pith. sign in

arxiv: 2509.03050 · v2 · submitted 2025-09-03 · 📊 stat.ME

Covariate Adjustment Cannot Hurt: Treatment Effect Estimation under Interference with Low-Order Outcome Interactions

Xinyi Wang , Shuangning Li This is my paper

Pith reviewed 2026-05-18 19:45 UTC · model grok-4.3

classification 📊 stat.ME
keywords covariate adjustmentinterferencetreatment effect estimationcausal inferencesparsityasymptotic variancerandomized experimentsneighborhood model
0
0 comments X

The pith

Under sparse interference networks, covariate-adjusted estimators for total treatment effects are asymptotically unbiased with variance no larger than the unadjusted estimator.

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

In randomized experiments where one unit's treatment can affect others' outcomes, direct use of covariates for adjustment has been tricky because of cross-unit dependencies. This paper shows that under a neighborhood interference model with low-order interactions and sparsity on the network, a class of covariate-adjusted estimators remains asymptotically unbiased for the total treatment effect. These estimators achieve a no-harm guarantee in which their asymptotic variance is at most as large as that of the unadjusted estimator, even when covariates can depend arbitrarily across units. The work also supplies an asymptotically conservative variance estimator that is less conservative than prior alternatives and therefore yields tighter confidence intervals in finite samples.

Core claim

Under the neighborhood interference model with low-order outcome interactions and sparsity conditions on the interference network, the proposed class of covariate-adjusted estimators is asymptotically unbiased for the total treatment effect and attains an asymptotic variance no larger than that of the unadjusted estimator, paralleling the classical no-interference result while permitting arbitrary covariate dependence.

What carries the argument

A class of covariate-adjusted estimators constructed by extending the base estimator of Cortez-Rodriguez et al. (2023) to incorporate covariates while preserving unbiasedness and the no-harm variance property under sparsity.

If this is right

  • Covariates can be used to reduce variance of total treatment effect estimates without introducing asymptotic bias under the stated conditions.
  • The developed variance estimator supports valid asymptotic inference and produces less conservative intervals than existing methods.
  • The no-harm property holds even when covariates exhibit arbitrary dependence across units.
  • The approach directly extends the classical result of Lin (2013) to settings with low-order interference.

Where Pith is reading between the lines

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

  • Similar no-harm guarantees might be provable for other interference models that admit a sparse representation.
  • Experiment designers facing social or spatial interference could adopt covariate adjustment more confidently when networks are known to be sparse.
  • Relaxing the low-order interaction assumption while retaining the variance guarantee would be a natural next direction.
  • The variance estimator's reduced conservativeness could improve power calculations for future experiments with interference.

Load-bearing premise

The data follow a neighborhood interference model with low-order interactions and the interference network satisfies the required sparsity conditions.

What would settle it

Simulate data under a dense interference network that violates sparsity, compute the asymptotic variances of the adjusted and unadjusted estimators, and observe whether the adjusted variance exceeds the unadjusted variance.

Figures

Figures reproduced from arXiv: 2509.03050 by Shuangning Li, Xinyi Wang.

Figure 1
Figure 1. Figure 1: An interference network with three units. [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An interference network with many groups of three units. [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Relative bias (top row) and mean squared error (MSE; bottom row) of Reg-SNIPE, VIM [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Relative bias (top row) and mean squared error (MSE; bottom row) of DM, Reg-SNIPE, [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relative bias (top row) and mean squared error (MSE; bottom row) of DM, Reg-SNIPE, [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Relative bias (top row) and mean squared error (MSE; bottom row) of DM, Reg-SNIPE, [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Relative bias (top row) and mean squared error (MSE; bottom row) of DM, Reg-SNIPE, [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
read the original abstract

