Chronic Exposure to Fine Particulate Matter Increases Mortality Through Pathways of Metabolic and Cardiovascular Disease: Insights From a Large Mediation Analysis

Abstract

Background

Long‐term exposure to outdoor fine particulate matter (PM_2.5) is the leading environmental risk factor for premature mortality worldwide. Characterizing important pathways through which PM_2.5 increases individuals' mortality risk can clarify the PM_2.5–mortality relationship and identify possible points of interventions. Recent evidence has linked PM_2.5 to the onset of diabetes and cardiovascular disease, but to what extent these associations contribute to the effect of PM_2.5 on mortality remains poorly understood.

Methods and Results

We conducted a population‐based cohort study to investigate how the effect of PM_2.5 on nonaccidental mortality is mediated by its impacts on incident diabetes, acute myocardial infarction, and stroke. Our study population comprised ≈200 000 individuals aged 20 to 90 years who participated in population‐based health surveys in Ontario, Canada, from 1996 to 2014. Follow‐up extended until December 2017. Using causal mediation analyses with Aalen additive hazards models, we decomposed the total effect of PM_2.5 on mortality into a direct effect and several path‐specific indirect effects mediated by diabetes, each cardiovascular event, or both combined. A series of sensitivity analyses were also conducted. After adjusting for various individual‐ and neighborhood‐level covariates, we estimated that for every 1000 adults, each 10 μg/m³ increase in PM_2.5 was associated with ≈2 incident cases of diabetes, ≈1 major cardiovascular event (acute myocardial infarction and stroke combined), and ≈2 deaths annually. Among PM_2.5‐related deaths, 31.7% (95% CI, 17.2%–53.2%) were attributable to diabetes and major cardiovascular events in relation to PM_2.5. Specifically, 4.5% were explained by PM_2.5‐induced diabetes, 22.8% by PM_2.5‐induced major cardiovascular events, and 4.5% through their interaction.

Conclusions

This study suggests that a significant portion of the estimated effect of long‐term exposure to PM_2.5 on deaths can be attributed to its effect on diabetes and cardiovascular diseases, highlighting the significance of PM_2.5 on deteriorating cardiovascular health. Our findings should raise awareness among professionals that improving metabolic and cardiovascular health may reduce mortality burden in areas with higher exposure to air pollution.

Nonstandard Abbreviations and Acronyms

NDE

natural direct effect

NIE

natural indirect effect

PM_2.5

fine particulate matter

Clinical Perspective.

What Is New?

• This is the first causal mediation study to date that empirically assessed pathways from long‐term exposure to outdoor fine particulate matter to nonaccidental deaths through the incidence of diabetes and cardiovascular diseases and thereby quantified the relative contributions of these potential mediators.

• A third of the estimated adverse effects of exposure to fine particulate matter on mortality operates through its influence on the development of diabetes and cardiovascular events.

What Are the Clinical Implications?

• Beyond interventions to reduce exposure to fine particulate matter, there is an opportunity in clinical practice to help mitigate the adverse effects of fine particulate matter before they become overwhelming and irreversible. This can be achieved by improving metabolic and cardiovascular health in the polluted areas.

Outdoor fine particulate matter (PM_2.5) is a leading global health concern. ¹ A recent international study estimated that 8.9 million deaths globally may be attributable to outdoor PM_2.5 in 2015. ² During the past decade, PM_2.5 exposure has also been increasingly linked to the incidence of diabetes ³ , ⁴ , ⁵ and major cardiovascular events including acute myocardial infarction (AMI) ⁶ , ⁷ and stroke. ⁸ , ⁹ , ¹⁰ For example, a recent multicountry cohort study in Europe has shown that even lower levels of air pollution (ie, concentrations lower than the existing World Health Organization guideline limits) were associated with increased incidences of stroke and coronary heart disease. ¹⁰ The biological mechanisms underlying the PM_2.5 effects on these incident outcomes and mortality were hypothesized to include aggravating oxidative stress and inflammation, altering endothelial function, inducing insulin resistance, increasing blood coagulability, accelerating atherosclerosis progression, impairing autonomic balance, and increasing sympathetic tone, all of which can precipitate the development of metabolic diseases such as diabetes and cardiovascular events, ultimately increasing the risk of premature death. ¹¹ , ¹² Despite the hypothesized mechanisms, there is a dearth of empirical evidence about the extent to which the effect of PM_2.5 on deaths is mediated by its effect on the development of these conditions (ie, the proportion mediated). In the face of continuing mortality burden attributed to PM_2.5, it is crucial to elucidate the mechanistic structure that may provide valuable insight on the reduction of PM_2.5‐related mortality (eg, intervention on mediators when it is not possible to intervene on the exposure in a timely manner). ¹³

Causal mediation analysis is a valuable tool for epidemiologic research to assess how an exposure can affect an outcome of interest through ≥1 intermediate variables (ie, mediators) on complex pathways. ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ It reveals new insight toward understanding disease pathogenesis using observational data when large clinical trials to deduce mechanistic pathways are not feasible. Such information can greatly improve our knowledge of the health effects of an exposure by allowing us to suggest and test hypotheses about underlying mechanisms, which will help prioritize intervention targets. Furthermore, recent developments in statistical methods for the mediation analysis allow to study more complex causal pathways, such as when multiple mediators affect each other or when interactions are present. ¹⁵ , ¹⁸ However, mediation analysis has not been widely applied in air‐health studies, especially for such complex causal structures. So far, relatively few of these studies have evaluated mediation effects in relation to air pollution. None of them has considered mediation by different diseases and disease states nor their interactions, ¹⁹ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ , ²⁵ therefore it remains elusive to prioritize each potential pathway. Given the pervasive air pollution exposure and the dynamic nature of cardiometabolic health during the life course, it is crucial to determine how air pollution shapes individuals' health trajectories from the onset of disease through its progression and eventually to death. This, however, has not been addressed by previous mediation analyses. ¹⁹ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ , ²⁵

We conducted a large, population‐based mediation analysis to investigate the onset of diabetes and cardiovascular events as potential mediating pathways between long‐term exposure to PM_2.5 and mortality. Specifically, we estimated the extent to which the effect of PM_2.5 on nonaccidental deaths was mediated by the following 3 sets of sequential time‐varying mediators: (1) incidence of diabetes and AMI, (2) incidence of diabetes and stroke, and (3) incidence of diabetes and major cardiovascular events (AMI and stroke combined). Our analysis improved on existing work by using causal mediation approaches that allow for causally ordered mediators to be a time‐varying feature in the context of a time‐to‐event outcome. In addition, the use of population‐based health surveys with rich data on key risk factors offers unique opportunities to elucidate these knowledge gaps.

