Neyman-Orthogonal Rank-Learner

• The paper introduces a novel two-stage algorithm that directly ranks individual treatment effects using a pairwise loss function enhanced by Neyman-orthogonality.
• It employs cross-fitted nuisance estimators and influence-function corrections to achieve robustness against errors in estimating treatment and outcome models.
• Empirical evaluations show improved ranking performance and error resilience compared to traditional CATE estimators, especially in noisy or limited data scenarios.

The Neyman-orthogonal Rank-Learner is a model-agnostic, two-stage algorithm for ranking individuals by their treatment effects using observational data. Unlike traditional approaches that focus on precise estimation of the conditional average treatment effect (CATE), Rank-Learner directly targets the ranking problem through a pairwise, orthogonalized loss. This construction provides robustness to nuisance parameter estimation errors via Neyman-orthogonality, facilitating improved performance in practical, data-limited, or noisy nuisance estimation scenarios (Arno et al., 3 Feb 2026).

1. Problem Formulation and Motivation

Given $n$ i.i.d. samples $W_i = (X_i, T_i, Y_i)$ where $X \in \mathcal{X} \subset \mathbb{R}^d$ are covariates, $T \in \{0,1\}$ denotes binary treatment, and $Y \in \mathbb{R}$ the observed outcome, the goal is to rank individuals by their potential treatment effect. The problem is specified in the potential-outcome framework: each unit has unobserved $Y(0)$ and $Y(1)$, with the conditional average treatment effect (CATE) defined as $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$.

Standard identification assumptions are imposed:

1. Consistency: $Y = Y(T)$.
2. Unconfoundedness: $\{Y(0), Y(1)\} \perp\!\!\!\perp T \mid X$.
3. Overlap: $W_i = (X_i, T_i, Y_i)$0 for all $W_i = (X_i, T_i, Y_i)$1, with $W_i = (X_i, T_i, Y_i)$2.

Under these, $W_i = (X_i, T_i, Y_i)$3 admits the identification $W_i = (X_i, T_i, Y_i)$4, where $W_i = (X_i, T_i, Y_i)$5.

The objective is to induce a real-valued score function $W_i = (X_i, T_i, Y_i)$6 such that $W_i = (X_i, T_i, Y_i)$7 whenever $W_i = (X_i, T_i, Y_i)$8. Any strictly increasing transformation $W_i = (X_i, T_i, Y_i)$9 suffices. In contrast to MSE-based CATE estimation, which enforces $X \in \mathcal{X} \subset \mathbb{R}^d$0 pointwise, ranking only requires correct orderings.

2. Pairwise Ranking Objective and Learning Strategy

To operationalize rank learning, the Rank-Learner employs a pairwise learning objective based on the following constructs:

• Model’s pairwise preference: $X \in \mathcal{X} \subset \mathbb{R}^d$1 where $X \in \mathcal{X} \subset \mathbb{R}^d$2 is the logistic sigmoid.
• True pairwise preference: $X \in \mathcal{X} \subset \mathbb{R}^d$3.

The corresponding population-level pairwise risk is

$X \in \mathcal{X} \subset \mathbb{R}^d$4

with $X \in \mathcal{X} \subset \mathbb{R}^d$5 denoting the binary cross-entropy. In practice, a smooth surrogate target is used: $X \in \mathcal{X} \subset \mathbb{R}^d$6 yielding the loss

$X \in \mathcal{X} \subset \mathbb{R}^d$7

Any $X \in \mathcal{X} \subset \mathbb{R}^d$8 minimizes $X \in \mathcal{X} \subset \mathbb{R}^d$9 arbitrarily well; thus, the method only requires order preservation, not exact value learning.

3. Neyman-Orthogonality and Nuisance Correction

The approach estimates the nuisance vector $T \in \{0,1\}$0 using cross-fitted machine learning models. Plug-in approaches—replacing $T \in \{0,1\}$1 with $T \in \{0,1\}$2—result in first-order sensitivity to nuisance estimation errors.

Rank-Learner overcomes this via influence-function correction. The orthogonal pairwise loss is: $T \in \{0,1\}$3 where

$T \in \{0,1\}$4

with

$T \in \{0,1\}$5

and

$T \in \{0,1\}$6

where the doubly robust score

$T \in \{0,1\}$7

A key result is Neyman-orthogonality: for all perturbations $T \in \{0,1\}$8 and $T \in \{0,1\}$9, $Y \in \mathbb{R}$0 (Theorem 1). This ensures first-order insensitivity of the ranking-stage loss to nuisance estimation errors.

The population minimizer (Theorem 2) takes the form $Y \in \mathbb{R}$1, preserving correct ranking.

