Research map / Probabilistic and scientific modeling / Treatment effects

A Neural Framework for Generalized Causal Sensitivity Analysis

ICLR 2024 · Probabilistic and scientific modeling › Treatment effects · approach: Proximal and deep causal effect estimation · 1 citations (OpenAlex)

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.

placeholder

Problem

Causal sensitivity analysis under unobserved confounding is only addressed for specific combinations of sensitivity model, treatment type, and causal query; no unified method covers all of them, notably multiple outcomes and f-sensitivity/Rosenbaum with continuous treatments.

Approach

Define generalized treatment sensitivity models (GTSMs) that constrain the shift between P(U|x) and P(U|x,a), and learn that shift with two separately trained conditional normalizing flows: stage 1 fits the observational outcome distribution, stage 2 learns a shifted latent distribution optimized under the sensitivity constraint via augmented Lagrangian.

Key results

• On synthetic data under the MSM, NeuralCSA bounds almost coincide with optimal closed-form bounds for Gamma in {2, 4, 10}, averaged over 5 runs.
• On semi-synthetic MIMIC-III data, all sensitivity models achieve at least 50% coverage; e.g., MSM with Gamma*=5.48 gives coverage 0.91 ± 0.03 and KL with Gamma*=0.25 gives 0.54 ± 0.07.
• A multi-outcome case study on MIMIC-III bounds P(Y1(1)>=115, Y2(1)>=90 | x) under an MSM, with the danger-area probability increasing from Gamma=2 to Gamma=4.

Measured effects reported in the paper

Surprising findings

• The MSM is conservative on semi-synthetic data, achieving much higher coverage and interval length than needed, because its constraint does not adapt well to the data-generating process.

Limitations

• Stage 2 optimization does not guarantee sensitivity constraints by design, so the augmented Lagrangian and epoch selection are needed.
• Results depend on the choice of sensitivity model, which can yield different bound tightness; the MSM can be overly conservative.

Experimental setup

Datasets: MIMIC-III synthetic dataset 1 (adapted from prior work) synthetic dataset 2 (adapted from prior work)

Benchmarks: synthetic datasets 1 and 2 MIMIC-III semi-synthetic MIMIC-III multi-outcome case study

Metrics: bound coverage interval length closed-form bound agreement

Baselines: closed-form MSM bounds (CF)

Builds on: marginal sensitivity model augmented Lagrangian method autoregressive neural spline flows

Closest papers by meaning

• Generalized Bayes for Causal Inference (NeurIPS 2026 · similarity 0.87)
• Causal Inference with Conditional Front-Door Adjustment and Identifiable Variational Autoencoder (ICLR 2024 · similarity 0.86)
• NESTER: An Adaptive Neurosymbolic Method for Causal Effect Estimation (AAAI 2024 · similarity 0.86)
• Towards Learning and Explaining Indirect Causal Effects in Neural Networks (AAAI 2024 · similarity 0.86)
• Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time (NeurIPS 2026 · similarity 0.86)
• CausalPFN: Amortized Causal Effect Estimation via In-Context Learning (NeurIPS 2025 · similarity 0.86)