Study subjects were participants in the DPP, a multicenter randomized clinical trial of individuals at high risk of developing type 2 diabetes, recruited between July of 1996 and May 1999. Inclusion criteria included at least 25 y of age, a body mass index (BMI) of 24 kg/m2 or greater (22 or higher for Asians), and a fasting glucose concentration of 95 to 125mg/dL and 140 to 199mg/dL 2 h after a 75-gram oral glucose load. The DPP trial recruited participants from 27 clinical centers across the United States and randomly assigned them to three study arms: a lifestyle intervention group receiving intensive training in diet, physical activity, and behavior modification; a pharmacological intervention group receiving 850mg of metformin (Glucophage); or a placebo-treated control group (Knowler et al. 2002). Participants assigned to the placebo and metformin arms of the trial received standard information about diet and exercise, but no intensive or motivational counseling. The original aims and design of the study have been extensively described elsewhere (Diabetes Prevention Program Research Group 1999). The study was terminated in May of 2001, 1 y early, due to efficacy on type 2 diabetes prevention of both the lifestyle intervention and pharmacological intervention arms of the trial.

The current study was restricted to individuals in the intensive lifestyle intervention and the placebo-treated control arm of the trial with available stored plasma samples collected at baseline. We did not include the metformin-treated group due to the potential interaction between PFAS and the pharmacological intervention on health outcomes and lack of experimental data on this issue. Furthermore, metformin’s influence on PFAS kinetics is unknown. A total of 957 participants were eligible for quantification of plasma PFAS. The institutional review board at each clinical center approved the protocol, and all participants provided written informed consent for DPP. For the current analyses, the Institutional Review Board of Harvard Pilgrim Health Care reviewed and approved all study protocols. The involvement of the Centers for Disease Control and Prevention (CDC) laboratory did not constitute engagement in human subjects research.

Blood was collected at baseline and plasma was stored at the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) central repository. Plasma samples were shipped from the NIDDK central repository to the CDC for analyses. Briefly, a modification of the online solid-phase extraction–high-performance liquid chromatography coupled to isotope dilution–tandem mass spectrometry approach described previously (Kato et al. 2011) was used to measure: linear PFOS (n-PFOS), sum of perfluoromethylheptane sulfonic acid isomers (Sm-PFOS), sum of perfluorodimethylhexane sulfonic acid isomers (Sm2-PFOS); linear PFOA (n-PFOA), sum of perfluoromethylheptanoic and perfluorodimethylhexanoic acids (Sb-PFOA); PFHxS, N-ethyl-perfluorooctane sulfonamido acetic acid (Et-PFOSA-AcOH; also known as EtFOSAA), N-methyl-perfluorooctane sulfonamido acetic acid (Me-PFOSA-AcOH; also known as MeFOSAA), and perfluorononanoic acid (PFNA). The limit of detection (LOD) was 0.1 ng/mL for all PFAS examined.

We summed branched and linear isomers of PFOS (n-PFOS, Sm-PFOS, Sm2-PFOS) and PFOA (n-PFOA, Sb-PFOA) to calculate total concentrations of PFOS and PFOA, respectively. Summing linear and branched isomers has been implemented in previous epidemiological studies as well as in the U.S. national report on human exposure to environmental chemicals, allowing for comparability with past and future studies (CDC 2017; Fleisch et al. 2017; Harris et al. 2017). The sum of isomers for PFOS and PFOA was performed prior to accounting for concentrations below the LOD. At least one of the isomers was always detectable for any given sample, and therefore, the sum of PFOS and PFOA isomers was 100% detectable across samples. All other PFAS concentrations below the LOD were replaced by the LOD/√2 (Hornung and Reed 1990).

