Abstract
Estimating causal effects from spatiotemporal observational data is a critical task in fields such as public health and environmental science. However, this task is complicated by three primary challenges: spatial interference, temporal carryover, and time-varying confounding (feedback loops). We introduce GST-UNet (G-computation Spatio-Temporal UNet), the first unified neural framework designed to address these challenges simultaneously within a single spatiotemporal trajectory. GST-UNet integrates a U-Net architecture with ConvLSTM and attention mechanisms to capture complex dependencies, utilizing iterative G-computation for causal adjustment. A novel curriculum training strategy is employed to stabilize the recursive learning process. Evaluation on synthetic simulations and 2018 California wildfire data shows that GST-UNet significantly outperforms traditional methods, providing physically consistent and accurate causal estimates.
Introduction
In fields like public health and environmental science, researchers often rely on observational data to track how variables like air quality regulations or natural disasters affect outcomes over time and across regions. Spatiotemporal causal inference is notoriously difficult due to three main factors:
- Interference: Actions or events in one region (e.g., a wildfire) directly affect neighboring areas.
- Temporal Carryover: The effects of an event today may manifest fully only in the future.
- Time-Varying Confounding: Feedback loops where past outcomes influence future treatments (e.g., high pollution triggering new regulations).
Existing tools often fail to address all three simultaneously, particularly when learning from a single "history" or trajectory rather than multiple independent samples.
Methods
The GST-UNet framework consists of two primary components: a spatiotemporal backbone and a causal engine.
Spatiotemporal Backbone
The model utilizes a U-Net architecture to capture spatial patterns, augmented with ConvLSTM layers to track temporal evolution. An Attention mechanism helps the model focus on relevant neighboring regions, ensuring that spatial context is preserved. Ablation studies indicate that removing this spatial context can cause performance to drop by 30-40%.
Causal Engine: Iterative G-computation
To handle time-varying confounding, GST-UNet performs recursive adjustment via iterative G-computation. The model works backward from a future horizon, predicting outcomes and using them as "pseudo-outcomes" for earlier time steps. To stabilize this recursive chain, the authors employ Curriculum Training, where the model first learns to predict the final step before gradually incorporating earlier steps.
Datasets
The model was validated using two distinct setups:
- Synthetic Advection-Diffusion Simulation: A 64x64 grid simulation mimicking pollutant transport, allowing for controlled testing of confounding strength.
- 2018 California Wildfires: A real-world dataset combining daily PM2.5 levels, weather variables, and respiratory hospitalization counts across California counties during the Carr and Camp Fires.
Results
GST-UNet demonstrated superior accuracy and robustness:
- Accuracy: In synthetic tests with high confounding, the model reduced RMSE by over 30% compared to standard U-Nets and existing spatiotemporal causal models.
- Real-World Impact: Applied to the 2018 Camp Fire, GST-UNet estimated approximately 4,650 excess respiratory hospitalizations attributable to smoke.
- Consistency: Unlike baseline methods like IPWUNet, which produced volatile estimates, GST-UNet’s results were physically consistent and aligned with epidemiological studies at a finer spatial resolution.
Conclusion
While GST-UNet marks a significant advancement, it currently requires data on a regular grid and assumes no unobserved confounders. Future work will explore adapting the framework to Graph Neural Networks for irregular boundaries and addressing potential regime shifts in system dynamics.
Summary updated