Mortality burden of air pollution — Health Impact Assessment

The problem

The institute needed a rigorous quantification of the mortality burden attributable to urban air pollution — specifically PM10, PM2.5, and NO₂ — under counterfactual traffic scenarios. The output had to be defensible to policy stakeholders and methodologically consistent with the epidemiological Health Impact Assessment (HIA) literature.

Approach

The HIA backbone

The standard HIA pipeline links exposure-response functions from cohort studies to population-level exposure distributions and baseline mortality rates, yielding attributable fractions and counterfactual mortality counts under scenario perturbations. I implemented this end-to-end with traceable assumptions — every concentration–response coefficient, every baseline rate, every population stratum could be inspected, swapped, and re-run.

Statistical modelling of pollutant effects

• Poisson regression for count outcomes — the workhorse for mortality counts at the area-period level, with overdispersion handled where the variance inflation justified it.
• Panel fixed-effects models exploiting within-unit variation across time, controlling for unobserved area-level confounders that would otherwise bias naive cross-sectional estimates.
• Traffic-scenario counterfactuals translated into pollutant exposure shifts via the institute's atmospheric dispersion outputs, then propagated through the attributable-fraction machinery.

Causal methods on the perception survey

Beyond the mortality work, the institute had a large-scale air-quality perception survey — what citizens believed about local air quality and how that belief drove behaviour. Naively regressing perception on measured pollutant levels conflates the genuine effect with confounders like media exposure, socioeconomic status, and area-level priors. I applied:

• Causal forests to recover heterogeneous treatment effects across respondent subgroups — surfacing where the link between objective pollution and perceived pollution diverged most strongly.
• Double/debiased Machine Learning (Double ML) for the average treatment effect under high-dimensional confounding controls — the orthogonalisation step letting flexible nuisance models (gradient boosting on both treatment and outcome residuals) coexist with valid inference on the effect of interest.

Outcome

The HIA delivered a defensible, scenario-conditional mortality burden estimate that the institute could put in front of policy stakeholders. The causal analysis of the perception survey added a complementary picture — how citizens read the air around them — that the mortality numbers alone did not capture.

Stack

Python · statsmodels · linearmodels (panel FE) · econml (causal forest, DoubleML) · R for HIA-specific tooling · spatial joins on dispersion outputs.