High-resolution mapping of traffic related air pollution with Google street view cars and incidence of cardiovascular events within neighborhoods in Oakland, CA

Study cohort

Kaiser Permanente Northern California (KPNC) is a large integrated health care system that provides comprehensive medical services to 4 million members. This retrospective cohort study included 41,869 KPNC adult members who lived in the geographic study region within Oakland, CA during January 1, 2010 to December 31, 2015.Entry into the cohort required one full year of prior KPNC membership to ascertain existing chronic conditions before starting follow-up. Continuous KPNC membership (defined as gaps of no more than 90 days) and residency within the study region was required to remain in the cohort. Subjects with a history of prior CVD (myocardial infarction, coronary revascularization, stroke, or congestive heart failure) were excluded.

Geographic study region

The geographic study region was defined by the three neighborhoods within Oakland, CA where mobile monitoring of TRAP was carried out [7]. The city of Oakland spans 202 km² and includes urban, residential, and industrial areas, and also includes both flat and hilly terrain. TRAP measurements were carried out on streets in West Oakland (10.4 km²), Downtown Oakland (4.5 km²) and East Oakland (15.4 km²). Additional file 1: Figure S1 in the Supplementary Material illustrates the study region within Oakland, showing the neighborhoods where TRAP measurements were collected. These neighborhoods consisted of a variety of land uses including urban offices and residences in high rise buildings, commercial areas, industrial areas and warehouses, mixed land use, and low and high density residential areas. Finally, two major highways run through these neighborhoods (Additional file 1: Figure S1), exposing many streets and residences to high volumes of traffic.

Cardiovascular events

Cardiovascular events were determined from electronic medical records based on ICD-9 and ICD-10 codes. The cardiovascular event outcomes for this study included myocardial infarction (MI; ICD-9 code 410.x, ICD-10 code I21.x to I22.x), coronary revascularization (ICD-9 code 36.x, ICD-10 procedure code 02.x), stroke (ICD-9 code 431.x to 434.x, and 436.0, ICD-10 code I60.x, I61.x, I63.x, and I64.x), death from coronary heart disease and death from cerebrovascular disease. Deaths of KPNC patients were collected from mortality files which combine KPNC data, California state death data, and Social Security Administration data to determine a patient’s vital status. The combinations of CVD event types were based on a previous study of long-term exposure to air pollution and incidence of cardiovascular events [2].

Air pollution exposure

The spatial mapping of NO, NO₂ and BC have been described in detail previously [7]. In brief, we partnered with Google Earth Outreach and outfitted two Google Street View cars (cars used by Google to take panoramic pictures of streets which are often featured in Google Maps and Google Earth), with fast-response analyzers and two independent GPS (Global Position System) receivers with ~ 3 m precision. NO was measured by chemiluminescence (Model CLD64, Eco Physics AG, Switzerland), NO₂ was measured using 450 nm cavity-attenuation phase-shift spectroscopy (Model T500 U, Teledyne Inc., San Diego, CA), and black carbon (BC) particles were measured using a photoacoustic extinctiometer (Droplet Measurement Technologies, Boulder, CO). Drive plans for the cars were designed to carry out a seasonally balanced, repeated measurement campaign that covered different times of day (8 am to 6 pm) across the workweek for every road in the selected neighborhoods (West Oakland, Downtown and East Oakland; 30 km² area). The measurement campaign collected an average of 31 days and 200 1-Hz measurements between May 2015 and May 2016 for each 30 m of road in the study design. The large resulting data set (3 × 10⁶ 1-Hz measurements; 24,000 total vehicle-km on 750 road-km) is unique with very high coverage density and repeat-visit frequency. Measurements were temporally corrected to a daily average using the ratio of the hourly average and the daily average from the West Oakland site of the Bay Air Quality Management District. For every 30 m road length, the median of all 1 Hz concentrations measured over the year was computed to provide a long term pollutant map of BC, NO and NO₂. The median long-term exposure to each traffic-related air pollutant was chosen as the exposure of interest to be more robust to stochastic extreme concentrations (outliers). Exposures at 30 m resolution were used in the main analyses. Sensitivity analyses also considered averaged median exposures within a buffer of 60 m radius and 120 m radius. Each subjects’ address of residence at entry into the study cohort was geocoded and linked to the location of the closest air pollution measurement. Subjects were censored at the date of a change of address using the address history dates obtained from KPNC membership databases.

