Research map / Probabilistic and scientific modeling / Treatment effects
GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding
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 GST-UNet, a U-Net/ConvLSTM encoder with iterative G-computation heads and curriculum training for estimating location-specific potential outcomes from a single spatiotemporal trajectory under time-varying confounding.
Problem
Estimating causal effects from a single spatiotemporal observational trajectory with interference, spatial confounding, temporal carryover, and time-varying confounding, where existing methods rely on strong structural assumptions or ignore one of these challenges.
Approach
Combines a representation-based time-invariance assumption and regression-based iterative G-computation with a U-Net + ConvLSTM + attention encoder and multiple G-heads trained jointly with a curriculum that gradually activates heads using detached pseudo-outcomes.
Key results
- At tau=5 with beta1=1.0, GST-UNet RMSE 0.40 vs UNet+ 0.54 (about 21.6% reduction over best baseline), but GST-UNet is worse at beta1=0 (+17.9%).
- On the Camp Fire application, GST-UNet estimates approximately 4,650 excess respiratory hospitalizations over Nov 8-17 (95% CI [1888, 6535]), versus UNet+ 3,981 and STCINet 88.
- Ablation shows removing curriculum training substantially degrades RMSE across horizons, and a 1x1 spatial kernel (no neighbor aggregation) increases RMSE from 0.33 to 0.53 at beta1=0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use iterative G-computation instead of direct outcome modeling | GST-UNet vs UNet+ · Synthetic, β1=0.0, τ=5 and τ=10 | RMSE: 0.28 ± 0.00 -> 0.33 ± 0.00; 0.28 ± 0.00 -> 0.38 ± 0.00 | Hurt |
| Use iterative G-computation instead of direct outcome modeling | GST-UNet vs strongest baseline · Synthetic, β1>0, τ=5 and τ=10 | RMSE: 0.36 ± 0.00 -> 0.35 ± 0.00; 0.61 ± 0.00 -> 0.55 ± 0.00; 0.54 ± 0.01 -> 0.40 ± 0.00; 1.18 ± 0.00 -> 0.68 ± 0.00; 0.59 ± 0 | Helped |
| Add attention (ablation) | GST-UNet · Synthetic, τ=5, β1=0.0–2.0 | RMSE: 0.50 ± 0.00 -> 0.33 ± 0.00; 0.46 ± 0.00 -> 0.35 ± 0.00; 0.51 ± 0.00 -> 0.40 ± 0.00; 0.45 ± 0.01 -> 0.44 ± 0.00; 0.47 ± 0 | Mixed |
| Add curriculum training (ablation) | GST-UNet · Synthetic, τ=5 and τ=10, β1=0.0–2.0 | RMSE: 0.69 ± 0.00 -> 0.33 ± 0.00; 0.64 ± 0.00 -> 0.35 ± 0.00; 0.63 ± 0.00 -> 0.40 ± 0.00; 0.61 ± 0.01 -> 0.44 ± 0.00; 0.61 ± 0 | Helped |
| Use 3×3 rather than 1×1 spatial kernel to retain neighbor aggregation (ablation) | GST-UNet · Synthetic, τ=5, β1=0.0–2.0 | RMSE: 0.53 ± 0.004 -> 0.33 ± 0.004; 0.55 ± 0.005 -> 0.35 ± 0.004; 0.54 ± 0.005 -> 0.40 ± 0.005; 0.60 ± 0.007 -> 0.44 ± 0.004; | Helped |
| Estimate Camp Fire effects with GST-UNet rather than UNet+ | Camp Fire counterfactual analysis · Camp Fire, Nov 8–17, 2018 | Excess respiratory hospitalizations: 3,981; [-899, 5202] vs approximately 4,650; [1888, 6535] | Mixed |
| Replace GST-UNet with IPWUNet | Camp Fire counterfactual analysis · Camp Fire, Nov 8–17, 2018 | Excess respiratory hospitalizations: approximately 4,650 vs ∼20{,}500 | Hurt |
Surprising findings
- UNet+ (no causal adjustment) performs best when there is no time-varying confounding (beta1=0), so G-computation adds error in that regime.
- IPWUNet is biased even at beta1=0 due to uncorrected interference.
Limitations
- Relies on untestable assumptions (sequential unconfoundedness, representation-based time invariance) and consistency theory that assumes equicontinuity conditions.
- Real-world application uses binary treatment defined by PM2.5 threshold and interpolated county data with small-population exclusions.
- Theoretical consistency is stated as o_p(1) with conditions that are hard to verify in practice.
Experimental setup
Datasets: synthetic DGP with advection-diffusion kernels California county-level daily PM2.5, hospitalizations and weather data (2018 Camp Fire)
Benchmarks: synthetic DGP with beta1 confounding sweep (tau=5, 10) 2018 California Camp Fire wildfire smoke application
Metrics: RMSE estimated excess hospitalizations with 95% bootstrap CI
Baselines: UNet+ STCINet IPWUNet GST-UNet w/o Attention GST-UNet w/o Curriculum
Builds on: iterative G-computation U-Net ConvLSTM attention gates
Closest papers by meaning
- SpaCE: The Spatial Confounding Environment (ICLR 2024 · similarity 0.88)
- Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach (NeurIPS 2025 · similarity 0.88)
- Discovering Latent Causal Graphs from Spatiotemporal Data (ICML 2025 · similarity 0.87)
- CausalPFN: Amortized Causal Effect Estimation via In-Context Learning (NeurIPS 2025 · similarity 0.86)
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.86)
- Generalizing Causal Effects from Randomized Controlled Trials to Target Populations across Diverse Environments (ICML 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.