Research map / Probabilistic and scientific modeling / Treatment effects
Telling Peer Direct Effects from Indirect Effects in Observational Network Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes gDIS, a causal framework and GNN-based estimator (attention, multi-layer GNN, HSIC regularization) that identifies and estimates group-level peer direct effects (PDE), peer indirect effects (PIE), and self-treatment effects (STE) from observational network data.
Problem
Estimating causal effects in observational network data is hard due to peer interference; existing methods either lump peer effects together or separate direct and indirect effects only for one-to-one transmission settings.
Approach
Applies causal mediation analysis tailored to network data to separate group-level peer direct and indirect effects from self-treatment effects, derives identification formulas via backdoor adjustment and do-calculus, and estimates them with attention-weighted multi-layer GNNs, KDE density estimates, and an HSIC regularizer.
Key results
- On BlogCatalog within-sample, gDIS achieves PEHE of 0.2093 for peer effect vs 0.6591 for the best baseline TNet.
- On Flickr out-of-sample, gDIS self-treatment PEHE 0.1502 vs 0.3409 for NetEst (best baseline).
- On the Coat case study, STE accounts for 65.20% of total effect while PIE (21.76%) exceeds PDE (13.04%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add HSIC regularization (ablation) | gDIS · BC and Flickr, within-sample and out-of-sample | PEHE: 0.2478±0.0197 -> 0.2093±0.014; 0.1456±0.0129 -> 0.1265±0.0126; 0.194±0.0018 -> 0.1358±0.0026; 0.0673±0.0025 -> 0.0588±0. | Helped |
| Increase HSIC regularization coefficient above 0.3 (ablation) | gDIS · BC and Flickr | PEHE | Hurt |
| Use HSIC regularization coefficient λ = 0.3 (ablation) | gDIS · BC and Flickr | PEHE | Helped |
| Increase treatment-flip rate (ablation) | gDIS · BC and Flickr | MSE | Hurt |
| Use gDIS instead of the compared baseline methods | Causal effect estimation on BC and Flickr · BC and Flickr, within-sample and out-of-sample | PEHE | Helped |
Surprising findings
- In the Coat recommendation case, peer indirect effects were larger than peer direct effects, suggesting users respond more to peers' purchases than to peers' ad exposure.
Limitations
- Relies on the network unconfoundedness assumption, which may be violated in practice with hidden confounders.
- Evaluation relies on semi-synthetic outcomes rather than real ground-truth causal effects.
Experimental setup
Datasets: BlogCatalog Flickr Coat Shopping
Benchmarks: BlogCatalog (semi-synthetic) Flickr (semi-synthetic) Coat (recommendation case study)
Metrics: PEHE MSE
Baselines: CFR+(N) TARNet+(N) NetDeconf 1-GNN NetEst TNet gDIS(-HSIC)
Builds on: NetEst TNet CFR (counterfactual regression) sequential ignorability VanderWeele et al. 2012 peer mediation
Closest papers by meaning
- Towards Learning and Explaining Indirect Causal Effects in Neural Networks (AAAI 2024 · similarity 0.87)
- Online Experimental Design With Estimation-Regret Trade-off Under Network Interference (NeurIPS 2025 · similarity 0.86)
- Discovering Heterogeneous Causal Effects in Relational Data (AAAI 2024 · similarity 0.85)
- Data Fusion for Partial Identification of Causal Effects (NeurIPS 2025 · similarity 0.85)
- Independent-Set Design of Experiments for Estimating Treatment and Spillover Effects under Network Interference (ICLR 2024 · similarity 0.85)
- On Online Experimentation without Device Identifiers (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.