METHODS

The data set from this study is held securely in coded form at ICES. Although legal data‐sharing agreements between ICES and data providers (eg, health care organizations and government) prohibit ICES from making the data set publicly available, access may be granted to those who meet prespecified criteria for confidential access, available at https://www.ices.on.ca/DAS (email: das@ices.on.ca).

Study Participants

In this retrospective, population‐based cohort study, our population was derived from the 1996 to 1997 cycle of the National Population Health Survey ²⁶ and the 2000 to 2001, 2003, 2005, 2007 to 2008, 2009 to 2010, 2011 to 2012, and 2013 to 2014 cycles of the Canadian Community Health Survey (Figure 1; n=254 965 individuals). ²⁷ Details about the methodology of these surveys have been described elsewhere (see Table S1 for time periods of data collection). ²⁶ , ²⁷ Briefly, the National Population Health Survey and Canadian Community Health Surveys are nationally representative surveys that collect self‐reported data related to health status and determinants of health (eg, sociodemographic factors) from a representative sample of people aged ≥12 years who are living in private dwellings in Canada.

Figure 1. Flowchart of participant selection.

AMI indicates acute myocardial infarction; CCHS, Canadian Community Health Surveys; and NPHS, National Population Health Survey.

We included the respondents who, at the time of survey, resided in the province of Ontario, were aged 20 to 90 years, were not pregnant, were eligible for Ontario's provincial health insurance plan, and provided informed consent to share and link their responses to provincial health administrative data (n=218 158). Eligible respondents were followed up from the time of survey and censored when reaching the end of follow‐up (December 31, 2017), becoming ineligible for provincial health insurance, or death.

The use of the data in this study is authorized under section 45 of Ontario's Personal Health Information Protection Act and does not require review by a research ethics board.

Outcome

Consistent with previous studies, ² we considered nonaccidental death as our outcome. Using unique encoded identifiers (encrypted health card numbers), we anonymously linked the National Population Health Survey and Canadian Community Health Survey respondents to Ontario health administrative databases. We ascertained nonaccidental deaths from the provincial death registry, Ontario's Registrar General Death File (see Data S1). The nonaccidental death was determined based on the underlying cause of death listed on the death certificates.

PM_{2 .5} Exposure Assessment

We estimated ground‐level PM_2.5 concentrations by relating satellite retrievals of aerosol optical depth to PM_2.5 using a global atmospheric chemistry transport model combined with a geographically weighted regression model. ²⁸ The PM_2.5 estimates were available on a grid with a spatial resolution of ≈1 km×1 km for each year between 2000 and 2016. Briefly, satellite retrievals of aerosol optical depth were related to near‐surface PM_2.5 concentrations using the geophysical relationship predicted by a chemical transport model and subsequently calibrated via geographically weighted regression. These results were validated against ground PM_2.5 concentrations measured at fixed monitoring stations (n=2312) across North America and showed good cross‐validated performance (R ² =0.70). Recent methodological modifications were made (eg, treating topographical changes and urban land cover as separate predictors), which further improved the performance (R ² =0.73).

Because the PM_2.5 data are only available for 2000 to 2016, we conducted annual calibration of the surfaces to relevant time periods during the study, similar to previous studies. ²⁹ , ³⁰ Briefly, we rescaled the annual estimate of PM_2.5 in 2000 to years from 1991 to 1999 by taking the ratio of the 2000 surface to the average concentration from 1991 to 1999 of PM_2.5 at all fixed‐site monitors across Ontario. We also estimated the concentrations of PM_2.5 in 2017 by scaling the data in 2016. We were thus able to assign annual estimates of PM_2.5 exposure from 1991 to 2017 to the centroid of each respondent's residential area determined by the annual 6‐character postal code in that year, therefore accounting for residential mobility and long‐term trends in exposure. In Canada, a 6‐character postal code represents a block face or a large building in urban areas, but much larger areas in rural areas. For each year of follow‐up of a participant, we estimated a 5‐year moving window of past exposures to PM_2.5 with a 1‐year lag, as was done in previous studies. ³¹ For example, a respondent's moving window of exposures for 2001 would be computed as the mean concentrations from 1996 to 2000.

Mediators

The mediators of interest in this study are the incidence of diabetes, AMI, and stroke and a composite of the incidence of AMI and stroke (referred to as major cardiovascular events). Using algorithms previously validated against patient charts, we ascertained the incidence of these conditions during follow‐up based on health administrative databases in Ontario including hospital discharge abstracts from the Canadian Institute for Health Information (see Data S2), physician service claims from the Ontario Health Insurance Plan database, or claims for prescription drugs from the Ontario Drug Benefit database. An incident case of diabetes was defined as an individual with ≥2 physician claims with a diabetes diagnostic code (250), ≥1 drug claim for diabetes, or ≥1 hospitalization for diabetes within 1 year (International Classification of Diseases, Ninth Revision [ICD‐9] code 250 and International Classification of Diseases, Tenth Revision [ICD‐10] codes E10–E14). This validated algorithm has a high sensitivity (90.0%) and specificity (97.7%). ³² Incident AMI was defined as having ≥1 hospitalization with AMI and having had no hospitalization for AMI in the previous 1 year (ICD‐9 code 410 and ICD‐10 code I21). A previous study has shown high accuracy of coding of AMI, with a sensitivity of 89% and a specificity of 93%. ³³ To ascertain incident stroke, we used an algorithm of ≥2 physician claims with diagnostic codes for stroke (436, 432) or transient ischemic attack (435) or ≥1 hospitalization for stroke or transient ischemic attack within 1 year (ICD‐9 codes 362.3, 430, 431, 434.x, 436, and 435.x and ICD‐10 codes I60.x, I61.x, I63.x [excluding I63.6 cerebral infarction attributed to central venous thrombosis], I64, H34.1, G45.x [excluding G45.4 transient global amnesia], and H34.0). ³⁴ This algorithm for the stroke definition has been found to have a sensitivity of 85% and a specificity of 97%. ³⁴ The databases have been also used to identify people who were diagnosed with diabetes, AMI, or stroke before the survey date.

