Local average treatment effects estimation via substantive model compatible multiple imputation.

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Causal Inference

Background:

• Nonadherence to assigned treatment is prevalent in randomized controlled trials (RCTs).
• Estimating the causal effects of treatment actually received, such as the local average treatment effect (LATE), is of growing interest.
• Traditional instrumental variables (IV) methods, including two-stage least squares (TSLS), face challenges with binary outcomes and causal odds ratios.

Purpose of the Study:

• To propose and evaluate a novel method for estimating causal treatment effects in RCTs with nonadherence.
• To address the limitations of existing methods, particularly for binary outcomes, by utilizing mixture models.
• To enhance the practical application of mixture models through a new multiple imputation technique.

Main Methods:

• Introduction of substantive model compatible multiple imputation (SMC MI) to impute latent compliance classes within mixture models.
• Application of maximum likelihood estimation to multiply imputed datasets, followed by combining results using Rubin's rules.
• Comparison of SMC MI performance against existing methods via simulations and reanalysis of a UK primary health RCT.

Main Results:

• SMC MI demonstrates superior performance compared to two-stage methods, particularly for binary outcomes.
• The proposed method shows increased efficiency over full Bayesian estimation when auxiliary variables are included.
• SMC MI effectively handles missing data in outcomes and other model variables.

Conclusions:

• SMC MI provides a more efficient and robust approach for estimating causal treatment effects in RCTs with nonadherence.
• This method overcomes practical limitations of mixture models, making them more accessible for researchers.
• The findings suggest SMC MI is a valuable tool for analyzing complex clinical trial data, especially with binary outcomes.