Neymanian Framework in Causal Inference

• The Neymanian framework is a potential outcome-based approach to causal inference that defines individual and average treatment effects.
• It relies on key assumptions—consistency, SUTVA, positivity, and ignorability—to enable robust estimation across randomized and observational designs.
• Extensions like Neyman orthogonality and semiparametric methods integrate machine learning to debias estimators and improve finite-sample inference.

The Neymanian framework, also known as the Neyman–Rubin potential outcome framework, is a foundational approach for causal inference grounded in hypothetical contrasts between potential outcomes that each unit would exhibit under different treatment regimes. It formalizes causality as a missing data problem, precisely quantifies exchangeability and irreducibility via randomization or ignorability assumptions, and supplies robust estimation theory via repeated-sampling arguments. Its scope extends from classical randomized trials to complex observational designs, multifactor experiments, and machine learning–integrated semiparametric settings.

1. Formal Structure: Potential Outcomes, Assumptions, and Estimands

At the core of the Neymanian framework is the definition of unit-level potential outcomes and their contrasts. For a binary treatment $A_i \in \{0,1\}$ on unit $i$, define

$$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$

Only one of $Y_i(1), Y_i(0)$ is ever observed per unit; the other is counterfactual. The fundamental estimands are

• Individual causal effect: $\delta_i = Y_i(1) - Y_i(0)$,
• Average treatment effect (ATE): $\tau = E[Y(1) - Y(0)]$.

Empirical identification of $\tau$ rests on four assumptions (Gorbach et al., 9 Dec 2025):

1. Consistency: $Y_i = Y_i(A_i)$.
2. Stable Unit Treatment Value Assumption (SUTVA): no hidden variations in treatments and no interference between units.
3. Positivity (Overlap): $0 < P(A_i=1 \mid C_i=c) < 1$ for all covariate values $c$ with positive density.
4. Ignorability (Unconfoundedness): $i$0; equivalently, $i$1.

Under these, the back-door or covariate-adjustment formula identifies the ATE: $i$2

2. Neymanian Inference in Randomized and Observational Designs

In randomized trials, ignorability and positivity are guaranteed by design. The difference-in-means estimator is unbiased for $i$3, and Neyman's finite-population variance decomposition applies. For an assignment vector $i$4, and $i$5 units, with $i$6 treated and $i$7 controls, the estimator and its variance are

$i$8

with $i$9 the finite-population variance for treatment $$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$0, and $$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$1 the variance of unit-level effects (Chattopadhyay et al., 2024, Ding et al., 2015).

In observational studies, ignorability may only hold after adjusting for pre-treatment covariates $$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$2. Assignment is commonly modeled, e.g., via a logistic propensity score, and ATE is estimated by regression, inverse probability weighting, or doubly robust approaches. Identification under the Neymanian formalism can strictly leverage the potential outcome machinery even when the covariate structure is undirected, cyclic, or deterministic, which can be problematic for graphical models (Gorbach et al., 9 Dec 2025).

3. Variance Estimation: Classical, Improved, and Partial Identification

Neyman's original variance estimator is conservative in general because $$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$3 in the decomposition is not identifiable unless treatment effects are strictly additive (constant across units). The standard estimator,

$$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$4

where $$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$5 are sample variances by treatment group, overestimates the true variance unless $$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$6 (Chattopadhyay et al., 2024).

Partial identification theory provides sharper lower bounds. For binary or factorial experiments, the sampling variance correction reduces conservatism by explicitly subtracting the sharp lower bound of $$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$7, yielding improved estimators that produce tighter Wald intervals while retaining nominal coverage (Lu, 2017, 1803.04503, Ding et al., 2015, Pashley et al., 13 Mar 2025). For example, in binary-treatment experiments: $$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$8 (Ding et al., 2015).

For complex experimental designs, variance can equivalently be estimated by either imputation (using potential outcomes, possibly under constant effect assumptions or estimated effects) or contrast substitution (swapping unidentifiable contrasts with observable ones), both providing conservative or unbiased estimators under specified conditions (Chattopadhyay et al., 2024). Extensions to general assignment mechanisms use Horvitz–Thompson-type estimators and minimax bias criteria for variance estimation under milder than strict additivity (Mukerjee et al., 2016).

4. Extensions: Multifactor Experiments, Randomization-Based Inference, and Partial Interference

The Neymanian approach generalizes to multifactorial ($$Y_i(1) = \text{outcome for } i \text{ if } A_i=1, \qquad Y_i(0) = \text{outcome for } i \text{ if } A_i=0.$$9) designs using model-matrix contrast vectors, and unbiased contrasts of potential outcomes as estimators (Pashley et al., 13 Mar 2025, 1803.04503). For example, for $Y_i(1), Y_i(0)$0 treatment arms and sorted contrast vectors $Y_i(1), Y_i(0)$1, the estimand and estimator for factorial effect $Y_i(1), Y_i(0)$2 are

$Y_i(1), Y_i(0)$3

with $Y_i(1), Y_i(0)$4 the observed proportions per arm.

Randomization-based inference exploits the finite-population CLT for asymptotics and applies equally to stratified, multi-arm, and peer-effect group-formation experiments, delivering asymptotically valid Wald intervals for complex estimands (Xu et al., 2021). Under partial interference and exposure mappings, Neymanian inference naturally extends to estimating exposure contrasts given group-randomized treatments with spillovers.

5. Neyman Orthogonality and Modern Semiparametric Inference

The principle of Neyman orthogonality—that moment equations' Gateaux derivatives with respect to nuisances vanish at the true parameter—is central to both classical and modern machine-learning–augmented causal inference. Under local product structure (independent ‘wiggling’ of target and nuisance parameters), Neyman orthogonality of moment equations is formally equivalent to semiparametric pathwise differentiability (Chen et al., 16 Mar 2026). Influence functions deriving from Neyman-orthogonal moments provide efficient and debiased estimators, furnishing root-$Y_i(1), Y_i(0)$5 consistency and asymptotic normality under minimal nuisance estimation rate assumptions, and underlie frameworks for double/debiased machine learning, targeted maximum likelihood, and Bayesian plug-in estimation (Kato, 27 Oct 2025, Sabbagh et al., 23 Feb 2026, Bonhomme et al., 2024).

Practical implication: when nuisance components (propensity scores, regression functions, etc.) are estimated flexibly (e.g., nonparametric or ML methods), Neyman orthogonality is required for valid frequentist or Bayesian inference on the causal parameter. Sample splitting and cross-fitting mitigate overfitting and ensure orthogonality holds in finite samples (Chen et al., 16 Mar 2026, Sabbagh et al., 23 Feb 2026).

Higher-order Neyman orthogonality, as developed for fixed effects/challenging panel data, systematically annihilates higher-degree bias terms, allowed by iterated projections and sample splitting (Bonhomme et al., 2024).

6. Comparison with Graphical Causal Models and Complementarity

Graphical frameworks (DAGs, d-separation, do-calculus) and the Neymanian potential-outcome framework have complementary strengths. Neymanian analysis remains unaffected by undirected, cyclic, or deterministic relations among covariates—contexts where faithfulness or acyclicity in DAGs are violated—being driven only by consistency, SUTVA, positivity, and ignorability (Gorbach et al., 9 Dec 2025). By contrast, methods grounded in graphical criteria efficiently expose the analytic consequences of conditioning and are essential for handling collider bias (M-bias), front-door/trapdoor identification, or multi-stage mediation with complex relationships.

Practitioners often synthesize both frameworks: DAGs articulate conditional independence and graphical identification, while the Neyman–Rubin formalism precisely defines potential-outcome estimands, articulates assumptions, and guides estimation strategies (Gorbach et al., 9 Dec 2025).

7. Limitations, Ongoing Developments, and Practical Guidelines

• Non-additivity/heterogeneity: The classic Neymanian variance estimator is only unbiased under strict additivity. Modern sharpened and partial identification estimators recover efficiency in finite-population inference.
• Complex designs: Extensions to factorials, composite experiments, network interactions, and spillovers require careful mapping of exposures and may demand customized variance expressions.
• Model-free validity: Neymanian inference is robust to mis-specification of outcome models but relies heavily on correct specification of assignment mechanisms and structural assumptions (e.g., ignorability, SUTVA).
• Neyman orthogonality in practice: Sample-splitting and higher-order corrections are essential for bias reduction when nuisance parameters are high-dimensional and only slowly estimable.
• Integration with graphical models: For cases where covariate structures violate DAG assumptions, Neymanian methods still yield valid identification, but graphical tools are indispensable for recognizing and appropriately adjusting for collider and trapdoor structures.

In summary, the Neymanian framework offers a minimalistic yet powerful foundation for causal inference, supporting rigorous identification, estimation, and uncertainty quantification across a broad array of experimental, observational, and complex designs. Advanced developments in semiparametric and debiased learning methods, as well as robust variance estimation, continue to extend its reach while maintaining its original commitment to randomization-based, model-free inference (Gorbach et al., 9 Dec 2025, Ding et al., 2015, Lu, 2017, 1803.04503, Mukerjee et al., 2016, Chen et al., 16 Mar 2026, Bonhomme et al., 2024, Sabbagh et al., 23 Feb 2026, Pashley et al., 13 Mar 2025, Chattopadhyay et al., 2024, Kato, 27 Oct 2025, Xu et al., 2021).