REVIEW 5 minor 1 cited by
Design-based causal inference in bipartite experiments
T0 review · 0 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Under three sparsity conditions on the bipartite graph, the paper proves that the Hájek estimator for the total treatment effect is consistent, asymptotically normal, admits a conservative variance estimator, and can be improved by…
desk verdict A solid and honest design-based theory for bipartite experiments under explicit sparsity; the proofs are thorough and the scope limitations are clearly acknowledged. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Exposure indicators $T_i = \prod_{k \in G_{+i}} Z_k$ and $C_i = \prod_{k \in G_{+i}} (1-Z_k)$ record whether all intervention units linked to outcome unit $i$ are treated or control; Hájek weighting with $p^{G_{+i}}$ and $(1-p)^{G_{+i}}$ turns them into ratio estimators. Consistency is proved by bounding the variance of the numerator with a graph-degree lemma. Asymptotic normality is carried by a general random-polynomial CLT, obtained by writing the estimator as a sum of centered Bernoulli products and applying a martingale central limit theorem. The variance formula involves the overlap matrices $(\Lambda_z)_{ij} = p_z^{-|G_{+i}\cap G_{+j}|}-1$ and $(\Lambda_\tau)_{ij} = \mathbf{1}\{G_{+i}\cap G_{+j}\neq\emptyset\}$, and the conservative variance estimator uses the elementary bound $\mathrm{cov}(\hat{\mu}_1,\hat{\mu}_0)^2 \le \mathrm{var}(\hat{\mu}_1)\mathrm{var}(\hat{\mu}_0)$ because $\Lambda_\tau$ need not be positive semidefinite in general graphs. Covariate adjustment is formulated as maximizing an estimable efficiency-gain quadratic whose solution is computed with the Moore–Penrose pseudoinverse.
What would settle it
Construct a dense bipartite graph—for instance, make every outcome unit connect to a fixed fraction of the intervention units so $\max_i G_{+i}$ grows with $m$—generate potential outcomes with substantial overlap dependence, and run the proposed estimator and its conservative interval at moderate $n$. If the estimated coverage of the nominal 95% intervals drifts away from 95% as the graph densifies, or if the ratio $\hat{v}_{n,\mathrm{UB}}/v_n$ fails to approach a stable bound, the sparsity-based central limit theorem is the part that breaks.
Extended reading notes
Core claim
The paper's central claim is that the Hájek estimator $\hat{\tau} = \hat{\mu}_1 - \hat{\mu}_0$ consistently estimates the total treatment effect $\tau = n^{-1}\sum_i \{Y_i(\mathbf{1}_m) - Y_i(\mathbf{0}_m)\}$ under Bernoulli assignment when the bipartite graph is sparse. Sparsity requires each outcome unit to connect to only a bounded number of intervention units, each intervention unit to connect to a vanishing fraction of outcome units, and each intervention unit to overlap with only a bounded number of other intervention units. Under these conditions the estimator is asymptotically normal with variance $v_n$ given by a quadratic form in the centered potential outcomes weighted by overlap counts $|G_{+i} \cap G_{+j}|$, and the plug-in variance estimator $\hat{v}_{n,\mathrm{UB}}$ converges to a sharp upper bound of $v_n$, so nominal Wald intervals are conservative. The paper also constructs a feasible covariate-adjusted estimator that is asymptotically no less efficient than the unadjusted one while retaining valid inference.
Load-bearing premise
The whole argument gives way if the bipartite graph is dense: when outcome units connect to a growing number of intervention units, or when intervention units share many outcome units, the consistency and normality proofs no longer hold, and the paper notes even its variance bounds may fail.
Editorial extensions
If this is right
- For sparse bipartite experiments, researchers can report point estimates and confidence intervals for the total treatment effect without specifying an outcome or exposure model.
- Classic Bernoulli randomization and cluster randomization appear as boundary cases of the variance formula, so the results unify two existing design-based literatures.
