Causal Effects of Air Pollution on Mortality Rate in Massachusetts

Abstract

Air pollution epidemiology studies have primarily investigated long- and short-term exposures separately, have used multiplicative models, and have been associational studies. Implementing a generalized propensity score adjustment approach with 3.8 billion person-days of follow-up, we simultaneously assessed causal associations of long-term (1-year moving average) and short-term (2-day moving average) exposure to particulate matter with an aerodynamic diameter less than or equal to 2.5 μm (PM_2.5), ozone, and nitrogen dioxide with all-cause mortality on an additive scale among Medicare beneficiaries in Massachusetts (2000–2012). We found that long- and short-term PM_2.5, ozone, and nitrogen dioxide exposures were all associated with increased mortality risk. Specifically, per 10 million person-days, each 1-μg/m³ increase in long- and short-term PM_2.5 exposure was associated with 35.4 (95% confidence interval (CI): 33.4, 37.6) and 3.04 (95% CI: 2.17, 3.94) excess deaths, respectively; each 1–part per billion (ppb) increase in long- and short-term ozone exposure was associated with 2.35 (95% CI: 1.08, 3.61) and 2.41 (95% CI: 1.81, 2.91) excess deaths, respectively; and each 1-ppb increase in long- and short-term nitrogen dioxide exposure was associated with 3.24 (95% CI: 2.75, 3.77) and 5.60 (95% CI: 5.24, 5.98) excess deaths, respectively. Mortality associated with long-term PM_2.5 and ozone exposure increased substantially at low levels. The findings suggested that air pollution was causally associated with mortality, even at levels below national standards.

Abbreviations

CI

confidence interval

GPS

generalized propensity score

OLS

ordinary least squares

PM_2.5

particulate matter with an aerodynamic diameter less than or equal to 2.5 μm

ppb

parts per billion

Epidemiologic researchers consistently report associations of exposure to ambient air pollutants such as particulate matter with an aerodynamic diameter less than or equal to 2.5 μm (PM_2.5), ozone, and nitrogen dioxide with adverse health outcomes (1–6). In general, health outcomes associated with long- and short-term exposure to air pollution have been assessed separately, leaving unanswered the question of whether the risks would be present after mutual adjustment and, if so, to what extent. Although a few studies have evaluated long- and short-term exposures simultaneously, they focused on PM_2.5 only, without adjusting for other pollutants (7, 8).

Multiplicative models are the most commonly used approach in air pollution epidemiology studies (9). For example, survival analysis has been widely used for studying chronic exposure to air pollution over a period of years (1, 2, 6). Studies of short-term exposure have been primarily based on a case-crossover design or Poisson log-linear regression for event counts (3, 7). These analyses estimate risks on a multiplicative scale, which is the ratio of the risk of the outcome in one group to the risk in another. Multiplicative models build in interactions among all covariates, which makes extrapolation of results to populations with different distributions of covariates difficult. Moreover, multiplicative models may obscure the magnitude of an exposure-outcome association (10). When the outcome event is rare, an extremely large multiplicative risk of the outcome does not necessarily indicate a high risk; conversely, a moderate multiplicative risk may not indicate a low risk if the outcome is more common. Additive models, which estimate the absolute risk difference between groups, can overcome these issues. Compared with multiplicative risk, additive risk is usually of greater public health importance and is more useful for assessing biological interactions (11, 12). For example, modeling a binary outcome on an additive probability scale can help one estimate directly how many outcome events would occur per increment of the exposure, which provides insight into the actual magnitude of the relationship.

In addition, causal modeling approaches have been rare in observational studies of air pollution. This has been cited by the Environmental Protection Agency’s scientific advisors as a major limitation, suggesting a lack of casual evidence regarding the health effects of ambient air pollutants (13). In randomized experiments, subjects are randomly assigned to treatment groups, which ensures that different groups are comparable with respect to any potential confounders. In the context of observational studies, however, whether an individual is exposed to high or low levels of pollutants is not determined at random but instead can be related to characteristics of that individual or population (14). If all necessary confounders can be adequately adjusted for so that the hypothetical treatment groups are comparable with respect to all of the measured confounders, then assignment to a high or low air pollution level can be constructed as an approximate randomized experiment (15).

