Pith. sign in

REVIEW 2 major objections 5 minor 109 references

Under network interference, prediction accuracy is not estimator efficiency, and a graph-weighted calibration step provably fixes the gap.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 12:10 UTC pith:JHSJLXG2

load-bearing objection A genuinely useful calibration idea for design-based inference under interference, with a real gap between the fixed-feature theory and the GNN implementation the paper actually runs. the 2 major comments →

arxiv 2607.19627 v2 pith:JHSJLXG2 submitted 2026-07-21 stat.ME

GAUGER: Generalized Regression Adjustment via Graph-Weighted Exposure-Level Residualization for Design-Based Inference Under Interference

Lei Shi , Rita Lyu , Sizhu Lu This is my paper
classification stat.ME
keywords causal inferenceinterferencedesign-based inferenceregression adjustmentgraph neural networksvariance reductionexposure mappingAIPW estimator
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

GAUGER shows that, under network interference, minimizing prediction error does not in general minimize the variance of an AIPW treatment-effect estimator. The reason is that estimator variance is a quadratic form in the residuals weighted by a design-induced matrix that encodes the network and the assignment mechanism. The paper therefore separates prediction from calibration: first learn any flexible outcome model, then apply Graph-weighted Exposure-level Residualization (GER) to re-weight residuals toward the variance objective. The main theorem proves that, for any fixed feature space, the GER-calibrated AIPW estimator is asymptotically normal and has the smallest asymptotic variance among all linear adjustments, so it is never worse than a purely prediction-based adjustment. Implemented with graph neural networks followed by GER, the pipeline yields substantial variance reductions in synthetic, semi-synthetic, and real-world network experiments.

Core claim

The paper's central claim is that the variance of design-based estimators under interference is governed by a graph-weighted structure, not by prediction accuracy: for the AIPW estimator, n² Var(τ̂) = (residual)ᵀ Λ (residual), where Λ is the covariance matrix of the reweighted exposure indicators. Consequently, minimizing MSE does not imply minimizing variance. The paper constructs an oracle coefficient β* that minimizes this quadratic form and proves that the GER calibration estimator converges to β* and that the resulting estimator is asymptotically normal with avar(τ̂_{F,β*}) ≤ avar(τ̂_{F,β}) for every β. This gives a formal efficiency guarantee for the calibrated adjustment relative to a

What carries the argument

The central object is the graph-weighted residualization objective: the calibration coefficients (β̂₁, β̂₂) minimize {ε̂(β)}ᵀ Δ {ε̂(β)}, where Δ is the dependency graph indicating which units' exposure assignments are dependent, and the regressors are exposure-level residualized features. This replaces the dense Λ matrix with a sparse graph surrogate and directly targets the variance quadratic form. The same machinery yields a conservative variance estimator that augments Λ to handle dependent units that are never jointly observed at a given exposure level.

Load-bearing premise

The theorems treat the feature matrix F as fixed in the asymptotic analysis, while the practical pipeline learns F from the same observed data used to compute the calibration; no theorem in the paper covers consistency or asymptotic normality for data-dependent features.

