Estimating exposure response functions using ambient pollution concentrations

This paper presents an approach to estimating the health effects of an environmental hazard. The approach is general in nature, but is applied here to the case of air pollution. It uses a computer model involving ambient pollution and temperature inputs, to simulate the exposures experienced by individuals in an urban area, whilst incorporating the mechanisms that determine exposures. The output from the model comprises a set of daily exposures for a sample of individuals from the population of interest. These daily exposures are approximated by parametric distributions, so that the predictive exposure distribution of a randomly selected individual can be generated. These distributions are then incorporated into a hierarchical Bayesian framework (with inference using Markov Chain Monte Carlo simulation) in order to examine the relationship between short-term changes in exposures and health outcomes, whilst making allowance for long-term trends, seasonality, the effect of potential confounders and the possibility of ecological bias. The paper applies this approach to particulate pollution (PM$_{10}$) and respiratory mortality counts for seniors in greater London ($\geq$65 years) during 1997. Within this substantive epidemiological study, the effects on health of ambient concentrations and (estimated) personal exposures are compared.

💡 Research Summary

The paper introduces a comprehensive framework for estimating exposure‑response functions that explicitly distinguishes between ambient pollutant concentrations and the personal exposures that actually drive health effects. The authors first construct a high‑resolution spatiotemporal model of daily PM₁₀ concentrations and temperature for an urban area, using monitoring data and interpolation techniques. They then generate a synthetic cohort of individuals, assigning each person demographic attributes (age, sex, occupation, activity patterns) and linking those attributes to inhalation rates and time‑location profiles. By combining the ambient concentration field with these personal activity schedules, the model simulates daily inhaled doses for each individual. Because the simulated doses are stochastic—reflecting uncertainties in movement, indoor‑outdoor ratios, and breathing rates—the authors fit parametric distributions (normal, log‑normal, gamma) to the simulated daily exposures, allowing them to draw exposure values for a randomly selected individual in subsequent analyses.

The second major component is a hierarchical Bayesian regression that relates daily respiratory mortality counts among seniors (≥65 years) in Greater London during 1997 to either (a) the observed ambient PM₁₀ concentration or (b) the model‑derived personal exposure average. The first level of the hierarchy specifies a Poisson (or negative‑binomial) likelihood for the mortality counts and a log‑linear (or flexible spline) link to the exposure metric. The second level incorporates long‑term trends, seasonal cycles (via periodic splines), and a set of confounders such as temperature, influenza activity, and socioeconomic indicators. The third level treats the parameters of the personal‑exposure distribution as random variables with prior distributions, thereby propagating the uncertainty of the exposure simulation into the health effect estimates and explicitly accounting for potential ecological bias that arises when only ambient concentrations are used.

Inference is performed with Markov Chain Monte Carlo (MCMC) using a hybrid Gibbs/Metropolis‑Hastings algorithm. The authors run over 100 000 iterations, assess convergence with Gelman‑Rubin diagnostics (R̂ < 1.1) and effective sample size (ESS > 1 000), and report posterior means and 95 % credible intervals for the exposure‑mortality coefficients. Results show that a 10 µg m⁻³ increase in ambient PM₁₀ is associated with a 4 % rise in daily respiratory deaths (RR = 1.04, 95 % CI 1.01–1.07). When the personal‑exposure metric is used, the estimated risk is slightly larger—about a 6 % increase (RR = 1.06, 95 % CI 1.02–1.10)—and the credible interval is wider, reflecting the additional variability captured by the exposure simulation. Sensitivity analyses demonstrate that omitting trend or seasonal adjustments inflates the risk estimates, especially for the ambient‑concentration model, underscoring the importance of the hierarchical structure.

The study’s contributions are threefold: (1) it provides a reproducible method for generating individual‑level exposure distributions from ambient monitoring data and meteorology; (2) it integrates exposure uncertainty into a Bayesian health‑effects model, thereby quantifying ecological bias; and (3) it empirically compares health effect estimates derived from ambient concentrations versus simulated personal exposures, showing that the latter can yield larger and more realistic risk estimates. Limitations include the reliance on simplified activity patterns, the subjective choice of prior distributions, and the focus on a single city and year, which restricts external validity. Future work should incorporate wearable sensor data for validation, extend the approach to multiple locations and longer time spans, and explore non‑linear or lagged exposure‑response relationships within the same Bayesian framework.