Individual-Pair Causal Estimands

• Individual-pair causal estimands are contrast-based parameters that compare outcomes between pairs of individuals to define granular causal effects beyond population averages.
• They enable robust estimation in settings with interference, cluster structure, and high-dimensional covariates by leveraging pairwise differences.
• These estimands improve efficiency and clarify direct, indirect, and spillover effects in networks, cluster-randomized trials, and infectious disease studies.

Individual-pair causal estimands are contrast-based causal parameters that quantify the effect of interventions or treatments by leveraging data at the level of pairs of individuals—not only to define more granular effects than population averages, but to rigorously identify and estimate effects in settings with interference, cluster structure, or high-dimensional covariate heterogeneity. Individual-pair estimands have emerged as central objects across contemporary causal inference, appearing in factorial analysis of observational data, intervention trials for infectious disease, cluster-randomized trials (CRTs) with informative cluster size, peer-effect network studies, and modern machine learning approaches for individual treatment effect estimation.

1. Formal Structure of Individual-Pair Estimands

At their core, individual-pair estimands quantify a treatment effect or contrast on the scale of pairs constructed from observed units. The general strategy is to consider, for any ordered or unordered pair $(i, j)$, a contrast between their covariates, treatments, and/or outcomes, and to use these pairwise differences to infer causal structure in three broad regimes:

1. Contrast of potential outcomes across units and treatment levels: For units $i$ and $j$, under possible assignments $a$ and $a'$, associate the contrast $h(Y_i(a), Y_j(a'))$ for a contrast function $h$, e.g. difference or ratio.
2. Pseudo-treatment induced by observed differences: In high-dimensional nonexperimental data, set $T_{ij} = X_i - X_j$ as a "treatment" and $\Delta Y_{ij} = Y_i - Y_j$ as the observed effect, viewing the pair $(i,j)$ as a small randomized experiment (Ribeiro et al., 2021).
3. Potential-outcome contrasts reflecting interference or spillover: For treatment assignment vectors $x$, define $Y_i(x)$, and then for two-person settings, compare $Y_i$ under different joint assignments, e.g., $Y_i(1,0)$ vs $Y_i(0,0)$ for susceptibility or $Y_j(0,1)$ vs $Y_j(0,0)$ for infectiousness (Cai et al., 2019, Cai et al., 2021, Luo et al., 8 Apr 2025).

Individual-pair estimands can formalize direct, indirect, and interaction effects in networked or clustered structures, average treatment effects (ATEs) reduced to pairwise differences, or high-resolution heterogeneous effects as in machine-learning frameworks.

2. Identification Conditions and Assumptions

Rigorous identification of individual-pair estimands requires precise structural and statistical assumptions tailored to the context:

• Pairwise Ignorability: For a pair $(i,j)$, conditional ignorability of the pairwise "treatment" with respect to potential outcomes, optionally conditioning on covariate differences $X_j \ominus X_i$ (Ribeiro et al., 2021), or on observed instruments and covariates in peer effect models (Luo et al., 8 Apr 2025).
• No Interference: For SUTVA-type settings, ensure $Y_{ij}(a)$ depends only on the assigned treatment for the involved clusters or individuals (Chen et al., 19 Jan 2026).
• Overlap / Positivity: Each "treatment contrast" or pair comparison of interest must occur with positive probability in the data.
• Exchangeability and Randomization: In CRTs, require cluster-randomization and i.i.d. sampling of clusters (Chen et al., 19 Jan 2026).
• Instrumental Variable (IV) Validity: For direct/spillover effects using IVs, dual IV assumptions must hold (unconfoundedness, exclusion restriction, relevance, no unmeasured effect modification) (Luo et al., 8 Apr 2025).
• Network/Contagion Structure: For infectious disease, assumptions on transmission, independence of exogenous hazards, and transparent DAG/SWIG graphs are used to build potential-outcome contrasts under interference (Cai et al., 2019, Cai et al., 2021).

3. Estimand Examples Across Domains

Individual-pair estimands arise in several modern applications, each with distinct formalization.

Context	Typical Individual-Pair Estimand	Reference
Nonexperimental “factorial”	$T_{ij}=X_i-X_j$, $\Delta Y_{ij}=Y_i-Y_j$; estimate $f(X_i\ominus X_j)$	(Ribeiro et al., 2021)
Infectious disease (pairs)	Controlled susceptibility: $SE_i^C = E[Y_i(1,x_{(i)}) - Y_i(0,x_{(i)})]$	(Cai et al., 2019)
Infectious disease (clusters)	Per-pair Cox model: infectivity/susceptibility $\beta_1$, $\beta_2$	(Cai et al., 2021)
Dyadic social/peer effects	Direct: $E[Y_1(1,d) - Y_1(0,d)]$; Spillover: $E[Y_1(d,1) - Y_1(d,0)]$	(Luo et al., 8 Apr 2025)
CRT, multi-outcome	$$\theta_{\mathrm{ind}} = E[h(Y_{ij}(1),Y_{k\ell}(0))]$$-weighted	(Chen et al., 19 Jan 2026)
Deep ITE estimation (PairNet)	Pairwise loss using $(y_i-y_j) - [\mu(x_i,t_i)-\mu(x_j,t_j)]$	(Nagalapatti et al., 2024)

Each setting leverages pairwise contrasts to achieve either improved statistical efficiency, explicit causal interpretation under interference, or computational/estimation advantages.

