Research map / Probabilistic and scientific modeling / Treatment effects
NESTER: An Adaptive Neurosymbolic Method for Causal Effect Estimation
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 NESTER, an adaptive neurosymbolic program synthesis method that uses a causal-inductive-bias DSL and A* search (with NEAR or dPads) to synthesize programs for causal effect estimation, outperforming baselines on IHDP, Twins and Jobs.
Problem
Estimating causal effects (ACE, ATE, ACT) from observational data, where confounding and selection bias make unbiased estimation hard and existing multi-head neural architectures each address only one aspect.
Approach
A DSL of differentiable primitives (if-then-else, mlp, shared mlp, subset, propensity, align, arithmetic) encoding causal inductive biases, with neurosymbolic program synthesis via A* (NEAR heuristic) or dPads that automatically composes multi-head architectures such as TARNet, CFR and Dragonnet as special cases.
Key results
- On IHDP, NESTER-dPads reached epsilon_ACE .05 in-sample and .05 out-of-sample, versus .09 (SNet) and .11 (CMGP) for the best baselines.
- On Twins out-of-sample epsilon_ACE, NESTER-dPads achieved .0028 versus .0040 for SNet.
- On Jobs in-sample epsilon_ACT, NESTER variants reached .01, tied with several baselines, while out-of-sample ACT was not best (.08 vs .06 for MHNET and GANITE).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NESTER-NEAR instead of SNet | Causal effect estimation · IHDP In-Sample | ε_ACE: .09$\pm$.10 -> .05$\pm$.04 | Helped |
| Use NESTER-dPads instead of SNet | Causal effect estimation · IHDP In-Sample | ε_ACE: .09$\pm$.10 -> .05$\pm$.01 | Helped |
| Use NESTER-NEAR instead of CMGP | Causal effect estimation · IHDP Out-of-Sample | ε_ACE: .13$\pm$.12 -> .05$\pm$.03 | Helped |
| Use NESTER-dPads instead of CMGP | Causal effect estimation · IHDP Out-of-Sample | ε_ACE: .13$\pm$.12 -> .05$\pm$.02 | Helped |
| Use NESTER-NEAR instead of k-NN | Causal effect estimation · Twins In-Sample | ε_ACE: .0028$\pm$.0021 -> .0034$\pm$.0005 | Hurt |
| Use NESTER-dPads instead of k-NN | Causal effect estimation · Twins In-Sample | ε_ACE: .0028$\pm$.0021 -> .0035$\pm$.0001 | Hurt |
| Use NESTER-NEAR instead of k-NN | Causal effect estimation · Twins Out-of-Sample | ε_ACE: .0051$\pm$.0039 -> .0039$\pm$.0006 | Helped |
| Use NESTER-dPads instead of k-NN | Causal effect estimation · Twins Out-of-Sample | ε_ACE: .0051$\pm$.0039 -> .0028$\pm$.0001 | Helped |
| Use NESTER-NEAR instead of OLS-1 | Causal effect estimation · Jobs In-Sample | ε_ACT: .01$\pm$.00 -> .01$\pm$.00 | No clear effect |
| Use NESTER-dPads instead of OLS-1 | Causal effect estimation · Jobs In-Sample | ε_ACT: .01$\pm$.00 -> .01$\pm$.00 | No clear effect |
| Use NESTER-NEAR instead of MHNET | Causal effect estimation · Jobs Out-of-Sample | ε_ACT: .06$\pm$.02 -> .08$\pm$.08 | Hurt |
| Use NESTER-dPads instead of MHNET | Causal effect estimation · Jobs Out-of-Sample | ε_ACT: .06$\pm$.02 -> .08$\pm$.08 | Hurt |
Surprising findings
- On Twins, the synthesized program reduced to a simple mlp(v), with simple linear-style models performing better than complex ones.
- Jobs synthesized program branched on a propensity-score mlp rather than the treatment variable, a structure not in the existing literature.
Limitations
- Assumes no latent confounding; relaxing this is left to future work.
- Run-time issues may arise as the DSL is enlarged with more primitives.
- Out-of-sample performance on Jobs is not best across methods.
Experimental setup
Datasets: IHDP Twins Jobs
Benchmarks: IHDP Twins Jobs
Metrics: epsilon_ACE epsilon_ACT sqrt(epsilon_PEHE)
Baselines: OLS-1 OLS-2 k-NN BLR BART Random Forest Causal Forest BNN TARNet MHNET GANITE CFR_WASS
Builds on: NEAR dPads TARNet CFR Dragonnet
Closest papers by meaning
- CausalPFN: Amortized Causal Effect Estimation via In-Context Learning (NeurIPS 2025 · similarity 0.87)
- Towards Learning and Explaining Indirect Causal Effects in Neural Networks (AAAI 2024 · similarity 0.86)
- A Neural Framework for Generalized Causal Sensitivity Analysis (ICLR 2024 · similarity 0.86)
- Doubly robust identification of treatment effects from multiple environments (ICLR 2025 · similarity 0.86)
- End-To-End Causal Effect Estimation from Unstructured Natural Language Data (NeurIPS 2024 · similarity 0.86)
- Data Fusion for Partial Identification of Causal Effects (NeurIPS 2025 · 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.