Pith. sign in

REVIEW 1 major objections 1 minor 35 references

Horvitz-Thompson estimators with three-fold splitting support valid causal inference for outcomes defined on network edges.

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 · grok-4.3

2026-06-28 17:02 UTC pith:3VNQSI6Z

load-bearing objection Three-fold splitting lets ML adjust edge-level network causal estimators without efficiency loss, but whether it fully breaks shared-unit dependence in directed graphs is the key open question. the 1 major comments →

arxiv 2606.00965 v1 pith:3VNQSI6Z submitted 2026-05-31 stat.ME

Design-based edge-level causal inference with machine learning assisted covariate adjustment

Haoyang Yu , Yilin Li , Lu Deng , Yong Wang , Xin Lu , Hanzhong Liu This is my paper
classification stat.ME
keywords causal inferencenetwork dataedge-level outcomesHorvitz-Thompson estimatorsample splittingmachine learningdyadic interferencecovariate adjustment
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.

This paper develops design-based methods for estimating causal effects when the outcomes sit on the directed edges of a network rather than on the nodes. Edge outcomes depend on the joint treatment assignments of pairs of units, creating dependence that invalidates standard node-level procedures. The authors construct Horvitz-Thompson estimators for a broad class of such effects, prove asymptotic normality under mild conditions, and supply variance estimators that use identifiable network dependence to produce tighter bounds. They further show how to adjust for covariates with linear or machine-learning methods by replacing the usual two-fold sample split with a three-fold scheme that restores the conditional independence needed for unbiased estimation, then add a calibration step that prevents any asymptotic efficiency loss relative to the unadjusted estimator.

Core claim

We construct Horvitz-Thompson estimators for a general class of edge-level causal effects in directed networks under dyadic interference and establish their asymptotic normality under mild regularity conditions. We develop variance estimators that exploit identifiable components of network dependence. To improve efficiency we introduce a three-fold sample splitting and cross-fitting procedure that restores the conditional independence required for unbiased covariate-adjusted estimation; under a stability condition the resulting estimator is asymptotically normal for both linear and flexible machine-learning adjustments, and a calibration step guarantees no asymptotic efficiency loss relative

What carries the argument

Horvitz-Thompson estimators for edge-level causal effects paired with three-fold sample splitting and cross-fitting for covariate adjustment

Load-bearing premise

The three-fold sample splitting scheme restores the conditional independence required for unbiased estimation of the covariate-adjusted estimator.

What would settle it

A Monte Carlo experiment in which the three-fold cross-fitting procedure is applied to data with known edge-level effects yet the covariate-adjusted estimator exhibits persistent finite-sample bias that the calibration step does not remove would falsify the independence-restoration claim.

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

If this is right

  • The estimators are asymptotically normal under mild regularity conditions.
  • Variance estimators that use network dependence components produce substantially less conservative bounds than classical approaches.
  • Covariate-adjusted estimators remain asymptotically normal when linear or machine-learning methods are used under the stated stability condition.
  • The calibration step ensures the adjusted estimator loses no asymptotic efficiency relative to the unadjusted version.
  • Simulation studies and a real-data application demonstrate the efficiency gains.

Where Pith is reading between the lines

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

  • The three-fold splitting logic may generalize to other settings with overlapping units, such as spatial or temporal data with local dependence.
  • Analysts evaluating interventions that affect pairwise interactions in social or biological networks could obtain more precise estimates than node-level methods allow.
  • If the stability condition holds in practice, the procedure offers a template for incorporating flexible machine-learning adjustments into other design-based network estimators without efficiency penalties.

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

1 major / 1 minor

Summary. The manuscript develops design-based Horvitz-Thompson estimators for a general class of edge-level causal effects in directed networks under dyadic interference. It establishes asymptotic normality of these estimators under mild regularity conditions, constructs variance estimators that exploit identifiable components of the network dependence structure, and introduces a three-fold sample splitting and cross-fitting scheme to enable covariate adjustment (including via machine learning) while restoring the conditional independence needed for unbiased estimation. A calibration step is proposed to ensure the adjusted estimator incurs no asymptotic efficiency loss relative to the unadjusted version. The results are illustrated via simulations and a real-data application.