4. Estimation Methodologies

Estimation strategies for individual-pair estimands differ by sampling regime and data structure:

• Weighted Pairwise (Clustered) U-Statistics: In CRTs, estimate $\theta_{\mathrm{ind}}$ nonparametrically using cluster-level data by solving moment equations over all cross-cluster pairs, with explicit weighting by cluster size (Chen et al., 19 Jan 2026).
• Bayesian and Penalized Approaches: In "factorial" analysis of observational data, estimation proceeds via a penalized Bayesian objective over all $O(n^2)$ pairs, incorporating penalties for treatment size and balance, and yielding an embedding $T_y(X)$ that induces individual-level and pairwise causal estimates (Ribeiro et al., 2021).
• Partial Likelihood & Semiparametric Models: For infectious disease, pairwise transmission models (Cox-type) are used, maximizing likelihoods over all transmission pairs to extract susceptibility and infectiousness effects (Cai et al., 2021).
• Instrumental Variables and Influence Functions: In peer-effects models, dual IVs for each dyad drive identification, with triply-robust, efficient influence function-based plug-in estimators targeting each direct or spillover effect (Luo et al., 8 Apr 2025).
• Pairwise Supervised Deep Learning: In PairNet, estimation of heterogeneous or individual treatment effects is performed by minimizing a factual pairwise loss—contrasting observed outcomes for near-neighbor pairs under differing treatments, ensuring consistency without outcome imputation (Nagalapatti et al., 2024).

These strategies can be adapted with covariate adjustment via doubly-robust or debiased (cross-fitted) machine learning methods, supporting asymptotic normality and semiparametric efficiency (Chen et al., 19 Jan 2026, Luo et al., 8 Apr 2025).

5. Interpretation and Analytical Properties

Individual-pair estimands offer several salient advantages and require nuanced interpretation:

• Efficiency and Information Expansion: By utilizing all possible pairs, either within or across clusters or covariate values, the information content is increased from $O(n)$ to $O(n^2)$, reducing variance and recovering fine-grained heterogeneity (Ribeiro et al., 2021).
• Heterogeneous Effect Recovery: Pairwise estimands enable recovery of heterogeneity not accessible to traditional mean-based ATE estimands. Iterative fitting procedures (e.g. iterative embeddings) perform implicit variable selection and can identify non-causal factors (Ribeiro et al., 2021, Nagalapatti et al., 2024).
• Clarity Under Interference: In settings with contagion, network spillover, or within-cluster effects, pairwise estimands clarify the disparate roles of susceptibility, infectiousness, and pure contagion components, often exposing biases in population-level direct or indirect effects commonly used in randomized trial analysis (Cai et al., 2021, Cai et al., 2019).
• Automatic Informative-Cluster Size Handling: The use of cross-cluster pairs with explicit $N_iN_k$ normalization in CRTs ensures unbiasedness even in the presence of informative cluster size (Chen et al., 19 Jan 2026).
• Identification of Direct and Indirect Effects: In dyads, fours estimands—ego direct, ego spillover, peer direct, and peer spillover—fully specify the landscape of pairwise causal effects, with identification via dual IVs and robust estimation (Luo et al., 8 Apr 2025).

6. Practical Applications and Comparative Performance

Empirical applications have demonstrated the superiority or complementary advantages of individual-pair estimands:

• In factorial analysis of observational data, pairwise approaches outperform standard regression, matching, IPW, and latent confounder approaches, particularly under heterogeneity and sample imbalance (Ribeiro et al., 2021).
• In CRTs, the individual-pair U-statistic estimand admits consistent, normal, and efficient estimation with analytic variance estimators, and is robust to high-dimensional adjustment via debiased ML (Chen et al., 19 Jan 2026).
• In deep learning for ITE estimation, pairwise training losses (PairNet) dominate baselines (including T-/R-/X-Learners, representation learning, and matching) across diverse tasks, especially under confounding or lack of strong overlap, and support extension to multi-armed and continuous treatments (Nagalapatti et al., 2024).
• In infectious disease and peer effect settings, individual-pair estimands clarify and unify competing definitions of direct, indirect, and overall effects, and expose situations where classical estimands are directionally biased or lack a meaningful causal interpretation (Cai et al., 2019, Cai et al., 2021, Luo et al., 8 Apr 2025).

7. Extensions and Open Problems

Ongoing research explores generalizations and limitations:

• Generalized Pairwise Contrasts: Extension of contrast functions $h$ to multivariate outcomes, prioritized combinations, or functional outcomes in CRT and network settings (Chen et al., 19 Jan 2026).
• Higher-Order and Multiway Matching: Extending beyond pairs to triplets or higher-order groupings to capture complex interaction effects (Nagalapatti et al., 2024).
• Pair Selection and Computation: Scalability concerns motivate approximate nearest neighbor algorithms and adaptive matching for large datasets (Nagalapatti et al., 2024).
• Robustness to Sparsity: Pairwise approaches depend on sufficient overlap/positivity; in sparse data settings, estimation may degrade as pairwise distances increase (Nagalapatti et al., 2024).
• Connections Across Domains: The underlying mathematical structure—contrast-based, symmetric, or directional—links work in factor analysis, potential outcomes, network/interference, and learning-theoretic generalization bounds, providing a rich ground for further theory (Ribeiro et al., 2021, Nagalapatti et al., 2024, Chen et al., 19 Jan 2026, Luo et al., 8 Apr 2025).

Individual-pair causal estimands thus form a unifying abstraction, enabling precise identification, interpretable inference, and robust estimation in the face of interference, high-dimensionality, and heterogeneity, and are central to ongoing innovation in both theoretical and applied causal inference.