What would settle it

Simulate a large network with known potential outcomes and Bernoulli assignment, then compare the GER-calibrated β̂ with the oracle β* that minimizes the Λ-weighted quadratic form using the true potential outcomes: if β̂ does not converge to β* as n grows, or if some fixed feature matrix F and coefficient β yield a smaller Monte Carlo variance than the GER-calibrated estimator over many randomizations, the core efficiency theorem fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • For a fixed feature space, no linear adjustment of the AIPW form can beat the GER-calibrated estimator in asymptotic variance; prediction-based adjustment is at best tied.
  • Model selection or hyperparameter tuning that minimizes prediction error can be misaligned with efficiency, so tuning should target the graph-weighted variance objective or the estimator's final variance.
  • The conservative variance estimator supports design-based confidence intervals and hypothesis tests, making the method usable for substantive inference in network experiments.
  • The framework is model-agnostic: any predictive model—linear, GNN, or other—can enter stage 1, and GER calibration works on top of it.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same graph-weighted calibration logic could be applied to other design-based estimators, such as Hájek-style ratio estimators or cluster-level contrasts, by deriving the analogous Λ for those designs.
  • Because the theory treats features as fixed, the practical GNN pipeline should adopt cross-fitting or extend the theory to data-dependent representations before the efficiency guarantee is fully rigorous for the end-to-end procedure.
  • A natural next step is to learn the exposure mapping itself inside the pipeline so that exposure definition and graph-weighted calibration are optimized jointly; the appendix sketches this but its identifiability is not established.
  • The accuracy–efficiency mismatch may also affect observational network settings where outcome models are selected by cross-validated prediction error and then used for adjustment; checking the variance decomposition directly would be a useful diagnostic.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes GAUGER, a two-stage framework for design-based causal inference under network interference. In the first stage, a flexible predictive model (e.g., a GNN) produces outcome predictions or representations; in the second stage, a calibration step called Graph-weighted Exposure-level Residualization (GER) reweights the linear adjustment to minimize an approximate variance quadratic form. The authors show that the variance of a linearly adjusted AIPW estimator is governed by a graph-weighted covariance matrix Λ, and that minimizing prediction error is not equivalent to minimizing this variance. They provide theorems for a fixed feature matrix F: unbiasedness, consistency of the GER coefficient to the variance-optimal oracle coefficient, asymptotic normality, and a conservative variance estimator. Numerical studies on synthetic, semi-synthetic, and real network data compare GNN-based GAUGER variants with existing estimators and report variance reductions.

Significance. If the practical GNN-assisted estimator were covered by the theory, the paper would make a useful contribution: it identifies a genuine mismatch between predictive accuracy and estimator efficiency under interference, and it proposes a calibration principle that directly targets the design-based variance. The fixed-F theoretical core is coherent and reasonably detailed: Theorem 3.1's variance formula is correct for fixed features, the GER coefficient is shown to converge to the Λ-optimal oracle coefficient, and the conservative variance estimator is a sensible design-based construction. The Monte Carlo approximation to Λ and the graph-based cross-fitting proposal are also useful components. However, the main theorems are conditional on a fixed, prespecified feature matrix F, while the implemented pipeline learns F from the same data used to form the estimator. This gap is load-bearing because the abstract claims consistency, asymptotic normality, and valid inference for 'the resulting estimator' in the GNN implementation. The paper's practical claims therefore currently outrun its formal guarantees.

