King's College London

Background: Various spatio-temporal models have been proposed for predicting ambient particulate exposure for inclusion in epidemiological analyses. We investigated the effect of measurement error in the prediction of particulate matter with diameter <10 μm (PM10) and <2.5μm (PM2.5) concentrations on the estimation of health effects. Methods: We sampled 1,000 small administrative areas in London, U.K., and simulated the “true” underlying daily exposure surfaces for PM10 and PM2.5 for 2009-2013 incorporating temporal variation and spatial covariance informed by the extensive London monitoring network. We added measurement error assessed by comparing measurements at fixed sites and predictions from spatio-temporal land use regression (LUR) models; dispersion models; models using satellite data and applying machine learning algorithms; and combinations of these methods through generalized additive models. Two health outcomes were simulated to assess whether the bias varies with the effect size. We applied multi-level Poisson regression to simultaneously model the effect of long- and short-term pollutant exposure. For each scenario, we ran 1,000 simulations to assess measurement error impact on health effect estimation. Results: For long-term exposure to particles we observed bias towards the null, except for traffic PM2.5 for which only LUR underestimated the effect. For short-term exposure results were variable between exposure models and bias ranged from -11% (underestimate) to 20% (overestimate) for PM10 and of -20% to 17% for PM2.5. Integration of models performed best in almost all cases. Conclusions: No single exposure model performed optimally across scenarios. In most cases measurement error resulted in attenuation of the effect estimate.