Research map / Probabilistic and scientific modeling / Treatment effects
A Dual-module Framework for Counterfactual Estimation over Time
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 ACTIN, a dual-module framework (adversarial distribution-based balancing module plus a Temporal Integration Predicting integrating module) for counterfactual estimation over time, instantiated with TCN or LSTM base models.
Problem
Estimating counterfactual outcomes over time from observational longitudinal data under time-varying confounding, where existing balancing methods are limited to categorical treatments and simple models struggle with long-range dependencies and temporal treatment interactions.
Approach
An adversarial, distribution-based balancing module that makes representations independent of treatment (joint vs. marginal discrimination) for any treatment type, combined with a TIP integrating module that feeds all balanced representations and past treatments up to the current time into the predictor rather than only the current step (DP).
Key results
- On TG one-step prediction across gamma 0-4, ACTIN (TCN-based) achieves lowest RMSE (e.g., 0.75 at gamma=0 vs. 0.78 for CT) and lowest RMSE on SS (0.15) and RW (4.56) one-step settings.
- On SS multi-step prediction (tau=11), ACTIN reaches RMSE 0.54 vs. 0.60 for CT.
- ACTIN's running time is roughly 10% of CT's on the MIMIC-III-derived RW and SS datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ACTIN instead of the strongest reported baseline | counterfactual prediction models · TG, one-step (γ=0–4) | RMSE: 0.78 -> 0.75; 0.80 -> 0.78; 0.87 -> 0.85; 1.02 -> 1.01; 1.33 -> 1.26 | Helped |
| Use ACTIN instead of the strongest reported baseline | counterfactual prediction models · RW and SS, one-step | RMSE: 4.60 -> 4.56; 0.20 -> 0.15 | Helped |
| Use ACTIN instead of CT | counterfactual prediction models · SS, τ=2–9 | RMSE: 0.37 -> 0.33; 0.45 -> 0.39; 0.49 -> 0.43; 0.51 -> 0.46; 0.53 -> 0.48; 0.55 -> 0.49; 0.56 -> 0.50; 0.58 -> 0.51 | Helped |
| Use ACTIN instead of CT | counterfactual prediction models · RW, τ=2–6 | RMSE: 8.99 -> 8.98; 9.59 -> 9.56; 9.91 -> 9.87; 10.14 -> 10.11; 10.34 -> 10.30 | Helped |
| Remove adversarial balancing (ablation) | ACTIN · TG, τ=6 (γ=0,2,4) | RMSE: −0.1%; +1.6%; +4.3% | Mixed |
| Remove adversarial balancing (ablation) | ACTIN · SS, τ=11 | RMSE: +1.9% | Hurt |
| Remove adversarial balancing (ablation) | ACTIN · CISD, τ=6 | RMSE: +9.4% | Hurt |
| Remove the integrating module (ablation) | ACTIN · TG, τ=6 (γ=0,2,4) | RMSE: +6.8%; +4.0%; +12.6% | Hurt |
| Remove the integrating module (ablation) | ACTIN · SS, τ=11 | RMSE: +41.2% | Hurt |
| Remove the integrating module (ablation) | ACTIN · CISD, τ=6 | RMSE: +17.1% | Hurt |
| Replace TIP with DP (ablation) | ACTIN integrating module · TG, τ=6 (γ=0,2,4) | RMSE: +5.4%; +6.6%; +6.2% | Hurt |
| Replace TIP with DP (ablation) | ACTIN integrating module · SS, τ=11 | RMSE: +5.8% | Hurt |
Surprising findings
- Replacing TIP with DP in the integrating module degrades performance similarly to removing the integrating module entirely, underscoring temporal treatment interactions.
- Adding the TIP integrating module to CT yields smaller gains than on simple TCN/LSTM base models, since CT already handles long-range dependencies.
Limitations
- Counterfactuals cannot be validated directly on real-world data, so real-data evaluation relies on semi-synthetic constructions.
- Gains are smaller when the integrating module is added to a strong base model such as CT.
Experimental setup
Datasets: Tumor Growth (TG) simulator MIMIC-III Semi-synthetic MIMIC-III (SS) Real-world MIMIC-III (RW) Continuous Interventions Synthetic Dataset (CISD)
Benchmarks: TG SS RW CISD
Metrics: RMSE normalized RMSE running time parameters MFLOPs
Baselines: RMSN CRN G-Net CT
Builds on: Causal Transformer CRN gradient reversal domain adaptation temporal convolutional network Exponential Moving Average
Closest papers by meaning
- Adversarially Balanced Representation for Continuous Treatment Effect Estimation (AAAI 2024 · similarity 0.89)
- Conditional Counterfactual Mean Embeddings: Doubly Robust Estimation and Learning Rates (NeurIPS 2026 · similarity 0.88)
- Counterfactual Contrastive Learning with Normalizing Flows for Robust Treatment Effect Estimation (ICML 2025 · similarity 0.86)
- Counterfactual Generative Modeling with Variational Causal Inference (ICLR 2025 · similarity 0.86)
- Towards Characterizing Domain Counterfactuals for Invertible Latent Causal Models (ICLR 2024 · similarity 0.86)
- LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching (NeurIPS 2025 · 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.