Multinational prediction of household and personal exposure to fine particulate matter (PM2.5) in the PURE cohort study

Abstract

Introduction: Use of polluting cooking fuels generates household air pollution (HAP) containing health-damaging levels of fine particulate matter (PM2.5). Many global epidemiological studies rely on categorical HAP exposure indicators, which are poor surrogates of measured PM2.5 levels. To quantitatively characterize HAP levels on a large scale, a multinational measurement campaign was leveraged to develop household and personal PM2.5 exposure models. Methods: The Prospective Urban and Rural Epidemiology (PURE)-AIR study included 48-hour monitoring of PM2.5 kitchen concentrations (n = 2,365) and male and/or female PM2.5 exposure monitoring (n = 910) in a subset of households in Bangladesh, Chile, China, Colombia, India, Pakistan, Tanzania and Zimbabwe. PURE-AIR measurements were combined with survey data on cooking environment characteristics in hierarchical Bayesian log-linear regression models. Model performance was evaluated using leave-one-out cross validation. Predictive models were applied to survey data from the larger PURE cohort (22,480 households; 33,554 individuals) to quantitatively estimate PM2.5 exposures. Results: The final models explained half (R2 = 54%) of the variation in kitchen PM2.5 measurements (root mean square error (RMSE) (log scale):2.22) and personal measurements (R2 = 48%; RMSE (log scale):2.08). Primary cooking fuel type, heating fuel type, country and season were highly predictive of PM2.5 kitchen concentrations. Average national PM2.5 kitchen concentrations varied nearly 3-fold among households primarily cooking with gas (20 μg/m3 (Chile); 55 μg/m3 (China)) and 12-fold among households primarily cooking with wood (36 μg/m3 (Chile)); 427 μg/m3 (Pakistan)). Average PM2.5 kitchen concentration, heating fuel type, season and secondhand smoke exposure were significant predictors of personal exposures. Modeled average PM2.5 female exposures were lower than male exposures in upper-middle/high-income countries (India, China, Colombia, Chile). Conclusion: Using survey data to estimate PM2.5 exposures on a multinational scale can cost-effectively scale up quantitative HAP measurements for disease burden assessments. The modeled PM2.5 exposures can be used in future epidemiological studies and inform policies targeting HAP reduction.

Keywords

• Household air pollution
• PM2.5
• Kitchen concentrations
• Personal exposures
• Predictive modeling
• Bayesian hierarchical modeling