- The conservative variance estimator guarantees asymptotically valid (at least nominal level) Wald intervals whenever the sparsity assumptions hold.
- The covariate-adjusted estimator is asymptotically at least as efficient as the unadjusted Hájek estimator, and its variance estimator remains conservative, so efficiency gains come without sacrificing inference validity.
- The sparsity conditions give practical guidance: if the bipartite graph is too dense for the method, the analysis signals the need for other estimands or additional structural assumptions.
Reading between the lines
- The rate condition linking $\max_k G_{k+}$, $m$, and $n$ could be turned into a diagnostic: practitioners can check whether their graph's density is within the regime where the confidence intervals are trustworthy.
- For dense graphs, the paper's own Amazon marketplace example suggests a route the authors leave implicit: coarsen the intervention units (e.g., cluster items) to restore sparsity, which connects this work to cluster-randomization design.
- The same random-polynomial CLT machinery should extend to unequal treatment probabilities or stratified Bernoulli assignment, because the martingale argument does not rely on identical $p$.
- The equality condition characterizing when the variance upper bound is tight could be tested in simulations by varying the correlation structure of potential outcomes, revealing how much power the conservative intervals sacrifice in practice.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops a design-based inference framework for bipartite experiments under Bernoulli randomization. Treatment is randomized over m intervention units, outcomes are measured on n outcome units, and a fixed bipartite graph encodes which intervention units affect which outcome units. The authors define the total treatment effect as the average difference between the all-treated and all-control potential outcomes, propose a Hájek-type estimator, and prove consistency under explicit sparsity conditions (Assumptions 1–4). They further prove asymptotic normality via a general random-polynomial central limit theorem (Theorem 3.2), construct a conservative variance estimator (Theorem 3.3), and develop a covariate-adjusted estimator that is asymptotically no less efficient than the unadjusted estimator (Theorem 4.1). The theory is supplemented with simulations on synthetic bipartite graphs and on a graph derived from a power-plant air-pollution study.
Significance. If the results hold, this is a valuable contribution to causal inference under interference. The paper provides a model-agnostic alternative to existing model-based analyses of bipartite experiments, gives explicit assumptions under which consistency and asymptotic normality hold, and reduces to classical results for Bernoulli and cluster randomization as special cases. The general random-polynomial CLT in Section A is a useful standalone tool. The variance estimator is conservative by construction, and the covariate adjustment strategy has a clean efficiency-gain interpretation. The proofs in the supplementary material are detailed and follow standard martingale/Chebyshev arguments, and the simulations support the finite-sample claims. The main limitation, that Assumptions 4 and 5 exclude dense bipartite graphs, is explicitly acknowledged by the authors in Sections 3.2 and 3.3 and is an applicability restriction rather than an internal inconsistency.
minor comments (5)
- [Section 3.3 / Remark 3.1] The displayed regularity condition in Remark 3.1 appears truncated, with the expression ending in "−∞" where it should presumably read "→ ∞"; this should be corrected for clarity.
- [Theorem 3.3(b)] Theorem 3.3(b) as stated does not exclude the degenerate case v_{n,UB} = 0, in which the ratio v̂_{n,UB}/v_{n,UB} is undefined. The proof implicitly assumes a non-degeneracy condition on the weighted covariance matrices. Adding an explicit condition such as v_{n,UB} > 0, or importing the non-degeneracy condition from Theorem 3.2, would make the statement precise.
- [Supplementary Material, Proposition C.1(b)] The proof of Proposition C.1(b) contains a factor-of-two slip in the displayed inequality: the bound on 2|(∑α)(∑β)| is used to justify inequalities involving unweighted sums of squares without the factor of 2. The positive-semidefiniteness claim is true, but the displayed argument should be corrected so that the derivation matches the claimed conclusion.
