In this study, we conducted a time-stratified case-crossover analysis for major air pollutants using spatially resolved exposure data, estimating the associations of short-term PM2.5, O3, and NO2 exposures with mortality for the entire population of seven US states at the individual level, which covered over 3 million deaths that occurred between 2000 to 2015. These estimates were not restricted to major cities but include smaller cities and rural areas. Exposure was assigned either as the concentration in the 1 km grid cell that contained the home address of the decedent, or the census tract of the decedent, which is a much finer spatial resolution that most preceding studies. Moreover, because we used spatio-temporal exposure models differences in the temporal pattern of exposure by geography were incorporated, which has not been the case for city-wide time series studies.
We found an independent and significant effect for PM2.5 and a marginal one for NO2, where a 10 μg/m3 and 10 ppb increase was significantly associated with a 0.73 and 0.19% increase in the risk of all-cause mortality, respectively. The association with O3 (0.20) was also marginally significant in the three-pollutant model. The association for PM2.5 remained significant when restricting the analysis to days with pollutant levels lower than the WHO AQG [31] (25 μg/m3), indicating that current standards are not sufficient to protect the general population. Importantly, we incorporated a double negative control strategy to protect against confounding by omitted variables. We controlled for negative exposure control (exposure after death) in the main analysis, which would capture any omitted covariate that was correlated with both air pollution before and after the death and mortality. In addition, we saw no association of any pollutant with the negative outcome control (mortality due to NAFLD), which would capture any time varying covariate that is associated with deaths from any cause (including NAFLD). Finally, we used a two-stage approach that treats the expected NAFLD cases as a surrogate for the omitted confounders, in our model for all-cause mortality. The effect size for PM2.5 was little changed, increased and became significant for O3, and decreased for NO2. The case-crossover design itself controls by matching for slowly varying individual and neighborhood covariates. Together, these suggest that the PM2.5 association is robust to control for other pollutants and omitted confounders, and the two-stage and negative control analyses strongly suggest a causal association. The NO2 and O3 results are more mixed with mostly marginal associations in multipollutant models and more indication of omitted confounding, albeit of unsure direction of bias. However, in the models with state specific temperature effects, both gaseous pollutants were significant.
Although other publications have investigated the effect of air pollutants utilizing a case-crossover design [32,33,34], none was on the scale in terms of area and age coverage comparable to the present study. In addition, our high exposure resolution has not yet been provided by existing literature. Case-crossover analyses have mostly assigned the same exposure to all inhabitants in a city or metropolitan area. In contrast our exposure was assigned at the individual address or census tract. Hence, our models greatly reduce exposure error. Of course, while the exposure models were very good, exposure error still remains, and can still induce bias in effect sizes. A recent simulation study of multipollutant measurement error reported that the bias was almost always toward the null [35]. A case-crossover study of Medicare participants by Di et al. used spatially resolved air pollution at the ZIP code level [22] which had a coarser resolution as compared to our census tract level exposure (about one-third of the population of a ZIP code) or 1 km exposure. Our effect estimates for PM2.5 and O3 were lower than that of Di’s (0.73 and 0.20% respectively), but we also adjusted an additional air pollutant NO2 as well as incorporated negative exposure controls, and negative outcome controls, and two-stage methods. The observed associations between PM2.5 and mortality were robust to adjustment by co-pollutants and weather variables. In addition, while Di restricted the study to the US Medicare population of people 65 years and older, our study included people of all ages, providing increased generalizability.
The effect of PM2.5 was in agreement with those obtained by a study across 112 US cities from 1999–2005, which reported a 0.98% (95% CI: 0.75–1.22%) increase in mortality with each 10 μg/m3 increase in PM2.5 [36]. Although our estimation for O3 was only marginally significant, the estimate was on par with that observed in a study of 48 US cities, which found a 0.3% (95% CI: 0.2–0.4%) increase in total mortality with each 10-ppb increase in O3 [15]. However, similar to other previous US studies [18, 37,38,39,40,41], those daily air pollutant exposure data were obtained from local ambient monitoring stations. As a result, all individuals residing in the metropolitan area were assigned the same exposure, leading to substantial measurement error. In comparison, the present study did not use central monitors, thereby providing a finer resolution and more accurate exposure data for all individuals, including individuals living in smaller cities, rural communities or unmonitored areas that would be misclassified or not included in earlier time-series studies. We observed a larger, although insignificantly different, effect for PM2.5 and NO2 in rural areas as compared to urban areas, suggesting the need for improved rural monitoring to contrast the adverse effect in urban versus rural regions, and the need to examine sources of rural vulnerability.