Covariates

Model covariates were chosen a priori and included age, sex, race, body mass index (BMI), smoking status, comorbidities (diabetes, COPD, hypertension, hyperlipidemia), hypertensive medication use, statin medication use, and neighborhood socioeconomic status (SES) at the block-group level. Covariates were selected on the basis of plausibility and previously published relationships [2, 12, 13]. Individual-level covariates and pharmacy data were obtained from electronic medical records. Each subject’s address of residence was linked to block-group level Census data. We computed the neighborhood deprivation index (NDI) as a measure of neighborhood SES using census block group level variables of income/poverty, education, employment characteristics, housing and occupation [14].

Statistical analysis

Cox proportional-hazards (PH) regression was used to model the time to the first cardiovascular event. Subjects were censored at the time of non-cardiovascular related death, disenrollment, at change of address, or at the end of follow-up on Dec 31, 2015, whichever occurred first. All covariates were included as baseline covariates in the regression models. We checked the proportional hazards assumption of the Cox model by assessing diagnostic plots and by testing for the statistical significance of interaction terms with time.

To model the shape of the association between each air pollutant and the incidence of cardiovascular disease, we first assessed the validity of assuming a linear relationship. We used restricted cubic splines to flexibly model the association between each air pollutant and the incidence of cardiovascular disease, and we formally tested for any evidence of a nonlinear relationship using a Wald test for the statistical significance of the spline terms. When no evidence of nonlinearity was found, we then included each air pollutant as a continuous measure and modeled its linear effect, estimating the hazard ratios associated with an increase of one standard deviation in the level of median long-term exposure.

Primary analyses examined the main effect of NO₂, NO and BC independently on cardiovascular events and cardiovascular mortality in the full cohort. Secondary analyses examined the relationship of NO₂, NO and BC with risk of cardiovascular outcomes in two potentially susceptible subgroups: elderly (age 65+) and diabetics. We used interaction terms to allow the estimated effects of TRAP to vary by group, and we tested for statistically significant differences in the estimated effects of TRAP by group.

We conducted several sensitivity analyses to better understand the impact of our modeling assumptions. One sensitivity analysis estimated the hazard ratios under an alternate specification of the Cox PH model using age as time scale with left censoring occurring at the age of inclusion. Another sensitivity analysis estimated the hazard ratios for alternate exposure calculations where median long-term TRAP exposure was calculated using a distance-weighted average within a buffer zone of 60 m and 120 m radius. A third sensitivity analysis examined whether the relationship of TRAP to risk of CVD events varied by neighborhood. For all statistical tests, a level of α = 0.05 was used to determine statistical significance. All analyses were conducted using SAS software, version 9.3 (SAS Institute).

Susceptible populations

Our study found that the elderly are more susceptible than non-elderly adults to the effects of TRAP. In previous research, the elderly have been considered a sensitive subpopulation to the effects of air pollution because of the established association between air pollution and mortality endpoints. Some studies of air pollution and CVD events only include older subjects, such as studies of the Medicare population [26, 27]; these studies are able to establish associations of air pollution on risk of CVD events among the elderly but cannot address whether the elderly are indeed more susceptible than the non-elderly because there is no non-elderly comparison group. However, evidence is mixed on whether there is effect modification for older ages in the effects of air pollution on risk of CVD events. Some studies of both long-term and short-term air pollution exposure and CVD mortality have reported effects that appear stronger in the elderly with suggestive evidence of effect modification that did not meet the threshold of statistical significance [20, 28]. In addition, some studies of long-term air pollution exposure and CVD mortality have found no evidence of effect modification by older age [15]. The stronger associations we observed among the elderly that represented statistically significant effect modification add to a growing body of evidence suggesting that elderly populations may have increased susceptibility to air pollution. We note that we performed 24 statistical tests for interaction by elderly age and 10 of them were statistically significant at level 0.05.

