidemiology

Background:

• Cluster randomized trials (CRTs) are susceptible to recruitment bias, necessitating adjustment for individual-level confounders.
• Traditional cluster-level analyses for binary outcomes may not adequately address confounding in CRTs.
• Estimating risk differences in CRTs requires methods that account for both cluster and individual-level factors.

Purpose of the Study:

• To compare the performance of various analytical methods for estimating risk differences in two-arm parallel CRTs with individual-level confounders.
• To evaluate the impact of adjusting for individual-level versus both individual- and cluster-level covariates.
• To identify the most robust and unbiased method for CRT analysis under different scenarios, particularly with small numbers of clusters.

Main Methods:

• A simulation study was conducted to compare analytical methods including unadjusted (UN), two-stage procedure (TSP), G-computation (GC), and targeted maximum likelihood estimation (TMLE).
• The simulation focused on a two-arm parallel CRT design with binary outcomes, incorporating individual-level confounders and cluster-level covariates.
• Performance was assessed using bias, type I error rate, coverage rate, and relative error of the standard error.

Main Results:

• The unadjusted (UN) method exhibited bias.
• Two-stage procedure (TSP) methods were biased when a treatment effect was present and the number of clusters per arm was small.
• G-computation (GC) and Targeted Maximum Likelihood Estimation (TMLE) methods provided unbiased estimates; TMLE demonstrated superior performance and robustness, especially in scenarios with few clusters, avoiding convergence issues seen with GC.

Conclusions:

• Targeted Maximum Likelihood Estimation (TMLE) is recommended for analyzing cluster randomized trials (CRTs) with individual-level confounders, particularly when the number of clusters is small.
• Adjustment using only individual-level covariates in GC and TMLE generally yielded better performance than adjusting for both individual- and cluster-level covariates.
• TMLE offers a reliable and unbiased approach for risk difference estimation in CRTs, outperforming TSP and GC in challenging scenarios with limited clusters.