England Environmental Justice & Health Inequalities Analysis

My Analytical Approach: Integrating Diverse Methods

To tackle this multi-faceted problem, I designed and implemented an integrated analytical pipeline primarily using Python and key libraries (Pandas, GeoPandas, PySAL, Scikit-learn, Statsmodels, SHAP). The innovation lies in combining spatial, machine learning, and causal inference perspectives for a more holistic understanding, with careful consideration of methodological choices:

• Spatial Statistics (GeoPandas, PySAL): Employed spatial autocorrelation techniques (Moran's I, LISA, Getis-Ord Gi*) to rigorously test if patterns of pollution and deprivation were clustered or random. Justification: Standard statistics often assume independence, which is violated by geographic data; spatial methods account for this. LISA specifically identifies statistically significant local clusters (hotspots/coldspots), crucial for targeted interventions. Queen contiguity was chosen for spatial weights, suitable for administrative areas sharing borders. Spatial lag models were used later to account for spillover effects between neighbouring LADs, addressing potential bias in standard regression (Ordinary Least Squares (OLS)) from spatial dependence (indicated by the model's rho coefficient).
• Machine Learning (Scikit-learn, SHAP):
  • Unsupervised Clustering (K-Means): Chosen for its computational efficiency on this dataset and its ability to partition LADs into distinct, interpretable typologies based on pollution, deprivation, and health profiles using cluster centroids. Silhouette scores guided cluster number selection. Justification: While alternatives like DBSCAN (density-based) or GMM (probabilistic) were considered (and explored in advanced_cluster_analysis.py), K-Means provided the clearest separation into relatively balanced, policy-relevant groups for this specific dataset and research question.
  • Predictive Modelling (Gradient Boosting Regressor (GBR)): Selected for policy simulation due to its strong predictive power, ability to model complex non-linear relationships and feature interactions (e.g., how deprivation modifies pollution impact), and robustness to outliers compared to linear models. Justification: GBR allowed simulating the potential impact of interventions (like PM2.5 reduction) on the respiratory health index, providing quantitative estimates for policy evaluation. Performance was validated using cross-validation (R²/MSE) to mitigate overfitting.
  • Interpretability (SHAP): Used with GBR to understand why the model makes certain predictions, identifying key drivers (e.g., specific IMD domains) and their impact, moving beyond black-box predictions.
• Causal Inference (Associational - Statsmodels): Applied PSM to estimate the associational effect of high PM2.5 exposure (treatment) on respiratory health (outcome) at the LAD level, controlling for observed confounding variables (IMD domains). Justification: PSM was chosen as a pragmatic approach given observational data limitations. It aims to create comparable groups (high vs. low PM2.5) based on observed characteristics (checking for sufficient overlap via propensity score distributions), approximating a quasi-experimental setup to isolate the association of interest. Diagnostic checks confirmed that propensity score matching successfully balanced observed covariates like IMD domains between the high and low PM2.5 groups (Standardised Mean Differences (SMDs) < 0.1 post-matching), strengthening the comparison. While acknowledging limitations (cannot control for unobserved confounders), rigorous diagnostics (sensitivity analysis via Rosenbaum bounds) were performed to assess the robustness of the association found (ATT).
• Data Integration & Index Construction (Pandas, GeoPandas): Merged complex datasets (ONS, DEFRA, DLUHC-IMD, NHS) across LSOA/LAD scales. Developed custom indices (env_justice_index, respiratory_health_index) with specific weighting rationales (see DATA_DICTIONARY.md) to capture the core concepts of 'double burden' and respiratory health burden effectively.

Impact & Actionable Policy Recommendations

This analysis translates directly into actionable strategies for policymakers and public health bodies, enabling a shift from broad strokes to targeted interventions: