- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The authors present a method (SiDE: Signed-exposure Doubly robust Estimator) for estimating causal effects under interference (aka peer effects) in which ties can be signed as either positive or negative. The authors describe and evaluate a doubly-robust estimator and evaluate it using a set of semi-synthetic data sets and demonstrate it on a real-world cases study.
Critical strengths
The authors appear to have identified an interesting gap in the existing literature on causal inference under interference. Specifically, they examine causal estimation when interference arises in two “flavors” (positive and negative ties between units). This gives rise to several complexities, including multiple effects (positive and negative ties can have different types of interference) and interaction effects (between the two types of interference).
The empirical evaluation is extensive and detailed. The authors provide results for a six signed networks and a large number of estimators.
Critical weaknesses
The paper is not entirely clear about its’ contribution. First, it seems unlikely that no prior work has considered causal estimation in the presence of multiple tie types, and the authors' work would seem to be a special case of this inference task. Second, the authors do not clearly state how their work builds on the current frontier.
The evaluation results (as opposed to the case study in section 5.5) are entirely synthetic. This is no surprise, given the difficulty of finding real-world data with known causal effect for causal inference without interference. It is, of course, even more difficult to find real-world data with known causal effect for causal inference problems with interference. Still, this limits the confidence that we can have in the results, given the substantial and known problems with evaluations on synthetic data.
Questions for authors
How does your work differ from the fairly substantial amount of work on causal inference under network interference described in Section 2? The paper says that you “build on these foundations” but does not clearly state the gap in prior work that is filled by your work.
How does the problem addressed in this paper differ from one in which there are two (or more) tie types that are entirely disjoint (e.g., family ties and non-family social ties)? The problem addressed here would just seem to be a special case of multiple (disjoint) tie types.
Why just binary? It seems likely that ties would be better described by strength (including both positive and negative values) rather than merely by sign (positive or negative). Why not treat them in this way?
What’s the intuition behind the sign-composition effect? The first three effects described in Section 3.2 are easy to understand. However, the fourth (“sign-composition effect”) is more difficult to understand. What’s a real world case in which this effect would arise independently of the first three?
4 comments
L5Ac — Part 2/3: W1/Q1 — Complete references (1–25)
References
- Michael G. Hudgens and M. Elizabeth Halloran. Toward causal inference with interference. JASA, 103(482):832–842, 2008.
- Peter M. Aronow and Cyrus Samii. Estimating average causal effects under general interference, with application to a social network experiment. The Annals of Applied Statistics, 11(4):1912–1947, 2017.
- Laura Forastiere, Edoardo M. Airoldi, and Fabrizia Mealli. Identification and estimation of treatment and interference effects in observational studies on networks. JASA, 116(534):901–918, 2021.
- Elizabeth L. Ogburn, Oleg Sofrygin, Iván Díaz, and Mark J. van der Laan. Causal inference for social network data. JASA, 119(545):597–611, 2024.
- Michael P. Leung and Pantelis Loupos. Graph neural networks for causal inference under network confounding. arXiv:2211.07823v5, 2025.
- Zhaonan Qu, Ruoxuan Xiong, Jizhou Liu, and Guido Imbens. Semiparametric estimation of treatment effects in observational studies with heterogeneous partial interference. Journal of Business & Economic Statistics, 44(4):1304–1316, 2026.
- Naoki Egami. Spillover effects in the presence of unobserved networks. Political Analysis, 29(3):287–316, 2021.
- Yunpu Ma and Volker Tresp. Causal inference under networked interference and intervention policy enhancement. In Proceedings of the 24th International Conference on Artificial Intelligence and Statistics, pages 3700–3708, 2021.
- Song Jiang and Yizhou Sun. Estimating causal effects on networked observational data via representation learning. In Proceedings of the 31st ACM International Conference on Information and Knowledge Management, pages 852–861, 2022.
- Ziyu Zhao, Yuqi Bai, Ruoxuan Xiong, Qingyu Cao, Chao Ma, Ning Jiang, Fei Wu, and Kun Kuang. Learning individual treatment effects under heterogeneous interference in networks. ACM Transactions on Knowledge Discovery from Data, 18(8):199, 2024.
- Xiaofeng Lin, Guoxi Zhang, Xiaotian Lu, Han Bao, Koh Takeuchi, and Hisashi Kashima. Estimating treatment effects under heterogeneous interference. In Machine Learning and Knowledge Discovery in Databases: Research Track, ECML PKDD 2023, Proceedings, Part I, volume 14169 of Lecture Notes in Computer Science, pages 576–592. Springer, 2023.