In the present study, we analyzed 3.8 billion person-days of follow-up among Medicare beneficiaries aged ≥65 years who resided in Massachusetts during the years 2000–2012. We implemented a parametric generalized propensity score (GPS) adjustment approach, which was motivated by simultaneous assessment of the causal associations of long- and short-term exposure to ambient PM_2.5, ozone, and nitrogen dioxide with all-cause mortality on an additive scale.

METHODS

This study was approved by the institutional review board of the Harvard T.H. Chan School of Public Health and was exempt from informed consent requirements as a study of previously collected administrative data.

Mortality data

The study outcome was all-cause mortality. We obtained the Medicare beneficiary denominator file from the Centers for Medicare and Medicaid Services. We constructed an open cohort with person-days of follow-up for Medicare beneficiaries aged 65 years or older living in Massachusetts from January 1, 2000, to December 31, 2012. For each beneficiary, we extracted demographic details, including age, race/ethnicity, sex, Medicaid eligibility, zip code of residence, and date of death. Beneficiaries who were alive on January 1, 2000, or enrolled in the Medicare program during the years 2000–2012 were followed up each day until death, loss to follow-up, or the end of the study, whichever occurred first. In total, we created a data set with 3.8 billion person-days of follow-up.

Exposure data

We predicted daily ambient PM_2.5, ozone, and nitrogen dioxide levels in each 1-km × 1-km grid cell across the contiguous United States using well-validated ensemble models (16, 17). Briefly, for each pollutant, we estimated daily concentrations in each grid cell by combining predictions from 3 machine learning algorithms (i.e., random forest, gradient boosting, and neural network) in a geographically weighted regression. The variables used in the learners included ground monitoring data, satellite-derived measurements of aerosol optical depth, meteorological conditions (e.g., daily air temperature, relative humidity, wind speed, and height of the planetary boundary layer), chemical transport model simulations, and land-use variables (e.g., distance to major roads, local pollutant emissions, and land-use patterns). Using the daily predictions for the 1-km × 1-km grid cells, we then estimated daily concentrations in specific zip codes. There are 2 major types of zip codes: a standard zip code, which is the code for an area surrounding a post office, and a zip code for a Post Office box, which is used only for a particular facility, such as a large office building, university, bank, etc. (18). For a standard zip code, we estimated its daily concentrations of pollutants by averaging the estimates from grid cells whose centroids fell within the boundary of that zip code; for a Post Office box, we estimated the daily concentrations by linking it to the nearest grid cell.

We considered 3 long-term exposures (long-term PM_2.5, ozone, and nitrogen dioxide) and 3 short-term exposures (short-term PM_2.5, ozone, and nitrogen dioxide). On each person-day, following the methods used in previous studies (2, 6), we defined a long-term exposure as a 1-year moving average of the exposure concentration (i.e., the average concentration on that day and the previous 364 days) and a short-term exposure as a 2-day moving average concentration (i.e., the average concentration on that day and the previous day). Most prior studies evaluating the association between short-term exposure to ambient ozone and excess mortality restricted analysis to the warm season (2, 4, 5). Using the same approach as Di et al. (2), we defined the warm season as the period from April 1 to September 30, which is the time window for evaluating the short-term effect of ozone on mortality. Evaluation of long-term ozone effects used a 1-year moving average, and hence the entire period. We assigned the long- and short-term PM_2.5, ozone, and nitrogen dioxide exposures to each person-day on the basis of that person’s zip code of residence and the date of follow-up.