Findings from this study were also consistent with the effect sizes of PM2.5 observed in other countries [42,43,44]. However, our estimates for PM2.5 were higher than the 0.22% increase in 272 Chinese cities [45] and the 0.55% increase in 10 Mediterranean metropolitan areas [46]. Those regions have higher PM2.5 concentrations, and the lower effect sizes may be due to a nonlinear dose-response, with lower slopes at high concentrations, which has been reported previously [47]. On the other hand, our estimates for NO2 were lower than the 0.9% increase in previously reported studies [19], although that study did not control for O3 and PM2.5. These discrepancies may also be partly explained by differences in population structure, the number of cities, age category, and air pollutant measurement method involved. The marginal insignificance of the O3 association when controlling for NO2 should be treated cautiously, since NO2 has a complex association with O3, serving as a driver of photochemistry but also a marker for NO quenching in more heavily trafficked areas. This can create a complex confounding pattern that can lead to effect transfer across the two pollutants.
The WHO AQG daily standards were until recently 25 μg/m3 for PM2.5, 50 ppb for O3, and 106.4 ppb for NO2. In comparison, the United States has a less restrictive standard for PM2.5 and NO2 (35 μg/m3 for PM2.5, 70 ppb for O3, and 100 ppb for NO2). When restricting the analysis to a PM2.5 concentration below the WHO standards, its effect size remained the same. The EPA recently proposed to maintain the current national particulate matter standards due to insufficient evidence for effect at lower concentrations [48]. Our findings showed that even at levels below the standards, PM2.5 pollution is significantly associated with an increase in daily mortality rates, including after incorporation of multiple causal modeling methods.
In addition to all-cause mortality, we also found a significant association with cardiovascular and respiratory mortality for PM2.5. Exposure to air pollution has been consistently associated with death due to chronic obstructive pulmonary disease (COPD), death due to pneumonia, as well as emergency room visits for asthma [14, 15, 19, 49], and our estimates for respiratory mortality are in line with previously reported estimates. Many studies have reported associations between exposure to PM2.5 and cardiovascular deaths [19, 50] and provided evidence that these disease processes can be mediated through a combination of inflammatory, autonomic, and vascular changes [51, 52].
Profound racial and socioeconomic disparities in PM2.5 exposure have been well documented in prior studies, where the burden of death associated with PM2.5 exposure was disproportionately borne by the elderly [38, 53] and people of races other than white [54, 55]. Our effect modification analysis suggested a slightly elevated, although insignificantly different, association between PM2.5 and all-cause mortality among females, people of lower educational attainment, those residing in rural areas, and people of Black race. This is in addition to the effects of higher exposure in minorities. Greater attention is needed to address the issue faced by minorities who might also be least equipped to deal with the adverse health consequences of air pollution.
Attention has recently focused on causal methods of analysis for observational data. Causal modeling seeks to mimic a randomized controlled trial by making exposure independent of all confounders but can fail if there are omitted confounders. Case-crossover analyses, by matching each person to themselves, on a nearby day without the event make exposure independent of all fixed or slowly changing individual covariates by design, and hence render exposure independent of many unmeasured confounders. In addition, we used negative exposure and outcome controls to capture omitted time-varying confounders, and a two-stage regression model to control for unmeasured, time-varying confounders. These methods strengthened the evidence for a causal association between air pollution and daily mortality.
This study has several limitations. First, there is a lack of data differentiating exposure at residence and exposure elsewhere. However, in this study, 77% of the deaths occurred in people over the age of 65 and we, therefore, expected little workplace or commuting exposure, and a higher relevance for residential exposure [56]. As a result, the extent of misclassification was reduced. Moreover, the National Human Activity Pattern Survey in the U.S. reported that U.S. adults spent 69% of their time at home and 8% of the time immediately outside their home [57]. Second, we did not have individual data on behavioral factors, medication, and specific health histories or treatments. By design, these cannot be confounders, but this limited our ability to investigate potential modifications by these characteristics. Third, we did not investigate potential confounding by other co-pollutants such as sulfur dioxide (SO2) and carbon monoxide (CO). However, the levels of SO2 and CO are low in the US [58, 59]. In addition, Dominici et al. [60] adjusted for all O3, NO2, SO2, and CO but found no change in the magnitude of the effect between particular matter and mortality, suggesting there is little evidence that the effect of particulate matter is confounded by the additional pollutants. Finally, while our exposure models were good, they were not perfect in estimating exposure at 1 km resolution. Further, the exposure error in the models varied spatially, which may account for the lack of finding of interactions with spatially varying effect modifiers.
Despite its limitations, the study adds to our understanding of the effect of short-term air pollution exposure. The most important strength of this study is the high resolution of exposure data covering the multiple states, even in areas without air monitoring stations. This provided accurate estimates of daily levels of air pollution and meteorological conditions, allowing us to examine the entire population of these states instead of only larger cities, and reduced exposure misclassification compared to prior studies with a central-monitor approach. Second, our analysis on the whole population of seven US states avoids potential selection bias and ensures the generalizability of the results. Finally, we used several causal modeling techniques, including negative exposure and negative outcome controls to increase the likelihood of a causal association.