Fine particulate matter (PM2.5) in air pollution, primarily from combustion sources, is recognized as an important risk factor for cardiovascular events but studies of workplace PM2.5 exposure are rare. We conducted a prospective study of exposure to PM2.5 and incidence of ischemic heart disease (IHD) in a cohort of 11,966 US aluminum workers. Incident IHD was identified from medical claims data from 1998 to 2008. Quantitative metrics were developed for recent exposure (within the last year) and cumulative exposure; however, we emphasize recent exposure in the absence of interpretable work histories prior to follow-up. IHD was modestly associated with recent PM2.5 overall. In analysis restricted to recent exposures estimated with the highest confidence, the hazard ratio (HR) increased to 1.78 (95%CI: 1.02, 3.11) in the second quartile and remained elevated. When the analysis was stratified by work process, the HR rose monotonically to 1.5 in both smelter and fabrication facilities, though exposure was almost an order of magnitude higher in smelters. The differential exposure-response may be due to differences in exposure composition or healthy worker survivor effect. These results are consistent with the air pollution and cigarette smoke literature; recent exposure to PM2.5 in the workplace appears to increase the risk of IHD incidence.
Fine particulate matter (PM2.5) in air pollution, primarily from combustion sources, is recognized as an important risk factor for cardiovascular events including hypertension,
In a recent meta-analysis of cardiovascular mortality,
In this report, we describe results from a prospective study of ischemic heart disease incidence (IHD) and PM2.5 in a cohort of almost 12,000 actively employed aluminum production workers. The analysis is based on quantitative metrics of PM2.5 generated from a variety of sources in aluminum smelting, fabrication and refining facilities. We identified incident cases of IHD based on medical insurance claims in hourly workers. Smoking histories and data on other potential confounders were available from employment records and routine health exams. We focused on the association between IHD and recent occupational exposure to PM2.5. Past exposure was given less attention due to limitations in available data. We stratified by manufacturing process to explore the role of composition and job placement practices.
Hourly workers enrolled in the primary insurance plan and employed for at least two years during follow up in eight U.S. aluminum facilities were considered for the cohort. Follow up began on Jan 1 1998 for six facilities and on Jan 1 2005 for two facilities subsequently acquired by the company. Actively employed workers were followed for incidence of IHD identified from health insurance claims through 2009 or until they left work (whichever occurred first). Prior to 2003, all workers were assumed to use the primary insurance plan. This assumption is supported by the fact that 97% of workers filed at least one claim in this system during this period. After 2003, insurance options increased with the acquisition of new facilities and health insurance enrollment in the primary plan was tracked on a monthly basis.
Health insurance claims for a relevant procedure (revascularization, angioplasty, or bypass), hospitalization for two or more days or a face-to-face visit with an ICD-9 code for IHD (410–414) comprised an IHD diagnosis. Admission codes only were recorded for hospitalization claims. Claims were available starting Jan 1 1996 and a two year “wash-out” was used to exclude cohort members with prevalent disease. The date of IHD incidence was the earliest date of the first applicable claim. All cohort members were therefore at work for two disease-free years after Jan 1 1996, prior to entering follow up. For example, a worker hired on Jan 1 1997 would enter follow up on Jan 1 1999 if no claims for heart disease had been filed during the two intermediate years.
Age, sex, race and job grade data were available from employment records for all workers. Additionally, smoking status, weight and height information was maintained at occupational clinics located at each facility and were made available to researchers via chart abstraction. Chart availability varied by facility with some retaining all charts and others retaining only the charts of actively employed workers. Multiple imputation was used for missing smoking and body mass index (BMI) data.
Workers in the eight study plants were primarily engaged in smelting aluminum or one of several fabricating processes involving aluminum and related products. The constituents of particulates in these divergent work environments, as well as the nature of the work environments generally, have been well characterized in the literature.
A job exposure matrix was first constructed for arithmetic mean total particulate matter (TPM) by distinct exposure group. The company had developed an industrial hygiene database of over 300,000 samples collected over the past 25 years. We included samples at the relevant facilities collected randomly for at least 70% of an employee’s shift and analyzed using gravimetric methods.
To estimate PM2.5, side-by-side personal size-selective sampling was conducted in 2010 and 2011 in eight facilities with the traditional closed face cassette paired with a Personal Modular Impactor (PMI). The PMI measured airborne particles in three size ranges (< 2.5μm (PM2.5), 2.5–10.0 μm and >10.0 μm). The percent PM2.5 for each sample was calculated by dividing the concentration of PM2.5 (from the PMI) by the concentration of paired TPM sample (cassette). The values within each distinct exposure group at each facility were averaged to percent PM2.5, which was multiplied by TPM to create PM2.5 concentration in the job exposure matrix.