4. Computational Procedure

The Rank-Learner algorithm proceeds in two explicit stages:

1. Nuisance Estimation (Stage 1): Apply cross-fitting over $Y \in \mathbb{R}$2 splits to estimate $Y \in \mathbb{R}$3, $Y \in \mathbb{R}$4, $Y \in \mathbb{R}$5 on held-out folds using flexible regressors (neural networks, trees, forests).
2. Orthogonal Ranking (Stage 2): Using cross-fitted nuisance estimators, initialize $Y \in \mathbb{R}$6 in a differentiable hypothesis class $Y \in \mathbb{R}$7. In each optimization epoch:
   • Randomly sample a subset of $Y \in \mathbb{R}$8 unit pairs ($Y \in \mathbb{R}$9, typically $Y(0)$0–$Y(0)$1).
   • For each pair, compute the model’s predicted pairwise preference, the soft target using $Y(0)$2, $Y(0)$3, and the doubly robust pseudo-label $Y(0)$4 as above.
   • Update $Y(0)$5 via gradient steps to minimize the average loss over the sampled pairs.

Inference is performed by applying $Y(0)$6 to new instances.

Stage	Description	Typical Tools
Nuisance Estimation	Cross-fit $Y(0)$7, $Y(0)$8 over $Y(0)$9 folds	Neural nets, trees
Orthogonal Ranking	Pairwise, loss-minimizing $Y(1)$0	Any autodiff learner

The pairwise objective's computational complexity per-epoch is $Y(1)$1, which motivates aggressive pair subsampling and mini-batching.

5. Theoretical Properties

Key assumptions include unconfoundedness, overlap, boundedness of $Y(1)$2, and fixed $Y(1)$3. The theoretical findings include:

• Neyman-Orthogonality (Theorem 1): The cross second derivative of $Y(1)$4 with respect to nuisance and ranking functions vanishes at truth, yielding first-order insensitivity to nuisance estimation error.
• Population Minimizer (Theorem 2): Any function $Y(1)$5 minimizes $Y(1)$6, ensuring correct ranking is preserved.
• Excess Risk Convergence: If $Y(1)$7 converges at rate $Y(1)$8 in $Y(1)$9 and $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$0 at $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$1, the excess orthogonal risk converges at $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$2. The ranking risk $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$3 can be bounded by $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$4.
• Sign Consistency and Ranking Error: With $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$5 or better nuisance estimation and controlled $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$6-class complexity, sign-consistency and fast error rate for ranking are attained.

A plausible implication is rapid, robust consistency of rankings even in imperfect nuisance learning regimes.

6. Empirical Evaluation

Benchmarks use both synthetic (10-dimensional normals, nonlinear CATE) and semi-synthetic real covariate datasets (MovieLens, MIMIC-III, CPS) with simulated outcomes. Baselines include T-learner, doubly robust DR-learner, non-orthogonal plug-in rankers, and tree-based rankers from prior work.

Metrics include:

• AUTOC (area under the targeting-operator curve, the principal evaluation metric),
• Kendall's $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$7, and
• Normalized DCG, as well as mean policy value.

Findings show Rank-Learner:

• Outperforms T-learner and DR-learner in small-sample ($$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$8) regimes,
• Demonstrates robustness over non-orthogonal plug-in rankers in high-nuisance-noise settings,
• Yields improvements across all semi-synthetic datasets,
• Never underperforms oracle as $$\tau(x) \coloneqq \mathbb{E}[Y(1) - Y(0) \mid X = x]$$9 grows large; all methods converge.

7. Implementation and Practical Considerations

Nuisance regression should leverage modern, flexible learners, with cross-fitting mandatory ($Y = Y(T)$0). For the ranking stage:

• Select $Y = Y(T)$1 to balance ranking fidelity and variance control, tuning via out-of-sample AUTOC.
• Initial pair-subsampling rate $Y = Y(T)$2–$Y = Y(T)$3 is effective.

General-purpose autodiff frameworks (PyTorch, TensorFlow) support gradient-based optimization and vectorized computation of pseudo-labels. Scalability considerations motivate efficient batching since pairwise objective evaluation is computationally intensive for large $Y = Y(T)$4.

The Rank-Learner directly targets the ranking of treatment effects, eschewing the harder MSE pointwise CATE estimation, delivers Neyman-orthogonality for robustness to nuisance misestimation, and applies flexibly across nonparametric base learners. Empirical evidence demonstrates uniform improvement in ranking metrics compared to CATE estimators and non-orthogonal rankers (Arno et al., 3 Feb 2026).