Significance. If the central technical claims hold, particularly the validity of the three-fold splitting procedure and the stability condition, the work would advance causal inference methodology for network data by extending design-based estimation to edge-level outcomes with complex dependence and by providing a practical route to efficiency gains via flexible ML adjustment without asymptotic penalty. The explicit design-based foundation (avoiding outcome model assumptions) and the calibration device are notable strengths that could be useful in applied network settings.

major comments (1)
  1. [section on three-fold sample splitting] The section describing the three-fold sample splitting and cross-fitting scheme (the paragraph addressing the failure of two-fold splitting due to shared units and the proposed solution): the claim that three folds suffice to restore the conditional independence between the nuisance estimator and the edge outcome for all pairs requires an explicit lemma or argument showing that, after the three-way partition, no two units connected by a directed edge can appear together in a fold used for nuisance estimation. Without this, the unbiasedness and subsequent asymptotic normality of the ML-adjusted estimator remain at risk in dense directed graphs, which is load-bearing for the efficiency claim.
minor comments (1)
  1. [abstract] The abstract invokes a 'stability condition' for asymptotic normality of the covariate-adjusted estimator but does not state its precise form; moving a brief definition or reference to the relevant assumption into the abstract would improve readability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. We respond to the single major comment below.

read point-by-point responses

  1. Referee: The section describing the three-fold sample splitting and cross-fitting scheme (the paragraph addressing the failure of two-fold splitting due to shared units and the proposed solution): the claim that three folds suffice to restore the conditional independence between the nuisance estimator and the edge outcome for all pairs requires an explicit lemma or argument showing that, after the three-way partition, no two units connected by a directed edge can appear together in a fold used for nuisance estimation. Without this, the unbiasedness and subsequent asymptotic normality of the ML-adjusted estimator remain at risk in dense directed graphs, which is load-bearing for the efficiency claim.

    Authors: We appreciate the referee highlighting this point. The manuscript describes the three-fold scheme and states that it restores the conditional independence needed for unbiased estimation after noting the failure of two-fold splitting. We agree that an explicit lemma would make the argument more rigorous and directly address concerns in dense directed graphs. In the revision we will add a lemma proving that a random three-way partition of the units ensures that, for any directed edge (i,j), units i and j do not co-occur in any fold used for nuisance estimation. This will formally establish the required separation and support the unbiasedness and asymptotic normality of the ML-adjusted estimator. The addition will be placed in the relevant section and will not alter the existing results or proofs. revision: yes

Circularity Check

0 steps flagged

No circularity: design-based HT estimators and three-fold splitting rest on known probabilities and stated independence restoration, not self-referential fits or citations.

full rationale

The paper's core claims rest on Horvitz-Thompson estimators constructed from known treatment assignment probabilities for edge-level outcomes, with asymptotic normality derived under explicit regularity conditions and a stability condition for the ML-adjusted version. The three-fold splitting is introduced as a procedural fix for dependence induced by shared units, with validity asserted directly from the partition restoring conditional independence rather than from any fitted parameter or self-citation chain. No equations or steps reduce by construction to their inputs, no predictions are statistically forced by prior fits, and no load-bearing uniqueness theorems or ansatzes are imported from overlapping prior work. The calibration step is presented as an independent efficiency guarantee. This is a standard self-contained design-based derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on mild regularity conditions for asymptotic normality and a stability condition for the covariate-adjusted estimator; these are standard domain assumptions in asymptotic network inference and are not derived within the paper.

axioms (2)
  • domain assumption Mild regularity conditions
    Invoked to establish asymptotic normality of the Horvitz-Thompson estimators and variance estimators.
  • domain assumption Stability condition
    Required for asymptotic normality and efficiency properties of the machine-learning-adjusted estimator after three-fold cross-fitting.

pith-pipeline@v0.9.1-grok · 5756 in / 1377 out tokens · 40790 ms · 2026-06-28T17:02:24.342054+00:00 · methodology

0 comments
read the original abstract

We study design-based causal inference for edge-level outcomes in directed networks under dyadic interference. In this setting, outcomes are defined on directed edges and depend on the joint treatment assignments of pairs of units, inducing a complex dependence structure that invalidates standard estimation and inference procedures developed for node-level data. We construct Horvitz--Thompson estimators for a general class of edge-level causal effects and establish their asymptotic normality under mild regularity conditions. To enable valid inference, we develop variance estimators that exploit identifiable components of network dependence, yielding substantially less conservative bounds than classical approaches. To improve efficiency, we incorporate auxiliary covariates through a sample splitting and cross-fitting procedure. A key technical challenge is that standard two-fold sample splitting fails in the presence of edge-level outcomes due to the dependence induced by shared units. To address this issue, we introduce a three-fold sample splitting and cross-fitting scheme that restores the conditional independence required for unbiased estimation. Under a stability condition, the resulting covariate-adjusted estimator is asymptotically normal and accommodates both linear adjustment and flexible machine learning methods. We further introduce a calibration step that guarantees no asymptotic efficiency loss relative to the unadjusted estimator. Simulation studies and a real-data application confirm the theoretical results and demonstrate substantial efficiency gains.