major comments (2)
  1. [§5.1–5.3 vs. Theorems 3.1, 4.1, 4.2; Appendix F.1] Theorems 3.1, 4.1, and 4.2 treat the feature matrix F as fixed or prespecified. In Section 5, however, F is data-dependent: the GNN is trained end-to-end on (Y, A, G) with an IPM balance loss that depends on the exposure D, and then F is set to (1, \hat Y) or (1, H). The proof of Theorem 3.1 (Appendix F.1) uses the equality E[(1(D_i=d)/π_i^{(d)})(Y_i - F_i^T β_k)] = Y_i^{(d)} - F_i^T β_k, which requires F_i to be non-random. If F_i depends on the same treatment assignments and outcomes used in the AIPW weights, this equality can fail, and the unbiasedness/variance cancellation no longer holds. Consequently, Theorems 4.1–4.2 and D.3 do not apply to the GNN-assisted estimator in Section 5. This is not a cosmetic issue: the claimed coverage and variance-reduction guarantees for the implemented pipeline are not supported by the stated theory.
  2. [Appendix D.5 and Section 6] The paper itself recognizes the data-dependence problem in Appendix D.5 and proposes graph-based cross-fitting (Algorithm 2) to break the dependence between the learned features/predictions and the estimation data. However, Algorithm 2 is not incorporated into Theorems 4.1–4.2 or Theorem D.3, and Section 6 does not state that the numerical protocol uses it. Thus the proposed remedy is not connected to the consistency, normality, or variance-estimation results, and the reported Monte Carlo coverage in Tables 1–3 cannot be formally justified by the paper's theorems. The authors should either extend the theory to a cross-fitted, data-dependent-F estimator and explicitly use that protocol in the simulations, or restrict all formal and empirical claims to fixed or externally constructed features.
minor comments (5)
  1. [Theorem 4.2, Eq. (4)] The inequality avar{τ̂_{F,β*}} ≤ avar{τ̂_{F,β}} is a direct restatement of the definition of β* as the minimizer of the Λ-weighted quadratic form in Eq. (4), rather than an independent efficiency theorem. The paper should present it as such and emphasize that the substantive content of Theorem 4.2 is the consistency of β̂ and the resulting plug-in asymptotic normality.
  2. [Eq. (5) and Section 5.2] The notation \hat Y^{(d_k)}, \hat F^{(d_k)}, and \hat 1^{(d_k)} is used without precise definition. In particular, it is unclear whether \hat Y^{(d_k)} denotes the observed outcome multiplied by inverse-probability indicators or some other residualized quantity. The formal development would benefit from explicit definitions of all hatted quantities.
  3. [Tables 3, 8, 9] The real-data analysis uses 'PNA-no-harm' as a headline estimator, but this term is never defined in Section 5, Section 6, or the appendices. Without a definition of this estimator and how it differs from PNA-linear and PNA-pred, the real-data efficiency claims are not reproducible.
  4. [Table 1] The 'Coverage' column should state the nominal confidence level and describe how Monte Carlo coverage is computed, especially for primitive estimators where the reported coverage (e.g., 0.60 for HT) is far from the nominal level.
  5. [Throughout] There are several typos and minor wording issues, including 'abd' in Section 2, 'desciption' in Section 6, 'unites' in Algorithm 1, and 'interger' in Assumption C.1(e). Figure 1 is referenced but not included as a figure in the text.

Circularity Check

1 steps flagged

Theorem 4.2's efficiency inequality restates the definition of β*; the substantive GER consistency result is proved independently, and the data-dependent-F issue is a gap rather than circularity.

specific steps
  1. self definitional [Section 4.1, Eq. (4); Section 4.2, Theorem 4.2]
    "Given features F, the variance-optimal coefficient β* is given by β* = (F^{(d1⊕d2)⊤}ΛF^{(d1⊕d2)})†F^{(d1⊕d2)⊤}ΛY^{(d1;d2)} ... Moreover, the asymptotic variance satisfies avar{τ̂^{(d1,d2)}_{F,β*}} ≤ avar{τ̂^{(d1,d2)}_{F,β}} for any β∈R^p."

    β* is defined by Eq. (4) as the minimizer of the Λ-weighted residual quadratic form, and Theorem 3.1 identifies that quadratic form exactly with n²Var(τ̂). Therefore the inequality in Theorem 4.2 is an immediate restatement of the definition of β*, not an independently derived result. The non-tautological content — that the feasible GER estimator from Eq. (5) is consistent for β* — is proved separately in Theorem 4.1, so this is a minor self-definitional presentation issue rather than a fully circular derivation.

