Epidemiological studies likely need to consider PM_2.5 composition even if total outdoor PM_2.5 mass concentration is the exposure of interest

Abstract

Background:

Outdoor fine particulate air pollution, <2.5 µm (PM_2.5) mass concentrations can be constructed through many different combinations of chemical components that have varying levels of toxicity. This poses a challenge for studies interested in estimating the health effects of total outdoor PM_2.5 (i.e., how much PM_2.5 mass is present in the air regardless of composition) because we must consider possible confounders of the version of treatment-outcome relationships.

Methods:

We evaluated the extent of possible bias in mortality hazard ratios for total outdoor PM_2.5 by examining models with and without adjustment for sulfate and nitrate in PM_2.5 as examples of potential confounders of version of treatment-outcome relationships. Our study included approximately 3 million Canadians and Cox proportional hazard models were used to estimate hazard ratios for total outdoor PM_2.5 adjusting for sulfate and/or nitrate and other relevant covariates.

Results:

Hazard ratios for total outdoor PM_2.5 and nonaccidental, cardiovascular, and respiratory mortality were overestimated due to the confounding version of treatment-outcome relationships, and associations for lung cancer mortality were underestimated. Sulfate was most strongly associated with nonaccidental, cardiovascular, and respiratory mortality suggesting that regulations targeting this specific component of outdoor PM_2.5 may have greater health benefits than interventions targeting total PM_2.5.

Conclusions:

Studies interested in estimating the health impacts of total outdoor PM_2.5 (i.e., how much PM_2.5 mass is present in the air) need to consider potential confounders of the version of treatment-outcome relationships. Otherwise, health risk estimates for total PM_2.5 will reflect some unknown combination of how much PM_2.5 mass is present in the air and the kind of PM_2.5 mass that is present.

What this study adds?

Outdoor fine particulate air pollution, <2.5 µm (PM_2.5) has an important impact on human health, but little attention has focused on the consequences of spatial/temporal differences in PM_2.5 composition in estimating the health impacts of total outdoor PM_2.5 mass concentrations. This is an important question, as current regulatory metrics are concerned only with how much PM_2.5 mass is present in the air and not what kind of PM_2.5 is present. We examined potential bias in health risk estimates for total outdoor PM_2.5 (i.e., how much PM_2.5 is present) after adjusting for potential confounders of the version of treatment-outcome relationships using sulfate and nitrate as two specific examples. Residual analysis was also conducted (i.e., removing variations in total outdoor PM_2.5 caused by sulfate/nitrate). Our findings suggest that studies interested in the health effects of total outdoor PM_2.5 need to consider the kind of PM_2.5 and potential confounders of the version of treatment-outcome relationships. Otherwise, health risk estimates for total outdoor PM_2.5 will reflect some unknown combination of how much PM_2.5 mass is present and the kind of PM_2.5 mass that is present.

Introduction

Outdoor fine particle mass concentrations (fine particulate air pollution, <2.5 µm [PM_2.5]) are regulated around the world owing to decades of epidemiological data supporting important impacts on human health.^1,2 While the term PM_2.5 suggests a single entity, this metric captures all airborne particles in the PM_2.5 size range and does not consider chemical composition. As a result, the primary exposure of interest in epidemiological studies designed to support regulatory interventions is total undifferentiated outdoor PM_2.5 mass concentration (i.e., how much PM_2.5 mass is present in the air without concern for composition). Expressed in terms of the consistency assumption under the potential outcomes framework,³ these studies implicitly assume that:

Y_j^obs = Y_j(x, k) if x = X_j, no matter the value of k.

This relationship states that we assume that the potential outcome observed for individual j (i.e., Y_j^obs) is a function of the exposure, x = X_j, administered by method k, where all k methods of administering exposure are equivalent. For outdoor PM_2.5 mass concentrations, this means that we are assuming that particle composition does not matter and that all versions (k) of a given outdoor PM_2.5 mass concentration are equally harmful. Put another way, we are assuming that two interventions that reduce outdoor PM_2.5 mass concentrations by 1 μg/m³ would have the same health benefits even if one intervention achieved this reduction by reducing sea salt in PM_2.5 and the other reduced fossil fuel/coal burning-related PM_2.5, which have been shown to be strongly associated with mortality.^4–6 Intuitively, it seems likely that epidemiological studies of outdoor PM_2.5 mass concentrations will violate the consistency assumption because some components are known to be more harmful than others. In essence, when we conduct cohort studies of total outdoor PM_2.5 mass concentrations and mortality, we are contrasting the survival experience of people who are exposed to the kind of PM_2.5 that is present at higher mass concentrations to the kind of PM_2.5 that is present at lower mass concentrations without ever specifying what “kind” means. The health risks estimated in these studies (e.g., mortality hazard ratios [HRs]) are generally attributed entirely to differences in outdoor PM_2.5 mass concentrations (i.e., the “how much” effect) but in reality, differences in health risks across gradients of total outdoor PM_2.5 mass concentrations reflect some unknown combination of differences both in how much PM_2.5 is present in the air as well as the kind of PM_2.5 that is present.