In randomized experiments, covariates are often used to reduce variance and improve the precision of treatment effect estimates. However, in many real-world settings, interference between units, where one unit's treatment affects another's outcome, complicates causal inference. This raises a key question: how can covariates be effectively used in the presence of interference? Addressing this challenge is nontrivial, as direct covariate adjustment, such as through regression, can increase variance due to dependencies across units. In this paper, we study covariate adjustment for estimating the total treatment effect under interference. We work under a neighborhood interference model with low-order interactions and build on the estimator of Cortez-Rodriguez et al. (2023). We propose a class of covariate-adjusted estimators and show that, under sparsity conditions on the interference network, they are asymptotically unbiased and achieve a no-harm guarantee: their asymptotic variance is no larger than that of the unadjusted estimator. This parallels the classical result of Lin (2013) under no interference, while allowing for arbitrary dependence in the covariates. We further develop a variance estimator for the proposed procedures and show that it is asymptotically conservative, enabling valid inference in the presence of interference. Compared with existing approaches, the proposed variance estimator is less conservative, leading to tighter confidence intervals in finite samples.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a class of covariate-adjusted estimators for the total treatment effect under a neighborhood interference model with low-order outcome interactions. Building on the base estimator of Cortez-Rodriguez et al. (2023), it claims that under sparsity conditions on the interference network these estimators are asymptotically unbiased and satisfy a no-harm property: their asymptotic variance is no larger than that of the unadjusted estimator. The paper also develops an asymptotically conservative variance estimator that is less conservative than existing alternatives, enabling valid inference while allowing arbitrary dependence among covariates.

Significance. If the no-harm variance guarantee holds under the stated sparsity conditions, the result usefully extends the classical Lin (2013) covariate-adjustment result to interference settings where naive regression adjustment can inflate variance due to cross-unit dependence. The explicit allowance for arbitrary covariate dependence and the construction of a tighter variance estimator are practical strengths. The work is grounded in a concrete model (neighborhood interference plus low-order interactions) rather than purely asymptotic abstractions.

major comments (2)
  1. [§4, Theorem 4.2] §4, Theorem 4.2 (variance comparison): The proof that Var(adjusted) - Var(unadjusted) ≤ 0 asymptotically expands the difference and invokes network sparsity to drive higher-order covariance terms to zero. However, the argument does not explicitly bound the cross term between the covariate projection and the interference-induced error component; under only marginal sparsity (e.g., max degree growing as log n), this cross term can remain positive and violate the no-harm inequality while preserving unbiasedness. A concrete counter-example or additional assumption ruling out this sign would strengthen the claim.
  2. [Assumption 3.3] Assumption 3.3 (sparsity): The condition is stated in terms of the maximum number of paths of length k growing slower than n^α for small α. It is unclear whether this is sufficient to control the covariance between the covariate-adjusted residuals and the network-dependent potential-outcome errors when covariates are allowed to depend arbitrarily on the network structure, as claimed in the abstract.
minor comments (2)
  1. [§2 and §5] The notation for the low-order interaction terms (e.g., the definition of the k-th order neighborhood effect) is introduced in §2 but reused with slightly different indexing in the variance estimator of §5; a single consolidated definition would improve readability.
  2. [Figure 2] Figure 2 (finite-sample coverage) lacks error bars or replication counts; adding these would make the comparison to existing variance estimators easier to interpret.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments on the variance comparison and sparsity assumption. We address each point below and will revise the manuscript to strengthen the presentation of the proofs.

read point-by-point responses

  1. Referee: [§4, Theorem 4.2] §4, Theorem 4.2 (variance comparison): The proof that Var(adjusted) - Var(unadjusted) ≤ 0 asymptotically expands the difference and invokes network sparsity to drive higher-order covariance terms to zero. However, the argument does not explicitly bound the cross term between the covariate projection and the interference-induced error component; under only marginal sparsity (e.g., max degree growing as log n), this cross term can remain positive and violate the no-harm inequality while preserving unbiasedness. A concrete counter-example or additional assumption ruling out this sign would strengthen the claim.

