A Self-Controlled Approach to Survival Analysis, with application to Air Pollution and Mortality

Abstract

Background:

Many studies have reported that long-term air pollution exposure is associated with increased mortality rates. These investigations have been criticized for failure to control for omitted, generally personal, confounders. Study designs that are robust to such confounders can address this issue.

Methods:

We used a self-controlled design for survival analysis. We stratified on each person in the Medicare cohort between 2000 and 2015 who died, and examined whether PM_2.5, O₃ and NO₂ exposures predicted in which follow-up period the death occurred. We used conditional logistic regression stratified on person and controlled for nonlinear terms in calendar year and age. By design slowly varying covariates such as smoking history, BMI, diabetes and other pre-existing conditions, usual alcohol consumption, sex, race, socioeconomic status, and green space were controlled by matching each person to themselves.

Results:

There were 6,452,618 deaths in the study population in the study period. We observed a 5.37% increase in the mortality rate (95% CI 4.67%, 6.08%) for every 5 μg/m³ increase in PM_2.5, a 1.98% (95% CI 1.61%, 2.36%) increase for 5 ppb increment in O₃, and a 2.10% decrease (95% CI 1.88%, 2.33%) for a 5ppb increase in NO₂. When restricted to persons whose PM_2.5 exposure never exceeded 12 μg/m³ in any year between 2000 and 2015, the effect size increased for PM_2.5 (12.71% (11.30, 14.15)), and the signs of O₃ and NO₂ reversed (−0.26% (−0.88, 0.35) for O₃ and 1.77% increase (1.40, 2.13) for NO₂). Effect sizes were larger for Blacks (e.g. 7.71% (5.46, 10.02) for PM_2.5).

Conclusion:

There is strong evidence that the association between annual exposure to PM_2.5 and mortality is not confounded by individual or neighborhood covariates, and continues below the standard. The effects of O₃ and NO₂ are difficult to disentangle.

1. Introduction

Multiple studies have reported associations of air pollution, particularly PM_2.5, and mortality and morbidity following long and short-term exposure(Abu Awad et al. 2019; Beelen et al. 2014; Crouse et al. 2015; Di et al. 2017; Hoek et al. 2013; Pinault et al. 2016; Pope et al. 2019; Vodonos et al. 2018). These have been undertaken by investigators in multiple countries, with over 50 cohorts in a recent PM_2.5 meta-analysis(Vodonos et al. 2018). The Global Burden of Disease ranks air pollution, mostly PM_2.5 and O₃, as among the most important mortality risk factors in the world(Collaborators et al. 2015). Most pollution studies used standard methods for examining the association of exposure with outcome controlling for covariates, and reported strong associations with PM_2.5(Vodonos et al. 2018). Fewer studies have examined effects of long-term exposure to O₃ and NO₂, and results have been more mixed, with generally positive association in North America for O₃, and weaker evidence in Europe, where more NO₂ results have been reported.

The major critiques of these studies are the possibility of confounding by unmeasured predictors of exposure and outcome, particularly individual characteristics and socioeconomic status(Cox and Popken 2015), or that exposure error or small sample sizes renders the observed concentration-response relationships at low concentrations unreliable for standard setting. The most commonly mentioned omitted confounders are individual SES and detailed smoking, diet, and behavioral history. The recent meetings of EPA’s Clean Air Scientific Advisory Committee have highlighted the importance of these issues(CASAC. 2019). Meanwhile, the increasing availability of routinely collected administrative records has provided the opportunity to conduct large population-based studies, including large numbers of persons exposed at lower concentrations (Cesaroni et al. 2013; Di et al. 2017). However, since the data is not collected for specific study aims more potential confounders remain unmeasured. On the other hand, such data allows one to study entire populations rather than selected samples in cohort studies, which generally oversample healthy individuals living in urban areas and are not representative of the population. For example, both the American Cancer Society cohort and the Nurse’s Health Study cohort examined populations with higher than average educational attainment(Krewski et al. 2009; Puett et al. 2009). Other cohorts only included city dwellers(Lepeule et al. 2012).

