Research map / Probabilistic and scientific modeling / Causal discovery
Learning Joint Interventional Effects from Single-Variable Interventions in Additive Models
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.
Shows that joint interventional effects of multiple actions can be identified from observational and single-variable intervention data under a nonlinear additive outcome mechanism with pairwise confounding, and proposes a practical estimator validated on synthetic SCMs.
Problem
Estimating the effect of joint interventions on multiple action variables when only observational data and single-variable interventional data are available, with unobserved confounders between actions and outcome.
Approach
Assume the outcome mechanism is additive over per-action nonlinear functions of each action and its own confounder; then the joint interventional mean decomposes into terms learnable from observational and single-intervention regressions, using a single estimator with indicator variables R_k marking which actions were intervened on.
Key results
- Intervention Generalization approach achieves RMSE close to the joint-interventional topline and clearly outperforms observational-only and pooled estimators on 100 random SCMs with 5 actions.
- Optimal ratio of single-intervention to observational samples tends to lie between 0.1 and 0.9.
- The method needs more than an order of magnitude more data than a model trained directly on joint interventions to reach similar accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use observational and single-intervention data to estimate joint-intervention effects instead of training directly on joint-intervention data | Intervention Generalization · Synthetic joint-intervention test data | RMSE | No clear effect |
| Use Intervention Generalization instead of an observational-only estimator | Joint-intervention effect estimation · Synthetic joint-intervention test data | RMSE | Helped |
| Use Intervention Generalization instead of an estimator trained on pooled observational and single-intervention data | Joint-intervention effect estimation · Synthetic joint-intervention test data | RMSE | Helped |
| Estimate from observational and single-intervention data rather than directly training on joint-intervention data | Intervention Generalization · Synthetic joint-intervention test data | Sample efficiency: over an order of magnitude more data | Hurt |
Surprising findings
- Sample efficiency is substantially worse than training directly on joint interventional data, requiring over 10x more samples for similar error.
Limitations
- Requires additivity of the outcome mechanism over action-confounder pairs, and pairwise confounding structure, which may not hold in practice.
- Lower sample efficiency than direct joint-intervention training.
- Evaluation only on synthetic data; no real-world experiments.
Experimental setup
Datasets: synthetic observational data synthetic single-intervention data synthetic joint-intervention data
Benchmarks: synthetic SCMs with 5 actions
Metrics: root mean squared error standard error of the mean
Baselines: Joint-interventional topline estimator Pooled observational and single-intervention estimator Observational-only estimator
Builds on: Structural causal models Generalized additive models Neural additive models
Closest papers by meaning
- An Analysis of Causal Effect Estimation using Outcome Invariant Data Augmentation (NeurIPS 2025 · similarity 0.88)
- Identifying Representations for Intervention Extrapolation (ICLR 2024 · similarity 0.87)
- Coupling Generative Modeling and an Autoencoder with the Causal Bridge (NeurIPS 2025 · similarity 0.87)
- Doubly robust identification of treatment effects from multiple environments (ICLR 2025 · similarity 0.87)
- Causal Inference with Conditional Front-Door Adjustment and Identifiable Variational Autoencoder (ICLR 2024 · similarity 0.87)
- Learning Mixtures of Unknown Causal Interventions (NeurIPS 2024 · similarity 0.86)
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.