    Authors: We thank the referee for highlighting the cross term in the variance expansion of Theorem 4.2. The proof relies on Assumption 3.3 to ensure that all covariance terms of order higher than the interaction model vanish asymptotically, which includes the cross term between the covariate projection and the interference-induced errors; the limited path counts under sparsity drive these terms to zero uniformly. To address the concern under marginal sparsity (e.g., max degree ~ log n), we will add an explicit lemma in the revision that isolates and bounds this cross term directly, showing it is non-positive under the stated conditions. We will also discuss the growth rates under which the no-harm property is guaranteed, and introduce a mild strengthening of the sparsity assumption if needed to rule out sign violations. No counter-example is expected under the paper's conditions, but the added lemma will make the argument fully transparent. revision: yes

  2. Referee: [Assumption 3.3] Assumption 3.3 (sparsity): The condition is stated in terms of the maximum number of paths of length k growing slower than n^α for small α. It is unclear whether this is sufficient to control the covariance between the covariate-adjusted residuals and the network-dependent potential-outcome errors when covariates are allowed to depend arbitrarily on the network structure, as claimed in the abstract.

    Authors: Assumption 3.3 controls network dependence via the sublinear growth of path counts, which directly bounds the covariances between any network-dependent error components (including potential-outcome errors) regardless of how the covariates are generated. Because the covariate adjustment is a linear projection and the residuals are constructed to be orthogonal to the covariates, the relevant cross-covariances with the interference errors remain governed solely by the path-count condition; arbitrary covariate-network dependence is absorbed into the fixed covariate vectors without altering the sparsity-driven decay. We will revise the manuscript to include an intermediate result (e.g., a new lemma) that explicitly derives the covariance bound between the adjusted residuals and the network errors under arbitrary covariate dependence, thereby clarifying why Assumption 3.3 suffices for both unbiasedness and the variance comparison. revision: yes

Circularity Check

0 steps flagged

No significant circularity; no-harm guarantee derived from model assumptions and sparsity

full rationale

The paper builds on the base estimator from Cortez-Rodriguez et al. (2023) but derives the central no-harm result (asymptotic variance of covariate-adjusted estimator no larger than unadjusted) explicitly from the neighborhood interference model with low-order interactions plus sparsity conditions on the network. This controls cross-unit covariances in the variance expansion, paralleling Lin (2013) without reducing to a fitted quantity or self-definition by construction. The citation is to external prior work and is not load-bearing for the new guarantee; the derivation remains self-contained against the stated assumptions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper rests on standard domain assumptions for interference models plus sparsity; no free parameters or invented entities are described in the abstract.

axioms (2)
  • domain assumption Neighborhood interference model with low-order interactions
    Invoked to define the interference structure and enable the no-harm derivation.
  • domain assumption Sparsity conditions on the interference network
    Required for asymptotic unbiasedness and variance comparison.

pith-pipeline@v0.9.0 · 5756 in / 1226 out tokens · 33203 ms · 2026-05-18T19:45:25.291444+00:00 · methodology

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

64 extracted references · 64 canonical work pages

  1. [1]

    Aronow, P. M. and Samii, C. (2017). Estimating average causal effects under general interference, with application to a social network experiment. The Annals of Applied Statistics , 11(4):1912--1947

    work page 2017

  2. [2]

    Athey, S., Eckles, D., and Imbens, G. W. (2018). Exact p-values for network interference. Journal of the American Statistical Association , 113(521):230--240

    work page 2018

  3. [3]

    G., Hudgens, M

    Barkley, B. G., Hudgens, M. G., Clemens, J. D., Ali, M., and Emch, M. E. (2020). Causal inference from observational studies with clustered interference, with application to a cholera vaccine study. The Annals of Applied Statistics , 14(3):1432--1448

    work page 2020

  4. [4]

    and Feller, A

    Basse, G. and Feller, A. (2018). Analyzing two-stage experiments in the presence of interference. Journal of the American Statistical Association , 113(521):41--55

    work page 2018

  5. [5]