Approaches that control for unmeasured confounders by design can address the omitted confounder issue for both administrative and traditional cohort studies. Using a self-controlled design, which stratifies on person, is one such approach, since by matching each person to themselves in a different follow-up period, it controls for constant or slowly varying individual factors such as cumulative smoking, SES, BMI, etc., and neighborhood factors such as green space, walkability etc., whether measured or not. In this study, we have used this approach to address the individual confounder issue, in a national study of the entire Medicare population, which covers smaller towns and rural areas, and has many observations at lower concentrations. A second analysis restricted to persons whose individual exposures never exceeded the National Ambient Air Quality Standard for particles during follow-up. There is no annual standard for O₃ and there are no areas in the U.S. that exceed the NO₂ ambient standard. We have included simultaneous assessment of the three major ambient air pollutants: PM_2.5, NO₂, and O₃.

This approach has similarities to the differences in differences approach which we have applied to the Northeastern and Mid-Atlantic states Medicare data(M. Yitshak-Sade et al. 2019). We have shown that difference in differences is a form of a negative outcome control(Sofer et al. 2016). In such analyses, one conditions on a geographic location and looks at changes over time in e.g. a location specific rate. The negative outcome control comes from contrasting the change over time in each location with changes over time in other locations, where the exposure contrast was different. This serves as a control for common time trends such as improvements in medical care. By conditioning on location, all confounders that vary by location are controlled, measured or unmeasured. Our approach is similar, but applies the analysis to an individual, and only analyzes individuals who died. We condition on person, not area, thereby controlling for all slowly-varying personal and small area potential confounders. We control for general time trends by controlling for year in the regression analysis and add control for individual age.

Regarding exposure error, we used exposure models that have shown excellent performance (R² on held out monitoring locations =0.89 for annual average PM_2.5, 0.84 for annual average NO₂, and 0.86 for annual O₃). Moreover, the slope between predicted and held out measured pollutant was 0.96 for PM_2.5, 0.99 for NO₂, and 0.99 for O₃. Regression calibration is a standard method for correcting health effects estimates for exposure error and involves fitting a calibration regression of exposure estimates against true measurements, and correcting the exposure-response using the slope of that calibration. Our out of sample models supply just that calibration, and the slope was essentially one for all three exposures, meaning the adjustment is already in the exposure estimates. However, we have taken as our gold standard exposure the ambient exposure in each participant’s neighborhood, and not their personal exposure. We do this for two reasons: first, because that is the regulatory exposure, and regulators want to know the association between what they regulate and health. Second, personal exposure deviates from ambient because of time spent in traffic, behavior, etc., and those factors are also associated with additional confounders such as stress, exercise, etc. which are difficult to measure and control for. Ambient exposure avoids this problem, and can serve as an instrumental variable for personal exposure(Weisskopf and Webster 2017).

We here apply this approach to estimate the effect of annual air pollution (PM_2.5, NO₂, O₃) on survival in the Medicare cohort in the contiguous U.S between 2000 and 2015.

2.0. Data and Methods

2.1. Medicare Cohort

We obtained the Medicare beneficiary denominator file, which contains information on all Medicare participants in the U.S., from the Center for Medicare and Medicaid Services(RESDAC 2018). We constructed an open cohort using all beneficiaries ≥65 years of age in the contiguous U.S. from 2000 to 2015 and examined survival of those beneficiaries as our outcome. Medicare insurance covers over 95% of the population ≥65 years of age in the United States. Medicare participants alive on January 1 of the year following their enrollment in Medicare entered the open cohort, and follow-up periods were calendar years. We restricted our analysis to participants whose age at the end of the first year of entry was 65 or 66 to avoid entering older participants (who became 65 and entered Medicare prior to 2000) into our study differentially in 2000 when pollution was higher; and to participants who died during follow-up. This restricted our analysis to 6,452,618 participants who died with 46% of those deaths among participants having PM_2.5 exposure below 12 μg/m³ in all years. This study was approved by the human subjects committee at the Harvard T. H. Chan School of Public Health.

