Pair-switching rerandomization

Area of Science:

• Statistics
• Clinical Trials
• Experimental Design

Background:

• Rerandomization improves estimation and inference efficiency by discarding unbalanced covariate assignments.
• Traditional rerandomization using acceptance-rejection sampling is computationally inefficient and time-consuming.
• This inefficiency hinders the generation of assignments needed for randomization tests and confidence intervals.

Purpose of the Study:

• To propose an efficient method for drawing balanced assignments in rerandomization.
• To validate the statistical properties and computational speed of the new method.
• To provide a faster approach for constructing confidence intervals via randomization tests.

Main Methods:

• Introduced a novel pair-switching rerandomization (PSRR) method for efficient generation of balanced assignments.
• Demonstrated the unbiasedness and variance reduction of the difference-in-means estimator under PSRR.
• Developed an exact inversion approach for Fisher randomization tests to confidence intervals.

Main Results:

• PSRR ensures the validity of Fisher randomization tests.
• The proposed confidence interval inversion method is faster than existing techniques.
• Simulation studies show PSRR is 3-23 times faster than classical rerandomization with comparable statistical power.
• PSRR was successfully applied to two clinical trial datasets, demonstrating its practical advantages.

Conclusions:

• Pair-switching rerandomization (PSRR) provides a computationally efficient and statistically valid alternative to classical rerandomization.
• The PSRR method accelerates the generation of balanced assignments, benefiting randomization tests and confidence interval construction.
• This approach is applicable to both sequential and non-sequential randomized experiments, enhancing its utility in diverse clinical trial settings.