Figures

Figures reproduced from arXiv: 2606.00965 by Hanzhong Liu, Haoyang Yu, Lu Deng, Xin Lu, Yilin Li, Yong Wang.

Figure 1
Figure 1. Figure 1: Comparison of two-fold and three-fold sample splitting. [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Bias and variance for DGP 1 in Scenario 1. [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Bias and variance for DGP 1 in Scenario 2. [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Empirical coverage probability (CP) and mean confidence interval length [PITH_FULL_IMAGE:figures/full_fig_p028_4.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

35 extracted references · 7 canonical work pages · 2 internal anchors

  1. [1]

    Aronow, P. M. & Samii, C. (2017), ‘Estimating average causal effects under general interfer- ence, with application to a social network experiment’,The Annals of Applied Statistics 11(4), 1912–1947

    2017

  2. [2]

    Balcan, D., Colizza, V., Gon¸ calves, B., Hu, H., Ramasco, J. J. & Vespignani, A. (2009), ‘Multiscale mobility networks and the spatial spreading of infectious diseases’,Proceed- ings of the National Academy of Sciences106(51), 21484–21489

    2009

  3. [3]

    Bloniarz, A., Liu, H., Zhang, C.-H., Sekhon, J. S. & Yu, B. (2016), ‘Lasso adjustments of treatment effect estimates in randomized experiments’,Proceedings of the National Academy of Sciences113(27), 7383–7390

    2016

  4. [4]

    & Aronow, P

    Carlson, J., Incerti, T. & Aronow, P. (2024), ‘Dyadic clustering in international relations’, Political Analysis32(2), 186–198

    2024

  5. [5]

    Cohen, P. L. & Fogarty, C. B. (2024), ‘No-harm calibration for generalized oaxaca–blinder estimators’,Biometrika111(1), 331–338

    2024

  6. [6]

    Cortez-Rodriguez, M., Eichhorn, M. & Yu, C. L. (2023), ‘Exploiting neighborhood inter- ference with low-order interactions under unit randomized design’,Journal of Causal Inference11(1), 20220051. 32 De Jong, P. (1990), ‘A central limit theorem for generalized multilinear forms’,Journal of Multivariate Analysis34(2), 275–289

    2023

  7. [7]

    & Chen, C

    Deng, L., Li, Y., Zhang, J., Wang, Y. & Chen, C. (2024), ‘Unbiased estimation for to- tal treatment effect under interference using aggregated dyadic data’. arXiv preprint arXiv:2402.12653

    work page arXiv 2024

  8. [8]

    it’s always about the people. enron is no different

    Diesner, J., Frantz, T. L. & Carley, K. M. (2005), ‘Communication networks from the enron email corpus “it’s always about the people. enron is no different”’,Computational & Mathematical Organization Theory11(3), 201–228

    2005

  9. [9]

    (2024),A First Course in Causal Inference, CRC Press

    Ding, P. (2024),A First Course in Causal Inference, CRC Press

    2024

  10. [10]

    Freedman, D. A. (2008), ‘On regression adjustments to experimental data’,Advances in Applied Mathematics40(2), 180–193

    2008

  11. [11]

    & Basse, G

    Guo, K. & Basse, G. (2023), ‘The generalized oaxaca–blinder estimator’,Journal of the American Statistical Association118(541), 524–536

    2023

  12. [12]

    Collis, A., Moehring, A., Sowrirajan, T. et al. (2020), ‘Interdependence and the cost of uncoordinated responses to covid-19’,Proceedings of the National Academy of Sciences 117(33), 19837–19843

    2020

  13. [13]

    Hudgens, M. G. & Halloran, M. E. (2008), ‘Toward causal inference with interference’, Journal of the American Statistical Association103(482), 832–842

    2008

  14. [14]

    Imbens, G. W. & Rubin, D. B. (2015),Causal Inference in Statistics, Social, and Biomedical Sciences: An Introduction, Cambridge university press. 33

    2015

  15. [15]

    (2023), ‘High-dimensional central limit theorems for homogeneous sums’,Journal of Theoretical Probability36(1), 1–45

    Koike, Y. (2023), ‘High-dimensional central limit theorems for homogeneous sums’,Journal of Theoretical Probability36(1), 1–45

    2023

  16. [16]

    & Ding, P

    Lei, L. & Ding, P. (2021), ‘Regression adjustment in completely randomized experiments with a diverging number of covariates’,Biometrika108(4), 815–828

    2021

  17. [17]

    Leung, M. P. (2020), ‘Treatment and spillover effects under network interference’,Review of Economics and Statistics102(2), 368–380