    Basse, G. W. and Airoldi, E. M. (2018). Model-assisted design of experiments in the presence of network-correlated outcomes. Biometrika , 105(4):849--858

    work page 2018

  6. [6]

    and Ridder, A

    Botev, Z. and Ridder, A. (2017). Variance reduction. Wiley statsRef: Statistics reference online , 136:476

    work page 2017

  7. [7]

    Cai, J., De Janvry, A., and Sadoulet, E. (2015). Social networks and the decision to insure. American Economic Journal: Applied Economics , 7(2):81--108

    work page 2015

  8. [8]

    A., and Aronow, P

    Chang, H., Middleton, J. A., and Aronow, P. (2024). Exact bias correction for linear adjustment of randomized controlled trials. Econometrica , 92(5):1503--1519

    work page 2024

  9. [9]

    Chin, A. (2019). Regression adjustments for estimating the global treatment effect in experiments with interference. Journal of Causal Inference , 7(2):20180026

    work page 2019

  10. [10]

    Cortez-Rodriguez, M., Eichhorn, M., and Yu, C. L. (2023). Exploiting neighborhood interference with low-order interactions under unit randomized design. Journal of Causal Inference , 11(1):20220051

    work page 2023

  11. [11]

    Ding, P. (2024). A first course in causal inference . Chapman and Hall/CRC

    work page 2024

  12. [12]

    Eckles, D., Karrer, B., and Ugander, J. (2017). Design and analysis of experiments in networks: Reducing bias from interference. Journal of Causal Inference , 5(1):20150021

    work page 2017

  13. [13]

    Eichhorn, M., Khan, S., Ugander, J., and Yu, C. L. (2024). Low-order outcomes and clustered designs: combining design and analysis for causal inference under network interference. arXiv preprint arXiv:2405.07979

    work page arXiv 2024

  14. [14]

    and R\'enyi, A

    Erd o s, P. and R\'enyi, A. (1959). On random graphs. i. Publicationes Mathematicae Debrecen , 6:290--297

    work page 1959

  15. [15]

    Fan, X., Leng, C., and Wu, W. (2025). Causal inference under interference: Regression adjustment and optimality. arXiv preprint arXiv:2502.06008

    work page arXiv 2025

  16. [16]

    Fisher, R. A. (1971). The design of experiments . Springer

    work page 1971

  17. [17]

    Fogarty, C. B. (2018). Regression-assisted inference for the average treatment effect in paired experiments. Biometrika , 105(4):994--1000

    work page 2018

  18. [18]

    M., and Mealli, F

    Forastiere, L., Airoldi, E. M., and Mealli, F. (2021). Identification and estimation of treatment and interference effects in observational studies on networks. Journal of the American Statistical Association , 116(534):901--918

    work page 2021

  19. [19]

    Freedman, D. A. (2008). On regression adjustments to experimental data. Advances in Applied Mathematics , 40(2):180--193

    work page 2008

  20. [20]

    Galeotti, A., Golub, B., and Goyal, S. (2020). Targeting interventions in networks. Econometrica , 88(6):2445--2471

    work page 2020

  21. [21]

    and Ding, P

    Gao, M. and Ding, P. (2023). Causal inference in network experiments: regression-based analysis and design-based properties. arXiv preprint arXiv:2309.07476

    work page arXiv 2023

  22. [22]

    Glasserman, P. (2004). Monte Carlo methods in financial engineering , volume 53. Springer

    work page 2004

  23. [23]

    Han, K., Li, S., Mao, J., and Wu, H. (2023). Detecting interference in online controlled experiments with increasing allocation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 661--672

    work page 2023

  24. [24]

    and Ugander, J

    Han, K. and Ugander, J. (2023). Model-based regression adjustment with model-free covariates for network interference. Journal of Causal Inference , 11(1):20230005

    work page 2023

  25. [25]

    Holland, P. W. (1986). Statistics and causal inference. Journal of the American statistical Association , 81(396):945--960

    work page 1986

  26. [26]

