Robustness of response-adaptive randomization.

Area of Science:

• Biostatistics
• Clinical Trial Design
• Statistical Modeling

Background:

• Response-adaptive randomization, like Doubly adaptive biased coin design (DBCD), adjusts subject allocation based on responses.
• Existing research on DBCD assumes correct model specification, but its performance under misspecification is less understood.

Purpose of the Study:

• To evaluate the robustness of Doubly adaptive biased coin design (DBCD) to both design and analysis model misspecification.
• To assess the impact of misspecified regression models on treatment effect estimation and inference within DBCD.

Main Methods:

• Assessed the theoretical properties of allocation proportions under design model misspecification.
• Investigated three linear regression models (difference-in-means, ANCOVA I, ANCOVA II) for treatment effect estimation with arbitrarily misspecified analysis models.
• Derived consistency and asymptotic normality for treatment effect estimators.

Main Results:

• Allocation proportions in DBCD maintain consistency and asymptotic normality even with design model misspecification.
• Consistency and asymptotic normality of treatment effect estimators are preserved across misspecified regression models.
• The ANCOVA II model, incorporating covariate-by-treatment interactions, provides the most statistically efficient estimator.

Conclusions:

• Doubly adaptive biased coin design (DBCD) demonstrates robustness to model misspecification in both design and analysis.
• The findings support the use of DBCD in real-world scenarios where model assumptions may not perfectly hold.
• ANCOVA II is recommended for its superior efficiency in treatment effect estimation within misspecified DBCD frameworks.