Causal inference in randomized trials with partial clustering

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Epidemiology

Background:

• Participant dependence, known as clustering, requires careful consideration in randomized trial analysis.
• Clustering can occur within one or multiple trial arms and may arise before or after randomization.
• This study examines three trial designs: fully clustered and two partially clustered variations.

Purpose of the Study:

• To develop and evaluate statistical methods for analyzing randomized trials with participant dependence.
• To introduce a novel implementation of targeted minimum loss-based estimation (TMLE) for clustered trial data.
• To compare the performance of TMLE against alternative methods in various clustered trial designs.

Main Methods:

• Utilized causal models to describe data generation and formalize dependence structures.
• Developed a novel targeted minimum loss-based estimation (TMLE) approach for analysis.
• Conducted simulation studies to assess finite-sample performance and applied methods to the SEARCH-IPT trial data.

Main Results:

• Identified identical dependence structures for the two partially clustered trial designs, allowing unified statistical methods.
• Demonstrated that TMLE, incorporating covariate adjustment and machine learning, enhances precision and estimates a broad range of causal effects.
• Simulations showed TMLE achieved comparable or superior statistical power compared to alternatives for partially clustered designs.
• Application to the SEARCH-IPT trial yielded 20%-57% efficiency gains, highlighting practical benefits.

Conclusions:

• Partially clustered trial analysis can be significantly improved using targeted minimum loss-based estimation (TMLE).
• Properly accounting for data dependence is essential for efficient and accurate causal effect estimation in clustered trials.