Covariates

Data on meteorological covariates, including daily surface air and dew point temperatures, were obtained from the National Centers for Environmental Prediction/National Center for Atmospheric Research Reanalysis Project with an approximately 32-km × 32-km spatial resolution and were linked to each person-day on the basis of zip code of residence and the date of follow-up (19). Information on individual-level covariates, including sex (male or female), race/ethnicity (White, Black, or other), age (3-year categories), and Medicaid eligibility (as a proxy for low socioeconomic status; yes or no), was obtained from the Medicare denominator file. Data on Zip Code Tabulation Area–level socioeconomic status, including annual median household income, median value of owner-occupied housing units, percentage of the population living in poverty, percentage of the population with less than a high school education, population density, and home ownership rate, were linearly extrapolated from the 2000 and 2010 US Census and the American Community Survey to account for the time-varying nature of these covariates. Data on county-level covariates, including annual percentage of ever smokers and percentage of obese people, were obtained from the Behavioral Risk Factor Surveillance System (20). The Zip Code Tabulation Area–level socioeconomic status and county-level covariates were linked to each person-day on the basis of that person’s zip code of residence and the date of follow-up.

A GPS adjustment approach

The propensity score, which is the conditional probability of being exposed given the observed baseline covariates, is increasingly being used for estimating causal effects of exposures in observational studies (14). It was originally developed for binary exposures or treatments and has been extended to continuous cases, referred to as a GPS (15). For a continuous exposure, the GPS represents the relative likelihood of being exposed to the observed exposure level given the covariates, which in practice can be determined by the conditional density at the observed exposure level. As in the binary case, under the assumption that the GPS model has been correctly specified, the estimated GPS provides a scalar summary of baseline covariates that can be used to eliminate confounding through adjustment. In the present study, we applied a parametric GPS adjustment approach. For each exposure, the analysis comprised 2 stages: 1) a design stage where the GPS was estimated to summarize all measured confounders and 2) an analysis stage where the treatment effect was estimated conditional on the estimated GPS (21).

We had a data set with person-days of follow-up, indexed by . In the design stage, for each exposure (long- or short-term exposure to one of the 3 pollutants), we fitted an ordinary least squares (OLS) model regressing the exposure against a linear combination of covariates, including the other 5 exposures, long-term (lag 0–364 days) and short-term (both lag 0–1 days and lag 2–6 days) exposures to meteorological factors, year of follow-up, and the set of individual-, Zip Code Tabulation Area-, and county-level covariates as mentioned above. Inclusion of the other 5 exposures in the GPS model controlled for joint confounding by the concurrent exposures; inclusion of long-term meteorological factors controlled for confounding by changing weather or climate (22); inclusion of short-term meteorological factors controlled for potential confounding by meteorological lags, which could be delayed up to a week; and inclusion of year controlled for confounding by long-term time trend. For short-term exposures, we additionally included month, to control for seasonality, and day of the week, to control for seasonal variations in air pollution and mortality rate within a week. To capture a potentially nonlinear confounding effect, we modeled each continuous covariate as a cubic polynomial.

For each person-day i, we assumed that the conditional density of the exposure at the observed level was normally distributed with mean and variance given the covariate matrix (23):

where the parameters , , and were estimated using OLS.

Then the conditional density could be evaluated by means of the normal density function:

This was the estimated GPS for person-day i at the observed exposure level (15).

In the analysis stage, for each exposure, we again used OLS regression to fit a linear probability model relating the binary outcome of death () with the observed exposure level () and the estimated GPS () (15, 24):

The coefficient is the estimated causal effect of air pollution on mortality, which can be interpreted as the average change in the probability of death if the exposure level increased by 1 unit for each observation. Such causal interpretation comes from the fact that the identity link function for OLS regression is collapsible; thus, the conditional and marginal effect estimates are numerically identical (25). Because of the heteroscedastic nature of the linear probability model’s error, for each estimated coefficient we bootstrapped a 95% confidence interval by resampling with replacement from the original data with the same sample size, performing the analysis among the resampled data, and repeating this routine 200 times (26).

For each exposure, we further evaluated its causal effects on mortality at low concentration levels by restricting the analysis to person-days with exposure levels below increasingly stringent cutpoints.

Sensitivity analyses

We performed sensitivity analyses to assess the robustness of the results with respect to different lag times for short-term exposure to meteorological factors (at lag 0–1 days or lag 0–4 days) and different seasonal adjustment strategies by including week-of-the-year and weekday/weekend indicators. We adjusted for history of hospitalizations for chronic obstructive pulmonary disease and cardiovascular disease among persons who were enrolled in the Medicare fee-for-service program in order to check whether disease history was a confounder (details are provided in the Web material). In addition, we evaluated robustness to the outcome model specification by modeling the GPS with a cubic polynomial, which permits more flexibility. We also conducted single-pollutant analysis for each exposure without adjustment for the other exposures.