Details are provided in a separate manuscript by Noth et al (under review at Journal of Exposure Science and Environmental Epidemiology).
We partitioned exposure into two time-varying exposure windows: recent exposure (mg/m3) in each year and cumulative exposure (mg/m3-years). Recent exposure was the level of PM2.5 estimated for the job held on January 1 of each year. Cumulative exposure was the sum of annual average exposures computed as a weighted average of exposure from all jobs held in each year. Company job codes were not readily translated into the distinct exposure groups before 1996. Thus to extend cumulative exposure all the way back to start of employment for subjects hired before 1996, we assumed workers held the same job (with associated exposure) from hire up to 1996. Thus estimates of cumulative exposure were less reliable than those for recent exposures which only occurred during follow-up.
Respirator use was not considered, other than to address extreme exposure values. For samples over 50 mg/m3 a respirator protection factor was applied (Noth et al, under review at JESEE). In a sensitivity analysis we considered an alternative job exposure matrix in which no respirator adjustment was applied.
Given that TPM samples were collected as industrial hygiene monitoring data, not for research, jobs with potential for exposures greater than 30% of occupational exposure limits were targeted for sampling. Confidence scores were developed for exposure estimates in which higher confidence was assigned to estimates based on direct measurements rather than extrapolation algorithms (for more details, Noth et al, under review at JESEE). Some of the results presented here are restricted to recent exposures estimated by direct measurements.
Cox proportional hazard models were fit to estimate the effects of recent exposure to PM2.5 on IHD incidence. All Cox models used age in each year as the time metric. Sex, race (white/non-white), calendar year of follow-up, smoking status (ever/never), BMI, job grade (above/below median in each facility), and manufacturing process (smelter/fabrication/refinery/other) were included in all models to adjust for potential confounding. Models for recent PM2.5 were additionally adjusted for cumulative exposure up until each year to control for potential confounding by exposure history. Variables plausibly on the causal pathway between exposure and IHD, such as hypertension, were not included in the models. Robust sandwich variance estimators were used to account for clustering across person years belonging to the same person in all models.
Multiple imputation procedures in SAS were used to impute missing data for smoking status and BMI. Smoking was missing for 60% of workers and BMI was missing for 62%. The imputation proceeded in two steps, first continuous BMI was imputed using the EM algorithm to create a monotone missingness pattern, and then categorical smoking status was imputed using the logistic regression method. Case status, time since hire, hypertension, diabetes, obesity, employment termination, and all the variables used in the main models were used to impute missing information, as these variables were assumed to be sufficient to satisfy the missing at random assumption. Five imputed data sets were created and all subsequent model fits and inferences were calculated using the multivariate extension to Rubin’s rules.
Models included recent exposure to PM2.5 defined as categorical variables. Results for recent PM2.5 are presented with two strategies for choosing cut points: 1) cases were divided equally by quintiles of exposure and 2) cases were divided equally by quartiles of exposure above the reference level of 0.05 mg/m3. The second strategy was guided by federal EPA air pollution standards; the 2006 daily standard for PM2.5 was 0.035 mg/m3. Since industrial levels of PM2.5 are higher than ambient levels and fewer exposure measurements were taken at the low levels, the lowest cut point we could set was 0.05 mg/m3. To assess the possibility that uncertainty in low exposures might introduce bias, we conducted analyses restricted to exposures with high confidence scores.
Manufacturing process (smelter or fabrication) was considered as a potential effect measure modifier. To take advantage of the continuous exposure metric, we added a penalized spline function of recent PM2.5 to stratified Cox models. Degrees of freedom were based on minimum Akaike’s Information Criterion (AIC) and biologic plausibility. R software (R Development Core Team, Vienna, Austria) was used for the spline analysis and SAS software (SAS Institute Inc, Cary, NC) was used for all other analyses.
There was less emphasis on models for cumulative exposure; exposures were less reliable prior to 1996 and we could not feasibly restrict to higher confidence estimates. Pooled and stratified models, however, were also fit to estimate the effects of cumulative exposure to PM2.5 on IHD incidence. Exposure was defined as a categorical variable with cases divided equally by quintiles of exposure. Smoking and BMI were imputed and controlled for as described above.