- Shishir Adhikari, Sourav Medya, and Elena Zheleva. Learning exposure mapping functions for inferring heterogeneous peer effects. In International Conference on Learning Representations, 2026.
- Daan Caljon, Jente Van Belle, and Wouter Verbeke. Estimating treatment effects in networks under unknown exposure mappings. arXiv:2510.21457v2, 2026.
- Dorwin Cartwright and Frank Harary. Structural balance: A generalization of Heider’s theory. Psychological Review, 63(5):277–293, 1956.
- Jure Leskovec, Daniel Huttenlocher, and Jon Kleinberg. Predicting positive and negative links in online social networks. In Proceedings of the 19th International Conference on World Wide Web, pages 641–650, 2010b.
- R. Guha, Ravi Kumar, Prabhakar Raghavan, and Andrew Tomkins. Propagation of trust and distrust. In Proceedings of the 13th International Conference on World Wide Web, pages 403–412, 2004.
- Tyler Derr, Yao Ma, and Jiliang Tang. Signed graph convolutional networks. In Proceedings of the 18th IEEE International Conference on Data Mining, pages 929–934, 2018.
- Jiliang Tang, Yi Chang, Charu Aggarwal, and Huan Liu. A survey of signed network mining in social media. ACM Computing Surveys, 49(3):42, 2016.
- Michael D. König, Dominic Rohner, Mathias Thoenig, and Fabrizio Zilibotti. Networks in conflict: Theory and evidence from the Great War of Africa. Econometrica, 85(4):1093–1132, 2017.
- Yanhua Li, Wei Chen, Yajun Wang, and Zhi-Li Zhang. Influence diffusion dynamics and influence maximization in social networks with friend and foe relationships. In Proceedings of the 6th ACM International Conference on Web Search and Data Mining, pages 657–666, 2013.
- Robert M. Bond, Christopher J. Fariss, Jason J. Jones, Adam D. I. Kramer, Cameron Marlow, Jaime E. Settle, and James H. Fowler. A 61-million-person experiment in social influence and political mobilization. Nature, 489(7415):295–298, 2012.
- Karthik Rajkumar, Guillaume Saint-Jacques, Iavor Bojinov, Erik Brynjolfsson, and Sinan Aral. A causal test of the strength of weak ties. Science, 377(6612):1304–1310, 2022.
- Amir Ghasemian and Nicholas A. Christakis. The structure and function of antagonistic ties in village social networks. Proceedings of the National Academy of Sciences, 121(26):e2401257121, 2024.
- Elisa Bellotti, Federico Bianchi, Francesco Renzini, Carinthia Balabet Nengnong, Anna Maria van Eijk, Anne Kessler, Rajiv Sarkar, Jane M. Carlton, and Sandra Albert. Negative ties offset complex contagion effect on malaria preventive measure adoption in a hard-to-reach population. SocArXiv preprint, 2026.
- Károly Takács, Andreas Flache, and Michael Mäs. Discrepancy and disliking do not induce negative opinion shifts. PLOS ONE, 11(6):e0157948, 2016.
L5Ac — Part 3/3: W2/Q2/Q3/Q4 — Evaluation, tie types, and sign composition
W2: The evaluation results (as opposed to the case study in section 5.5) are entirely synthetic. This is no surprise, given the difficulty of finding real-world data with known causal effect for causal inference without interference. It is, of course, even more difficult to find real-world data with known causal effect for causal inference problems with interference. Still, this limits the confidence that we can have in the results, given the substantial and known problems with evaluations on synthetic data.
As you note, evaluating causal inference methods is challenging because ground-truth causal effects are generally unavailable in real data. Our benchmarks retain observed signed network structures to make the synthetic data sets as close to real-world data as possible, while simulating treatments and outcomes so that estimation error can be measured. The school application analyzes a real-world randomized intervention: invitations to a school program were randomized, and wristband wearing (outcome) was observed.
Q2: How does the problem addressed in this paper differ from one in which there are two (or more) tie types that are entirely disjoint (e.g., family ties and non-family social ties)? The problem addressed here would just seem to be a special case of multiple (disjoint) tie types.