RESULTS

During the years 2000–2012 in Massachusetts, the Medicare cohort comprised 1,503,572 beneficiaries, with a total of 3,874,869,248 person-days of follow-up. Among these individuals, 561,193 deaths occurred. The cohort comprised mostly females (57.5%), White persons (92.2%), and persons aged 65–69 years at study entry (54.2%) (Table 1). Among all zip codes of residence, long-term PM_2.5 concentrations ranged from 3.3 μg/m³ to 16.4 μg/m³, long-term ozone concentrations ranged from 25.7 parts per billion (ppb) to 47.1 ppb, and long-term nitrogen dioxide concentrations ranged from 3.2 ppb to 64.6 ppb. For short-term concentrations, PM_2.5 ranged from 0.1 μg/m³ to 65.3 μg/m³, ozone ranged from 3.0 ppb to 116.0 ppb during the warm season, and nitrogen dioxide ranged from 0.2 ppb to 119.0 ppb. Overall, for each pollutant, the long-term exposure had the same mean value as the short-term exposure but lower variability (Tables 2 and 3).

Table 1.

Characteristics of Medicare Beneficiaries Residing in Massachusetts, 2000–2012

Characteristic	No.	%
Person-days of follow-up	3,874,869,248
Population size	1,503,572	100.0
Average duration of follow-up, days	2,577
Death	561,193	37.3
Male sex	638,620	42.5
Race/ethnicity
White	1,386,883	92.2
Black	51,978	3.5
Other	64,711	4.3
Age at study entry, years
65–69	814,279	54.2
70–74	222,885	14.8
75–79	195,788	13.0
80–84	139,401	9.3
≥85	131,219	8.7
Medicaid eligibility	255,008	17.0

Table 2.

Estimated Ambient Concentrations of Air Pollutants in Massachusetts, 2000–2012

Exposure Period and Pollutant	Mean (SD)	Range
Long-term PM2.5, μg/m3a	9.0 (1.9)	3.3–16.4
Short-term PM2.5, μg/m3b	8.9 (5.4)	0.1–65.3
Long-term ozone, ppba	37.5 (3.0)	25.7–47.1
Short-term ozone, ppbb,c	37.4 (11.1)	3.0–116.0
Long-term nitrogen dioxide, ppba	20.4 (8.3)	3.2–64.6
Short-term nitrogen dioxide, ppbb	20.5 (11.6)	0.2–119.0

Abbreviations: PM_2.5, particulate matter with an aerodynamic diameter less than or equal to 2.5 μm; ppb, parts per billion; SD, standard deviation.

^a Long-term exposure to air pollution was defined as a 1-year moving average of the exposure level (lag 0–364 days).

^b Short-term exposure to air pollution was defined as a 2-day moving average of the exposure level (lag 0–1 days).

^c Short-term ozone level was estimated during the warm season (April 1–September 30).

Table 3.

Percentile Rankings for Estimated Ambient Concentrations of Air Pollutants in Massachusetts, 2000–2012

	Exposure Period and Pollutant
Percentile Rank	Long-Term PM 2.5 , μg/m3a	Short-Term PM 2.5 , μg/m3b	Long-Term Ozone, ppb a	Short-Term Ozone, ppb b , c	Long-Term Nitrogen Dioxide, ppb a	Short-Term Nitrogen Dioxide, ppb b
5	5.8	3.0	32.2	21.5	8.6	5.3
10	6.5	3.7	33.6	24.4	10.4	7.1
30	7.9	5.6	36.1	30.9	15.2	12.7
50	9.0	7.5	37.7	36.4	19.2	18.7
70	9.9	10.3	39.2	42.7	24.7	25.8
90	11.3	15.9	41.2	51.7	31.9	36.4
95	12.1	19.4	42.1	56.8	35.0	41.6

Abbreviations: PM_2.5, particulate matter with an aerodynamic diameter less than or equal to 2.5 μm; ppb, parts per billion.