This issue is related to the so-called “no-multiple-versions-of-treatment” assumption in the potential outcomes framework for causal inference,⁷ and can result in bias in health risk estimates for outdoor PM_2.5 if there are unmeasured confounders of “version of treatment”-outcome relationships.⁸ While several studies have applied causal inference methods to describe and address the issue of confounding bias in air pollution epidemiology,^9–11 much less attention has focused on the implications of violations of the consistency assumption with respect to estimating the health impacts of total outdoor PM_2.5 (i.e., how much PM_2.5 mass is present without concern for composition—the current regulatory metric). This is an important question given that PM_2.5 composition and toxicity likely vary spatially, temporally, and across the PM_2.5 distribution. We recently explored the possibility of confounding bias caused by violations of the “no-multiple-versions-of-treatment assumption” in cohort studies of outdoor PM_2.5 mass concentrations and mortality using simulations.⁸ Here we extend this work to a cohort study using the Canadian Census Health and Environment Cohort (CanCHEC) including estimates of total outdoor PM_2.5 mass concentrations as well as two specific chemical components estimated in outdoor PM_2.5 (sulfate and nitrate) on a national scale.¹² These components were selected because: (1) They reflect specific chemical compounds in PM_2.5 (i.e., versions of treatment); (2) Model predictions for these components are strongly correlated with ground-level measurements on a national scale;¹² and (3) These components have been independently associated with mortality in previous studies across the United States.¹³ Unlike many previous studies, in this analysis, our goal was not to identify the specific PM_2.5 components that are most strongly associated with health outcomes. Recent studies have examined this question in detail and confirm that heterogeneity exists in the strengths of associations between specific components in outdoor PM_2.5 and mortality, including sulfate and nitrate.¹³ Instead, our aim was to estimate the magnitude and direction of potential bias in mortality HRs for total outdoor PM_2.5 mass concentrations (i.e., how much PM_2.5 mass is present in the air, the regulatory metric of interest) by adjusting for possible confounders of “version-of-treatment”-outcome relationships using the two specific PM_2.5 components noted above as indicators of “version of treatment.” This question is of direct regulatory significance as policy interventions addressing the current regulatory definition of outdoor PM_2.5 require estimates of the health impacts of total outdoor PM_2.5 (i.e., how much PM_2.5 mass is present) that are not entangled with health impacts of the kind of PM_2.5 mass that is present.

Methods

Study population

Our study population was based on the 2006 CanCHEC. This cohort is made up of respondents to the 2006 long-form census questionnaire, which captures individual and household sociodemographic data that was subsequently linked to longitudinal vital statistics and tax records (which are used for geocoding residential location). Noninstitutionalized respondents to the long-form questionnaire who lived in Canada were considered in scope for linkage.¹⁴ Record linkage and data sharing at Statistics Canada are governed by the Directive on Microdata Linkage and were performed through the Social Data Linkage Environment through deterministic or probabilistic linkages.¹⁵ The 2006 census asked respondents to consent to record linkage with tax files, which allows researchers to attach environmental covariates to a dynamic postal code history. Of the 5,871,337 persons included in the 2006 CanCHEC, 3,776,376 consented to linkage. Consent was less likely in people who were living in the territories of Nunavut, Northwest Territories, and Yukon or reporting Inuit or Indigenous identity. We further excluded persons who were younger than 25 or older than 89 or immigrated to Canada less than 10 years before the census year, which left 3,027,399 people.

