Methods
Cohort Recruitment and Survey Administration
This study included a community cohort and a worker cohort, which were combined for some analyses and also considered separately. The methods for cohort recruitment and survey administration have been previously described (Winquist et al. 2013). Community cohort participants were recruited among 40,145 C8 Health Project participants who were ≥ 20 years of age and consented to be contacted by the C8 Science Panel for further studies. Worker cohort participants were recruited from an occupational cohort, formed for previous mortality studies (Leonard et al. 2008; Sakr et al. 2009), that included 6,026 people who worked at the chemical plant during 1948–2002 (of whom 2,090 were also in the community cohort). Study participants were asked to complete questionnaires during two survey rounds (2008–2010 and 2010–2011; see Supplemental Material, Figure S1 http://ehp.niehs.nih.gov/wp-content/uploads/122/12/ehp.1307943.s001.508.pdf), covering demographics, health-related behaviors, and lifetime personal history of various medical diagnoses. Some participants completed a questionnaire during both survey rounds, and some during only one round. Proxy respondents were sought for people who were deceased or too ill to respond themselves. Participants gave informed consent before study participation. This study was approved by the Emory University Institutional Review Board.
At least one study questionnaire was completed during 2008–2011 for 32,712 (81.5%) of the 40,145 people in the community cohort target population (of whom 4,152 were excluded from the community cohort because they worked at the plant), and for 4,391 (72.9%) of the 6,026 people in the worker cohort target population. People in the community cohort also in the worker cohort were included only in the worker cohort. Retrospective modeled serum estimates were available for 28,541 (99.9%) community cohort participants and 3,713 (84.6%) worker cohort participants.
Case Definitions
On the 2008–2011 surveys, participants were asked whether they had ever been told by a doctor or other health professional that they had high blood pressure, high cholesterol, or heart disease. Women were told to report hypertension only outside of pregnancy. Type of heart disease was also asked; choices included angina, arrhythmia, valve disease, heart attack, coronary artery disease, or "other" (asked to specify in a free text field). Responses in "other" fields were reviewed and classified. Reports (through selection of a listed option or through the text field) of angina, coronary atherosclerosis, coronary artery disease (including blocked arteries, bypass surgery, or stents), heart attack, or cardiac ischemia were considered to be reports of coronary artery disease for this analysis. For each reported condition, participants were asked the age at diagnosis. Participants reporting hypertension and high cholesterol were asked whether they were currently taking prescription medication for these conditions. Participants who reported hypertension or high cholesterol without medication during the first survey round were asked again about current medication use during the second survey round. If a proxy respondent reported hypertension or high cholesterol, the proxy was asked whether the person took prescription medication for the condition. Participants reporting heart disease were asked to consent for review of their medical records. Medical records were requested from the identified providers and were reviewed by trained medical record abstractors.
For this analysis, we analyzed incident hypercholesterolemia or hypertension cases reporting current prescription medication on the 2008–2011 surveys. We restricted cases of hypercholesterolemia and hypertension to those with prescription medication use as a way of identifying clinically important conditions. The onset age used in the analysis for these conditions was the date of first diagnosis, not the date of first medication use. People who reported hypertension or hypercholesterolemia without prescription medication use were excluded from the analysis. People diagnosed with persistent hypercholesterolemia or hypertension typically will take medication over their lifetime. However, we recognize that our case definition could have excluded some people who stopped medication before the survey, or who had been diagnosed but had not yet started medication use by the time of the survey. We think it is likely that these exclusions would be few, and there is no a priori reason why they would be associated with PFOA exposure.
Analyses for heart disease were restricted to cases for whom the medical record documented angina, coronary atherosclerosis, coronary artery disease, heart attack, or cardiac ischemia. We also accepted as valid any self-reported coronary artery disease cases (reported on the 2008–2011 surveys) that had been previously validated in the C8 Health Project survey (2005/2006).
Exposure Estimation
Details of the exposure modeling are described elsewhere (Winquist et al. 2013). Briefly, an environmental fate and transport model was used to generate yearly estimates of PFOA concentrations in local air, surface water, and groundwater (Shin et al. 2011a). These concentrations were used, in combination with survey information relating to residential history, drinking-water sources, and water consumption rates, in a residential exposure model to estimate yearly PFOA intake rates (Shin et al. 2011b). Finally, the yearly intake estimates were used in a pharmacokinetic model to generate yearly PFOA serum concentration estimates (Shin et al. 2011b). For people in the worker cohort, job- and department-specific yearly PFOA serum concentration estimates were generated in an occupational model based on historical serum PFOA measurements, participants' work histories, and knowledge of plant processes (Woskie et al. 2012). For people in the worker cohort, occupational exposure model estimates were used for the years when they worked at the plant if they were higher than the residential model estimates; if they were lower, the residential estimates were used. For years after a person stopped working at the plant, serum estimates were decayed 18% per year (based on a half-life of 3.5 years) (Olsen et al. 2007), until they reached a level predicted by the residential model. The Spearman's rank correlation between modeled serum concentration estimates and serum concentrations measured in 2005–2006 (among 30,303 people with serum concentration measurements) was 0.71 (Winquist et al. 2013). For prospective analyses (starting 1 year after the age at the time of the C8 Health Project), estimates were calibrated to the measured serum PFOA concentrations using Bayesian calibration, with measurements weighted more heavily for estimates closer in time to the measurements. Retrospective analyses used uncalibrated estimates because of uncertainty about whether calibration would improve the accuracy of estimates for years far removed from the measurements.
