On exact randomization-based covariate-adjusted confidence intervals.

Area of Science:

• Statistics
• Biostatistics
• Experimental Design

Background:

• Randomization-based inference and Fisher randomization tests are valuable for small experiments with non-normal outcomes.
• Calculating confidence intervals via test inversion is computationally intensive, hindering practical application.
• Existing methods for closed-form confidence intervals exist for simple difference-in-means but not for covariate-adjusted analyses.

Purpose of the Study:

• To develop a closed-form expression for randomization-based covariate-adjusted confidence intervals.
• To provide a verifiable condition ensuring correct coverage for these confidence intervals.
• To overcome the computational burden associated with traditional methods for covariate-adjusted randomization inference.

Main Methods:

• Extending the work of Zhu and Liu to derive a closed-form expression for covariate-adjusted confidence intervals.
• Developing a sufficiency condition for correct coverage, checkable with observed data.
• Conducting simulations to evaluate the performance and robustness of the proposed method.
• Applying the method to re-analyze Phase I clinical trial data.

Main Results:

• The proposed method yields randomization-based covariate-adjusted confidence intervals with correct coverage.
• These intervals are robust to violations of normality assumptions.
• The computational time is comparable to calculating Fisher-exact P-values, significantly faster than test inversion.
• The method was successfully demonstrated on real-world clinical trial data.

Conclusions:

• A computationally feasible method for randomization-based covariate-adjusted confidence intervals is now available.
• This significantly lowers the barrier for applying robust statistical inference in small, randomized studies.
• The approach enhances the utility of randomization tests, particularly in biostatistical and experimental settings.