2.2. Covariates.

From the Medicare denominator file for each calendar year, we obtained the age, sex, race, ZIP code of residence for that year, and date of death (or censoring) of each participant. Age and ZIP code were updated annually. Race and sex were self-reported at enrollment. This file is publicly available from the Centers for Medicare and Medicaid Services(RESDAC 2018).

2.3. Exposure Assessment

Mean annual exposure to air pollution for each year for each decedent was estimated at his/her residential ZIP Code for each year between 2000 and 2015. We estimated exposure using validated prediction models calibrated to measurements at almost 2000 monitoring stations using an ensemble of machine learners that provided daily estimates for a 1km grid of the contiguous U.S.(Di et al. 2019; Di et al. 2020; Requia et al. 2020) across all the years. In brief, the models used data from multiple sources including predictions of the chemical transport models GEOS-Chem, CMAQ, CAMS, and MERRA-2, meteorological data, land-use terms, and satellite-based measures of aerosol optical depth, NO₂, O₃, surface reflectance, and absorbing aerosol index. These variables were used to train a neural network, a random forest, and a gradient boosting machine to United States Environmental Protection Agency (EPA) Air Quality System monitoring data from the continental United States to generate daily predictions on a 1×1 km grid using all the data for years 2000-2015. The three predictions for each pollutant were combined in a nonlinear geographically weighted regression. The models showed good performance with ten-fold cross validation on held out monitoring sites yielding an out of sample R²’s of 0.89, 0.84 and 0.86 respectively for annual average predictions of PM_2.5, NO₂, and O₃. Penalized splines showed linear relationships between observed and predicted PM_2.5 from 0 to 60 μg/m³, between observed and predicted NO₂ from 0 to 150 ppb, and between observed and predicted O₃ from 0 to 120 ppb. Moreover, exposure error was similar at both low and high concentrations as illustrated for PM_2.5 in Figure 1. The predictions were all for the EPA targeted exposure averages: O₃ predictions were for the 8-hour maximum, NO₂ was for the 1-hour maximum, and PM_2.5 predictions were for the daily mean. Daily predictions were averaged for each year to generate annual averages. Predictions for all grid cells whose centroids were inside the Zip code boundary were averaged for each year and assigned to participants in that Zip code in that year.

Figure 1.

Annual average Predicted Vs Measured PM_2.5 (μg/m³) at Monitoring Locations in the U.S. The error around the line is similar at 5 and 12 μg/m³.

2.4. Statistical Analysis

Cox’s proportionate hazard model defines a failure time variable t_i for each subject i, in our case age. It stratifies data by time-periods with events (i.e. with non-censored failure time), and asks, conditional on an event, does exposure predict which person in the risk set at that failure time had the event. Here the risk set is all persons who survived up until failure time t. We propose an alternative, which stratifies on people who had events, and asks, conditional on an event in that person, does exposure predict in which time-period the event occurred. In this case the risk set is all the time periods the person who failed could have failed in. This is like a case-time control analysis but tailored for survival data. An advantage of the approach is that since each person is their own control, time invariant or slowly varying individual or neighborhood characteristics are controlled by matching each person to themselves, even if unmeasured.

