Conditional Monte Carlo randomization tests for regression models

Area of Science:

• Biostatistics
• Clinical Trial Design
• Statistical Inference

Background:

• Randomization tests are crucial for clinical trials to ensure valid statistical inference.
• Existing methods often rely on specific assumptions about the outcome distribution.
• The seminal work by Gail, Tan, and Piantadosi (1988) laid the foundation for randomization tests in regression settings.

Purpose of the Study:

• To describe and compute design-based randomization tests for clinical trials with regression-based outcomes.
• To extend randomization test methodology to longitudinal data analysis.
• To evaluate the performance of these tests under model misspecification.

Main Methods:

• Utilizing Monte Carlo generation of randomization sequences for design-based tests.
• Applying techniques for conditional randomization tests using residuals from generalized linear models and survival models.
• Developing a novel method for longitudinal data using generalized linear mixed models to predict the rate of change.

Main Results:

• The proposed Monte Carlo procedure generates design-based randomization tests that incorporate the specific randomization scheme.
• A new technique simplifies the computation of conditional randomization tests.
• Randomization tests for longitudinal data based on predicted rates of change were successfully developed.
• Simulations demonstrated that these randomization tests effectively preserve statistical size and power, even when the underlying statistical model is misspecified.

Conclusions:

• Design-based randomization tests offer a robust approach for analyzing clinical trial data, particularly when regression models are employed.
• The developed methods extend the applicability of randomization tests to longitudinal data, addressing a key limitation.
• These randomization tests are reliable and perform well under various conditions, including model misspecification, enhancing their utility in clinical research.