The CanCHEC dataset was further customized to create an analytical file for this work, following methods used in past studies.^16,17 This process involved: (1) imputing postal codes, (2) removing business postal codes, (3) recoding census variables, (4) attaching air pollutant concentrations, (5) calculating 10-year moving averages with a 1-year lag for all air pollution estimates, (6) attaching ecological covariates, (7) transforming the file from a person-file to a person-year file, and (8) excluding person-years that did not meet study criteria. Specifically, we performed a postal code imputation when there were gaps in postal code history preceded and proceeded by postal codes that had shared characters.¹⁸ For the 3,027,399 persons included in the preliminary dataset, 92.54% of the person-years (from years 2006–2019) had a complete postal code, 6.56% had no postal code, and 0.90% had a partial (imputed) postal code. Person-years were then excluded if they did not have a full or partial postal code. We also removed postal codes that were associated with a business rather than a residence.¹⁹ Additional person-years were excluded if age during follow-up exceeded 89 years; if they occurred after a person’s death; or if postal codes could not be matched to an air pollution estimate, a Canadian Marginalization Index value, or airshed. Person-years with PM_2.5 values in the 0.005 and 99.5th centile or invalid air pollution estimates were deleted. After performing the exclusions listed above, 3,007,441 unique persons and 35,436,475 person-years were available for analysis.

Outdoor air pollution concentrations

Outdoor air pollution concentrations were assigned as 10-year moving averages with a 1-year lag based on previous analysis examining different time windows of PM_2.5 exposure.²⁰ Outdoor PM_2.5 mass concentrations were assigned to years 1996 and 2019, and estimates for PM_2.5 components (i.e., sulfate and nitrate) were assigned to years 2000–2016 as these were the years of data available (2016 values were assigned to years 2017–2019). At the start of the follow-up in 2006, the 10-year moving average was informed by 10 years of PM_2.5 data and 5 years of component data, with subsequent years having additional datapoints to inform the average. PM_2.5 mass concentrations and PM_2.5 component fraction data reflect data versions V4.NA.02.MAPLE and V4.NA.03 available from Washington University in St. Louis (https://sites.wustl.edu/acag/datasets/surface-pm2-5/). These datasets combine multiple satellite retrievals, chemical transport model output, and ground-based observations to predict component and total PM_2.5 concentrations over North America at a monthly timescale, gridded at approximately 1 km² resolution.¹² Component data were also available for ammonium in PM_2.5 but the ammonium content of outdoor PM_2.5 was highly correlated with both sulfate (r = 0.89) and nitrate (r = 0.87) and thus was excluded from the analysis.

Statistical analysis

Cox proportional hazard models²¹ were used to estimate HRs for total outdoor PM_2.5 mass concentrations and mortality, adjusting for covariates described below. Individuals were followed from the census date of 2006 until either reaching the age of 89, the year of death, or the end of follow-up in 2019. Mortality outcomes considered were nonaccidental mortality (international classification of diseases-10 [ICD-10] codes A–R), cardiovascular (ICD-10 codes I10–I69), nonmalignant respiratory disease (ICD-10 codes J00–J99), and lung cancer (ICD-10 codes C33–C34). Multiple mortality outcomes were examined as it is possible that the magnitude and direction of potential confounding bias varies by outcome depending on the magnitude/direction of association between sulfate/nitrate in PM_2.5 and a given outcome.

Three main models were examined: (1) models including total outdoor PM_2.5 mass concentrations alone; (2) models including total outdoor PM_2.5 and sulfate or nitrate separately (as potential confounders of the version of treatment-outcome relationships); and (3) models including total outdoor PM_2.5 with both components together. As sensitivity analyses, additional models were examined, including PM_2.5, both components, and O_x (redox weighed average of NO₂ and O₃), which has been associated with mortality in previous studies.¹⁷ In addition, models were examined replacing total outdoor PM_2.5 mass concentrations with the residuals of linear models regressing outdoor PM_2.5 concentrations on concentrations of sulfate, nitrate, or both (i.e., to capture variation in outdoor PM_2.5 concentrations not explained by sulfate or nitrate alone or together). All models were stratified by age (5-year age groups), sex, and immigrant status. Models were adjusted for individual-level covariates derived from census respondents (income quintile, visible minority status, Indigenous identity, educational attainment, labor force status, marital status, and occupation) and ecological covariates assigned using census geographies or postal codes (community size, airshed, urban form, material deprivation, residential instability, dependency, and ethnic concentration).^22–24 The directed acyclical graph shown in Supplemental Figure 1; http://links.lww.com/EE/A286 depicts the assumed structural relationships between variables included in our models (using the sulfate example). All HRs were scaled by the interquartile range of each pollutant: PM_2.5 (3.54 µg/m³), sulfate (1.13 µg/m³), nitrate (0.675 µg/m³).