Covariates

Covariates were identified as potential confounders on the basis of a priori assumptions of their relationships with exposure (PM_2.5) and the mediators (diabetes, AMI, or stroke) and outcome (nonaccidental death) under investigation. ³⁵ , ³⁶ , ³⁷ We considered the following 3 types of confounders: exposure–outcome, exposure–mediators and mediators–outcome. ¹⁵ The covariates were considered at both individual‐ and neighborhood area–levels. Individual‐level covariates were identified at baseline only. Sociodemographic characteristics included age, sex, marital status (married or common law, never married, separated, widowed, or divorced), ethnicity (White/non‐White race [including Black, Korean, Filipino, Japanese, Chinese, South Asian, South East Asian, Arab, West Asian, Latin American, and others]), nativity (immigrants/nonimmigrants), education (less than high school, high school, some postsecondary, and postsecondary graduation), and household income level (≤$29 999, $30 000–$79 999, and ≥$80 000). Lifestyle factors included smoking status (never smoker, daily smoker, occasional smoker, always occasional smoker, former daily smoker, and former occasional smoker), smoking pack years (available in the Canadian Community Health Survey only), body mass index, and physical activity (based on energy expenditure).

We also created a series of time‐varying contextual variables at census division and dissemination levels, respectively, by linking the survey data to the 1996, 2001, 2006, and 2016 Canadian Census data based on the nearest census year: percentage of the population aged ≥15 years with less than a high school education, percentage of the population aged ≥15 years who are recent immigrants, income quintile (a measure of relative household income accounting for household size and community), and unemployment rate. ³⁸ In addition, we created 2 time‐varying geographic indicators of residence (ie, northern/southern Ontario and urban/rural areas) based on the participants' yearly postal code information to control for possible regional differences in the outcome and identification of mediators. Furthermore, we included survey cycles as a covariate.

We obtained information on several additional covariates at the individual‐level such as alcohol consumption (regular, occasional, former, and never drinkers), working status (having a job last week, no job last week, and unable/permanent disability), and sense of belonging to local community (very strong, somewhat strong, somewhat weak, and very weak). A summary of all covariates is provided in Table S2.

All data sets were linked using unique encoded identifiers and analyzed at ICES. ICES is an independent, nonprofit research institute whose legal status under Ontario's health information privacy law allows it to collect and analyze health care and demographic data without consent for health system evaluation and improvement.

Statistical Analysis

We conducted a causal mediation analysis to quantify to what extent the estimated effect of PM_2.5 on nonaccidental deaths might be mediated through its effect on 2 causally related intermediates (and their interaction) during the follow‐up (Figure 2). We excluded participants with missing exposure data at baseline (Figure 1; n=3643) (see Data S3). To capture the first‐ever incidences of mediators during follow‐up, which would facilitate temporality, we excluded respondents who had previous physician diagnoses of diabetes, AMI, and/or stroke at the time of entry (see Data S4). For example, for the analysis with incidence of diabetes and AMI as mediators, respondents with prior diagnoses of diabetes (n=22 394) and AMI (n=3051) were excluded by definition (Figure 1). In addition, we a priori considered the incidence of diabetes as the first mediator and the incidence of AMI, stroke, or major cardiovascular events (ie, AMI and stroke combined) as the second mediator, respectively, because diabetes is known to increase the risk for cardiovascular diseases. ³⁹ Thus, we excluded a small number of respondents who had the second mediator (ie, AMI and/or stroke) diagnosed before the diagnosis of the first mediator (ie, diabetes). This led to an exclusion of 351 (≈0.2%) individuals in the analysis where diabetes and AMI were 2 mediators, for example.

Figure 2. A diagram illustrating the relationships between exposure (PM_2.5), the first mediator, the second mediator, covariates, and outcome (nonaccidental deaths).

The first mediator was incidence of diabetes. The second mediator was incidence of AMI, stroke, or cardiovascular events (AMI and stroke combined). The confounders included individual‐level risk factors at baseline and time‐varying area‐level factors. AMI indicates acute myocardial infarction; and PM_2.5, fine particulate matter.

Although the first mediator would causally affect the second mediator, it is also plausible that there could be interactions between the 2 mediators on the mediated effect of PM_2.5 on mortality. Therefore, for each pair of causally ordered mediators, we created a time‐varying joint mediator (a categorical variable) (eg, 1, free of both diabetes and AMI diagnoses during follow‐up; 2, developed diabetes only; 3, developed AMI only; and 4, developed both diabetes and AMI). This joint‐mediator approach allows us to assess the individual mediation effects as well as their joint effect, which in turn helps elucidate the relative contributions of the mediators (eg, developed either diabetes or AMI) and their interaction (eg, developed both diabetes and AMI). ¹⁸ In addition, this approach ensured that none of the confounders for the second mediator–mortality association was affected by the exposure (eg, diabetes).

For the mediation analysis, we applied the method for estimating causal effects described by VanderWeele et al. ¹⁵ , ¹⁸ With the identifiability assumptions (eg, no confounding for the exposure–outcome, exposure–mediator, and mediator–outcome relationships), this approach allowed us to decompose the total effect of PM_2.5 on deaths into the natural direct effect (NDE) (ie, the estimated effect of PM_2.5 on deaths via pathways that do not involve mediators) and natural indirect effects (NIEs) (ie, the estimated effect of PM_2.5 on deaths attributed to the effect of PM_2.5 on the mediators) (Figure S1). The NIEs were measured through the effect mediated through the first mediator alone, the second mediator alone, and the interactive effect between the 2 mediators. The proportion mediated was calculated as the ratio between the NIE and the total effect.