The primary outcome of the trial, as well as for the present study, was diabetes incidence by the end of the DPP follow-up period. Diabetes was prospectively diagnosed using either fasting plasma glucose levels at semiannual visits or an OGTT at the scheduled annual visits. Diabetes was defined as fasting glucose ≥140mg/dL or 2-h postchallenge glucose ≥200mg/dL for visits through 23 June 1997, and fasting glucose ≥126mg/dL or 2-h postchallenge glucose ≥200mg/dL for visits after 23 June 1997. If a participant had elevated glucose levels at the annual visit (fasting or 2-h glucose post-OGTT) or at midyear (fasting glucose only), diabetes was confirmed in a subsequent follow-up visit within 6 wk using the same method as the trigger visit (i.e., either OGTT or fasting glucose for annual and semiannual visits, respectively). A few participants had elevated glucose levels at a scheduled study visit, but diabetes was not confirmed at the subsequent follow-up visit; we referred to this event as first-fasting hyperglycemia.

The following laboratory measures related to metabolic function and glycemia were included in baseline and longitudinal analyses: fasting proinsulin (pM), fasting insulin (μU/mL), 30-min post-OGTT insulin (μU/mL), fasting plasma glucose (mg/dL), 30-min post-OGTT glucose (mg/dL), 2-h post-OGTT glucose (mg/dL), % glycated hemoglobin (HbA1C), adiponectin (μg/mL), HOMA-IR=[(fasting insulin×fasting glucose)/18.01]/22.5, HOMA-β=(20×fasting insulin)/fasting glucose−3.5, corrected insulin response=(100×30-min insulin)/[30-min glucose×(30-min glucose−70mg/dL)], insulinogenic index=[(30-min insulin−fasting insulin)]/[(30-min glucose−fasting glucose)]×100, and BMI measured at baseline. Laboratory measurements were evaluated annually from baseline, with the exception of adiponectin, which was only measured at baseline and at the first annual visit; fasting plasma glucose, which was measured semiannually; and HbA1C, which was measured at baseline, 6 mo after baseline, and annually thereafter. Laboratory measures were taken prospectively at the scheduled visits until a participant received a diagnosis of diabetes or until the study was terminated in May of 2001. Analytical glycemic measurements were performed at a central laboratory (University of Washington, Seattle, WA) as previously described (Diabetes Prevention Program Research Group 2000).

Data were obtained from the NIDDK central database repository and linked to plasma concentrations using unique identifiers provided by the NIDDK repository. We extracted demographic characteristics such as participant sex, race/ethnicity, age, BMI, education, smoking history, marital status, and treatment assignment. To protect participants’ identities, age was only available in 5-y age groups with truncation of those <40 and ≥65 y of age. We calculated BMI in kg/m2 for every subjects using the average of two or three weight measures reported in the initial data release and subsequently linking it to the average of two or three height measures from the NIDDK repository. We report educational attainment into <high school, high school/GED, college graduate, and graduate school and greater. Self-reported smoking was obtained from baseline questionnaires and classified as never, former, and current smoker. Covariates were selected for adjustment a priori, and follow-up time was recorded as the number of days from randomization to the scheduled visit and transformed to years to ease interpretability.

We calculated geometric means (GMs) and interquartile ranges for all PFAS. For two of the most prevalent PFAS, PFOS and PFOA, we report their distributions across participant characteristics. We used a nonparametric Wilcoxon rank sum test and a Kruskal-Wallis test a to evaluate unadjusted exposure differences across participants’ characteristics. We calculated Spearman correlation coefficients among PFAS plasma concentrations.