    Hu, Y., Li, S., and Wager, S. (2022). Average direct and indirect causal effects under interference. Biometrika , 109(4):1165--1172

    work page 2022

  27. [27]

    Hu, Y., Li, S., and Wager, S. (2025). Optimal targeting in dynamic systems. arXiv preprint arXiv:2507.00312

    work page arXiv 2025

  28. [28]

    Hudgens, M. G. and Halloran, M. E. (2008). Toward causal inference with interference. Journal of the american statistical association , 103(482):832--842

    work page 2008

  29. [29]

    Imbens, G. W. and Rubin, D. B. (2015). Causal inference in statistics, social, and biomedical sciences . Cambridge university press

    work page 2015

  30. [30]

    and Wang, G

    Kitagawa, T. and Wang, G. (2023). Who should get vaccinated? individualized allocation of vaccines over sir network. Journal of Econometrics , 232(1):109--131

    work page 2023

  31. [31]

    Krueger, A. B. (1999). Experimental estimates of education production functions. The quarterly journal of economics , 114(2):497--532

    work page 1999

  32. [32]

    Lemieux, C. (2014). Control variates. Wiley StatsRef: Statistics Reference Online , pages 1--8

    work page 2014

  33. [33]

    Leung, M. P. (2020). Treatment and spillover effects under network interference. Review of Economics and Statistics , 102(2):368--380

    work page 2020

  34. [34]

    Leung, M. P. (2022). Causal inference under approximate neighborhood interference. Econometrica , 90(1):267--293

    work page 2022

  35. [35]

    and Wager, S

    Li, S. and Wager, S. (2022). Random graph asymptotics for treatment effect estimation under network interference. The Annals of Statistics , 50(4):2334--2358

    work page 2022

  36. [36]

    Lin, W. (2013). Agnostic notes on regression adjustments to experimental data: Reexamining Freedman’s critique . The Annals of Applied Statistics , 7(1):295 -- 318

    work page 2013

  37. [37]

    G., Saul, B., Clemens, J

    Liu, L., Hudgens, M. G., Saul, B., Clemens, J. D., Ali, M., and Emch, M. E. (2019). Doubly robust estimation in observational studies with partial interference. Stat , 8(1):e214

    work page 2019

  38. [38]

    Lu, X., Wang, Y., and Zhang, Z. (2024). Adjusting auxiliary variables under approximate neighborhood interference. arXiv preprint arXiv:2411.19789

    work page arXiv 2024

  39. [39]

    and Wooldridge, J

    Negi, A. and Wooldridge, J. M. (2021). Revisiting regression adjustment in experiments with heterogeneous treatment effects. Econometric Reviews , 40(5):504--534

    work page 2021

  40. [40]

    Nelson, B. L. (1990). Control variate remedies. Operations Research , 38(6):974--992

    work page 1990

  41. [41]

    Park, C., Chen, G., Yu, M., and Kang, H. (2024). Minimum resource threshold policy under partial interference. Journal of the American Statistical Association , 119(548):2881--2894

    work page 2024

  42. [42]

    Penrose, M. (2003). Random Geometric Graphs , volume 5 of Oxford Studies in Probability . Oxford University Press

    work page 2003

  43. [43]

    Rosenbaum, P. R. (2002). Covariance adjustment in randomized experiments and observational studies. Statistical Science , 17(3):286--327

    work page 2002

  44. [44]

    Rosenbaum, P. R. (2007). Interference between units in randomized experiments. Journal of the american statistical association , 102(477):191--200

    work page 2007

  45. [45]

    E., Anderson, G

    Rossouw, J. E., Anderson, G. L., Prentice, R. L., LaCroix, A. Z., Kooperberg, C., Stefanick, M. L., Jackson, R. D., Beresford, S. A., Howard, B. V., Johnson, K. C., et al. (2002). Risks and benefits of estrogen plus progestin in healthy postmenopausal women: principal results from the women's health initiative randomized controlled trial. Jama , 288(3):321--333

    work page 2002

  46. [46]

    Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of educational Psychology , 66(5):688

    work page 1974

  47. [47]

    Sacerdote, B. (2001). Peer effects with random assignment: Results for D artmouth roommates. The Quarterly Journal of Economics , 116(2):681--704

    work page 2001

  48. [48]

    S \"a vje, F., Aronow, P., and Hudgens, M. (2021). Average treatment effects in the presence of unknown interference. Annals of statistics , 49(2):673

    work page 2021

  49. [49]

    C., Hagenbuchner, M., and Monfardini, G

    Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. (2008). The graph neural network model. IEEE transactions on neural networks , 20(1):61--80

    work page 2008

  50. [50]

    Z., Burghardt, J., and McConnell, S

    Schochet, P. Z., Burghardt, J., and McConnell, S. (2008). Does job corps work? impact findings from the national job corps study. American economic review , 98(5):1864--1886

    work page 2008

  51. [51]

    Sobel, M. E. (2006). What do randomized studies of housing mobility demonstrate? causal inference in the face of interference. Journal of the American Statistical Association , 101(476):1398--1407

    work page 2006

  52. [52]

    and Ding, P

    Su, F. and Ding, P. (2021). Model-assisted analyses of cluster-randomized experiments. Journal of the Royal Statistical Society Series B: Statistical Methodology , 83(5):994--1015

    work page 2021

  53. [53]

    Tchetgen, E. J. T. and VanderWeele, T. J. (2012). On causal inference in the presence of interference. Statistical methods in medical research , 21(1):55--75

    work page 2012

  54. [54]

    and Kao, E

    Toulis, P. and Kao, E. (2013). Estimation of causal peer influence effects. In International conference on machine learning , pages 1489--1497. PMLR

    work page 2013

  55. [55]

    and Yin, H

    Ugander, J. and Yin, H. (2023). Randomized graph cluster randomization. Journal of Causal Inference , 11(1):20220014

    work page 2023

  56. [56]

    Viviano, D. (2020). Experimental design under network interference. arXiv preprint arXiv:2003.08421

    work page arXiv 2020

  57. [57]

    Viviano, D. (2025). Policy targeting under network interference. Review of Economic Studies , 92(2):1257--1292

    work page 2025

  58. [58]

    and Rudder, J

    Viviano, D. and Rudder, J. (2024). Policy design in experiments with unknown interference. arXiv preprint arXiv:2011.08174 , 4

    work page arXiv 2024

  59. [59]

    S., and Li, F

    Wang, B., Park, C., Small, D. S., and Li, F. (2024). Model-robust and efficient covariate adjustment for cluster-randomized experiments. Journal of the American Statistical Association , 119(548):2959--2971

    work page 2024

  60. [60]

    Wang, B., Susukida, R., Mojtabai, R., Amin-Esmaeili, M., and Rosenblum, M. (2023). Model-robust inference for clinical trials that improve precision by stratified randomization and covariate adjustment. Journal of the American Statistical Association , 118(542):1152--1163

    work page 2023

  61. [61]

    L., Airoldi, E

    Yu, C. L., Airoldi, E. M., Borgs, C., and Chayes, J. T. (2022). Estimating the total treatment effect in randomized experiments with unknown network structure. Proceedings of the National Academy of Sciences , 119(44):e2208975119

    work page 2022

  62. [62]

    and Imai, K

    Zhang, Y. and Imai, K. (2023). Individualized policy evaluation and learning under clustered network interference. arXiv preprint arXiv:2311.02467

    work page arXiv 2023

  63. [63]

    and Ding, P

    Zhao, A. and Ding, P. (2022). Reconciling design-based and model-based causal inferences for split-plot experiments. The Annals of Statistics , 50(2):1170--1192

    work page 2022

  64. [64]

    Zhao, A., Ding, P., and Li, F. (2024). Covariate adjustment in randomized experiments with missing outcomes and covariates. Biometrika , 111(4):1413--1420

    work page 2024