There were 697 IHD cases identified from a cohort of 11,966 aluminum workers. The IHD cases were more likely to be male, older and hired prior to the start of follow up (
The Hazard Ratios (HRs) for incident IHD were higher when the reference level was set at 0.05 mg/m3, rather than quintiles, and even higher still when exposures were restricted to those assessed with high confidence (
To examine effect measure modification, exposure-response was modeled as a smooth function of recent PM2.5 exposure separately in fabrication and smelters. Stratification with splines avoids the selection of a common reference group – a particular challenge in this analysis because PM2.5 exposures in the smelters were several times higher than in fabrication. The splines also allow us to take advantage of the continuous data without parametric assumptions (
In a sensitivity analysis, models with categorical exposure were fit in the each of the two work strata. In fabrication, the HR rose to 1.98 (95% CI: 1.06, 3.68) in the third exposure category and then dipped to 1.37 (95% CI 0.74, 2.51) in the top category (>0.38mg/m3) compared to a reference group of ≤ 0.05mg/m3. In smelters, the HR rose from 1.13 (95% CI 0.48, 2.65) in the first exposure category to 1.25 (95% CI 0.52, 2.99) in the top category (>5.23mg/m3) compared to a reference group of ≤ 0.075mg/m3 (data not shown).
Additional sensitivity analyses were conducted on a job exposure matrix in which no respirator adjustment was applied. The pooled and fabrication-specific results were unchanged. The spline in smelters remained linear, increased to a maximum HR of 1.4, and was statistically significant around the mean of exposure. However, without any respirator adjustment, the maximum exposure increased to 32mg/m3, thus the slope of the exposure-response curve was less steep than the slope presented in
In contrast with recent exposure, there was no indication that cumulative PM2.5 exposure increased IHD risk overall; all point estimates were below one (
Results provide evidence that recent occupational exposure to PM2.5 is associated with IHD incidence in a cohort of actively employed hourly workers. Increased risk of IHD was strongest at the relatively low PM2.5 exposure levels found in the fabricating process. By contrast, we saw no evidence that cumulative PM2.5 increases risk of IHD incidence, though the exposure estimates were less reliable and results more subject to survivor bias (as elaborated below). These findings are consistent with the literature on cardiovascular disease and ambient air pollution.
One striking feature of our PM2.5 and IHD incidence results is the shape of the exposure-response curve in the pooled analysis. Risk peaked in the second category of exposure and leveled off in higher categories. This pattern was evident in the main analysis (
In choosing the lowest possible cut point for the reference group, we incorporated the wealth of existing knowledge and public health policy regarding ambient PM2.5 exposure. The ideal cut point would have been 0.015 mg/m3, the recent US EPA standard for annual ambient PM2.5. In this occupational cohort, however, no cases were exposed at or below this level. We then considered 0.035 mg/m3, the standard for daily ambient PM2.5. The daily standard can be exceeded only three days a year in the ambient environment but offered a more realistic cut-point in this industrial setting. However, only 13 cases were exposed at less than 0.035 mg/m3 and none of the exposure estimates had high confidence scores. Thus, 0.05 mg/m3 was determined to be the lowest level at which we had adequate power for categorical analysis. We also presented analysis based on a “naïve” reference group based on the 20th%-tile of the cases. Despite the increased power in the “naïve” categorization, the hazard ratios were attenuated likely due to increased baseline risk.
In categorical exposure models, there is a common reference group for all higher categories of exposure. However, the mean PM2.5 is considerably lower in fabrication than in smelters with little overlap in levels. Thus the reference group in the pooled models comprised mostly workers from fabrication whereas the higher categories comprised mostly workers from smelters. There are other relevant differences between smelters and fabrication jobs in addition to PM2.5 exposure levels and constituents. First, jobs in smelters are the most physically demanding jobs in the aluminum industry. Some smelting jobs also involve exposure to high levels of heat that can be hazardous for workers with heart disease risk factors. As a result of these hazards, there is a heart disease risk factor screening program for placement and retention in many smelter jobs. Furthermore, ongoing screening occurs for almost all smelter workers in the context of mandatory respiratory fitness and work in hot environments. Thus it is possible that the plateau in the highest two categories reflects a phenomenon analogous to the “healthy hire” effect that plagues occupational studies with external reference groups. If workers in the higher exposure categories (smelter workers) are screened and thus at less risk for heart disease than those in the reference group (fabrication workers), bias may occur even in our pooled internal analysis (
In addition to selection bias, the plateau may also arise from misclassification of the highest exposures in the smelters. The presence of well established risk factors for occupational asthma
To explore the impact of exposure misclassification on the steep increase in risk over lower exposure categories of our pooled analysis, we restricted our analysis to recent exposures estimated with a high confidence score. Despite a marked reduction of cases and person-years in the reference group, the shape of the exposure-response was robust and the HRs increased more sharply.