- [Section 3.3, Assumption 5] The text following Assumption 5 notes that the proof accommodates a value of B that grows polynomially in n, while the formal assumption fixes an absolute constant B. The relation between the formal assumption and this asserted relaxation should be stated explicitly, for example by marking the relaxation as a remark or by restating the assumption in the supplementary material.
- [Section 5, Tables 2 and 3] The simulation tables report bias, standard error, estimated standard error, coverage, and power, but no Monte Carlo standard errors are given. Reporting Monte Carlo standard errors for the estimated coverage and power, or at least noting their magnitude, would help readers assess the precision of the simulation results.
Circularity Check
No significant circularity: the main results are conditional theorems proved from stated assumptions with self-contained arguments.
full rationale
The paper's central claims are conditional asymptotic results: consistency of the Hájek estimator under Assumptions 1–4, asymptotic normality under Assumptions 1–5 plus a non-degeneracy condition, and conservativeness of the variance estimator. These are not disguised restatements of inputs. The consistency proof decomposes the estimation error into a centered Horvitz–Thompson term and a denominator that converges to 1, then bounds the variance using Lemma B.1, which is a combinatorial graph-degree bound rather than an assumed conclusion. The central limit theorem is derived by representing the estimator as a random polynomial and verifying the conditions of the paper's own Theorem A.1, whose proof uses the external martingale CLT of Hall and Heyde; the sparsity conditions in Assumptions 4 and 5 are explicitly used to verify bounded coefficients and limited overlap, and the non-degeneracy condition is stated separately. The variance upper bound v_n,UB is obtained from the Cauchy–Schwarz inequality, and the estimator v̂_n,UB is shown consistent via Lemma B.2; the unidentifiable cross term is not assumed away but conservatively bounded. The covariate adjustment efficiency gain is a mathematical consequence of maximizing L(β1, β0), which is defined as the difference between the unadjusted and adjusted asymptotic variances, so the claim that the oracle adjusted variance is no larger is an optimization identity rather than a circular prediction; the feasible estimator is then shown asymptotically equivalent to the oracle solution. Self-citations such as Su and Ding (2021) and Li and Ding (2017) appear only as special-case comparisons or background, and no load-bearing uniqueness theorem is imported from the authors' prior work. The restrictive sparsity assumptions are explicitly stated and their limitations, including the dense Amazon marketplace example, are acknowledged by the authors, which is an applicability concern rather than circularity. The derivation chain is therefore self-contained and no circular step is present.
Assumptions & free parameters
assumptions (7)
- domain assumption Assumption 1: Yi(z) = Yi(z_{G_{+i}}), i.e., potential outcomes of outcome unit i depend only on the treatment status of the intervention units connected to i.
- domain assumption Assumption 2: Z1,...,Zm are i.i.d. Bernoulli(p).
- domain assumption Assumption 3: Potential outcomes and covariates are bounded.
- domain assumption Assumption 4: max_i |G_{+i}| = O(1) and max_k |G_{k+}| = o(n).
- domain assumption Assumption 5: The number of intervention units connected to any given intervention unit via shared outcome units is bounded by B.
- domain assumption Assumption 6: Weighted covariance matrices of potential outcomes and covariates converge to limits at rate m^{1+δ}.
- standard math Martingale CLT (Hall and Heyde 2014) and standard probability results.
Cite this review
Pith. "Pith review of Design-based causal inference in bipartite experiments." pith.science (2026). https://pith.science/paper/RPAEIH7Z
@misc{pith2026250109844,
author = {Pith},
title = {Pith review of: Design-based causal inference in bipartite experiments},
year = {2026},
howpublished = {\url{https://pith.science/paper/RPAEIH7Z}},
note = {Machine review of arXiv:2501.09844}
}
read the original abstract
Bipartite experiments arise in various fields, in which the treatments are randomized over one set of units, while the outcomes are measured over another separate set of units. However, existing methods often rely on strong model assumptions about the data-generating process. Under the potential outcomes formulation, we explore design-based causal inference in bipartite experiments under weak assumptions by leveraging the sparsity structure of the bipartite graph that connects the treatment units and outcome units. We make several contributions. First, we formulate the causal inference problem under the design-based framework that can account for the bipartite interference. Second, we propose a consistent point estimator for the total treatment effect, a policy-relevant parameter that measures the difference in the outcome means if all treatment units receive the treatment or control. Third, we establish a central limit theorem for the estimator and propose a conservative variance estimator for statistical inference. Fourth, we discuss a covariate adjustment strategy to enhance estimation efficiency.