Formally, we begin with the potential outcomes framework of the Rubin Causal Model(Rubin 1991). Let Y_it^A=a be the potential outcome in person i if exposed to A=a in year t, and let Y_it^A=a’ be the potential outcome under the alternative exposure a’. We would like to estimate E(Y_t^A=a)/E(Y_t^A=a’). We assume the potential outcome depends on predictors in the following manner:

$$\mathrm{Log}(\mathrm{E}({\mathrm{Y}}_{\text{i,t}}^{\mathrm{a}}))={\beta }_0+{\beta }_1{\mathrm{a}}_{\mathrm{t}}+{\beta }_2{\mathrm{Z}}_{\mathrm{i}}+{\beta }_3{\mathrm{W}}_{\text{i,t}}$$ [1]

where Z_i are the slowly varying location or individual level confounders, such as SES, smoking history, blood pressure, diet etc., and W_i,t the confounders that vary across time as well as subject. If we stratify on person in our analysis than it no longer involves Z, and only involves the time variation of confounders W. That is, there is no possible confounding by variables that only differ by person (e.g. sex, race, SES, smoking status, diet). If we control for year of follow-up, then all W_i that vary similarly by year for all subjects are likewise removed, measured or unmeasured. The only remaining confounders are those that vary over time differentially by person.

The likelihood for the proportionate hazard model is

$$\begin{gathered} L(\beta )={\prod }_1^n(\lambda (t\mid X_i,\beta {)}^{c_i}S(t\mid X_i,\beta )\text{where} \\ \lambda (t\mid X_i,\beta )={\lambda }_0(t)e^{{\beta }_1{X}_1+..+{\beta }_p{X}_p} \end{gathered}$$

Where i is subject, ci is an indicator for whether the failure time is observed, and t is time.

Cox proposed simplifying this by stratification. Given that an event occurred, and a risk set for that event, the conditional probability that the jth observation in the risk set is the one that failed is simply

$$\frac{e^{{\beta }_1{X}_{i1}..{\beta }_p{X}_{ip}}}{\sum _{jinriskset}{e}^{{\beta }_1{X}_{j1}..{\beta }_p{X}_{jp}}}$$

because the λ₀(t) and S(t,X,β) cancels out. And so, combining over strata, we have

$$L_p(\beta )=\prod _{i=1}^n{\left(\frac{e^{X\beta }}{\sum _{riskset}{e}^{X\beta }}\right)}^{c_i}$$

Nothing in this formulation restricts the strata to being defined by time. They merely have to contain an event, and a risk set for that event. The requirement that the sum of over all observations in the risk set of the probabilities of being the observation with the event equaling one is what is required to reduce to the partial likelihood. While it is traditional to stratify on times with events, it is certainly possible to stratify on people with events, with all observations with that person as the risk set, as we have done. Hence in our case the likelihood for each strata is:

$${}^{e^{\beta 1a_t+\beta 3W_{it}}}/_{{\sum }_t{e}^{\beta 1a_t+\beta 3W_{it}}}$$ [2]

This can be fit using conditional logistic regression, which we used to model the association of the three air pollutants with death, stratifying on each participant who died, and controlling for linear and quadratic terms for calendar year, and linear and quadratic terms for age. In addition, we controlled for time varying median household income and percent of black residents in the ZIP code to capture any time varying confounding by socio-economic factors that might change differentially over time by neighborhood.

In addition, stratifying on individual controls for exposure prior to the start of follow-up, since, for each individual, this pre-follow-up exposure was the same in each follow-up period, and therefore cannot predict in which period the death occurred. Other variables that change over time (or with age) were addressed by the nonlinear terms for calendar year and age and by median household income and percentage of residents who are black. For computational reasons, the data was randomly divided into 10 groups of participants, and models were run separately for each group, and then combined using fixed effects meta-analysis. All analyses were done in R(R_Core_Team 2013).

3.0. Results:

Of the 6,452,618 decedents, 79% were white, 55% male, and 20% of them were also eligible for Medicaid. The average PM_2.5 concentration was 10 μg/m³, the average NO₂ concentration was 18.7 ppb, and the average O₃ concentration was 38.8 ppb. Details are shown in Table 1. During the 16 years of follow-up, 46% of the participants never had annual PM_2.5 exceed 12 μg/m³ in any year.

