Robust extraction of covariate information to improve estimation efficiency in randomized trials

Area of Science:

• Biostatistics
• Clinical Trials
• Epidemiology

Background:

• Randomized trials often use unadjusted outcome estimates for causal inference.
• Covariate adjustment can enhance efficiency but faces reluctance due to precision concerns and potential manipulation.
• Existing methods like conditional logistic regression may provide subgroup-level rather than population-level effects.

Purpose of the Study:

• To propose a general methodology for covariate adjustment in two-arm randomized trials using targeted maximum likelihood estimation (TMLE).
• To address concerns regarding precision and model selection manipulation in covariate adjustment.
• To provide a criterion for assessing efficiency gains from covariate adjustment.

Main Methods:

• Utilized targeted maximum likelihood estimation (TMLE) for covariate adjustment in two-arm trials with 50% treatment probability.
• Compared TMLE-based estimates with conditional logistic regression models.
• Developed a criterion to determine efficiency gains over unadjusted methods.
• Illustrated the methodology with a resampled clinical trial dataset.

Main Results:

• The proposed TMLE methodology yields population-level treatment effect estimates.
• Demonstrated that covariate adjustment can lead to efficiency gains and increased statistical power, even for binary outcomes.
• Showcased improved precision compared to unadjusted methods in a real-world clinical trial example.

Conclusions:

• The proposed TMLE framework offers a robust method for covariate adjustment in randomized trials.
• Covariate adjustment, when appropriately applied using TMLE, enhances statistical power and provides more precise causal inferences.
• This approach overcomes previous limitations, enabling more efficient estimation of population-level treatment effects.