Reproduce “Orthogonal machine learning for conditional odds and risk ratios” on OpenResearch

About this paper

Jiacheng Ge, Iván Díaz

Researchers from NYU Grossman School of Medicine developed new orthogonal machine learning estimators for conditional odds and risk ratios, generalizing DR-learner and R-learner frameworks to these multiplicative causal effects using efficient influence functions. Through extensive simulations and a real-world application, these methods demonstrated improved accuracy in complex settings with heterogeneous effects, often outperforming traditional parametric and plug-in approaches.

Key insights

• The derived pseudo-outcomes for conditional OR and RR satisfy conditional-mean identities with second-order remainder properties, formally establishing the Neyman orthogonality of their associated squared losses.
• Orthogonal machine learning estimators (DR-LOR, DR-LRR) offer improved accuracy and robustness, particularly by reducing bias, in complex data-generating scenarios with high treatment effect heterogeneity and sufficient sample size.
• While beneficial in complex settings, the variance penalty for these debiased estimators can make simpler plug-in methods (like SuperLearner) more favorable in simpler or data-limited regimes.

Summary by alphaXiv.