New Q-Learner method tackles ratio-based treatment effects with doubly robust guarantees

Fuchs and Kreiss (arXiv:2605.26288) address a critical gap in causal inference: estimating ratio-based treatment effects (e.g., relative risk) as conditional average treatment effects (CATE). While existing approaches either force log-linear parametric forms or apply generic regression without robustness guarantees for the ratio functional, the authors propose the Q-Learner, which decomposes the ratio-CATE τ(x) into a product of two odds ratios, reducing the problem for binary outcomes to two propensity classification tasks. This avoids imbalanced regression issues that plague outcome-based estimators, particularly in low-conversion regimes.

To extend robustness, the paper derives doubly robust (DR) augmentations for both standard S/T-learners and the new Q-style ratio learners. Each DR variant has distinct robustness properties: the DR learners for S/T learners protect against misspecification of either outcome or propensity models, while the Q-Learner's DR version adds a second line of defense. Benchmarks on 7 RCT datasets show the Q-Learner is consistently the most competitive method in low-conversion scenarios. On 4 observational datasets where propensity must be estimated and confounding is present, the proposed DR learners decisively outperform all alternatives, making them a natural default for practitioners working with confounded observational data.