Necessity of personal sampling for exposure assessment on specific constituents of PM2.5: Results of a panel study in Shanghai, China

Abstract

Many epidemiological studies have evaluated the health risks of ambient fine particulate matter (PM_2.5). However, few studies have investigated the potential exposure misclassification caused by using ambient PM_2.5 concentrations as proxy for individual exposure to PM_2.5 in regions with high-level of air pollution. This study aimed to compare the differences between personal and ambient PM_2.5 constituent concentrations, and to predict the personal exposure of sixteen PM_2.5 constituents. We collected 141 72-h personal exposure filter samples from a panel of 36 healthy non-smoking college students in Shanghai, China. We then used the liner mixed effects models to predict personal constituent-specific exposure using ambient observations and several possible influencing factors including time-activity patterns, temporal variables, and meteorological conditions. The final model of each component was further evaluated by determination coefficient (R²) and root mean square error (RMSE) from leave-one-out-cross-validation (LOOCV). We observed ambient concentrations were higher than personal concentrations for all PM_2.5 components except for Mn, Fe, Ca, and V. Especially, ambient NH₄⁺, As, and NO₃^- concentrations were 3.65, 5.65 and 7.33-fold higher than their corresponding personal concentrations, respectively. The ambient level was the strongest predictor of their corresponding personal PM_2.5 components with the highest marginal R² (R_M²: 0.081 ~ 0.901), meteorological conditions (R_M²: 0.000 ~ 0.357), time-activity pattern (R_M²: 0.000 ~ 0.083) and temporal indicators (R_M²: 0.031 ~ 0.562) were also important predictors. Our final models predicted at least 50% of the variance of all personal PM_2.5 constituents and even over 90% for K, Pb, and SO₄^2-. LOOCV analysis showed that R² and RMSE ranged from 0.251 to 0.907 and 0.000 to 0.092 μg/m³, respectively. Our results showed that ambient concentration of most PM_2.5 constituents along with time-activity patterns, temporal variables, and meteorological conditions, could adequately predict personal exposure concentration. Prediction models of individual PM_2.5 constituent may help to improve the accuracy of exposure measurement in future epidemiological studies.