Adjusted Win Ratio with Stratification: Calculation Methods and Interpretation

TL;DR

This paper develops a unified framework for estimating and interpreting the win ratio in stratified and adjusted settings, comparing it with traditional nonparametric tests for group differences.

Contribution

It introduces a comprehensive theory for win ratio estimation with stratification and adjustment, linking it to established nonparametric tests and enhancing interpretability.

Findings

Win ratio provides an interpretable treatment effect measure.

The proposed methods compare favorably with classical nonparametric tests.

The approach requires minimal distributional assumptions.

Abstract

The win ratio is a general method of comparing locations of distributions of two independent, ordinal random variables, and it can be estimated without distributional assumptions. In this paper we provide a unified theory of win ratio estimation in the presence of stratification and adjustment by a numeric variable. Building step by step on the estimate of the crude win ratio we compare corresponding tests with well known nonparametric tests of group difference (Wilcoxon rank-sum test, Fligner-Plicello test, Cochran-Mantel-Haenszel test, test based on the regression on ranks and the rank ANCOVA test). We show that the win ratio gives an interpretable treatment effect measure with corresponding test to detect treatment effect difference under minimal assumptions.