To characterize the shape of associations between outdoor PM_2.5 components, total outdoor PM_2.5 mass concentrations, and mortality, we fit restricted cubic splines (RCS) defined by the number of knots in SAS EG (SAS Institute Inc, Cary, NC). We tested 3–15 knots and selected the minimum AIC from these results. Using the RCS parameter estimates and the covariance matrix from this best-fitting model, we simulated realizations of the RCS at all concentrations between the minimum to the maximum by increments of 0.1 µg/m³. We estimated the shape of the relationship between the components of interest (sulfate or nitrate) and mortality outcomes while including total outdoor PM_2.5 mass concentrations in the models (along with the other covariates described above).

A note on interpretation

In this analysis, our goal was to estimate the average effect of long-term exposure to total outdoor PM_2.5 mass concentrations on mortality (using HRs) taking into consideration the fact that all outdoor PM_2.5 mass does not have the same composition and toxicity. If there is no bias caused by confounders of the version of the treatment-outcome relationship (and confounders of the treatment-outcome relationship are also controlled for), differences in mortality rates across levels of total outdoor PM_2.5 mass concentrations (i.e., how much mass is present in the air) are interpreted as reflecting randomly assigned versions of treatment from the distributions of versions of treatment available for total outdoor PM_2.5 occurring in the study population.²⁵ In this example, because we are including PM_2.5 sulfate and nitrate as possible confounders of the version of treatment-outcome relationships, HRs for total outdoor PM_2.5 mass concentrations reflect health risks at fixed levels of sulfate and nitrate. This interpretation is more consistent with the current regulatory definition of outdoor PM_2.5 which is concerned only with how much mass is present (and not composition), even if this is not always explicitly stated. If we do not adequately control for confounders of the version of treatment-outcome relationships, differences in mortality rates across levels of total outdoor PM_2.5 mass concentrations reflect some combination of how much PM_2.5 mass is present in the air and what kind of PM_2.5 mass is present across the gradient.

Results

Descriptive data for cohort characteristics are shown in Supplemental Table 1; http://links.lww.com/EE/A286 and outdoor air pollution concentrations are shown in Table 1. In total, 324,548 nonaccidental, 89,700 cardiovascular, 30,009 respiratory, and 34,100 lung cancer deaths were observed during follow-up. Long-term average total outdoor PM_2.5 concentrations ranged from 2.49 to 14.47 µg/m³ whereas sulfate concentrations ranged from 0.19 to 5.32 µg/m³ and nitrate concentrations ranged from 0.0 to 3.45 µg/m³. Mass proportions of sulfate and nitrate in total outdoor PM_2.5 are shown in Figure 1 and varied both within levels of total outdoor PM_2.5 mass concentrations as well as across levels of total outdoor PM_2.5. Nitrate in PM_2.5 was more strongly associated with total PM_2.5 mass concentrations (r = 0.80) than sulfate (r = 0.74). The correlation between PM_2.5 and O_x was 0.76.

Table 1.

Descriptive data for total outdoor PM2.5 mass concentrations, nitrate, and sulfate (µg/m³)

Pollutant	Min	1%	5%	10%	25%	Median	75%	90%	95%	99%	Max	Mean	SD
PM2.5	2.49	2.89	3.54	4.11	5.58	7.30	9.12	10.33	10.78	11.98	14.47	7.30	2.28
Nitrate	0.00	0.03	0.10	0.18	0.54	0.90	1.22	1.45	1.62	1.93	3.45	0.88	0.47
Sulfate	0.19	0.50	0.62	0.72	0.99	1.56	2.12	2.64	2.96	3.61	5.32	1.62	0.75

Pollutant values reflect 10-year moving averages for each person year.

Figure 1.

Mass proportions (%) of sulfate (A) and nitrate (B) in outdoor PM_2.5 across the range of total outdoor PM_2.5 concentrations. The lines are smoothers demonstrating trends in component mass proportions across the range of total outdoor PM_2.5 concentrations.