Data Analysis
We examined associations between PFOA serum concentration and the outcomes of interest using Cox proportional hazard models with age as the time scale and time-varying PFOA exposure measures. Retrospective analyses started at the later of age 20 years (to consider only adult disease) or the age in 1952 (the year after PFOA production started at the plant). Prospective analyses started at the participant's age 1 year after enrollment in the C8 Health Project (2005–2006) or the age 1 year after 1 August 2006. Analyses ended at the earliest of the age at diagnosis of the condition of interest, death, or the last study questionnaire. Both types of analyses excluded those with a diagnosis of the condition of interest before the analysis start age and those missing a diagnosis age. We excluded people born before 1920 (n = 173) because of uncertain reliability of disease reporting in this group. Models for hypercholesterolemia and hypertension excluded people who reported the condition without current medication (n = 5,916 for hypercholesterolemia, and 2,470 for hypertension). Models for coronary artery disease excluded people who reported heart disease that was not validated coronary artery disease (597 reported coronary artery disease that was not validated, and 3,728 reported heart disease other than coronary artery disease). Models for coronary artery disease also excluded people without validated coronary artery disease but with coronary artery disease identified in mortality data (n = 75). All analyses were done using SAS version 9.2 (SAS Institute Inc., Cary, NC).
Because hypercholesterolemia, hypertension and coronary artery disease are chronic conditions that likely develop over time, our primary exposure metric was a measure of cumulative PFOA exposure, calculated as the sum of all yearly serum concentration estimates for a person through a given age. Because current PFOA serum concentrations could affect these outcomes more acutely, we also used the yearly serum PFOA concentration estimates (at time of case diagnosis or the corresponding age for noncases) in secondary analyses. Primary analyses considered the exposure by quintiles of the exposure metric, defined among the exposure estimates for cases at their diagnosis age. As a test for trend, we considered the log of the cumulative or yearly serum concentration estimates as a continuous variable.
There was some evidence of violation of the proportional hazards assumption for the cumulative and yearly exposure metrics in retrospective hypercholesterolemia analyses and in prospective hypertension analyses. Therefore, all analyses were performed both across all ages and separately by age strata (20–39, 40–59, and 60–79 years). All models were stratified by single-year birth year to tightly control for birth year. Models either included a term for the interaction between sex and age or considered sexes separately. Models also controlled for years of schooling (not time-varying; < 12 years, high school diploma/GED (General Educational Development), some college, or bachelor's degree or higher), race (white vs. nonwhite or missing; Hispanic ethnicity was not considered because there were only 42 self-reported Hispanics in the study population), smoking (time-varying; current, former, none), smoking duration (time varying), smoking pack-years (time-varying linear term created by multiplying the self-reported number of packs smoked per day by the smoking duration to that point), regular alcohol consumption (time-varying; current, former, none), body mass index (BMI; at time of first study survey; underweight, normal, overweight, obese), and self-reported type 2 diabetes (time-varying according to reported age at diagnosis). Sensitivity analyses that did not control for self-reported type 2 diabetes gave very similar results (data not shown). Family history of coronary artery disease did not change estimates of exposure effect (data not shown) and was not included in models. Models for coronary artery disease did not control for hypercholesterolemia or hypertension because these outcomes were considered to potentially be in the causal pathway.
The primary analyses considered the combined cohorts (88% community, 12% workers). To assess the impact of combining the two cohorts (which likely had different exposure levels, different overall health status, and possibly different exposures to other environmental factors), we considered sensitivity analyses confined to the community cohort. To examine the extent to which results may have been affected by low exposures occurring before a person lived in the study area, and possible differential migration into the area by disease status, additional sensitivity analyses started at the first age at which each person was known to have qualified for one of the cohorts (by living in the study area for at least 1 year or working at the plant, referred to as the "qualifying year"). Additional sensitivity analyses considered myocardial infarction as an outcome (a subset of coronary artery disease cases), or were restricted to nonsmokers (to consider a group without another strong cardiovascular disease risk factor). To assess possible effect modification by calendar time, we considered models that ended the cohort follow-up (and hence the analysis) in varying calendar years, in 3-year intervals back to 1987 (before peak exposure).