The distribution of all PFAS plasma concentrations were right skewed, and therefore, we log2-transformed them for statistical analyses (log base 2). We used adjusted linear regression models to estimate associations of individual baseline PFAS plasma concentrations with baseline metabolic and glycemic clinical measurements adjusting for sex, race/ethnicity, BMI (continuous), age (categorical), marital status, education, and smoking history. Additionally, we used longitudinal mixed-effects regression models with random intercepts and slopes to estimate the association of baseline plasma PFAS concentrations with prospective metabolic and glycemic clinical measurements to test if concentrations at baseline influenced glycemic or metabolic changes over the study period. Longitudinal mixed models included the fixed effects of plasma PFAS independently, sex, race/ethnicity, BMI (continuous), age (categorical), marital status, education, and smoking history as well as treatment arm and follow-up time in years. Visual inspection of scatterplots for the glycemic outcomes over the study period suggested nonlinear time trends. Therefore, we also included the fixed effects of follow-up time squared as well as the interaction between treatment and follow-up time and treatment and follow-up time squared as necessary. We report the longitudinal association for each PFAS separately with glycemic outcomes using a two-way interaction between follow-up time in years and log2-PFAS plasma concentrations measured at baseline, using repeated measures models. A second longitudinal model tested the two-way interaction between PFAS at baseline and treatment assignment to evaluate effect modification by treatment over time. Lastly, a third longitudinal regression model tested the prospective association between baseline plasma PFAS concentrations over time by treatment using a three-way interaction of each individual PFAS, follow-up time, and treatment assignment whenever there was a significant interaction between treatment and follow-up time. Additionally, longitudinal models for HOMA-IR, HOMA-β, fasting insulin, fasting glucose, and HbA1C included two-way interactions between follow-up time and treatment as well as follow-up time squared and treatment. The fasting proinsulin model only required an interaction between follow-up time and treatment. The corrected insulin response model did not require any interactions between follow-up time and treatment. Changes in adiponectin from baseline to the first-year follow-up were modeled using linear regression adjusted for baseline covariates. Longitudinal regression model equations used are specified in Appendix A of the Supplemental Material.

We used Cox proportional hazards models to estimate hazard ratios (HRs) and 95% confidence intervals (95% CI) for the risk of developing diabetes relative to plasma baseline log2-PFAS concentrations adjusted for participant’s sex, race/ethnicity, BMI (continuous), age (categorical), marital status, education, smoking history, and treatment assignment. We evaluated effect modification by treatment assignment by stratifying on treatment group as well as testing for statistical interaction of each PFAS and treatment. Kaplan-Meier survival plots were generated for PFAS observed to be significantly associated with diabetes incidence. We also conducted a secondary analysis for time to first-fasting hyperglycemia using Cox proportional hazards models for participants for which a trigger visit for elevated glucose was not confirmed during the subsequent follow-up visit (nevents=47). The proportional hazards assumption was evaluated by inspecting the Schoenfeld residuals and using a global test of the proportional hazards assumption.

We evaluated effect modification between baseline PFAS plasma concentrations and treatment assignment by testing the interaction term in longitudinal models. We also considered three-way interactions between each individual PFAS, time squared, and treatment arm in longitudinal models (results not shown). Data management and analyses were performed using R (version 3.3.0; R Development Core Team).

From the 2,054 total participants randomized to the lifestyle intervention (n=1,024) or the placebo-control group (n=1,030) in the DPP, 957 (46.6%) had sufficient stored plasma for PFAS quantification and thus were included in this study. Most participants included in our study were female (65.3%), Caucasian (57.7%), obese or overweight (97.7%), with a college or graduate education (74.4%), nonsmokers (56.8%), between 40 and 64 y of age (76.5%), and who reported being married or cohabitating (67.6%). In this sample, participants were equally assigned to the intensive lifestyle intervention arm of the trial (50.3%) or the placebo-treated control group (49.7%), and a total of 204 (21.3%) developed diabetes by the end of the DPP trial period. These include 47 participants (4.9%) who had a hyperglycemic glucose level that was not confirmed during the immediate follow-up visit, although all were eventually diagnosed as diabetic and included in the time-to-diabetes analysis. DPP participants excluded from this analysis had similar incidence of diabetes (21.5%) and did not differ by sex, age, or BMI distribution. However, our sample included a lower proportion of individuals from other races (4.4%) compared to excluded DPP participants (7.3%).