Table 1.

Characteristics of Medicare Decedents 2000-2015. Results shown are the means over all person-years of follow-up.

Variable	25th %tile	50th %tile	75th %tile	Mean
	All Observations
Age (yrs)	68	71	74	71.1
Year of Study	2006	2009	2013	2009
Percent Female	NA	NA	NA	45
Percent on Medicaid	NA	NA	NA	20
Percent Black	NA	NA	NA	14.1
Percent Other Race	NA	NA	NA	5.5
PM2.5 (μg/m3)	8.2	9.8	11.7	10
NO2 (ppb)	10.9	16.5	24.7	18.7
O3 (ppb)	36.5	38.8	41	38.8
	Only Persons Never above 12 μg/m3
PM2.5 (μg/m3)	6.5	8.0	9.4	7.9
NO2 (ppb)	10.1	14.2	20.1	16.1
O3 (ppb)	36.4	38.7	41.4	39
Age (yrs)	67	70	73	70.5
Year of Study	2007	2010	2013	2010
Percent Female	NA	NA	NA	43.8
Percent on Medicaid	NA	NA	NA	19.2
Percent Black	NA	NA	NA	7.3
Percent Other Race	NA	NA	NA	4.7

The correlation structure of the pollutants is shown in Table 2. The correlation between annual average exposures was modest. However, when stratified by whether or not people ever had exposure exceeding 12 μg/m³, a different pattern emerged. NO₂ and O₃ were negatively associated in the higher exposure beneficiaries, but positively associated in the lower exposed group.

Table 2.

Correlation of the Annual Exposures

All Observations
PM2.5	NO2	O3
PM2.5	0.40	−0.04
NO2	1	−0.094
O3		1
Persons whose PM2.5 exposure never exceeded 12 μg/m3
PM2.5	0.244	−0.122
NO2	1	0.150
O3		1
Person whose PM2.5 exposure sometimes exceeded 12 μg/m3
PM2.5	0.387	0.030
NO2		−0.327
O3		1

By stratifying on person we eliminate potential confounders that vary between person; necessarily we also eliminate exposure variations between people, reducing power. When we compare the residual variation in air pollution after conditioning out individuals to the total variation in annual air pollution, we find the following. For PM2.5, 38%, for NO2, 16%, and for O3, 20% of the total variation remains.

In multipollutant models we observed a 5.37% increase in the mortality rate (95% CI 4.67%, 6.08%) for every 5 μg/m³ increase in PM_2.5, a 2.10% decrease in mortality rate (95% CI −1.88%, −2.33%) for each 5ppb increase in NO₂, and a 1.98% increase in mortality rate (95% CI 1.61%, 2.36%) for each 5ppb increase in O₃. Among blacks the effect size was larger (7.71%, 95% CI 10.02%, 5.46%) than for whites (4.22%, 95%CI 4.98%, 3.46% ) for PM_2.5 and O₃ (3.12%, 95% CI 4.17%, 2.08% Vs 1.88% 95% CI 2.31%, 1.45%), and more negative for NO₂ (−5.23% 95% CI −4.29, −6.15 Vs −1.80%, 95% CI −1.55%, −2.05%).

When restricted to people whose PM_2.5 exposure never exceeded 12 μg/m³ the effect size for PM_2.5 increased to 12.71% (95% CI 11.30%, 14.15%), became positive and significant for NO₂ at 1.77% (95% CI 1.40%, 2.13%), and became negative and non-significant for O₃ at −0.26% (95% CI −0.88%, +0.35%)