HRs describing associations between total outdoor PM_2.5 mass concentrations and mortality with and without adjusting for possible confounding of the version of treatment-outcome relationships by sulfate and nitrate are shown in Tables 2 and 3. For nonaccidental, cardiovascular, and respiratory mortality, HRs for total outdoor PM_2.5 mass concentrations were confounded by both sulfate and nitrate with inverse associations observed for total outdoor PM_2.5 when models were adjusted for potential confounding of the version of treatment-outcome relationships. Sulfate and nitrate components were each independently associated with nonaccidental, cardiovascular, and respiratory mortality with stronger associations observed for sulfate. Evidence of confounding was also observed in the relationship between total outdoor PM_2.5 mass concentrations and lung cancer mortality, but in this case, adjusting for potential confounding of the version of treatment-outcome relationships resulted in elevated risk estimates for total outdoor PM_2.5 given the inverse associations between both components and lung cancer mortality. HRs for sulfate and nitrate alone (i.e., without total PM_2.5 in the model) are shown in Supplemental Table 2; http://links.lww.com/EE/A286 and are similar to models including total PM_2.5 with the exception of lung cancer mortality, where inverse associations were observed for both sulfate and nitrate when PM_2.5 was included in the model compared with weak/null associations when examined individually.

Table 2.

Hazard ratios (95% confidence intervals) describing the association between total outdoor PM_2.5 mass concentrations and nonaccidental and cardiovascular mortality with and without adjustment for sulfate and nitrate

Outdoor PM2.5 component	Hazard ratio (95% CI)
Nonaccidental mortality
PM2.5 only
Total PM2.5	1.063 (1.053, 1.072)
PM2.5 + sulfate
Total PM2.5	0.998 (0.986, 1.010)
Sulfate	1.109 (1.095, 1.122)
PM2.5 + nitrate
Total PM2.5	0.996 (0.981, 1.010)
Nitrate	1.072 (1.059, 1.086)
PM2.5 + both components
Total PM2.5	0.981 (0.966, 0.995)
Sulfate	1.096 (1.081, 1.111)
Nitrate	1.027 (1.012, 1.041)
Cardiovascular mortality
PM2.5 only
Total PM2.5	1.090 (1.072, 1.110)
PM2.5 + sulfate
Total PM2.5	0.983 (0.961, 1.005)
Sulfate	1.183 (1.156, 1.211)
PM2.5 + nitrate
Total PM2.5	0.961 (0.935, 0.989)
Nitrate	1.144 (1.117, 1.171)
PM2.5 + both components
Total PM2.5	0.937 (0.911, 0.964)
Sulfate	1.146 (1.117, 1.176)
Nitrate	1.074 (1.046, 1.102)

All hazard ratios are scaled by the interquartile range of each pollutant: PM_2.5 (3.54 µg/m³), sulfate (1.13 µg/m³), nitrate (0.675 µg/m³).

All models were stratified by age (5-year age groups), sex, immigrant status, and census year. Models were adjusted for individual-level covariates derived from census respondents (income quintile, visible minority status, Indigenous identity, educational attainment, labor force status, marital status, and occupation) and ecological covariates assigned using census geographies or postal codes (community size, airshed, urban form, material deprivation, residential instability, dependency, and ethnic concentration).

Table 3.

Hazard ratios (95% confidence intervals) describing the association between total outdoor PM_2.5 mass concentrations and respiratory and lung cancer mortality with and without adjustment for sulfate and nitrate

Outdoor PM2.5 component	Hazard ratio (95% CI)
Respiratory mortality
PM2.5 only
Total PM2.5	1.073 (1.041, 1.106)
PM2.5 + sulfate
Total PM2.5	0.981 (0.943, 1.019)
Sulfate	1.162 (1.116, 1.210)
PM2.5 + nitrate
Total PM2.5	0.951 (0.906, 0.998)
Nitrate	1.139 (1.093, 1.186)
PM2.5 + both components
Total PM2.5	0.934 (0.890, 0.981)
Sulfate	1.122 (1.072, 1.175)
Nitrate	1.077 (1.029, 1.128)
Lung cancer mortality
PM2.5 only
Total PM2.5	1.074 (1.044, 1.104)
PM2.5 + sulfate
Total PM2.5	1.110 (1.071, 1.151)
Sulfate	0.947 (0.912, 0.983)
PM2.5 + nitrate
Total PM2.5	1.172 (1.120, 1.227)
Nitrate	0.909 (0.981, 1.010)
PM2.5 + both components
Total PM2.5	1.175 (1.122, 1.230)
Sulfate	0.985 (0.944, 1.028)
Nitrate	0.916 (0.876, 0.957)

All hazard ratios are scaled by the interquartile range of each pollutant: PM_2.5 (3.54 µg/m³), sulfate (1.13 µg/m³), nitrate (0.675 µg/m³).

All models were stratified by age (5-year age groups), sex, immigrant status, and census year. Models were adjusted for individual-level covariates derived from census respondents (income quintile, visible minority status, Indigenous identity, educational attainment, labor force status, marital status, and occupation) and ecological covariates assigned using census geographies or postal codes (community size, airshed, urban form, material deprivation, residential instability, dependency, and ethnic concentration).