Plasma PFOS and PFOA concentrations differed by sex, race/ethnicity, and education, but were similar across BMI classification, smoking history, and treatment assignment. The distribution of PFOA but not PFOS concentration was significantly different across age groups and marital status (Table 1). At least one PFOS and PFOA isomer was detected in 100% of the samples, and both PFOS and PFOA had the highest total concentrations of all compounds measured. Among branched isomers, Sm2-PFOS was undetectable in 58.9% of the samples and Sb-PFOA in 16.6% of samples. PFHxS was below the LOD for one participant (0.1%), while Et-PFOSA-AcOH, Me-PFOSA-AcOH, and PFNA were undetectable in 3.3%, 2.6%, and 6.8% of participants, respectively (Table 2). All plasma PFAS concentrations were positively correlated (Figure 1). GMs of plasma PFAS concentrations measured at baseline in DPP (1996–1999) were similar to those measured in NHANES in 1999–2000, but higher compared to more recent serum PFAS concentrations as reported by NHANES in 2013–2014 (CDC 2017), except for PFNA concentrations that were higher compared to DPP (Table S1 and Figure S1).

We observed cross-sectional associations between several PFAS and glycemia measurements (Table 3). In adjusted linear regression models, a doubling in plasma PFOS concentration was associated with 1.37μU/mL higher fasting insulin (95% CI: 0.41, 2.34), 4.63μU/mL higher 30-min post-OGTT insulin levels (95% CI: 0.89, 8.36), 1.37 pM higher fasting proinsulin (95% CI: 0.50, 2.25), 0.55mg/dL higher fasting glucose (95% CI: 0.03, 1.06), and 0.03% higher HbA1C (95% CI: 0.002, 0.07). Consistent with the insulin and glucose results, a doubling in plasma PFOS concentration was associated with 0.39 higher HOMA-IR (95% CI: 0.13, 0.66) and 9.62 higher HOMA-β (95% CI: 1.55, 17.70). No associations were observed for plasma PFOS and post-OGTT glucose levels.

Similarly, a doubling in plasma PFOA concentration was associated with 1.71 pM higher fasting proinsulin (95% CI: 0.72, 2.71), 2.26μU/mL higher fasting insulin (95% CI: 1.16, 3.35), 7.85μU/mL higher 30-min post-OGTT insulin (95% CI: 3.63, 12.07), 0.66mg/dL higher fasting glucose (95% CI: 0.07, 1.24), 0.04 higher corrected insulin response (95% CI: 0.001, 0.07), 0.08 higher insulinogenic index (95% CI: 0.01, 0.15), 0.04% higher HbA1C (95% CI: 0.001, 0.07), and 0.29μg/mL lower adiponectin levels (95% CI: −0.54, −0.04). Consistent with the insulin and glucose results, a doubling in plasma PFOA was associated with 0.64 higher HOMA-IR (95% CI: 0.34, 0.94) and 15.93 higher HOMA-β (95% CI: 6.78, 25.08). No associations were observed between PFOA and post-OGTT glucose levels. Results for all glycemia endpoints for PFOA and PFOS are shown in Table 3. Associations among individual isomers of PFOS (n-PFOS, Sm-PFOS, Sm2-PFOS) and PFOA (n-PFOA and Sb-PFOA) were similar (Table S2).

A doubling in plasma PFHxS concentration was associated with 0.34 higher HOMA-IR (95% CI: 0.12, 0.55) and 7.96 higher HOMA-β (95% CI: 1.46, 14.47). Plasma PFHxS concentrations were also positively associated with fasting insulin, 30-min post-OGTT insulin levels, fasting proinsulin, corrected insulin response, and insulinogenic index. Interestingly, PFHxS concentrations were not associated with fasting glucose, but a doubling in exposure was associated with 0.64mg/dL higher 30-min post-OGTT glucose (95% CI: 0.05, 2.79) and 1.26mg/dL lower 2-h post-OGTT glucose (95% CI: −2.23, −0.29).