4.0. Discussion

To our knowledge, this is the first study to use a self-controlled design to examine the association of annual air pollution with death rates in a national sample. Our findings, which control for invariant or slowly varying individual and small area covariates (e.g. smoking history, BMI, diabetes and other pre-existing conditions, usual alcohol consumption, sex, race, socioeconomic status, green space) in the design stage, provide considerable evidence that the associations are not due to omission of any such confounders. Our findings of an association when restricted to people who never were exposed to PM_2.5 concentrations exceeding 12 μg/m³ during the follow-up period also provides strong evidence that the effect of PM_2.5 on mortality continues well below the current U.S. EPA standard. In addition, stratification on each person assured that exposure before the follow-up began cannot confound the association with more recent exposure, since it is the same in all follow-up years for that person. Hence, our results support other recent reports that indicate that the effect of PM_2.5 is not spread out over many years, but primarily reflect recent annual exposure(Abu Awad et al. 2019; Puett et al. 2009). This is important because EPA regulatory impact assessments and other risk assessments assume that the health impacts of PM_2.5 exposure are spread out over many years, and because some have argued that associations at low concentrations may reflect earlier exposure at higher concentrations.

Our findings regarding NO₂ and O₃ are less clear. In the full dataset, we find a protective effect of NO₂ and a harmful effect of O₃. However, when restricted to PM_2.5 concentrations below 12 μg/m³ the NO₂ effect becomes harmful and the O₃ effect becomes protective but not significant. This suggests that there are different patterns of confounding between NO₂ and O₃, pushing results in opposite directions, for high Vs more moderate levels of PM_2.5. This is supported by the change in sign of the correlation of annual NO₂ with O₃, which was negative in the full data set, more strongly negative in the higher exposed beneficiaries, but positive in the restricted dataset. In addition, this may derive from the smaller amount of variation in NO₂ and O₃ than in PM_2.5 after conditioning on individual, reducing power too much for stable results.

The relationship between NO₂ and O₃ is complex. NO₂ is a precursor of O₃, participating in the photochemical reactions that produce it. However, NOx quenches O₃, and hence near emission sources, a negative correlation between NO₂ and O₃ is common. Contra wise, near high VOC sources, a positive association is found. This may explain the change in sign of correlation and health effects estimates we found.

For PM_2.5 the effect size we report is similar to those from standard cohort analyses at similar concentrations, and is very close to the estimated effect size at 10 μg/m³ (the mean concentration in this study) reported in a recent meta-analysis of all of then extant cohort studies (1.29% per μg/m³ increase in PM2.5, 95% CI 1.09%, 1.50%)(Vodonos et al. 2018). Other cohort studies done at lower concentrations have reported larger effect sizes(Pinault et al. 2016). This provides considerable evidence that those studies were not confounded by unmeasured individual or neighborhood covariates that change slowly. The low exposure error and calibration slope near one in our exposure model also provides assurance that exposure error is unlikely to be biasing the association with ambient pollution.

Any remaining potential confounder can only confound if it varies over time in a way not captured by the nonlinear time trend and age terms, and not due to changes over time in median income or percent of population that is Black. Such unobserved confounders must vary with time differently for different decedents in a way that is correlated with the variation of air pollution over time differently for different decedents. Changes in SES are possible, but individual SES changes very slowly after age 65, and we controlled for changes in median income. There is little reason to believe that such changes are correlated with the change in air pollution concentrations over the same time for each person. For PM_2.5 the changes were primarily driven by the reductions caused by the Cross State Air Pollution Rule and year to year variations in the back trajectories of air masses. That rule required retrofit of SO₂ and NOx controls on some, but not all coal burning electric generating units. These gases result in secondary formation of sulfate, nitrate, and secondary organic carbon particles, which impacted Zip codes hundreds of miles downwind. These distant exposure changes affected different Zip codes differently, but in a way driven by wind patterns that are unlikely to be correlated with changes in SES, smoking, diet, etc. Ozone and NO₂ changes were also driven by regulatory changes such as the NOx controls cited above, or stricter controls on motor vehicles, which were applied uniformly across the country.

Of note, 95% of the Medicare population are non-smokers, so stratification on individual should eliminate that as a confounder.