To implement this approach, we first evaluated the associations between the exposure, the outcome, and the mediators using Aalen additive hazards models (see Data S5 and Data S6), which has been increasingly used in recent mediation studies. ¹⁶ , ⁴⁰ An advantage of using Aalen additive hazards models is that they can be used with relatively common mediators and outcomes, unlike Cox proportional hazards models. ⁴⁰ In addition, the Aalen additive hazards models give rise to effect measures on the additive scale rate (ie, rate difference), which is more interpretable from the public health perspective. In all models, we included exposure to PM_2.5 as a time‐varying variable (continuous) and adjusted for the selected individual‐level covariates measured at baseline (ie, age, sex, marital status, education, immigration status, household income adequacy, smoking status, smoking pack years, physical activity, and body mass index) and time‐varying area‐level variables (ie, education, income, percentage of unemployment, percentage of immigrants, indicators for rural/urban and north/south). Then, using the coefficients derived from these models, we calculated NDE, NIEs, and the proportions mediated based on the product coefficient method. ¹⁵ We performed bootstrapping based on 500 replications to derive CIs for NDE, NIEs, and the proportions mediated. We reported percentile‐based CIs as demonstrated previously. ⁴¹

Sensitivity analyses were performed to test the robustness of these estimates by (1) further incrementally adjusting for additional individual risk factors, including alcohol consumption, working status, and sense of belonging to local community; (2) using the restricted cubic splines of age in the models to allow for the nonlinear effect of age, ⁴² and (3) further adjusting for a linear term for time to account for potential changes in the diagnosis and disease incidence over time. Because of computational constraints, we only calculated the point estimates of the proportion mediated in these sensitivity analyses.

We also conducted an additional sensitivity analysis to estimate the proportion mediated using an alternative approach developed by Aalen et al. ⁴³ With the framework of dynamic path analysis, Aalen et al combines the g‐formula (ie, a causal inference technique) ⁴⁴ with the additive hazards model and a sequential linear model for the mediator process. It also allows for multiple mediators to be a time‐varying feature in the context of survival analysis. Using this method, we derived the proportion mediated by calculating the ratio of the mean cumulative indirect and total effects. Because this method is computationally demanding, we were unable to perform bootstrapping to derive CIs for the point estimates (each bootstrapping replication takes ≈3 days with our large cohort size).

Lastly, to make a comparison with previous studies, we estimated the association between the long‐term exposure to PM_2.5 and nonaccidental deaths using the Cox proportional hazards model with adjustments for all covariates and mediators. To investigate the potential impact of the exclusion of respondents with prior diagnoses of diabetes and cardiovascular events, we also estimated the association between PM_2.5 and nonaccidental deaths in the cohort with these respondents added back using the Aalen additive model and Cox proportional hazards model without mediators, respectively.

The study cohorts were created using SAS version 9.4 (SAS Institute, Cary, NC); all analyses were done using RStudio and the timereg package.

RESULTS

Participant Characteristics

A total of 192 362 participants were included in the analytical cohort with incidence of diabetes and AMI considered as mediators (referred to as the diabetes–AMI cohort); 190 052 participants were included the analytical cohort with incidence of diabetes and stroke considered as mediators (referred to as the diabetes–stroke cohort), and 187 809 participants were included in the analytical cohort with incidence of diabetes and cardiovascular events (AMI and stroke combined) considered as mediators (referred to as the diabetes–cardiovascular event cohort) (Figure 1). Table 1 summarizes characteristics of the participants in the diabetes–cardiovascular event cohort (characteristics of the participants in the diabetes–AMI and diabetes–stroke cohorts are included in Table S3). For the diabetes–cardiovascular event cohort, the total follow‐up time is 2016152.6 person‐years. The annual average mortality was ≈10 per 1000 people. For all 3 cohorts, the 5‐year moving average exposure to PM_2.5 at respondents' baseline residences was 8.5 μg/m³ (SD, 2.47 μg/m³). There had been a decreasing trend in PM_2.5 during the study period.

Table 1.

Demographic Characteristics of the Study Cohort With Incidence of Diabetes and Cardiovascular Events (AMI and Stroke Combined) Analyzed as Mediators