full rationale

The only place where a headline claim reduces to its own construction is the efficiency half of Theorem 4.2. Eq. (4) defines β* as the minimizer of the Λ-weighted quadratic form, and Theorem 3.1 shows that form is exactly the estimator variance, so avar{τ̂_{F,β*}} ≤ avar{τ̂_{F,β}} is true by definition rather than by an independent argument. This makes the abstract's 'provable variance reduction' partly tautological. However, the central estimation claim — consistency of the feasible GER coefficient to the oracle β* — is established by a separate variance calculation in Theorem 4.1 and Appendix F.7, where the GER Gram matrix is shown to converge to F^TΛF. That part is not circular. The paper's own Appendix D.5 acknowledges in substance that the main theorems treat the feature matrix F as fixed while Section 5 learns F from the same data; the proposed cross-fitting extension is not incorporated into Theorems 4.1–4.2. This is an inference-coverage gap, not a circular-reasoning step, so it does not increase the circularity score. Self-citations (Lu et al. 2025a,b; Shi and Ding 2025) appear only as literature pointers and are not load-bearing. Overall score 2 reflects the minor self-definitional efficiency claim; the derivation is otherwise self-contained.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central theoretical claim rests on standard design-based regularity conditions plus one unstated premise: F fixed. The empirical claims rest on tuned GNN hyperparameters, exposure-mapping choices, and an undefined 'PNA-no-harm' variant. No new physical or causal entities are introduced.

free parameters (4)
  • GNN hyperparameters (GAT/PNA) = e.g., GAT layers [64,32], representation dim 16, λ=0; PNA per Tables 4-5
    Selected by validation performance; real-data wage-bill PNA uses seed 50039, so numerical efficiency gains are conditional on model-selection choices.
  • IPM regularization weight λ = 0.0, 0.1, 1.0 (searched)
    Tuned on validation; enters the prediction loss but not the variance objective, so it is not derived from the design-based efficiency criterion.
  • Exposure bin definitions / thresholds = k=3 bins for treated-neighbor proportion; real-data S_i=1 if >1/3 neighbors treated
    Exposure mapping and discretization are chosen by the authors; the estimand and the matrix Λ change with these choices.
  • Spatial network construction in real application = nearest 5 villages within 8 km; degree capped at 8
    The network is a modeling choice that defines the interference structure and hence Λ and Δ.
axioms (6)
  • domain assumption Exposure mapping summarizes all interference (Assumption 2.1).
    Potential outcomes depend on the full treatment vector only through a scalar exposure; if this is wrong, the estimand is a projection of the true effect rather than a physical effect.
  • domain assumption Bernoulli randomization with known or estimable positivity (Assumption C.1(a)-(b)).
    Design-based inference requires first- and second-order exposure probabilities bounded away from zero.
  • domain assumption Stability of graph-weighted second moments (Assumption C.1(e)).
    Requires normalized Σε, ΣY, ΣF to be constant order with κ_n at least constant; this is needed for the CLT and pseudo-inverse perturbation argument.
  • ad hoc to paper Feature matrix F is fixed or non-data-dependent in Theorems 4.1/4.2.
    Section 5 uses GNN representations learned from the same data; no theorem covers this case, and cross-fitting is only proposed in the appendix.
  • domain assumption Δ = 1(I_n + G + G^2 > 0) correctly encodes the exposure-dependence graph.
    Correct for Bernoulli randomization and 1-hop exposure mappings; for other exposure mappings the user must construct the correct dependence graph.
  • standard math Monte Carlo approximation of Λ is consistent (Prop D.1, Aronow and Samii 2017).
    The paper uses Monte Carlo to approximate joint inclusion probabilities and Λ; this relies on the cited consistency result.

pith-pipeline@v1.3.0-alltime-deepseek · 26523 in / 15598 out tokens · 163833 ms · 2026-08-01T12:10:51.586811+00:00 · methodology

0 comments
read the original abstract

Estimating causal effects under interference is a common problem in social science and economics. However, it is challenging due to the complex dependency structure induced by network connections. In this paper, we propose GAUGER, a Generalized regression Adjustment framework via Graph-weighted Exposure-level Residualization for design-based causal inference under general interference. We first reveal a surprising mismatch between accuracy and efficiency in this setting: model adjustments that minimize prediction error (e.g., MSE) do not necessarily lead to the most variance reduction of the treatment effect estimator. To address this mismatch, we propose a two-step approach: (1) leveraging a strong prediction model to learn outcome patterns from the network and covariates, and (2) applying a novel calibration scheme called Graph-weighted Exposure-level Residualization (GER) that directly targets variance reduction. The resulting estimator is consistent for target causal parameters, enjoys provable variance reduction, and is asymptotically normal with a conservative variance estimator for valid statistical inference. As a practical implementation of the pipeline, we present a scheme that leverages Graph Neural Networks (GNNs) to construct the prediction model and use GER to steer the model adjustment for better variance reduction. Numerical studies show substantial efficiency gains over existing methods.

