MOVER confidence intervals for a difference or ratio effect parameter under stratified sampling

Area of Science:

• Biostatistics
• Clinical Trials
• Statistical Methods

Background:

• Stratification is crucial in clinical trials to minimize covariate imbalance and enhance treatment effect precision.
• Existing methods for confidence intervals (CIs) in stratified analyses may lack optimal precision or applicability across diverse endpoints.

Purpose of the Study:

• To propose a general framework for constructing confidence intervals (CIs) for difference or ratio effect parameters under stratified sampling using the method of variance estimates recovery (MOVER).
• To adapt MOVER for both additive variance and additive CI approaches for differences, and for ratio parameters using Fieller and log-ratio methods.
• To demonstrate the applicability of MOVER CIs across various endpoints, including binary and time-to-event outcomes.

Main Methods:

• Developed a general framework using the method of variance estimates recovery (MOVER) for stratified sampling.
• Applied additive variance and additive CI approaches for difference parameters, and Fieller and log-ratio methods for ratio parameters.
• Illustrated methods with real-world examples for binary outcomes (risk difference, risk ratio) and time-to-event outcomes (restricted mean survival time, milestone survival).

Main Results:

• Proposed MOVER confidence intervals (CIs) generally outperform standard large sample CIs in stratified analyses.
• The additive CI approach demonstrated superior performance compared to the additive variance approach.
• MOVER methods require only point estimates, CIs, and variance estimates from stratum-specific data, facilitating broad application.

Conclusions:

• The MOVER framework provides a flexible and effective approach for constructing confidence intervals in stratified clinical trials.
• MOVER CIs offer improved precision and performance over traditional methods, particularly the additive CI approach.
• The proposed methods are broadly applicable to various endpoints and facilitate robust statistical inference in stratified analyses.