Our study did not find differences between diabetic and non-diabetic subjects in susceptibility to the effects of air pollution. Previous studies of air pollution and cardiovascular health suggest diabetic individuals may experience a heightened risk to the effects of air pollution exposure [29, 30]. Diabetes has been examined as a potential effect modifier in previous studies of TRAP exposure with cardiovascular outcomes including heart rate variability, ventricular arrhythmias, blood pressure, hospital admissions and ED visits for IHD and MI, but the evidence has been inconsistent as to whether diabetes modifies TRAP-related cardiovascular effects [31–34]. Studies of long-term NO₂ exposure with stroke and with total mortality did not find evidence of effect modification by diabetes [21, 35], while a study of short-term NO₂ exposure with total mortality did find that subjects with diabetes experienced a stronger susceptibility to the effects of NO₂ [36]. We found no evidence of effect modification by diabetes in our study. We did not have sufficient statistical power to look specifically at whether diabetes may influence susceptibility among the elderly only, which would have required a three-way interaction term.

Mechanisms

Several recent review papers have summarized the hypothesized mechanisms that underlie the relationship between air pollution and CVD [22, 24]. Air pollution exposure, including diesel exposure, has been shown to induce oxidative stress and endothelial dysfunction, which is followed by alteration of circulating lipids [22]. Some evidence indicates that respiratory irritation in response to NO or NO₂ exposure leads to activation of the autonomic nervous system, generation of vasoactive molecules and reactive oxidant species entering the systemic circulation, which in turn cause endothelial dysfunction, systemic inflammation and altered heart rhythm [4]. All these are important biological pathways in the pathophysiology of cardiovascular disease. Evidence from the Normative Aging Study has shown that short and long term BC exposures are associated with several subclinical cardiovascular health endpoints, including Heart Rate Variability [37], carotid intima-media thickness [38], systolic and diastolic blood pressure [39], circulating biomarkers of inflammation and endothelial response [40], suggesting that inflammation, endothelial dysfunction and the autonomic nervous system play a role in the etiology of its effects on CVD. In addition, traffic intensity, a measure of long-term exposure to traffic-related air pollution, has been associated carotid intima-media thickness [41]. However, toxicological and controlled human chamber studies have not been able to tease apart the effects of PM and BC [42].

Strengths and limitations

This study has a number of strengths and weaknesses. A key innovation of our study is the use of an intensive mobile monitoring campaign to obtain TRAP measurements at 30 m resolution on every street in the Oakland neighborhoods where the study population lived. This measurement approach captured the variability of these pollutants at a higher resolution than has been possible in previous studies, enabling us to better examine differences in TRAP in the urban environment. We expect that our high-resolution measurements provided a more spatially resolved map with lower measurement error than previous studies using only land-use regression. Land-use regression is thought to induce mainly Berkson error [43, 44], which does not induce bias in effect estimates but does reduce precision, making effects more difficult to detect. We expect that the main advantage in using mobile monitors as compared to land-use regression models would be a reduction in Berkson error and a resulting increase in precision. This increased precision allows us to study differences at smaller neighborhood scales than can be studied using existing land use regression methods. There is also a classical measurement error component in land-use regression models [44], which is thought to bias effects toward the null. Certain cases of spatial measurement error due to model misspecification may lead to bias away from the null [45]. In addition, all studies that examine exposures at residential addresses have a component of measurement error due to limited information on time-activity patterns [46].

Our study assumed spatial stability in the median air pollution exposure patterns over the follow-up period. Although we could have built our study cohort with a start date prior to 2010, we wanted to balance the need for having a long enough follow-up time to observe events while also upholding the assumption of spatial stability in the TRAP exposure patterns over the follow-up. A period of 6 years fell within the range of previous studies of long-term air pollution effects. For example, Miller et al. 2007 followed subjects from 1994 to 2002 with air pollution measured during the year 2000 [2] and Hart et al. 2013 followed subjects from 1990 to 2008 and used modeled NO₂ exposures for the year 2000 [47]. Furthermore, a previous study of NO₂ over 20 years found that the spatial distributions of NO2 did not change appreciably over the follow-up period [8]. TRAP measurements in this study were only taken during the daytime and on weekdays, and only during 1 year; thus these median air pollution exposure patterns may not represent the patterns at night or on weekends, and do not reflect any changes in the overall pattern from year to year. We represented long-term exposures by calculating the median over repeated mobile measurements at each location, which was found to be spatially stable as the number of repeated measurements increased [7]. The typical exposure used in studies of long-term effects of air pollution is the one-year or multi-year mean of the 24-h average exposure, although most etiologically relevant metric of long-term exposure to TRAP with respect to long-term risk of CVD events has not been established.

