A comparison of imputation strategies in cluster randomized trials with missing binary outcomes

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Epidemiology

Background:

• Cluster randomized trials (CRTs) involve randomizing groups, not individuals, posing unique challenges for data analysis.
• Missing outcome data is a common issue in CRTs, potentially biasing results.
• Effective strategies are needed to handle missing data in CRTs with binary outcomes.

Purpose of the Study:

• To evaluate different methods for handling missing binary outcome data in cluster randomized trials.
• To compare the bias and coverage rates of population-averaged intervention effect estimates across various imputation strategies.
• To assess the impact of these strategies on intracluster correlation coefficient (ICC) estimation.

Main Methods:

• Simulation study comparing complete case analysis, adjusted complete case analysis, simple imputation, and multiple imputation.
• Analysis of binary outcomes using logistic regression models (classical and random-effects).
• Evaluation of bias and coverage properties of intervention effect estimates and ICC estimates.

Main Results:

• Multiple imputation, using either random-effects or classical logistic regression, yielded unbiased intervention effect estimates.
• Both multiple imputation strategies demonstrated good coverage rates, with a slight advantage for the random-effects model.
• The random-effects logistic regression approach resulted in a less negatively biased ICC estimate compared to classical logistic regression.

Conclusions:

• Multiple imputation is a robust strategy for handling missing binary outcome data in cluster randomized trials.
• Random-effects logistic regression with multiple imputation provides reliable intervention effect estimates and improved ICC estimation.
• These findings are applicable to real-world trials, such as those investigating treatments for head lice.