^a Long-term exposure to air pollution was defined as a 1-year moving average of the exposure level (lag 0–364 days).

^b Short-term exposure to air pollution was defined as a 2-day moving average of the exposure level (lag 0–1 days).

^c Short-term ozone level was estimated during the warm season (April 1–September 30).

We found that long- and short-term exposures to PM_2.5, ozone, and nitrogen dioxide were all statistically significantly and independently associated with an increased risk of mortality. Specifically, per 10 million person-days, each 1-μg/m³ increase in long- and short-term exposure to PM_2.5 was associated with 35.4 (95% confidence interval (CI): 33.4, 37.6) and 3.04 (95% CI: 2.17, 3.94) excess deaths, respectively; each 1-ppb increase in long- and short-term exposure to ozone was associated with 2.35 (95% CI: 1.08, 3.61) and 2.41 (95% CI: 1.81, 2.91) excess deaths, respectively; and each 1-ppb increase in long- and short-term exposure to nitrogen dioxide was associated with 3.24 (95% CI: 2.75, 3.77) and 5.60 (95% CI: 5.24, 5.98) excess deaths, respectively (Figure 1). When restricting the analysis to person-days with exposure levels below increasingly stringent thresholds, we found substantially increasingly larger effects of long-term exposures to PM_2.5 and ozone on mortality. Specifically, per 10 million person-days, the number of excess deaths associated with a 1-μg/m³ increase in long-term PM_2.5 exposure increased from 35.5 (at levels ≤14 μg/m³; 95% CI: 33.4, 37.7) to 60.7 (at levels ≤7 μg/m³; 95% CI: 47.9, 73.9), and the number of excess deaths associated with a 1-ppb increase in long-term ozone exposure increased from 1.89 (at levels ≤42 ppb; 95% CI: 0.33, 3.29) to 47.4 (at levels ≤32 ppb; 95% CI: 31.9, 63.7). The effects of short-term ozone exposure on mortality varied at lower concentration levels. The effects of short-term nitrogen dioxide exposure remained stable across different concentration levels (Figure 1). Numerical results are presented in Web Table 1 (available at https://doi.org/10.1093/aje/kwaa098).

Figure 1.

Absolute increase in the probability of death per person-day among Medicare beneficiaries for a 1-unit increase in each of 3 air pollutants (particulate matter with an aerodynamic diameter less than or equal to 2.5 μm (PM_2.5), ozone, and nitrogen dioxide) when analysis was restricted to person-days with pollutant concentrations below increasingly stringent cutpoints, Massachusetts, 2000–2012. The notation “<∞” indicates that the probabilities of death were estimated using the full data set. A) Long-term PM_2.5 exposure; B) short-term PM_2.5 exposure; C) long-term ozone exposure; D) short-term ozone exposure; E) long-term nitrogen dioxide exposure; F) short-term nitrogen dioxide exposure. Bars, 95% confidence intervals. ppb, parts per billion.

In the sensitivity analyses, the estimated effects on mortality remained robust after adjustment for short-term exposures to meteorological factors at different time lags, seasonal fluctuations with different categorizing strategies, or disease history or modeling of the GPS with a cubic polynomial in the outcome regression (Web Tables 2–4). In the single-pollutant analysis, the effects were slightly larger when exposures were not mutually adjusted for (Web Table 5).

DISCUSSION

In the present study, we analyzed 1.5 million Medicare beneficiaries living in Massachusetts with 3.8 billion person-days of follow-up. We implemented a GPS adjustment approach, which accounted for joint confounding by the concurrent exposures, individual- and area-level covariates, meteorological covariates, seasonal variations, and long-term time trends, and we simultaneously evaluated causal effects of long- and short-term exposures to primary air pollutants on mortality. By modeling the binary outcome of death with a linear probability model, we estimated the potential number of deaths that would occur per unit increase in each air pollutant exposure. We found that long- and short-term exposures to PM_2.5, ozone, and nitrogen dioxide were all positively associated with increased risk of death. In addition, the risk difference associated with long-term exposures to PM_2.5 and ozone increased substantially at increasingly stringent concentration levels, including those well below the current National Ambient Air Quality Standards (27) and the World Health Organization air quality guidelines (28). These estimates can be interpreted as the average change in the probability of death per unit increase in the air pollution level within the range of the exposure levels.