There are two sources of bias that might lead to attenuation of the HRs for cumulative exposure: left-truncation
By using medical claims, we were able to identify incident IHD as well as conditions such as hypertension. The use of claims data has been found to be highly specific for hypertension in a validation study for this population.
The increased risk of IHD observed at low levels of recent exposure in this study is consistent with the existing PM2.5 literature. The cardiovascular effects of ambient PM2.5 and second hand smoke have been established at levels below our reference cut point. In many US cities PM2.5 in air pollution rarely rises above 0.03 mg/m3 and second hand smoke ranges from 0.02 to 0.05 mg/m.3
In the heart disease and PM2.5 research in occupational cohorts, internal comparisons are rare. Moreover, most studies lack detail in the exposure assessment and have limited data on potential confounders. A recent meta analysis of four studies comparing high (or any) versus low (or no) PM exposure found an RR of 1.15 (95% CI 1.06–1.14) for IHD mortality, although none of the studies included a minimally acceptable set of potential confounders.
Data on many important covariates are available in this dataset, however we had to impute missing data for smoking and BMI. If the missing at random assumption was not met, then we may have residual confounding by these variables. Results presented here were similar to results from models run with a category for missingness for BMI and smoking, but slightly stronger than results unadjusted for BMI or smoking, especially in the smelters.
Three major pathophysiologic pathways by which particulate matter may cause heart disease have been proposed.
Other heart disease risk factors such as noise
We did not control for hypertension in our models because it might be on the causal pathway between exposure and disease. Diagnosis with this condition might result in a change of exposure, via transferring jobs or reducing hours, thus including it in a person-year Cox model could introduce bias. We plan to apply marginal structural models to deal with time varying confounding by this condition and to address censoring due to leaving work.
This is the first report of incident IHD and recent PM2.5 in an occupationally exposed US cohort of active aluminum production workers. We found evidence of increased risk of IHD even at the low end of the recent exposure range. The exposure-response was most striking when stratifying workers by manufacturing process. At the same recent exposure level, workers in the smelters have lower risk of IHD compared to the fabrication workers, possibly an unintended benefit of the workplace programs in place for heat and respiratory fitness. However, the risk does increase in both work environments, suggesting that composition of PM2.5 may be etiologically relevant for cardiovascular risk.
This work was supported by National Institutes of Health, Institute of Aging (R01 AG026291–01) and Center for Disease Control and Prevention, National Institute of Occupational Safety and Health (R01OH009939-01) and by Alcoa, Inc.
NIA Data Sharing
As an alternative to providing a de-identified data set to the public domain, we allow access for the purpose of re-analyses or appropriate “follow-on” analyses by any qualified investigator willing to sign a contractual covenant with the host Institution limiting use of data to a specific agreed upon purpose and observing the same restrictions as are limited in our contract with Alcoa, such as 60-day manuscript review for compliance purposes.
Conflicts of Interest: Drs Costello, Brown, Noth and Eisen do not have any conflicts of interest to declare. Ms Cantley and Ms. Tessier-Sherman receive salary support from Alcoa, Inc. through contracts with Yale University. Dr. Cullen receives salary support from Alcoa, Inc. through contracts with Stanford University. Dr. Hammond receives compensation as a member of the scientific advisory board for Alcoa, Inc and has also consulted on exposure assessment and received compensation.
Percent of Person-years in Recent (within the last year) PM2.5 Exposure Categories shown in Overlapping Distributions for Fabrication and Smelting Processes in an Aluminum Manufacturing Cohort followed from 1998 through 2008. (Exposures above 5 mg/m3 lumped into top category.)
Figure 2a: Penalized Spline of the Adjusted Hazard Ratio for Ischemic Heart Disease and Recent (within the last year) PM2.5 in a Cox Model for Males working in Fabrication Facilities in an Aluminum Manufacturing Cohort in the United States, 1998–2008. Restricted to Exposures Measured with the Highest Confidence. Graph truncated at 97th percentile of exposure.
Figure 2b: Penalized Spline of the Adjusted Hazard Ratio for Ischemic Heart Disease and Recent (within the last year) PM2.5 in a Cox Model for Males in Smelting Facilities in an Aluminum Manufacturing Cohort in the United States, 1998–2008. Restricted to Exposures \ Measured with the Highest Confidence.