The study cohort members were all members of Kaiser Permanente Northern California who had equal access to health care, which limits confounding by health care access and increases the internal validity of our study. We controlled for a number of key confounders, however there still may be residual confounding in some variables such as SES, which was controlled for by area-level SES measures rather than individual-level SES, and smoking, which was controlled for by smoking status but not by packyears or number of year smoked. The average duration of follow-up was 2.5 years, which may be too short to fully capture the association between long-term exposure to TRAP and time to CVD event. The generalizability of our study is limited because the small geographic region which includes subjects with lower SES on average than the general KPNC population and is comprised of a high proportion of African Americans. Our exclusion criteria are consistent with previous studies that limit the study population to subjects without history of CVD, focusing on incident cardiovascular events in relation to long-term air pollution exposure [2, 48, 49]. We excluded subjects with CHF who are typically analyzed as a separate group in air pollution studies to examine recurrent CVD events [50–52] since those subjects experience much higher rates of CVD events than the general population and may be more susceptible to cardiac effects of air pollution than the general population. In addition, we are capturing only the insured population with access to health care within that region.

We controlled for hypertension as a comorbidity as well as hypertension medication use at baseline to account for the greater risk of CVD events among subjects whose hypertension is not controlled by medication as compared to those with hypertension whose blood pressure is controlled by medication. Approximately 70% of subjects with hypertension were taking hypertensive medications (Table 1). The ability to adjust for the effects of hypertension on the risk of CVD events at this level of detail is an advantage of our study over many previous studies that do not have such detailed medical record and pharmacy data available. All covariates were included as baseline covariates in the regression models. Notably, recent studies and a recent meta-analysis have found that air pollution exposures may contribute to the incidence of hypertension [53]. Thus, our rationale for not adjusting for changes in comorbidity status and medication use during follow-up was that worsening disease status during follow-up may reflect a mediating pathway between air pollution exposure and cardiovascular events.

Our main analyses used time on study as the time scale for the Cox PH model, adjusting for baseline age as a model covariate. Time on study (i.e. follow-up time) has often been used as the time scale for the Cox PH model, although interest in alternative time scales such as age as the time scale has grown with the increase in large scale epidemiologic cohort studies [54]. When using time on study as the time scale, an assumption is needed that the baseline hazard is an exponential function of age to ensure unbiased estimation of parameters representing covariate effects. When this assumption is met, then either time scale will yield unbiased estimates of covariate relative hazards [55]. However, violation of this assumption combined with substantial correlation between the exposure of interest and age can results in substantial bias in parameter estimates [55]. The main advantage of using age as the time scale is flexible control for age effects where any non-linear effects of age are handled implicitly; this avoids the need to directly model the effect of age that satisfies the proportional hazard assumption underlying the Cox model [54]. In this study, we found no violation of assumptions in the distribution of age in the hazard function. We further conducted a sensitivity analysis using age as time scale, and we found that point estimates and CIs were extremely similar for either time scale (Additional file 1: Table S2, Supplementary Material).

One limitation of this study was low power to detect small effect sizes. Although we were able to capture fine-scale differences in exposure to TRAP within neighborhoods, we were limited by the number of events in our study due to small the geographic coverage of our study area, and the duration of follow up. Further, the small geographic region restricted the range in exposures, since many people lived on streets with similar levels of exposure with fewer people living in areas of high exposure. All of the above reduced the power of the analyses. It is plausible that in a larger study we would be able to detect statistically significant associations in the general population, and effect modification by other susceptibility factors.

Using mobile monitoring measurements of NO₂, NO, and BC collected at 30 m resolution on each street represents an innovation in the study of TRAP and cardiovascular health. While many previous studies have assigned the same air pollution exposure to people in one neighborhood, this study moves the field of TRAP research forward by demonstrating that street-level variation in TRAP exposure within urban neighborhoods may contribute to increased risk of cardiovascular events. The results of this study suggest that mobile monitoring measurements of TRAP may be useful in studying the relationship of intra-urban differences in air pollution exposure and risk of CVD. Future studies of TRAP using spatially-resolved mobile monitoring that cover larger areas and assess exposures at multiple times of day throughout the year would be beneficial to our understanding of the long-term health effects related to intra-urban differences in TRAP.