	Diabetes–cardiovascular events cohort
Characteristics	All participants	Participants who were diagnosed with diabetes during follow‐up	Participants who were diagnosed with AMI during follow‐up	Participants who died as a result of nonaccidental causes during follow‐up
No.	187 048	17 630	11 473	19 267
PM2.5 exposure at entry	8.45±2.47	8.70±2.67	8.67±2.76	8.71±2.79
Individual‐level risk factors
Age at entry, y	49.66±17.53	54.80±14.42	64.79±14.30	69.46±12.9
Male sex	44.7	49.0	47.9	54.2
Immigrant	20.2	25.1	23.2	22.6
Non‐White race*	11.4	13.1	5.8	4.3
Living in an urban area	86.3	74.0	84.2	72.3
Living in southern region	74.1	84.9	72.7	86.4
Marital status
Married or common law	59.6	61.7	55.7	48.5
Single (never married)	19.4	13.1	8.8	8.7
Separate, widowed, or divorced	21.0	25.2	35.6	42.9
Education
Less than high school	16.2	24.8	31.7	37.7
High school graduation	19.5	19.2	18.2	18.3
Some form of postsecondary education	8.0	8.4	7.6	7.5
Postsecondary graduation	56.3	47.6	42.5	36.6
Family income
≤$29 999	19.9	26.5	32.7	41.0
$30 000 to $79 999	43.2	44.6	41.4	37.4
≥$80 000	29.8	19.7	15.0	9.1
Missing	7.1	9.1	10.9	12.6
Smoking (the NPHS)
Never smoker	40.6	37.4	39.3	35.2
Daily smoker	25.0	24.9	23.2	25.1
Occasional smoker	2.0	1.9	1.5	1.7
Always occasional smoker	1.5	1.6	0.9	1.0
Former daily smoker	24.0	28.7	29.8	31.4
Former occasional smoker	6.8	5.6	5.3	5.6
Smoking (the CCHS)
Never smoker	34.0	30.9	29.0	26.9
Current smoker of <10 pack‐year	8.5	4.9	2.8	2.1
Current smoker of 10 to 20 pack‐year	5.2	4.6	4.6	3.5
Current smoker of ≥20 pack‐year	9.1	12.8	14.8	17.5
Current smoker with missing pack‐year	0.9	1.0	0.0	1.2
Former smoker who quit within 5 y	6.8	7.5	5.1	6.5
Former smoker who quit >5 y ago	22.2	27.3	32.2	33.2
Missing	13.4	11.1	10.5	9.0
BMI, kg/m2
18.5 to 25.0	43.1	18.6	35.8	37.0
25.0 to 30.0	33.8	36.5	34.5	29.1
30.0 to 35.0	11.9	24.0	12.1	10.1
≥35.0	4.3	12.9	3.7	3.8
<18.5	2.1	0.6	1.8	3.2
Missing	4.8	7.4	12.1	16.7
Physical activity
Active	24.5	18.2	19.4	16.1
Moderate active	25.6	23.0	23.7	20.4
Inactive	49.9	58.7	56.8	63.5
Area‐level risk factor†
Percentage of recent immigrants	2.56±2.92	2.56±2.99	2.18±2.69	2.17±2.68
Percentage of the population aged ≥15 y without employment	6.98±1.44	7.00±1.42	7.05±1.44	7.04±1.42
Percentage of the population aged ≥15 y with less than a high school education	26.82±4.43	26.77±4.36	27.26±4.10	27.25±4.06
Income quintiles
Lowest	18.3	21.8	22.4	24.6
Lower	19.6	21.1	20.9	21.4
Middle	19.8	20.6	19.6	19.2
Upper	22.2	19.9	19.4	18.5
Uppermost	20.0	16.6	17.7	16.3

Data are provided as number, mean±SD, or percentage. AMI indicates acute myocardial infarction; BMI, body mass index; CCHS, Canadian Community Health Surveys; NPHS, National Population Health Survey; and PM_2.5, fine particulate matter.

^* Non‐white includes Black, Korean, Filipino, Japanese, Chinese, South Asian, South East Asian, Arab, West Asian, Latin American, and others.

^† From Canadian Census 2001, at the dissemination‐area level.

At baseline, participant characteristics were broadly similar across the 3 cohorts: the mean ages were ≈50 years, 45% were men, 20% were immigrants, 89% were White race, 86% were urban residents, and 16% had less than a high school diploma. Of the participants, ≈23% to 28% were smokers depending on the surveys. The neighborhood‐level socioeconomic status of the participants was also similar across the cohorts (Table 1). During follow‐up, roughly 10% of participants were diagnosed with diabetes, ≈2% developed AMI, ≈5% had a stroke, and ≈11% died from nonaccidental causes. Overall, compared with the entire cohorts, participants who were diagnosed with any of the 3 conditions or who died during follow‐up were older and were more likely to be immigrants, have low education, have a history of smoking, and live in the lowest income neighborhoods.

Main Analysis

The associations between PM_2.5, mediators, and nonaccidental deaths estimated by rate differences using Aalen additive hazard models are presented in Table 2. After adjusting for all covariates, every 10 μg/m³ increase in exposure to PM_2.5 was associated with 1.51 to 1.76 excess deaths for every 1000 person‐years, depending on the analytical cohort investigated.

Table 2.

Rate Differences and 95% CIs for the Associations Between PM_2.5, Nonaccidental Deaths, and Selected Mediators

Model*	Rate difference (95% CI)†
Mediators: incidence of diabetes and AMI
Outcome model
Association between PM2.5 and deaths	1.76 (0.88 to 2.64)
Association between incident diabetes and deaths	4.32 (3.39 to 5.25)
Association between incident AMI and deaths	40.04 (35.93 to 44.15)
Association of the interaction between diabetes and AMI with deaths	59.53 (45.67 to 73.40)
Mediator model
Association between PM2.5 and incident diabetes	1.95 (1.06 to 2.84)
Association between PM2.5 and incident AMI	0.55 (0.18 to 0.93)
Association between PM2.5 and the interaction between diabetes and AMI	0.10 (−0.05 to 0.24)
Mediators: incidence of diabetes and stroke
Outcome model
Association between PM2.5 and deaths	1.61 (0.75 to 2.47)
Association between incident diabetes and deaths	4.67 (3.76 to 5.58)
Association between incident stroke and deaths	54.02 (51.10 to 57.31)
Association of the interaction between diabetes and stroke with deaths	59.50 (49.28 to 69.35)
Mediator model
Association between PM2.5 and incident diabetes	2.12 (1.23 to 3.02)
Association between PM2.5 and incident stroke	0.51 (−0.07 to 1.10)
Association between PM2.5 and the interaction between diabetes and stroke	0.16 (−0.03 to 0.34)
Mediators: incidence of diabetes and cardiovascular events‡
Outcome model
Association between PM2.5 and deaths	1.51 (0.65 to 2.37)
Association between incident diabetes and deaths	4.33 (3.44 to 5.22)
Association between incident cardiovascular events and deaths	47.23 (44.90 to 49.64)
Association of the interaction between diabetes and cardiovascular events with deaths	53.40 (45.26 to 61.69)
Mediator model
Association between PM2.5 and incident diabetes	1.98 (1.09 to 2.87)
Association between PM2.5 and incident cardiovascular events	0.97 (0.29 to 1.65)
Association between PM2.5 and the interaction between diabetes and cardiovascular events	0.16 (−0.06 to 0.38)

AMI indicates acute myocardial infarction; and PM_2.5, fine particulate matter.

^*Aalen additive hazards models adjusted for the selected individual‐level covariates measured at baseline (ie, age, sex, marital status, education, immigration status, household income adequacy, smoking status, smoking pack years, type of drinker, daily consumption of total fruits and vegetables, physical activity, and body mass index) and time‐varying area‐level variables (ie, education, income, percentage of unemployment, percentage of immigrants, indicators for rural/urban and north/south).