An important attribute of this study is that we examined the entire Medicare population, including persons living in smaller cities, exurbs, small towns, and rural areas, who are less represented in many air pollution studies. This enhances the generalizability of the results.

Another important finding of this study is the larger effects on Black than on White Medicare participants (Asian and others were too small a proportion to analyze separately). These results indicate that environmental justice concerns apply not merely higher exposures, but also differential susceptibility to air pollution.

The results for O₃ and NO₂ are puzzling. In the subsample with higher PM_2.5 exposure, we see negative effect sizes for NO₂ and positive effect sizes for O₃, whereas for lower particle exposure the signs reverse, with positive effect sizes for NO₂ and negative (but insignificant) for O₃. The correlation between the two pollutants also flips sign between the two subsamples. This suggests that there is some effect of collinearity going on, with one pollutant getting a positive effect size and the other a negative, which reverses when the sign of the correlation changes. Which is the true toxic agent is not clear from this analysis.

There is considerable support for an ozone association from other studies, however. Turner et al. reported a significant effect of ozone in the ACS cohort(Turner et al. 2015). The CanCHEC study(Crouse et al. 2015), examined a cohort of 2.5 million Canadians, and reported a similar effect size on all-cause mortality. Both these effect sizes are similar to ours in the full dataset analysis. NO₂ associations have also been reported in multiple cohort studies. Faustini et al. reviewed 14 studies examining annual NO₂ exposure and all-cause mortality, controlling for particle concentrations and reported a significant positive association. The CANCHEC and ACS studies discussed above also found positive effects. They differed from this study in that we did not find all three pollutants positive predictors of mortality, and we found different results in persons with higher or lower PM_2.5 exposure. In an examination of acute effects PM_2.5 and NO₂ we reported that positivity violation was a substantial issue with NO₂, in that conditional on covariates it was almost impossible for a high NO₂ day to be a low NO₂ day(J Schwartz et al. 2018). Positivity violation is a problem for standard studies as well as ones using propensity scores, and that may be an issue for annual exposures as well.

Support for these results comes from the toxicology and controlled exposure field, which has produced hundreds of studies elucidating mechanisms and effects. These are strongest for PM_2.5, which has been shown to increase atherosclerosis and decrease the stability of atherosclerotic plaque(Araujo et al. 2008; Soares et al. 2009; Sun et al. 2005; Tzeng et al. 2007), to increase blood pressure(Bartoli et al. 2009; Chuang et al. 2017), to increase systemic inflammation and endothelial activation(Kleinman et al. 2008), to increase oxidative stress(Chuang et al. 2017; Lundback et al. 2009; Sun et al. 2008), produce pro-arrhythmic changes in ECGs(Hemmingsen et al. 2015), to worsen the response to ischemia(Wellenius et al. 2003), and impair lung clearance(Sigaud et al. 2007). Mice exposed to outdoor air with an average PM_2.5 concentrations of 16.8 μg/m³ had lower lung function than mice in the same location but with filtered air(Mauad et al. 2008), so these effects occur at concentrations near those seen in our study.

For ozone these results are supported by a cohort study reporting an association between long term ozone exposure and factor VII coagulant activity(Green et al. 2015), a chamber study reporting that ozone affected fibrinolytic activity(Kahle et al. 2015), and a toxicology study reporting that following ozone exposure, isolated coronary vessels exhibited greater basal tone, enhanced susceptibility to serotonin stimulation, and impaired response to acetylcholine(Paffett et al. 2015). A review of toxicological studies on ozone found decreased heart rate, metabolism, blood pressure, and cardiac output when rats were exposed to typical concentrations of ozone, and that these effects may be at least partially mediated via the parasympathetic nervous system.^{(Watkinson et al. 2001)}

Toxicological support for the NO₂ mortality associations is weaker. Studies of exacerbation of COPD by NO₂ have been mixed with some reporting no association(Gong et al. 2005; Morrow et al. 1992) and others reporting small increases, but only at high exposures (Linn et al. 1985; Vagaggini et al. 1996). In contrast, there is substantial toxicological support for NO₂ increasing susceptibility and response to respiratory infections(Parker et al. 1989).