Models including total outdoor PM_2.5, sulfate, nitrate, and O_x are shown in Supplemental Table 4; http://links.lww.com/EE/A286. For nonaccidental mortality, additionally adjusting for O_x attenuated associations for nitrate, whereas HRs for sulfate remained elevated and the HR for total PM_2.5 remained largely unchanged. A similar pattern was observed for respiratory and cardiovascular mortality, and none of the lung cancer mortality results were meaningfully changed after adjusting for O_x. Finally, models including outdoor PM_2.5 as the residuals of linear models regressing outdoor PM_2.5 concentrations on sulfate, nitrate, or both are shown in Supplemental Table 4; http://links.lww.com/EE/A286. The results of these analyses are consistent (and similar in magnitude) with those above, with positive associations observed between sulfate and nitrate and nonaccidental, cardiovascular, and respiratory mortality and no association (or weak inverse associations) observed for variations in outdoor PM_2.5 mass concentrations not explained by these two components. Likewise, results for lung cancer mortality were similar to those above, with inverse associations observed for sulfate and nitrate and positive associations observed for PM_2.5 mass concentrations not explained by these two components. For lung cancer mortality, a slightly weaker association was observed for residual PM_2.5 when both sulfate and nitrate were included in the model (HR = 1.118, 95% confidence interval [CI] = 1.080, 1.157) compared with the model including total outdoor PM2.5 mass concentrations (HR = 1.175, 95% CI = 1.122, 1.230).

Concentration-response curves for nonaccidental mortality and sulfate, nitrate, and total outdoor PM_2.5 mass concentrations adjusted for sulfate or nitrate are shown in Figure 2. Clear positive associations were apparent for both sulfate and nitrate with nonaccidental mortality whereas weaker associations were observed for total outdoor PM_2.5 mass concentrations with 95% CI generally including the null (or always including the null in the case of models adjusted for sulfate). Concentration-response curves for sulfate and nitrate with cardiovascular and respiratory mortality are shown in Supplemental Figures 2 and 3; http://links.lww.com/EE/A286 and also display clear positive associations.

Figure 2.

Concentration-response relationships between nonaccidental mortality and sulfate (A), nitrate (B), and total outdoor PM_2.5 mass concentrations adjusted for sulfate (C) or nitrate (D). Dashed lines indicate 95% confidence intervals. A, Sulfate. B, Nitrate. C, PM_2.5 (adjusted for sulfate). D, PM_2.5 (adjusted for nitrate).

Discussion

Overall, our findings suggest that studies interested in estimating the health impacts of total outdoor PM_2.5 mass concentrations (i.e., how much mass is present in the air) need to consider PM_2.5 composition (i.e., the kind of mass that is present in the air) to avoid bias caused by confounding of version of treatment-outcome relationships. In this study, total outdoor PM_2.5 mass concentrations were not positively associated with nonaccidental, cardiovascular, or respiratory mortality after adjusting for sulfate or nitrate (potential confounders of version of treatment-outcome relationships), which were each positively associated with all three outcomes (i.e., suggesting that interventions targeting these specific kinds of PM_2.5 may be more efficient in terms of health benefits than actions targeting total PM_2.5). Conversely, the strength of the association between total outdoor PM_2.5 mass concentration and lung cancer mortality increased when sulfate/nitrate was included in the model because of inverse associations between these components and lung cancer mortality. Our analyses including the residuals of linear models regressing total outdoor PM_2.5 mass concentrations on sulfate and nitrate concentrations support these results, and other studies⁵ have also reported decreases in HRs (and in some cases inverse associations similar to those we observed) between total outdoor PM_2.5 mass concentrations and all-cause, respiratory, and cardiovascular mortality when variability from sulfate and nitrate were removed from total outdoor PM_2.5. However, Kazemiparkouhi et al⁵ reported a decrease in HRs between outdoor PM_2.5 and lung cancer when variations from sulfate and nitrate were removed, and this contrasts with our findings. Reasons for this discrepancy are not clear, but both studies suggest that the kind of PM_2.5 mass matters and thus needs to be considered to obtain unbiased estimates of the health risks of total undifferentiated outdoor PM_2.5 mass concentrations (i.e., the “how much” effect), even if the health impacts of specific PM_2.5 components are not of primary interest.