Figures

Figures reproduced from arXiv: 2607.19627 by Lei Shi, Rita Lyu, Sizhu Lu.

Figure 1
Figure 1. Figure 1: Overview of the GNN encoder for potential outcome prediction given exposure mapping. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Propensity score distributions by exposure level. Blue bars show the distribution of [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

109 extracted references · 1 canonical work pages

  1. [1]

    The Bell System Technical Journal , volume=

    An efficient heuristic procedure for partitioning graphs , author=. The Bell System Technical Journal , volume=. 1970 , publisher=

    1970

  2. [2]

    Czechoslovak Mathematical Journal , volume=

    Algebraic connectivity of graphs , author=. Czechoslovak Mathematical Journal , volume=

  3. [3]

    International Conference on Learning Representations , year=

    Semi-supervised classification with graph convolutional networks , author=. International Conference on Learning Representations , year=

  4. [4]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

    Masoero, Lorenzo and Vijaykumar, Suhas and Richardson, Thomas S and McQueen, James and Rosen, Ido and Burdick, Brian and Bajari, Pat and Imbens, Guido , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

  5. [5]

    International Conference on Learning Representations , year=

    Graph attention networks , author=. International Conference on Learning Representations , year=

  6. [6]

    International Conference on Learning Representations , year=

    How powerful are graph neural networks? , author=. International Conference on Learning Representations , year=

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    Principal neighbourhood aggregation for graph nets , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    Categorical reparameterization with

    Jang, Eric and Gu, Shixiang and Poole, Ben , booktitle=. Categorical reparameterization with

  9. [9]

    International Conference on Learning Representations , year=

    The concrete distribution: A continuous relaxation of discrete random variables , author=. International Conference on Learning Representations , year=

  10. [10]

    Journal of the American Statistical Association , volume=

    Toward causal inference with interference , author=. Journal of the American Statistical Association , volume=. 2008 , publisher=

    2008

  11. [11]

    Journal of the American Statistical Association , volume=

    The generalized oaxaca-blinder estimator , author=. Journal of the American Statistical Association , volume=. 2023 , publisher=

    2023

  12. [12]

    arXiv preprint arXiv:2311.06891 , year=

    Design-based estimation theory for complex experiments , author=. arXiv preprint arXiv:2311.06891 , year=

    Pith/arXiv arXiv

  13. [13]

    Biometrika , volume=

    Regression-based causal inference with factorial experiments: estimands, model specifications and design-based properties , author=. Biometrika , volume=. 2022 , publisher=

    2022

  14. [14]

    arXiv preprint arXiv:2508.15664 , year=

    Conditional cross-fitting for unbiased machine-learning-assisted covariate adjustment in randomized experiments , author=. arXiv preprint arXiv:2508.15664 , year=

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2501.09844 , year=

    Design-based causal inference in bipartite experiments , author=. arXiv preprint arXiv:2501.09844 , year=

    Pith/arXiv arXiv

  16. [16]

    Econometrica , volume=

    Causal inference under approximate neighborhood interference , author=. Econometrica , volume=. 2022 , publisher=

    2022

  17. [17]

    arXiv preprint arXiv:2507.20231 , year=

    Causal Inference when Intervention Units and Outcome Units Differ , author=. arXiv preprint arXiv:2507.20231 , year=

    Pith/arXiv arXiv

  18. [18]