^† The effects are presented as number of additional cases per 1000 person‐years for each 10 μg/m³ increase in PM_2.5.

^‡ A composite indicator of incident acute myocardial infarction and stroke (whichever occurred first).

As expected, having diabetes or cardiovascular events (AMI or stroke) was associated with an increased risk for death (Table 2). For every 1000 person‐years, the impact on mortality estimated by the rate difference is 4.33 (95% CI, 3.44–5.22) and 47.23 (95% CI, 44.90–49.64) for incident diabetes and cardiovascular events, respectively. Patients who were diagnosed with both of diabetes and cardiovascular events during follow‐up were at the highest risk (53.40 [95% CI, 45.26–61.69] excess deaths per 1000 person‐years). These estimates were in dependent of the impact of PM_2.5 on death.

In the models for the associations between PM_2.5 and mediators, we found that every 10 μg/m³ increase in exposure to PM_2.5 was associated with ≈2 excess cases of diabetes, 0.5 excess cases of AMI, 0.5 excess cases of stroke, and 1 excess case of combined cardiovascular events for 1000 person‐years, adjusting for all covariates (Table 2).

Table 3 shows the NDE, NIEs through mediators, total effect of long‐term exposure to PM_2.5 on nonaccidental deaths, and the proportion mediated. In the analysis that decomposed the PM_2.5–mortality association via a pathway involving the incidence of diabetes and AMI, we estimated that the proportion mediated through incident diabetes and AMI jointly was 18.6% (95% CI, 8.8%–32.1%): 4.2% (95% CI, 1.9%–7.3%) through diabetes alone, 11.5% (95% CI, 4.0%–22.1%) through AMI alone, and 2.9% (95% CI, −1.7% to 9.1%) through the interaction of diabetes and AMI. When considering the incidence of diabetes and stroke as mediators, 24.0% (95% CI, 6.6%–43.7%) of the estimated effect of PM_2.5 could be attributed to the pathways through these 2 mediators jointly: 5.3% (95% CI, 2.5%–10.4%) through diabetes alone, 13.9% (95% CI, −2.1% to 30.7%) through stroke alone, and 4.8% (95% CI, −1.0% to 11.2%) through the interaction of diabetes and stroke. Lastly, we found that the pathways through the incidence of diabetes and cardiovascular events (AMI and stroke combined) explained 31.7% (95% CI, 17.2%–53.2%) of the total effect of PM_2.5: 4.5% (95% CI, 1.9%–8.4%) through diabetes alone, 22.8% (95% CI, 9.4%–40.9%) through cardiovascular events alone, and 4.5% (95% CI, −1.5% to 11.7%) through the interaction of diabetes and cardiovascular events.

Table 3.

Estimated Total Effects, Natural Direct Effects, and Natural Indirect Effects and Proportion Mediated Through Selected Mediators for the Association Between Long‐Term Exposure to PM_2.5 and Nonaccidental Deaths

	Main analysis
Path‐specific effect*	Estimate† (95% CI)	Percentage of effect‡ (95% CI)
Mediators: incidence of diabetes and AMI
Direct effect§	18.6 (9.25 to 27.0)	81.4 (67.9 to 91.2)
Total indirect effect||	4.0 (2.2 to 6.2)	18.6 (8.8 to 32.1)
Indirect effect via diabetes¶	0.9 (0.5 to 1.4)	4.2 (1.9 to 7.3)
Indirect effect via AMI#	2.5 (0.9 to 4.2)	11.5 (4.0 to 22.1)
Indirect effect via the interaction between diabetes and AMI**	0.6 (−0.4 to 1.7)	2.9 (−1.7 to 9.1)
Total effect	22.6 (13.4 to 31.8)	…
Mediators: incidence of diabetes and stroke
Direct effect	16.7 (8.3 to 25.7)	76.0 (56.3 to 93.4)
Total indirect effect	5.1 (1.2 to 9.1)	24.0 (6.6 to 43.7)
Indirect effect via diabetes	1.1 (0.6 to 1.6)	5.3 (2.5 to 10.4)
Indirect effect via stroke	3.0 (−0.4 to 6.4)	13.9 (−2.1 to 30.7)
Indirect effect via the interaction between diabetes and stroke	1.0 (−0.2 to 2.1)	4.8 (−1.0 to 11.2)
Total effect	21.8 (12.5 to 31.6)	…
Mediators: incidence of diabetes and cardiovascular events††
Direct effect	15.7 (6.2 to 24.9)	68.3 (46.8 to 82.8)
Total indirect effect	7.0 (3.3 to 10.8)	31.7 (17.2 to 53.2)
Indirect effect via diabetes	1.0 (0.5 to 1.4)	4.5 (1.9 to 8.4)
Indirect effect via cardiovascular events	5.1 (1.7 to 8.6)	22.8 (9.4 to 40.9)
Indirect effect via the interaction between diabetes and cardiovascular events	1.0 (−0.3 to 2.3)	4.5 (−1.5 to 11.7)
Total effect	22.7 (11.9 to 33.2)	…

AMI indicates acute myocardial infarction; and PM_2.5, fine particulate matter.

^*Adjusted for the selected individual‐level covariates measured at baseline (ie, age, sex, marital status, education, immigration status, household income adequacy, smoking status, smoking pack years, type of drinker, daily consumption of total fruits and vegetables, physical activity, and body mass index) and time‐varying area‐level variables (ie, education, income, percentage of unemployment, percentage of immigrants, indicators for rural/urban and north/south). These effects were estimated using Aalen additive hazards models and scaled by the mean number of follow‐up days (3796 days).

^† These effects are presented as number of additional cases per 1000 person‐years for every 10 μg/m³ increase in PM_2.5 exposure.

^‡ The proportions of direct effects and indirect effects through selected mediators.

^§ The effect of PM_2.5 on nonaccidental deaths that was not mediated by selected mediators.

^|| The total effect of PM_2.5 on nonaccidental deaths mediated by selected mediators.