Our results raise several interesting questions with respect to possible interventions that may reduce the population health impacts of total outdoor PM_2.5 mass concentrations. First, because there are always multiple versions of treatment for total outdoor PM_2.5 mass concentrations, the interpretation of health risk estimates across a gradient of exposure (e.g., between locations with long-term average outdoor concentrations of 5 µg/m³ vs. 15 µg/m³) is more subtle than generally appreciated. For example, in the best-case scenario of no confounding, contrasts across levels of total outdoor PM_2.5 mass concentrations are interpreted as reflecting randomly assigned versions of treatment from the distributions of versions of treatment available in the study population (in our example, this would mean versions of treatment available at 5 µg/m³and 15 µg/m³). Therefore, in most cases, we are not simply comparing health outcomes across levels of how much mass is present, but we are also comparing different kinds of PM_2.5 because the composition of total outdoor PM_2.5 can vary across the range of outdoor PM_2.5 mass concentrations (as shown for sulfate and nitrate in Figure 2). This makes the planning and implementation of effective and efficient interventions more challenging because the underlying exposures being contrasted are opaque and it is not clear how much of a given health impact for total outdoor PM_2.5 is attributable to the “how much mass effect” and how much is attributable to the “what kind of mass effect.” Moreover, if the underlying components most relevant to the adverse health effects of total outdoor PM_2.5 mass concentrations are not evenly distributed across space/time, we should not expect homogeneous benefits to health with mandated reductions targeting total outdoor PM_2.5 mass concentrations.

If we recognize that existing concentration-response curves for outdoor PM_2.5 typically reflect some combination of the health impacts of how much PM_2.5 is present in the air and the kind of PM_2.5 is present, this also raises another interesting thought. Specifically, perhaps there is a continuum across the global PM_2.5 distribution where the “how much mass effect” is dominant at higher PM_2.5 concentrations and the “what kind of mass effect” is dominant at lower PM_2.5 concentrations. In turn, as these effects may have different slopes, the interplay between these two components of the total PM_2.5 health impact could contribute to observed heterogeneity in the shapes of concentration-response curves²⁶ and PM_2.5-related mortality rates in places with similar total outdoor PM_2.5 mass concentrations.²⁷ Further studies are needed to explore this question in depth. In particular, studies of spatial/temporal contrasts in outdoor PM_2.5 mass concentrations of similar composition in low pollution areas would be helpful in disentangling the “how much mass effect” from the “what kind of mass effect.”

One question that may arise from our analysis is why sulfate and nitrate may be independently associated with mortality. For sulfate, one possibility is that increasing sulfate content increases particle acidity, which makes co-occurring transition metals more biologically available.²⁸ Indeed, previous studies have observed stronger associations between outdoor PM_2.5 and acute cardiovascular events when both sulfate and metal content are elevated.²⁹ Moreover, previous studies have observed positive associations between nitrate and biomarkers of inflammation in adults³⁰ and both components have been independently associated with mortality in the United States.¹³ With respect to lung cancer, sulfate, and nitrate might be inversely correlated with the kinds of PM_2.5 that cause lung cancer, thus explaining the inverse associations observed between lung cancer mortality and these components. Indeed, sulfate and nitrate might just be reasonable surrogates for sources that emit versions of PM_2.5 that are relevant to health owing to the nature of their specific chemical mixtures. Moreover, it is important to note that these components were selected as examples of potential confounders of the version of treatment-outcome relationships, and there are likely other examples as well. For example, the polycyclic aromatic hydrocarbon (PAH) content of outdoor PM_2.5 could be a relevant confounder of the version of the treatment-outcome relationship in studies of total outdoor PM_2.5 and lung cancer because PAH content is related to the version of treatment and PAHs are known human carcinogens.³¹ If this is true, the hazard ratio presented for total outdoor PM_2.5 and lung cancer could be confounded by this source of bias and reflects some unknown combination of the effect of total outdoor PM_2.5 mass concentration (i.e., how much mass is present) and the effect of the version of treatment (i.e., the kind of mass that is present) on lung cancer mortality.

One point of confusion in our analysis may be related to the framing of the problem in terms of confounding rather than effect modification. We argue that the confounding approach and the multiple versions of the treatment framework are most appropriate for several reasons. First, we reiterate that total outdoor PM_2.5 mass concentration is the exposure of interest and that in the best-case scenario of no confounding, we are contrasting mortality experiences across randomly assigned versions of treatment from the distributions of versions of treatment available in the study population. Effect modification typically refers to the situation where some third variable modifies the strength of association between the main exposure of interest and the outcome. In the case of effect modification by PM_2.5 composition, the “third variable” (i.e., variations in PM_2.5 composition) actually creates systematically different versions of the treatment. In that context, examining the health effects of total PM_2.5 mass concentrations across strata of varying PM_2.5 composition is really a comparison of completely different exposures rather than a single version (or randomly assigned version) of the same exposure across levels of an effect modifier.