    Journal of Causal Inference , volume=

    Some theoretical foundations for the design and analysis of randomized experiments , author=. Journal of Causal Inference , volume=. 2024 , publisher=

    2024

  19. [19]

    International conference on machine learning , pages=

    Learning representations for counterfactual inference , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  20. [20]

    IEEE transactions on neural networks and learning systems , volume=

    A comprehensive survey on graph neural networks , author=. IEEE transactions on neural networks and learning systems , volume=. 2020 , publisher=

    2020

  21. [21]

    arXiv preprint arXiv:2211.07823 , year=

    Graph neural networks for causal inference under network confounding , author=. arXiv preprint arXiv:2211.07823 , year=

    arXiv

  22. [22]

    arXiv preprint arXiv:2309.07476 , year=

    Causal inference in network experiments: regression-based analysis and design-based properties , author=. arXiv preprint arXiv:2309.07476 , year=

    Pith/arXiv arXiv

  23. [23]

    Biometrika , volume=

    The role of the propensity score in estimating dose-response functions , author=. Biometrika , volume=. 2000 , publisher=

    2000

  24. [24]

    Journal of the American Statistical Association , volume=

    Analyzing two-stage experiments in the presence of interference , author=. Journal of the American Statistical Association , volume=. 2018 , publisher=

    2018

  25. [25]

    The Annals of Statistics , number =

    Anqi Zhao and Peng Ding and Rahul Mukerjee and Tirthankar Dasgupta , title =. The Annals of Statistics , number =. 2018 , doi =

    2018

  26. [26]

    Statistical Science , volume=

    Bipartite Causal Inference with Interference , author=. Statistical Science , volume=

  27. [27]

    Econometrica , volume=

    General equilibrium effects of cash transfers: experimental evidence from Kenya , author=. Econometrica , volume=. 2022 , publisher=

    2022

  28. [28]

    Aronow and Cyrus Samii , title =

    Peter M. Aronow and Cyrus Samii , title =. The Annals of Applied Statistics , number =. 2017 , doi =

    2017

  29. [29]

    2014 , publisher=

    Martingale Limit Theory and Its Application , author=. 2014 , publisher=

    2014

  30. [30]

    2013 , publisher=

    Theory of U-statistics , author=. 2013 , publisher=

    2013

  31. [31]

    Econometric Theory , volume=

    A limit theorem for quadratic forms and its applications , author=. Econometric Theory , volume=. 2007 , publisher=

    2007

  32. [32]

    Probability Theory and Related Fields , volume=

    A central limit theorem for generalized quadratic forms , author=. Probability Theory and Related Fields , volume=. 1987 , publisher=

    1987

  33. [33]

    Freedman , title =

    David A. Freedman , title =. The Annals of Applied Statistics , number =. 2008 , doi =

    2008

  34. [34]

    2004 , publisher=

    Convex optimization , author=. 2004 , publisher=

    2004

  35. [35]

    The Annals of Applied Statistics , volume=

    Agnostic notes on regression adjustments to experimental data: Reexamining Freedman’s critique , author=. The Annals of Applied Statistics , volume=. 2013 , publisher=

    2013

  36. [36]

    2024 , publisher=

    A First Course in Causal Inference , author=. 2024 , publisher=

    2024

  37. [37]

    Biostatistics , volume=

    Adjusting for unmeasured spatial confounding with distance adjusted propensity score matching , author=. Biostatistics , volume=. 2019 , publisher=

    2019

  38. [38]

    2015 , publisher=

    Causal Inference in Statistics, Social, and Biomedical Sciences , author=. 2015 , publisher=

    2015

  39. [39]

    , date-added =

    Neyman, J. , date-added =. Statistical Science , pages =

  40. [40]

    arXiv preprint arXiv:2010.02108 , year=

    Causal inference with bipartite designs , author=. arXiv preprint arXiv:2010.02108 , year=

