Yaguang Wei1, Brent Coull2, Petros Koutrakis3, Jiabei Yang4, Longxiang Li3, Antonella Zanobetti3, Joel Schwartz3,5. 1. Department of Environmental Health, Harvard T.H. Chan School of Public Health, Landmark Center 4th West, 401 Park Drive, Boston, MA, 02215, USA. weiyg@g.harvard.edu. 2. Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA, USA. 3. Department of Environmental Health, Harvard T.H. Chan School of Public Health, Landmark Center 4th West, 401 Park Drive, Boston, MA, 02215, USA. 4. Department of Biostatistics, School of Public Health, Brown University, Providence, RI, USA. 5. Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, MA, USA.
Abstract
BACKGROUND: We previously found additive effects of long- and short-term exposures to fine particulate matter (PM2.5), ozone (O3), and nitrogen dioxide (NO2) on all-cause mortality rate using a generalized propensity score (GPS) adjustment approach. The study addressed an important question of how many early deaths were caused by each exposure. However, the study was computationally expensive, did not capture possible interactions and high-order nonlinearities, and omitted potential confounders. METHODS: We proposed two new methods and reconducted the analysis using the same cohort of Medicare beneficiaries in Massachusetts during 2000-2012, which consisted of 1.5 million individuals with 3.8 billion person-days of follow-up. The first method, weighted least squares (WLS), leveraged large volume of data by aggregating person-days, which gave equivalent results to the linear probability model (LPM) method in the previous analysis but significantly reduced computational burden. The second method, m-out-of-n random forests (moonRF), implemented scaling random forests that captured all possible interactions and nonlinearities in the GPS model. To minimize confounding bias, we additionally controlled relative humidity and health care utilizations that were not included previously. Further, we performed low-level analysis by restricting to person-days with exposure levels below increasingly stringent thresholds. RESULTS: We found consistent results between LPM/WLS and moonRF: all exposures were positively associated with mortality rate, even at low levels. For long-term PM2.5 and O3, the effect estimates became larger at lower levels. Long-term exposure to PM2.5 posed the highest risk: 1 μg/m3 increase in long-term PM2.5 was associated with 1053 (95% confidence interval [CI]: 984, 1122; based on LPM/WLS methods) or 1058 (95% CI: 988, 1127; based on moonRF method) early deaths each year among the Medicare population in Massachusetts. CONCLUSIONS: This study provides more rigorous causal evidence between PM2.5, O3, and NO2 exposures and mortality, even at low levels. The largest effect estimate for long-term PM2.5 suggests that reducing PM2.5 could gain the most substantial benefits. The consistency between LPM/WLS and moonRF suggests that there were not many interactions and high-order nonlinearities. In the big data context, the proposed methods will be useful for future scientific work in estimating causality on an additive scale.
BACKGROUND: We previously found additive effects of long- and short-term exposures to fine particulate matter (PM2.5), ozone (O3), and nitrogen dioxide (NO2) on all-cause mortality rate using a generalized propensity score (GPS) adjustment approach. The study addressed an important question of how many early deaths were caused by each exposure. However, the study was computationally expensive, did not capture possible interactions and high-order nonlinearities, and omitted potential confounders. METHODS: We proposed two new methods and reconducted the analysis using the same cohort of Medicare beneficiaries in Massachusetts during 2000-2012, which consisted of 1.5 million individuals with 3.8 billion person-days of follow-up. The first method, weighted least squares (WLS), leveraged large volume of data by aggregating person-days, which gave equivalent results to the linear probability model (LPM) method in the previous analysis but significantly reduced computational burden. The second method, m-out-of-n random forests (moonRF), implemented scaling random forests that captured all possible interactions and nonlinearities in the GPS model. To minimize confounding bias, we additionally controlled relative humidity and health care utilizations that were not included previously. Further, we performed low-level analysis by restricting to person-days with exposure levels below increasingly stringent thresholds. RESULTS: We found consistent results between LPM/WLS and moonRF: all exposures were positively associated with mortality rate, even at low levels. For long-term PM2.5 and O3, the effect estimates became larger at lower levels. Long-term exposure to PM2.5 posed the highest risk: 1 μg/m3 increase in long-term PM2.5 was associated with 1053 (95% confidence interval [CI]: 984, 1122; based on LPM/WLS methods) or 1058 (95% CI: 988, 1127; based on moonRF method) early deaths each year among the Medicare population in Massachusetts. CONCLUSIONS: This study provides more rigorous causal evidence between PM2.5, O3, and NO2 exposures and mortality, even at low levels. The largest effect estimate for long-term PM2.5 suggests that reducing PM2.5 could gain the most substantial benefits. The consistency between LPM/WLS and moonRF suggests that there were not many interactions and high-order nonlinearities. In the big data context, the proposed methods will be useful for future scientific work in estimating causality on an additive scale.
Entities:
Keywords:
Air pollution; Causality; Data aggregation; Generalized propensity score; Scaling random forests
Authors: Yan Wang; Itai Kloog; Brent A Coull; Anna Kosheleva; Antonella Zanobetti; Joel D Schwartz Journal: Environ Health Perspect Date: 2016-04-15 Impact factor: 9.031
Authors: Barbara K Butland; Evangelia Samoli; Richard W Atkinson; Benjamin Barratt; Klea Katsouyanni Journal: Environ Health Date: 2019-02-14 Impact factor: 5.984