With respect to cardiovascular outcomes, controlled human exposure studies have not found increases in blood pressure or changes in cardiac output following NO₂ exposure(Folinsbee et al. 1978; Huang et al. 2012). A controlled human exposure study reported an effect of 2-hour exposure to NO₂ on heart rate variability, cholesterol level, and that in combination with PM_2.5 exposure produced increased inflammatory markers in lung lining fluid not seen with either exposure alone(Huang et al. 2012).

Our study has some limitations. While our exposure models are good, they are not perfect, and do not account for personal exposure to outdoor air pollution. However, Weisskopf has pointed out that personal exposure is correlated with many potential confounders (e.g. stress while driving) that are not associated with outdoor concentration, and hence outdoor exposures can act as instrumental variables for the personal exposure(Weisskopf and Webster 2017). Taking ambient exposure in the neighborhood as the goal, our models, fit to all years, showed high R² in predicting annual PM_2.5 concentrations. We clearly were unable to disentangle the effects of NO₂ from the effects of ozone, possibly due to the small remaining variation in these pollutants after conditioning on individual. Other study designs may be more appropriate for those pollutants. Finally, our ability to control for time varying confounders was limited. General medical advances and other common trends were controlled by use of year as a proxy, however, if other variables changed differentially by Zip code in a way correlated with differential changes in air pollution by Zip code, confounding could still exist. We think such patterns are unlikely but cannot rule them out. Other studies, however, have measured time varying covariates and our goal was to ensure that unmeasured individual and neighborhood level covariates were controlled by matching.

In conclusion, the large number of existing studies reporting associations, including ones using causal modeling(JD Schwartz et al. 2018; Wang et al. 2016; Wang et al. 2017; Maayan Yitshak-Sade et al. 2019), along with the ability of this study to control for unmeasured confounders, and the strong toxicological support, indicates that the association of PM_2.5 with mortality is likely causal, and continues well below current air quality standards. In contrast, our study provides weaker support for the previously reported associations between O₃ and NO₂ and mortality, because of the contradictory findings by level of PM_2.5.

Table 3.

Percent increase in Deaths for given increases in Air Pollutants

	All Observations
Pollutant	Percent Increase in Deaths	95% Confidence Interval
PM2.5 per 5 μg/m3	5.37	4.67, 6.08
NO2 per 5 ppb	−2.10	−1.88, −2.33
03 per 5 ppb	1.98	1.61, 2.36
	Blacks
PM2.5 per 5 μg/m3	7.71	10.02, 5.46
NO2 per 5 ppb	−5.23	−4.29, −6.15
O3 per 5 ppb	3.12	4.17, 2.08
	Whites
PM2.5 per 5 μg/m3	4.22	4.98, 3.46
NO2 per 5 ppb	−1.8	−1.55, −2.05
O3 per 5 ppb	1.88	2.31, 1.45
	PM2.5 < 12 μg/m3
PM2.5 per 5 μg/m3	12.71	14.15, 11.30
NO2 per 5 ppb	1.77	2.13, 1.40
O3 per 5 ppb	−0.26	0.35, −0.88

Highlights:

• Matching people to themselves at different times controls for stable individual and neighborhood level potential confounders

• In survival analysis this provides a good way to avoid potentially omitted confounders

• We introduce this approach in survival analysis and apply it to assessing the association of multiple air pollutants with mortality in a large cohort of older adults

• We find strong evidence of an association with PM_2.5 and mixed evidence for O₃ and NO₂

• By design, these effect estimates cannot reflect each person’s exposure prior to 2000 since that is controlled by matching.

• The effect size in blacks was considerably higher than in whites.