Particle composition varies across space and time, and if we are interested in estimating the health impacts of total undifferentiated outdoor PM_2.5 mass concentrations (i.e., the “how much mass effect”) and regulating PM_2.5 in this form, we need to do so in a manner that minimizes potential sources of bias. The fact that PM_2.5 is a mixture with multiple versions of treatment greatly complicates how we estimate (and interpret) health risks for total outdoor PM_2.5 mass concentrations. In light of this, we have several options. The first (and likely the easiest) option is to continue with the status quo and ignore the version of treatment/component information recognizing that health risk estimates obtained for total outdoor PM_2.5 mass concentrations will reflect some unknown combination of the health impacts of how much PM_2.5 mass is present in the air a given time/location and the kind of PM_2.5 mass that is present. If we do this, we also need to acknowledge that these health risk estimates do not perfectly align with the technical definition of the current regulatory metric which is focused only on total outdoor PM_2.5 mass (i.e., how much mass) and not the kind of PM_2.5 mass. In addition, we should not expect homogeneous health benefits from mandated reductions in total outdoor PM_2.5 mass concentrations, as the underlying components of health relevance are likely not evenly distributed across space, time, or the distribution of total outdoor PM_2.5 mass concentrations. The second option is to measure PM_2.5 components that are likely to act as confounders of the version of treatment-outcome relationships and include these in analyses aimed at quantifying the health impacts of total outdoor PM_2.5 (i.e., the “how much mass effect”). While we recognize that we do not currently understand all of the components relevant to the health impacts of total outdoor PM_2.5, existing evidence suggests that some components (such as sulfate and nitrate)¹³ are more important than others and thus including these should increase the validity of health risk estimates for total PM_2.5 mass concentrations.

While this study had many strengths, including a large population-based cohort, national-level exposure information, and updating of exposures for residential mobility it is important to note several limitations. First, as noted above, sulfate and nitrate were selected as two examples of possible confounders of the version of treatment-outcome relationships, and we cannot rule out other components that may contribute to bias in HRs for total PM_2.5 in a similar manner. In addition, the air pollution exposures used in this study were measured with an error that likely contained components of both classical (e.g., predicted pollutant concentrations at 1-km resolution likely varied around true long-term outdoor values on the same scale) and Berkson-type error (e.g., within each 1-km grid, true outdoor pollutant concentrations for each cohort member are likely distributed around the predicted value for the group). These sources of error likely resulted in a loss of precision, potential bias in health risk estimates for total PM_2.5 mass concentrations and components (the expectation being toward the null as the classical error component was likely non-differential), and possible residual confounding in health risk estimates for total PM_2.5 mass concentrations because of measurement error in component estimates used for covariate adjustment. Finally, not everyone who completed the 2006 long-form census was retained in our study population, and people who were living in the territories of Nunavut, Northwest Territories, and Yukon or reporting Inuit or Indigenous identity were less likely to consent to data linkage. This likely impacts the generalizability of our results but is not a serious threat to internal validity (i.e., selection bias), as selection into the study was not systematically related to both exposure and the outcome (or causes of the outcomes given than the outcomes had not yet occurred at the start of follow-up).

In summary, the heterogeneous nature of PM_2.5 raises several subtle complications related to the estimation and interpretation of the health risks associated with total outdoor PM_2.5 mass concentrations (i.e., the health effects of how much PM_2.5 mass is present in the air). In particular, as total outdoor PM_2.5 varies in composition and toxicity, we need to account for potential confounders of the version of treatment-outcome relationships to avoid bias in health risk estimates for total outdoor PM_2.5. This can be achieved by considering PM_2.5 components in our analyses that may be confounders of the version of treatment-outcome relationships or by conducting studies of total outdoor PM_2.5 across regions with similar composition. Otherwise, health risk estimates for total outdoor PM_2.5 will reflect some unknown combination of how much PM_2.5 mass is present and the kind of PM_2.5 mass that is present, which is somewhat misaligned with the current regulatory metric, which is focused only on how much PM_2.5 mass is present in outdoor air.