Research map / Probabilistic and scientific modeling / Molecules and proteins
How Your Location Relates to Health: Variable Importance and Interpretable Machine Learning for Environmental and Sociodemographic Data
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.
Uses knockoffs plus an ensemble of four variable importance metrics (permutation, SHAP, LOCO, CMR) to select top predictors from the MEDSAT England dataset, then analyzes them globally with GAMs and locally with MGWR and per-LAD GAMs.
Problem
Identifying which environmental and sociodemographic variables are most informative for health outcomes (prescription-based proxies such as asthma, anxiety, hypertension, diabetes, depression) in England, and where and when their effects vary, across pre- and post-COVID periods.
Approach
Combines a knockoff-based FDR-controlled filter with mean-rank aggregation of four complementary variable importance metrics, then pairs global GAMs with spatially varying MGWR and local per-district GAMs to reveal both global and regional effects.
Key results
- NO2 is identified as a global predictor (negative correlation, narrow CIs not overlapping zero) for asthma, hypertension, and anxiety prescriptions across pre- and post-COVID periods.
- Environmental variables such as NO2, solar and thermal radiation, and PM2.5 rank higher than sociodemographic variables in the anxiety variable importance analysis, with NO2 the only top predictor shared with the MEDSAT paper.
- Local GAMs for London show the negative NO2-asthma relationship is concentrated in ULEZ districts such as Westminster and Lambeth; PM2.5 effects in northwest England and hypertension in Solihull/Worcester/Wychavon shift around COVID.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the top 10 ranked variables per outcome in downstream models (ablation) | MEDSAT health-outcome models · MEDSAT | predictive performance | No clear effect |
| Reinsert population density, age, and ethnicity into the anxiety model (ablation) | Anxiety model · Anxiety | top-variable rankings | No clear effect |
| Remove spatial smoothing from local GAMs (ablation) | Local GAMs for London LADs · Asthma in London | local NO2 relationship | Helped |
| Include spatial smoothing in global GAMs and MGWR | Global GAMs and MGWR · London asthma | localization of the NO2 relationship | Hurt |
| Use environmental predictors such as NO2 | Global GAMs · Asthma, hypertension, and anxiety | global predictor association | Helped |
| Model NO2 locally rather than relying only on global relationships | GAM and MGWR analysis · Asthma in London | interpretability of local NO2 associations | Helped |
| Analyze PM2.5 coefficients across regions and years | MGWR · Diabetes | regional PM2.5 association: ≈15% decrease | Mixed |
| Analyze PM2.5 coefficients across regions and years | MGWR · Hypertension | regional PM2.5 coefficient: ≤ -0.5 to ≥ 0.5 | Mixed |
| Use local rather than only global analysis of solar radiation | MGWR · Anxiety | regional solar-radiation association | Mixed |
| Use skilled-trades occupation as a predictor | Global GAMs · Diabetes | association with prescriptions | Helped |
Surprising findings
- NO2 shows a negative correlation with asthma, hypertension and anxiety prescriptions despite known links between air pollution and respiratory disease.
- White ethnicity ranks high under SHAP but low under LOCO and CMR, likely because it can be inferred from other variables.
Limitations
- Negative NO2 finding is partly attributed to spatial smoothing in MGWR and global GAMs spreading ULEZ effects to neighboring areas, so local results may be confounded.
- Prescription quantity is used as a proxy for health outcomes, which may not reflect true disease prevalence.
- Lacks quantitative predictive evaluation of the GAM and MGWR models; several explanations of COVID-era shifts are speculative.
Experimental setup
Datasets: MEDSAT
Metrics: R^2 generalized cross validation score variable importance rank
Baselines: LightGBM model from MEDSAT paper
Builds on: MEDSAT knockoffs SHAP multiscale geographically weighted regression generalized additive models
Closest papers by meaning
- SpaCE: The Spatial Confounding Environment (ICLR 2024 · similarity 0.85)
- Smooth Sailing: Lipschitz-Driven Uncertainty Quantification for Spatial Associations (NeurIPS 2025 · similarity 0.84)
- GST-UNet: A Neural Framework for Spatiotemporal Causal Inference with Time-Varying Confounding (NeurIPS 2025 · similarity 0.83)
- Generalizing Causal Effects from Randomized Controlled Trials to Target Populations across Diverse Environments (ICML 2025 · similarity 0.83)
- Measuring Variable Importance in Heterogeneous Treatment Effects with Confidence (ICML 2025 · similarity 0.80)
- Automating the Selection of Proxy Variables of Unmeasured Confounders (ICML 2024 · similarity 0.79)
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.