Figures
Forward citations
Cited by 1 Pith paper
-
GAUGER: Generalized Regression Adjustment via Graph-Weighted Exposure-Level Residualization for Design-Based Inference Under Interference
GAUGER calibrates outcome predictions against the design-induced graph-weighted variance structure to yield a variance-optimal AIPW estimator under network interference.
Reference graph
Works this paper leans on
-
[1]
Aronow, P. M. and Samii, C. (2017). Estimating average causal effects under general interference, with application to a social network experiment. Annals of Applied Statistics , 11(4):1912--1947
2017
-
[2]
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
2024
-
[3]
De Jong, P. (1987). A central limit theorem for generalized quadratic forms. Probability Theory and Related Fields , 75(2):261--277
work page 1987
-
[4]
Ding, P. (2024). A First Course in Causal Inference . London: Chapman and Hall
work page 2024
-
[5]
Donner, A. (1998). Some aspects of the design and analysis of cluster randomization trials. Journal of the Royal Statistical Society: Series C (Applied Statistics) , 47(1):95--113
work page 1998
-
[6]
Donner, A., Klar, N., and Klar, N. S. (2000). Design and analysis of cluster randomization trials in health research . London: Arnold
work page 2000
-
[7]
Doudchenko, N., Zhang, M., Drynkin, E., Airoldi, E., Mirrokni, V., and Pouget-Abadie, J. (2020). Causal inference with bipartite designs. arXiv preprint arXiv:2010.02108
arXiv 2020
-
[8]
Fisher, R. A. (1925). Statistical Methods for Research Workers . Oliver and Boyd, Edinburgh, 1st edition
work page 1925
Show all 42 references
-
[9]
Fogarty, C. B. (2018). Regression-assisted inference for the average treatment effect in paired experiments. Biometrika , 105(4):994--1000
2018
-
[10]
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
2021
-
[11]
Freedman, D. A. (2008). On regression adjustments to experimental data. Advances in Applied Mathematics , 40(2):180--193
2008
-
[12]
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
2023 arXiv
-
[13]
H \'a jek, J. (1960). Limiting distributions in simple random sampling from a finite population. A Magyar Tudom \'a nyos Akad \'e mia Matematikai Kutat \'o Int \'e zet \'e nek k \"o zlemenyei , 5(3):361--374
1960
-
[14]
and Heyde, C
Hall, P. and Heyde, C. C. (2014). Martingale Limit Theory and Its Application . New York: Academic Press
2014
-
[15]
Harshaw, C., S \"a vje, F., Eisenstat, D., Mirrokni, V., and Pouget-Abadie, J. (2023). Design and analysis of bipartite experiments under a linear exposure-response model. Electronic Journal of Statistics , 17(1):464--518
2023
-
[16]
Imai, K. (2008). Variance identification and efficiency analysis in randomized experiments under the matched-pair design. Statistics in Medicine , 27(24):4857--4873
2008
-
[17]
Imbens, G. W. and Rubin, D. B. (2015). Causal Inference in Statistics, Social, and Biomedical Sciences . Cambridge: Cambridge University Press
2015
-
[18]
Korolyuk, V. S. and Borovskich, Y. V. (2013). Theory of U-statistics , volume 273. Springer Science & Business Media
2013
-
[19]
and Ding, P
Li, X. and Ding, P. (2017). General forms of finite population central limit theorems with applications to causal inference. Journal of the American Statistical Association , 112(520):1759--1769
2017
-
[20]
and Ding, P
Li, X. and Ding, P. (2020). Rerandomization and regression adjustment. Journal of the Royal Statistical Society Series B: Statistical Methodology , 82(1):241--268
2020
-
[21]
and Recht, B
Liang, T. and Recht, B. (2025). Randomization inference when n equals one. Biometrika , page asaf013