^¶ The effect mediated through the first mediator (ie, incidence of diabetes) alone.

^# The effect mediated through the second mediator (ie, incidence of AMI, stroke, or combined) alone.

^**The effect mediated through the interaction between the 2 mediators.

^†† A composite indicator of incident AMI and stroke (whichever occurred first).

Sensitivity Analysis

In the sensitivity analyses, the proportions mediated overall remained consistent with those observed in the main analysis (Table S4). For example, the sensitivity analyses with incidence of diabetes and cardiovascular events evaluated as mediators showed that the proportions mediated by these 2 conditions ranged from 30% to 34%. However, the proportion mediated by diabetes was slightly attenuated with the further adjustment for a linear term for time.

Using the approach of dynamic path analysis, ⁴³ the total effect, NDE, and NIEs through mediators (measured by mean cumulative effect estimates over follow‐up) were overall consistent with the effect estimates derived from the primary analysis (Table S4). Despite some differences in indirect effects via individual mediators, the proportions of total indirect effects through mediators in all 3 cohorts were similar to and slightly higher than those estimated using the primary approach. For example, the total proportion mediated by diabetes and cardiovascular events was 38% using this approach compared with 32% using the primary approach.

Using the Cox proportional hazards model, we estimated that the hazard ratio for nonaccidental deaths corresponding to each 10 μg/m³ increase in PM_2.5 exposure was 1.13 (95% CI, 1.05–1.22). The association between PM_2.5 and deaths derived based on the cohort with people with a prior history of diabetes and cardiovascular events added back were slightly strengthened (Table S5).

DISCUSSION

In this population‐based cohort study, we observed a substantial increased risk of nonaccidental deaths in association with long‐term exposure to PM_2.5. In addition, PM_2.5 was positively associated with the incidence of diabetes, AMI, and stroke, 3 major risk factors for premature mortality. Importantly, we found that incidence of diabetes, AMI, and stroke mediated the relationship between PM_2.5 and nonaccidental deaths, collectively accounting for roughly a third of the total effect of PM_2.5. Outdoor PM_2.5 has been recognized as the leading environmental risk factor for premature deaths worldwide. ¹ , ² Our findings demonstrate that a significant fraction of the estimated adverse effect of PM_2.5 on mortality operates through its influence on the development of diabetes and cardiovascular events.

There is consistent evidence that people with higher exposure to PM_2.5 were more likely to die prematurely. ¹ , ² , ⁴⁵ , ⁴⁶ We estimated that each 10 μg/m³ increase in PM_2.5 may give rise to an estimated ≈1.5 excess deaths among every 1000 Canadians in each year. This estimate can translate to ≈16 000 excess deaths in Canada each year attributed to PM_2.5 exposure given that the population‐weighted national average concentration of PM_2.5 is ≈6.1 μg/m³, the background concentration of PM_2.5 is ≈1.8 μg/m³ _, ⁴³ and the total population of Canada is 27.3 million aged >20 years according to the 2016 Census. ⁴⁷ This is broadly consistent with a recent report of 14 600 excess deaths in Canada attributed to PM_2.5 exposure. ⁴⁶ As well, we observed positive associations of PM_2.5 with the onset of diabetes, AMI, and stroke. Similar findings were also reported in recent population‐based studies. ³ , ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰

The estimated adverse effects of diabetes and cardiovascular diseases on increasing mortality are well known. In this study, we also estimated that for every 1000 person‐years, the impacts of these conditions on mortality estimated by the rate difference are ≈5 for diabetes, 40 for AMI, and 54 for stroke (Table 2). These estimates are in line with those reported elsewhere. ⁴⁸ , ⁴⁹ , ⁵⁰ For example, in a recent cohort study comprising ≈1 million adults receiving care in the US Veterans Affairs Healthcare System, it was estimated that for every 1000 individuals with diabetes, there were 7 excess deaths in each year compared with those without diabetes. ⁴⁸ In addition, a recent Danish national cohort study reported an estimated 53 excess deaths per 1000 people per year in those with nonfatal AMI compared with the general population. ⁴⁹ Another Danish study estimated that the excess annual mortality attributed to stroke was 57 per 1000 people among participants aged ≥25 years. ⁵⁰

A minimal amount of epidemiological evidence exists about how air pollution may affect the trajectory of human health. As previously mentioned, causal mediation analysis (or effect decomposition) is a valuable epidemiological method to assess the role of intermediates that lie along the paths from the exposure to the outcome. To our knowledge, this is the first causal mediation study that empirically assessed pathways from long‐term exposure to PM_2.5 to nonaccidental deaths through the incidence of diabetes and cardiovascular diseases and therefore quantified the relative contributions of these potential mediators. In this study, people with higher exposure to PM_2.5 were more likely to die in part because they had developed PM_2.5‐induced diabetes and cardiovascular diseases (31.7% in total including 4.5% mediated through diabetes, 22.8% through cardiovascular events, and 4.5% through the interaction of diabetes and cardiovascular events). The higher proportion mediated through cardiovascular events than that through diabetes may indicate that cardiovascular events are the main mediating pathway between PM_2.5 and mortality. On the other hand, the increased proportion of the indirect effect mediated by AMI and stroke combined (22.8%) than individually (11.5% and 13.9%, respectively) could be explained by a stronger impact of PM_2.5 on these 2 conditions combined than each of them alone (Table 2). Furthermore, the proportion mediated through the interaction between diabetes and cardiovascular events was not found statistically significant. This was likely attributed to smaller numbers of joint mediators because the follow‐up period might not be long enough for us to detect more participants developing both conditions.