Over the study period, each unit increase in short-term PM_2.5 and ozone exposure was associated with 3.04 and 2.41 excess deaths per 10 million person-days, respectively. Consistent with our findings, in a national case-crossover analysis, Di et al. (2) found that a unit increase in short-term exposure to PM_2.5 and ozone was associated with 1.42 and 0.66 deaths per 10 million person-days, respectively. In addition, our findings showed 35.4 excess deaths per 10 million person-days associated with long-term exposure to PM_2.5. Similarly, Wang et al. (11) proposed a doubly robust additive hazard model and estimated that per 10 million person-days, each unit increase in long-term PM_2.5 exposure was associated with approximately 22.7 excess deaths. The consistency across study designs in the Medicare population builds on the existing body of literature on the additive effects of air pollution exposure on total mortality. Very few studies have evaluated long- and short-term exposures simultaneously. Shi et al. (7) and Liang et al. (8) found larger associations of long-term PM_2.5 exposure than of short-term exposure with mortality. We consistently found that the risk difference associated with a unit increase in long-term PM_2.5 exposure was about 10 times larger than that for short-term PM_2.5. A key gap in air pollution epidemiology has been the lack of studies comparing the health effects of various pollutants. Our findings suggest that PM_2.5 is the most deadly air pollutant and that long-term exposure to PM_2.5 is of greater public health concern. Given the number of person-days in the study period, we estimated that each 1-μg/m³ reduction in long-term PM_2.5 exposure could save 1,055 lives every year in Massachusetts.

Assuming that each GPS was correctly specified, we identified causal associations between 3 major air pollutants and total mortality. In April 2018, the Environmental Protection Agency announced that it plans to bar regulators from considering studies that have not made their underlying data public (29). As Thorp et al. (30) pointed out, peer-reviewed studies that use publicly available data could sidestep this rule and can and should inform public policy. After controlling for the confounding effects of both long- and short-term temperature, we have demonstrated substantially increased effects of long-term PM_2.5 and ozone exposure at levels below US national standards and the World Health Organization air quality guidelines. This would seem to address the principal reasons the Clean Air Scientific Advisory Committee has cited for not recommending stricter standards.

The use of OLS regression at both stages of the analysis provided several advantages. First, because the data set was too big to fit into the computer’s memory, use of OLS regression allowed us to break the data into a set of manageable chunks, fit a model with 1 small data chunk, and update it when a new chunk was added. The small outcome regression model made constructing bootstrap confidence intervals substantially efficient. Second, because the identity link function for OLS is collapsible, the conditional and marginal effect estimates were numerically the same (25). Therefore, the coefficients can be considered to represent the population average effects of a 1-unit increase in air pollutant levels, which do not depend on the distribution of the covariates and which provide succinct and easy interpretations for a broad audience. Third, it estimated directly how many deaths would occur in association with multiple air pollution exposures, which is of greater public health importance and provides clear insight into the relative magnitude of their impacts.

This study also had limitations. First, we made a homogeneity assumption that there were no interactions among the effects of the air pollution exposures and confounders. Second, although exposure assessment models show excellent predictive accuracy, the exposures were subject to measurement error because zip code of residence (rather than home address) was the finest geographical unit we could use to link air pollution levels with each individual. Third, the causal conclusion relies on the assumption that all of the confounders were adequately adjusted for. Although residual confounding can never be ruled out, the consistent results across different study designs provide some reassurance that our positive findings were not substantially biased from residual confounding.

We concluded that long- and short-term exposures to PM_2.5, ozone, and nitrogen dioxide were all causally associated with increased risk of mortality. The effects of long-term exposure to PM_2.5 and ozone increased substantially at low concentration levels. Further, the estimates made on the additive probability scale provide clear insight into the relative magnitudes of air pollution effects.