    2020

  18. [18]

    & Wager, S

    Li, S. & Wager, S. (2022), ‘Random graph asymptotics for treatment effect estimation under network interference’,The Annals of Statistics50(4), 2334–2358

    2022

  19. [19]

    Causal inference with dyadic data in randomized experiments

    Li, Y., Deng, L., Wang, Y. & Miao, W. (2025), ‘Causal inference with dyadic data in randomized experiments’. arXiv preprint arXiv:2505.20780

    work page internal anchor Pith review arXiv 2025

  20. [20]

    (2013), ‘Agnostic notes on regression adjustments to experimental data: Reexam- ining freedman’s critique’,The Annals of Applied Statistics7(1), 295–318

    Lin, W. (2013), ‘Agnostic notes on regression adjustments to experimental data: Reexam- ining freedman’s critique’,The Annals of Applied Statistics7(1), 295–318

    2013

  21. [21]

    & Yang, Y

    Liu, H. & Yang, Y. (2020), ‘Regression-adjusted average treatment effect estimates in stratified randomized experiments’,Biometrika107(4), 935–948

    2020

  22. [22]

    & Hudgens, M

    Liu, L. & Hudgens, M. G. (2014), ‘Large sample randomization inference of causal ef- fects in the presence of interference’,Journal of the American Statistical Association 109(505), 288–301

    2014

  23. [23]

    & Liu, H

    Lu, X., Li, H. & Liu, H. (2026), ‘Causal inference under uniformly bounded neighbourhood interference’,Journal of the Royal Statistical Society Series B: Statistical Methodology in press. 34

    2026

  24. [24]

    & Ding, P

    Lu, X., Shi, L., Liu, H. & Ding, P. (2025), ‘Conditional cross-fitting for unbiased machine- learning-assisted covariate adjustment in randomized experiments’. arXiv preprint arXiv:2508.15664

    work page arXiv 2025

  25. [25]

    & Wang, Y

    Lu, X., Yang, F. & Wang, Y. (2025), ‘Debiased regression adjustment in completely randomized experiments with moderately high-dimensional covariates’,The Annals of Statistics53(4), 1535–1558

    2025

  26. [26]

    & Zhang, Q

    Muris, C., Pakel, C. & Zhang, Q. (2025), ‘Dyadic data with ordered outcome variables’. arXiv preprint arXiv:2507.16689

    work page arXiv 2025

  27. [27]

    (1923), ‘On the application of probability theory to agricultural experiments

    Neyman, J. (1923), ‘On the application of probability theory to agricultural experiments. essay on principles. section 9’,Statistical Science5(4), 465–472

    1923

  28. [28]

    Rubin, D. B. (1974), ‘Estimating causal effects of treatments in randomized and nonran- domized studies’,Journal of Educational Psychology66(5), 688–701

    1974

  29. [29]

    Rubin, D. B. (1980), ‘Randomization analysis of experimental data: The fisher randomiza- tion test comment’,Journal of the American Statistical Association75(371), 591–593. S¨ avje, F., Aronow, P. & Hudgens, M. (2021), ‘Average treatment effects in the presence of unknown interference’,The Annals of Statistics49(2), 673–701

    1980

  30. [30]

    arXiv preprint arXiv:2411.00947 , year=

    Shi, L. & Ding, P. (2024), ‘Asymptotic theory of the quadratic assignment procedure for dyadic data analysis’. arXiv preprint arXiv:2411.00947

    work page arXiv 2024

  31. [31]

    & Wainwright, M

    Su, F., Mou, W., Ding, P. & Wainwright, M. J. (2023), ‘A decorrelation method for general regression adjustment in randomized experiments’. arXiv preprint arXiv:2311.10076

    work page arXiv 2023

  32. [32]

    Sussman, D. L. & Airoldi, E. M. (2017), ‘Elements of estimation theory for causal effects in the presence of network interference’. arXiv preprint arXiv:1702.03578. 35

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  33. [33]

    Tchetgen, E. J. T. & VanderWeele, T. J. (2012), ‘On causal inference in the presence of interference’,Statistical Methods in Medical Research21(1), 55–75

    2012

  34. [34]

    L., Airoldi, E

    Yu, C. L., Airoldi, E. M., Borgs, C. & Chayes, J. T. (2022), ‘Estimating the total treatment effect in randomized experiments with unknown network structure’,Proceedings of the National Academy of Sciences119(44), e2208975119

    2022

  35. [35]

    Zhao, A., Ding, P. & Li, F. (2024), ‘Covariate adjustment in randomized experiments with missing outcomes and covariates’,Biometrika111(4), 1413–1420. 36

    2024