    Pith/arXiv arXiv 2010

  41. [41]

    The Annals of Probability , volume=

    Normal approximation under local dependence , author=. The Annals of Probability , volume=

  42. [42]

    and Sekhon, Jasjeet S

    Miratrix, Luke W. and Sekhon, Jasjeet S. and Yu, Bin , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume =. 2012 , month =. doi:10.1111/j.1467-9868.2012.01048.x , url =

    arXiv 2012

  43. [43]

    arXiv preprint arXiv:2408.17205 , year=

    Estimation and inference of average treatment effects under heterogeneous additive treatment effect model , author=. arXiv preprint arXiv:2408.17205 , year=

    Pith/arXiv arXiv

  44. [44]

    1925 , edition =

    Statistical Methods for Research Workers , author =. 1925 , edition =

    1925

  45. [45]

    Advances in Applied Mathematics , volume=

    On regression adjustments to experimental data , author=. Advances in Applied Mathematics , volume=. 2008 , publisher=

    2008

  46. [46]

    2000 , publisher=

    Design and analysis of cluster randomization trials in health research , author=. 2000 , publisher=

    2000

  47. [47]

    Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=

    Graph cluster randomization: Network exposure to multiple universes , author=. Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages=

  48. [48]

    The Annals of Mathematical Statistics , volume=

    A bound on tail probabilities for quadratic forms in independent random variables , author=. The Annals of Mathematical Statistics , volume=. 1971 , publisher=

    1971

  49. [49]

    Journal of the Royal Statistical Society: Series C (Applied Statistics) , volume=

    Some aspects of the design and analysis of cluster randomization trials , author=. Journal of the Royal Statistical Society: Series C (Applied Statistics) , volume=. 1998 , publisher=

    1998

  50. [50]

    Annual review of statistics and its application , volume=

    Randomized experiments in education, with implications for multilevel causal inference , author=. Annual review of statistics and its application , volume=. 2020 , publisher=

    2020

  51. [51]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Model-assisted analyses of cluster-randomized experiments , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2021 , publisher=

    2021

  52. [52]

    Journal of the American Statistical Association , volume=

    Identification and estimation of treatment and interference effects in observational studies on networks , author=. Journal of the American Statistical Association , volume=. 2021 , publisher=

    2021

  53. [53]

    arXiv preprint arXiv:2404.13442 , year=

    Difference-in-Differences under Bipartite Network Interference: A Framework for Quasi-Experimental Assessment of the Effects of Environmental Policies on Health , author=. arXiv preprint arXiv:2404.13442 , year=

    Pith/arXiv arXiv

  54. [54]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

    Randomized experimental design via geographic clustering , author=. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages=

  55. [55]

    Zigler and Georgia Papadogeorgou , title =

    Corwin M. Zigler and Georgia Papadogeorgou , title =. Statistical Science , number =. 2021 , doi =

    2021

  56. [56]

    Shankar, Shiv and Sinha, Ritwik and Mitra, Saayan and Fiterau, Madalina , year=

  57. [57]

    arXiv preprint arXiv:2402.11070 , year=

    Scalable Analysis of Bipartite Experiments , author=. arXiv preprint arXiv:2402.11070 , year=

    arXiv

  58. [58]

    Electronic Journal of Statistics , volume=

    Design and analysis of bipartite experiments under a linear exposure-response model , author=. Electronic Journal of Statistics , volume=. 2023 , publisher=

    2023

  59. [59]

    arXiv preprint arXiv:2404.04775 , year=

    Bipartite causal inference with interference, time series data, and a random network , author=. arXiv preprint arXiv:2404.04775 , year=

    Pith/arXiv arXiv

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Cluster randomized designs for one-sided bipartite experiments , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Variance reduction in bipartite experiments through correlation clustering , author=. Advances in Neural Information Processing Systems , volume=

  62. [62]