2025
-
[22]
Lin, W. (2013). Agnostic notes on regression adjustments to experimental data: Reexamining freedman’s critique. The Annals of Applied Statistics , 7(1):295--318
2013
-
[23]
and Yang, Y
Liu, H. and Yang, Y. (2020). Regression-adjusted average treatment effect estimates in stratified randomized experiments. Biometrika , 107(4):935--948
2020
-
[24]
Lu, X., Li, H., and Liu, H. (2024). Estimation and inference of average treatment effects under heterogeneous additive treatment effect model. arXiv preprint arXiv:2408.17205
2024 arXiv
-
[25]
Lu, X., Yang, F., and Wang, Y. (2025). Debiased regression adjustment in completely randomized experiments with moderately high-dimensional covariates. The Annals of Statistics , page in press
2025
-
[26]
W., Sekhon, J
Miratrix, L. W., Sekhon, J. S., and Yu, B. (2012). Adjusting treatment effect estimates by post-stratification in randomized experiments. Journal of the Royal Statistical Society Series B: Statistical Methodology , 75(2):369--396
2012
-
[27]
Mukerjee, R., Dasgupta, T., and Rubin, D. B. (2018). Using standard tools from finite population sampling to improve causal inference for complex experiments. Journal of the American Statistical Association , 113(522):868--881
2018
-
[28]
(1923/1990)
Neyman, J. (1923/1990). On the application of probability theory to agricultural experiments. Essay on principles. Section 9. Statistical Science , pages 465--472
1923
-
[29]
Papadogeorgou, G., Choirat, C., and Zigler, C. M. (2019). Adjusting for unmeasured spatial confounding with distance adjusted propensity score matching. Biostatistics , 20(2):256--272
2019
-
[30]
Pouget-Abadie, J., Aydin, K., Schudy, W., Brodersen, K., and Mirrokni, V. (2019). Variance reduction in bipartite experiments through correlation clustering. Advances in Neural Information Processing Systems , 32
2019
-
[31]
Rosenbaum, P. R. and Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika , 70(1):41--55
1983
-
[32]
Rubin, D. B. (1980). Randomization analysis of experimental data: The F isher randomization test comment. Journal of the American Statistical Association , 75(371):591--593
1980
-
[33]
Shi, L., Bakhitov, E., Hung, K., Karrer, B., Walker, C., Bhole, M., and Schrijvers, O. (2024). Scalable analysis of bipartite experiments. arXiv preprint arXiv:2402.11070
2024
-
[34]
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
2021
-
[35]
Ugander, J., Karrer, B., Backstrom, L., and Kleinberg, J. (2013). Graph cluster randomization: Network exposure to multiple universes. In Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages 329--337
2013
-
[36]
Viviano, D., Lei, L., Imbens, G., Karrer, B., Schrijvers, O., and Shi, L. (2023). Causal clustering: design of cluster experiments under network interference. arXiv preprint arXiv:2310.14983
2023 arXiv
-
[37]
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
2024
-
[38]
Wu, W. B. and Shao, X. (2007). A limit theorem for quadratic forms and its applications. Econometric Theory , 23(5):930--951
2007
-
[39]
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
2022
-
[40]
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
2022
-
[41]
Zigler, C., Liu, V., Mealli, F., and Forastiere, L. (2025). Bipartite interference and air pollution transport: Estimating health effects of power plant interventions. Biostatistics , 26(1):kxae051
2025
-
[42]
Zigler, C. M. and Papadogeorgou, G. (2021). Bipartite causal inference with interference. Statistical Science , 36(1):109--123
2021
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.