We study positive and negative ties as two disjoint neighbor types and define and identify four causal contrasts: positive-tie effects, negative-tie effects, their interaction, and the sign-composition effect. The composition effect addresses a fixed-budget allocation question: how does the outcome change when one treatment is transferred from a negative-tie neighbor to a positive-tie neighbor, holding own treatment and the total number of treated neighbors fixed? These allocations have the same unsigned exposure, so a total-count exposure mapping cannot distinguish them.
Proposition 2 establishes what pooling the two treated-neighbor counts measures. Under Assumptions 1–4, fixed signed degrees and sign-blind assignment, it decomposes the unsigned increment into composition-averaged positive- and negative-tie increments, with weights equal to the proportions of the two tie types. The identity holds for nonlinear responses and interactions and gives the exact condition under which nonzero signed increments cancel. SiDE estimates the four signed contrasts on their respective common reachable populations.
Q3: Why just binary? It seems likely that ties would be better described by strength (including both positive and negative values) rather than merely by sign (positive or negative). Why not treat them in this way?
Signed counts make the treatment changes being compared explicit, including treating one more neighbor of either type and reallocating one treatment at a fixed total.
For each ego, Assumption 1 specifies that the potential outcome depends on treatment assignment through own treatment and the two treated-neighbor counts. This makes a treatment slot the common unit in the fixed-total composition contrast. Incorporating continuous tie strength would require weighted exposures, their state probabilities and corresponding causal contrasts. We plan to consider this extension in future work.
Q4: What’s the intuition behind the sign-composition effect? The first three effects described in Section 3.2 are easy to understand. However, the fourth (“sign-composition effect”) is more difficult to understand. What’s a real world case in which this effect would arise independently of the first three?
The sign-composition effect measures the outcome change from reallocating treatment between tie types while keeping the total number of treated neighbors fixed. In Figure 1's health-program example, four treatment slots can be allocated to three trusted neighbors and one avoided neighbor, or to four trusted neighbors. With the ego's own treatment fixed, the sign-composition effect measures the difference in mean outcomes between these two allocations. On the same population, it equals the positive-tie increment minus the negative-tie increment from a common baseline, as stated in Definition 4.
Follow-up on the contribution and interpretation of signed effects
Dear Reviewer L5Ac,
Thank you for your helpful comments. Our responses clarify how the four signed causal contrasts, Proposition 2, and SiDE build on the related work discussed in Section 2. We also explain the practical meaning of the sign-composition effect using Figure 1’s example of reallocating treatment between negative- and positive-tie neighbors while keeping the total number of treated neighbors fixed.
We would appreciate your feedback on whether these explanations clarify the contribution and address your questions.
L5Ac — Part 1/3: W1/Q1 — Contributions and complete related-work comparison
Dear Reviewer L5Ac,
Thank you for your time and constructive comments.
There is substantial prior work on heterogeneity in peer effects. Qu et al. (2026) distinguish interference across neighbor types, Lin et al. (2023) model interference across multiple relation types, and Caljon et al. (2026) learn neighborhood representations that capture heterogeneous interference without a prespecified exposure mapping. Our contribution combines the definition and identification of signed causal contrasts, an exact decomposition of the unsigned increment, and estimation and inference for these contrasts through the Signed-exposure Doubly robust Estimator (SiDE).
Proposition 2 gives an exact decomposition of the unsigned increment into positive- and negative-tie increments. Under Assumptions 1–4, fixed signed degrees and sign-blind assignment, the weights are the proportions of the two tie types, with each signed increment averaged over its corresponding treatment compositions. The identity holds for nonlinear responses and interactions and specifies the exact cancellation condition. It therefore shows how a zero unsigned increment can conceal nonzero effects through both tie types. This provides a precise account of what pooling tie types measures and when opposing effects become invisible in that comparison.
We define and identify four causal contrasts: positive-tie effects, negative-tie effects, their interaction, and the sign-composition effect. The sign-composition effect addresses a fixed-budget allocation question. Holding own treatment and the total number of treated neighbors fixed, it measures the outcome change when one treatment is reallocated from a negative-tie neighbor to a positive-tie neighbor. The two allocations have the same unsigned exposure, so a total-count exposure mapping cannot distinguish them. We propose SiDE to estimate these contrasts using sign-specific neighborhood representations and joint exposure probabilities derived from individual treatment probabilities. We establish double robustness of its score under the conditions in Proposition 3 and construct approximate intervals that account for overlapping neighborhoods.
The table below places these contributions in the context of the related work discussed in Section 2.