    Linear Algebra and its Applications , volume=

    Graphs for which the least eigenvalue is minimal, I , author=. Linear Algebra and its Applications , volume=. 2008 , publisher=

    2008

  63. [63]

    Causal Inference for Dyadic Outcomes in Social Network Analysis , author=

  64. [64]

    Proceedings of the ACM Web Conference 2022 , pages=

    Interference, bias, and variance in two-sided marketplace experimentation: Guidance for platforms , author=. Proceedings of the ACM Web Conference 2022 , pages=

    2022

  65. [65]

    Matching, and the Role of Digital Marketplace Design in Enabling Trade: Evidence from Airbnb (March 21, 2017) , year=

    Search, matching, and the role of digital marketplace design in enabling trade: Evidence from airbnb , author=. Matching, and the Role of Digital Marketplace Design in Enabling Trade: Evidence from Airbnb (March 21, 2017) , year=

    2017

  66. [66]

    Journal of the American Statistical Association , volume=

    General forms of finite population central limit theorems with applications to causal inference , author=. Journal of the American Statistical Association , volume=. 2017 , publisher=

    2017

  67. [67]

    Statistics in Medicine , volume=

    Variance identification and efficiency analysis in randomized experiments under the matched-pair design , author=. Statistics in Medicine , volume=. 2008 , publisher=

    2008

  68. [68]

    Statistics in medicine , volume=

    Estimating equations for association structures , author=. Statistics in medicine , volume=. 2004 , publisher=

    2004

  69. [69]

    Biometrics , pages=

    Estimating equations for parameters in means and covariances of multivariate discrete and continuous responses , author=. Biometrics , pages=. 1991 , publisher=

    1991

  70. [70]

    Proceedings of the National Academy of Sciences , volume=

    Estimating the total treatment effect in randomized experiments with unknown network structure , author=. Proceedings of the National Academy of Sciences , volume=. 2022 , publisher=

    2022

  71. [71]

    A Magyar Tudom

    Limiting distributions in simple random sampling from a finite population , author=. A Magyar Tudom. 1960 , publisher=

    1960

  72. [72]

    Biostatistics , volume=

    Bipartite interference and air pollution transport: Estimating health effects of power plant interventions , author=. Biostatistics , volume=. 2025 , publisher=

    2025

  73. [73]

    arXiv preprint arXiv:2310.14983 , year=

    Causal clustering: design of cluster experiments under network interference , author=. arXiv preprint arXiv:2310.14983 , year=

    Pith/arXiv arXiv

  74. [74]

    Biometrika , volume=

    Regression-adjusted average treatment effect estimates in stratified randomized experiments , author=. Biometrika , volume=. 2020 , publisher=

    2020

  75. [75]

    Biometrika , volume=

    Regression-assisted inference for the average treatment effect in paired experiments , author=. Biometrika , volume=. 2018 , publisher=

    2018

  76. [76]

    and Li, Fan , title =

    Wang, Bingkai and Park, Chan and Small, Dylan S. and Li, Fan , title =. Journal of the American Statistical Association , volume =. 2024 , publisher =

    2024

  77. [77]

    Randomization analysis of experimental data: The

    Rubin, Donald B , journal=. Randomization analysis of experimental data: The. 1980 , publisher=

    1980

  78. [78]

    The Annals of Statistics , volume=

    Reconciling design-based and model-based causal inferences for split-plot experiments , author=. The Annals of Statistics , volume=. 2022 , publisher=

    2022

  79. [79]

    arXiv preprint arXiv:2401.01264 , year=

    Multiple randomization designs: Estimation and inference with interference , author=. arXiv preprint arXiv:2401.01264 , year=

    arXiv

  80. [80]

    Journal of the American Statistical Association , volume=

    Using standard tools from finite population sampling to improve causal inference for complex experiments , author=. Journal of the American Statistical Association , volume=. 2018 , publisher=

    2018

Showing first 80 references.