Research map / Probabilistic and scientific modeling / Causal discovery

Controllable Generative Sandbox for Causal Inference

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Counterfactual identification and bounds · 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.

Proposes CausalMix, a conditional VAE with a Bayesian Gaussian-mixture latent prior and explicit overlap, treatment-effect, and confounding controls, used to generate calibrated synthetic mixed-type data for benchmarking CATE estimators, tuning causal forests, and power analysis in an mCRPC claims study.

Problem

Synthetic data for causal inference must be both realistic for mixed-type tabular data and give explicit design-time control over overlap, unmeasured confounding, and treatment effect heterogeneity; existing simulators trade one off against the other.

Approach

Couples conditional VAEs with data-type-specific decoder heads, a post-hoc Dirichlet-process Bayesian Gaussian mixture latent prior, and penalties on a log density-ratio (overlap), CATE, and confounding functions so these causal quantities can be set independently.

Key results

• In Scenario 3, BGMM prior improved normalized energy distance to 0.011 vs 0.080 for Gaussian and C2ST AUC complement 0.625 vs 0.333.
• Causal forest tuning: min node size 80 cut PEHE from ~0.036 to ~0.023 but reduced CATE coverage from ~0.87-0.91 to ~0.75-0.76.
• Power analysis for detecting CVD effect modification reached 0.94 power at n=2000 and 0.67 at n=1000.

Surprising findings

• Gaussian prior gave stronger record-level privacy (higher DCR protection fraction) than BGMM, a realism-privacy trade-off.
• Strong ATE accuracy did not imply reliable CATE coverage; lasso-based DML/DR learners had CATE coverage of 0.37-0.40.

Limitations

• Evaluation on a single real clinical dataset with simulated outcomes and only three causal scenarios.
• Confounding and overlap control rely on user-specified functions; residual confounding bias persisted for all estimators.
• Monte Carlo replications (50-100) limit precision of coverage estimates.

Experimental setup

Datasets: Optum Clinformatics Data Mart (mCRPC cohort, 4,098 patients)

Benchmarks: mCRPC claims cohort (Scenarios 1-3)

Metrics: normalized Wasserstein KSComplement TVComplement correlation similarity MMD normalized energy distance C2ST CATE correlation ATE error overlap coefficient

Baselines: Gaussian prior CausalMix Credence RealCause Frengression WGAN-based generators

Builds on: conditional VAE Credence RealCause causal forest Bayesian Gaussian mixture

Closest papers by meaning

• Improving the Generation and Evaluation of Synthetic Data for Downstream Medical Causal Inference (NeurIPS 2025 · similarity 0.88)
• 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)
• Counterfactual Generative Modeling with Variational Causal Inference (ICLR 2025 · similarity 0.87)
• Generative Intervention Models for Causal Perturbation Modeling (ICML 2025 · similarity 0.87)