Collectively, our findings provide mechanistic insight from the epidemiological perspective and corroborate past experimental evidence that links exposure to PM_2.5 to changes in cardiometabolic biomarkers and measures. ⁵¹ , ⁵² Human and animal studies have shown that repetitive exposure to PM_2.5 results in alterations in endothelial function, which may precede changes in insulin resistance by reducing insulin‐dependent glucose uptake and then increase the risk of diabetes. ⁵¹ In addition, PM exposure may impair β cell function, resulting in reduced insulin secretion and metabolic dysfunction. As well, several mechanistic pathways between air pollution and cardiovascular disease have been described previously. ⁵² One pathway involves the initiation of pulmonary and systemic oxidative stress and inflammation, which could contribute to thrombosis, cardiac dysrhythmias, acute vascular dysfunction, plaque instability, the development or progression of atherosclerosis, and cardiovascular events. ⁵² Another pathway may involve disturbances of the cardiac autonomic nervous system through activation of pulmonary neural reflexes, which might also contribute to instability of vascular plaques and trigger cardiovascular events. ⁵² According to the reported causes of death, ≈15% of all deaths were attributable to AMI, stroke, and diabetes in Canada each year. ⁵³ By contrast, the present study estimated that ≈32% of PM_2.5‐related mortality was attributed to the 3 conditions.

Interventions on reducing PM_2.5 exposure and on mitigating intermediate conditions are both critical to reduce the harmful effect of PM_2.5 on mortality. ⁵⁴ , ⁵⁵ Population‐level strategies such as policies to reduce the amount of fuel burned for electricity generation, industrial production, and transportation are essential to reduce the total emissions of air pollutants. ⁵⁴ , ⁵⁵ Individual behavior change (such as staying indoors during periods of poor air quality and using personal filtration systems) can be also helpful in avoiding exposure. Clinicians and health professionals may also play an important role to communicate and advocate health‐protective actions and behaviors at the individual level. ⁵⁴ Beyond interventions to reduce exposure, by enhancing the mechanistic understanding of air pollution effects, our study further suggests that there is an opportunity in clinical practice to help mitigate the adverse effects of PM_2.5 before they become overwhelming and irreversible. This can be achieved by improving metabolic and cardiovascular health, even in the polluted areas. For example, for individuals at elevated risk of diabetes and cardiovascular disease, clinicians could offer a variety of tailored recommendations and interventions to reduce the overall risk of these conditions (eg, eating healthy foods, exercising more but being mindful of high pollution days or locations, and controlling blood pressure and cholesterol levels). ⁵⁴ , ⁵⁵ In addition, effective community‐level interventions may include targeted screening for diabetes and cardiovascular diseases and improved access to health care for the prevention and treatment of these conditions in the most polluted areas. ⁵⁴

Strengths of this study include the large cohort sizes and a long follow‐up period (up to 17 years), which allowed us to obtain valid mediation effect estimates with good statistical power. Further strengths are our ability to ascertain mediators (incident cases) using province‐wide registries and algorithms with high sensitivity and specificity, which greatly reduced the possibility of misclassification; the use of satellite‐based PM_2.5 exposure estimates at finer spatial scales (1 km×1 km); the ability to account for participants' residential mobility by assigning exposures to their residential postal code for each year of follow‐up; and the application of analytic methods that allow multiple, correlated, time‐varying mediators in the context of censored survival outcomes. Our study also benefits from the adjustment for extensive information on important risk factors of cardiometabolic diseases and deaths at both the individual and neighborhood levels to ensure better control of potential confounders.

Nevertheless, we also acknowledge several limitations. First, information on socioeconomic level and lifestyle was self‐reported and was only measured at baseline, thus measurement errors are inevitable. Second, the Ontario Diabetes Database does not distinguish between type 1 and type 2 diabetes. However, given that 90% to 95% of the entire population with diabetes in Ontario has type 2 diabetes and our study participants were aged ≥20 years at cohort entry, the vast majority of incident cases of diabetes in this analysis should be type 2 diabetes. Third, the incidence rate of diabetes in our cohort may be underestimated, although the proportion of diabetes that remained undiagnosed in Ontario has dropped significantly as a result of increasing screening. ⁵⁶ In addition, our cohorts could be subject to selection effects, biased toward including healthier respondents who had no prior history of diabetes and selected cardiovascular diseases, although the proportion of excluded participants was small (≈9%). As an attempt to investigate the extent of the potential bias, we estimated the association between PM_2.5 and deaths based on the cohort with people with prior history added back and found similar estimates with those derived from the cohort with these people excluded (Table S5). Furthermore, unmeasured confounding between exposure, outcomes, and mediators may have affected our findings. As an attempt to assess the extent of possible residual confounding, we conducted a series of sensitivity analyses (eg, adjusting for additional risk factors) and found no appreciable change in our observed indirect effect sizes. Our analysis may also be subject to exposure misclassification. The PM_2.5 estimates were available at the postal‐code level, which may not fully reflect each participant's complete personal exposure as it may be influenced by indoor exposures, occupations, and lifestyles.

CONCLUSIONS

In this large, population‐based cohort study in Ontario, Canada, we found that the association between chronic exposure to PM_2.5 and nonaccidental death is considerably (≈one‐third) mediated through the incidence of PM_2.5‐associated diabetes and cardiovascular events (AMI and stroke individually or combined). Our findings signify the influence of PM_2.5 on deteriorating cardiovascular health, highlighting the importance of improving metabolic and cardiovascular health in reducing the enormous burden of PM_2.5 on mortality, even in countries with low air pollution levels.

Sources of Funding

This study was funded by the Addressing Air Pollution Horizontal Initiative of the Government of Canada. This study was also supported by Public Health Ontario and ICES, which is funded by an annual grant from the Ontario Ministry of Health and the Ministry of Long‐Term Care. Parts of this material are based on data and information compiled and provided by Canadian Institute for Health Information (CIHI). Parts of this study are based on Ontario Registrar General information on deaths, the original source of which is Service Ontario. The opinions, results, and conclusions reported in this article do not necessarily represent the views of the ICES, Ontario Registrar General, Ontario Ministry of Health, Ministry of Long‐Term Care, or Canadian Institute for Health Information. No endorsement by ICES, Ontario Ministry of Health, or Ministry of Long‐Term Care is intended or should be inferred. Martin acknowledges support from NASA Health Air Quality Applied Sciences Team HAQAST (80NSSC21K0508).

Disclosures

None.

Supporting information

Data S1–S6

Tables S1–S5

Figure S1