Demographic Characteristics of an Aluminum Manufacturing Cohort in the United States 1998–2008
| Cohort | IHD Cases | |
|---|---|---|
| No. Workers | 11,966 | 697 |
| Person-Years | 68,848 | - |
| Male no. (%) | 10,049 (84) | 655 (94) |
| White no. (%) | 9,919 (83) | 591 (85) |
| Hired ≥Follow-up no. (%) | 3,665 (31) | 81 (12) |
| Year of Hire Median (IQR) | 1984 (1974–1996) | 1977 (1970–1988) |
| Year of Birth Median (IQR) | 1955 (1948–1962) | 1948 (1944–1953) |
| Manufacturing Process | ||
| Smelter no. (%) | 2194 (18) | 152 (22) |
| Fabrication no. (%) | 8290 (69) | 449 (64) |
| Refinery no. (%) | 960 (8) | 56(8) |
| Other facility type no. (%) | 522 (5) | 40 (6) |
| Recent | ||
| PM2.5 median (IQR) | 0.29 (0.12–1.04) | 0.26 (0.12–1.04) |
| Cumulative PM (mg/m3-yrs) | ||
| PM2.5 median (IQR) | 5.01 (1.74–18.41) | 6.95 (2.39–26.46) |
Recent exposure is defined as exposure within the last year.
Hazard Ratios for Ischemic Heart Disease According to Two Categorizations of Recent (within the last year) PM2.5, Each With All Exposures and When Restricted to Exposures Measured with High Confidence, in an Aluminum Manufacturing Cohort in the United States, 1998–2008.
| Recent PM2.5 (mg/m3) | No. cases | Person- years | Hazard Ratios | 95% Confidence Interval | Recent PM2.5 (mg/m3) | No. cases | Person- years | Hazard Ratios | 95% Confidence Interval |
|---|---|---|---|---|---|---|---|---|---|
| ≤ 0.11 | 135 | 11,490 | 1 | ≤ 0.12 | 129 | 10,891 | 1 | ||
| >0.11–0.22 | 143 | 14,272 | 1.05 | 0.81, 1.35 | >0.12–0.23 | 102 | 10,379 | 1.05 | 0.80, 1.39 |
| >0.22–0.45 | 140 | 11,708 | 1.23 | 0.92, 1.63 | >0.23–0.50 | 114 | 10,310 | 1.04 | 0.77, 1.41 |
| >0.45–1.47 | 130 | 11,971 | 1.06 | 0.78, 1.45 | >0.50–1.63 | 109 | 8,899 | 1.29 | 0.88, 1.88 |
| >1.47 | 149 | 14,321 | 1.09 | 0.73, 1.62 | >1.63 | 123 | 11,853 | 1.21 | 0.78, 1.88 |
| Recent PM2.5 (mg/m3) | No. cases | Person- years | Hazard Ratios | 95% Confidence Interval | Recent PM2.5 (mg/m3) | No. cases | Person- years | Hazard Ratios | 95% Confidence Interval |
| ≤ 0.05 | 61 | 6,151 | 1 | ≤ 0.05 | 15 | 1,452 | 1 | ||
| >0.05–0.16 | 156 | 13,770 | 1.26 | 0.92, 1.71 | >0.05–0.16 | 132 | 11,757 | 1.58 | 0.88, 2.63 |
| >0.16–0.34 | 166 | 13,706 | 1.48 | 1.07, 2.05 | >0.16–0.37 | 148 | 12,359 | 1.78 | 1.02, 3.11 |
| >0.34–1.15 | 155 | 14,975 | 1.15 | 0.82, 1.63 | >0.37–1.47 | 139 | 13,074 | 1.48 | 0.83, 2.66 |
| >1.15 | 159 | 15,159 | 1.18 | 0.77, 1.82 | >1.47 | 143 | 13,690 | 1.48 | 0.77, 2.85 |
Hazard Ratios adjusted for age, race, gender, calendar year, smoking, facility type, bmi, job grade, and past exposure
Hazard Ratios for Ischemic Heart Disease According to Cumulative PM2.5 (mg/m3-years) in an Aluminum Manufacturing Cohort in the United States, 1998–2008.
| Cumulative PM2.5 | No. cases | Person- years | Hazard Ratios | 95% Confidence Interval |
|---|---|---|---|---|
| ≤ 1.89 | 140 | 16,198 | 1 | |
| >1.89–4.52 | 139 | 13,392 | 0.89 | 0.70,1.12 |
| >4.52–10.51 | 140 | 11,855 | 0.81 | 0.64,1.03 |
| >10.51–35.58 | 139 | 11,648 | 0.82 | 0.63, 1.07 |
| >35.58 | 139 | 10,669 | 0.80 | 0.59, 1.07 |
Hazard Ratios adjusted for age, race, gender, calendar year, smoking, bmi, job grade, and facility type