Both Et-PFOSA-AcOH and Me-PFOSA-AcOH were associated with higher HOMA-IR and fasting insulin and lower adiponectin levels. Only Et-PFOSA-AcOH was associated with higher HOMA-β. Plasma PFNA concentrations were associated with higher fasting proinsulin and fasting glucose levels.

There was no evidence that baseline plasma PFAS concentrations were associated with changes in HOMA-IR, fasting insulin, fasting glucose, HbA1c, fasting proinsulin, corrected insulin response, insulinogenic index, or adiponectin over time or by treatment group (Figure 2 and Table S3).

A doubling in plasma Et-PFOSA-AcOH concentration was associated with a 1.96 decrease in HOMA-β over the study period (95% CI: −3.81, −0.11) after adjusting for sex, race/ethnicity, BMI, age, marital status, education, smoking history, treatment assignment, and the fixed effect of the interaction of treatment over time. For fasting-glucose levels, there was a significant statistical interaction between treatment assignment and plasma PFHxS concentrations over the study period (pinteraction=0.04). Adjusted analyses stratified by treatment were marginal for the association of PFHxS with longitudinal changes in fasting glucose among the placebo (βplacebo=−0.25; 95% CI: −0.59, 0.08; p=0.14) and the intensive lifestyle intervention (βLifestyle=0.15; 95% CI: −0.10, 0.41; p=0.24) arm. Lastly, there was a significant statistical interaction between Me-PFOSA-AcOH plasma concentrations and treatment arm for the change in adiponectin levels from baseline to the first-year study visit (pinteraction=0.05). Namely, in stratified analyses, a doubling in plasma Me-PFOSA-AcOH concentration was associated with a marginal decrease for the difference in adiponectin levels from baseline to year 1 among the placebo group (β=−0.09μg/mL; 95% CI: −0.18, 0.01; p=0.08), but not in the intensive lifestyle intervention arm (β=0.02μg/mL; 95% CI: −0.12, 0.17; p=0.72).

Median follow-up time was 2.9 y (range 0.2 to 4.6 y), and 204 participants (21.3%) developed diabetes by the end of the study. Baseline plasma concentrations of Sb-PFOA were associated with diabetes risk after adjusting for sex, race/ethnicity, BMI, age, marital status, education, smoking history, and treatment assignment; for every doubling in Sb-PFOA plasma concentration, the hazard of developing diabetes increased by 11% (HR=1.11; 95% CI: 1.00, 1.23). In stratified analyses, the incidence of diabetes among participants with detectable concentrations of Sb-PFOA was 22.6% (180/798) compared to 15.1% (24/159) for participants with undetectable Sb-PFOA concentrations (log-rank P=0.04), Figure 3. All other PFAS were not associated with diabetes incidence (Table 4). In a secondary analysis of time to first-fasting hyperglycemia (i.e., measured elevated glucose levels at the scheduled visit, but diabetes was not confirmed at the immediate follow-up visit), a doubling in plasma Sb-PFOA concentration was associated with a 29% increase in the hazard of developing hyperglycemia during the study period (HR=1.29; 95% CI: 1.03, 1.63) (Table S4).

We observed a significant statistical interaction (pinteraction=0.036) between Me-PFOSA-AcOH concentration and treatment assignment. Stratified analyses suggested that higher plasma Me-PFOSA-AcOH concentration might be inversely associated with diabetes incidence among participants assigned to the placebo group (HR=0.89, 95% CI: 0.77, 1.03), but positively associated with diabetes for individuals in the intensive lifestyle intervention group (HR=1.14, 95% CI: 0.91, 1.42). All analyses stratified by treatment assignment were not statistically significant within either the placebo or lifestyle intervention arms of the trial, and therefore, these results should be cautiously interpreted (Table S5).

Lastly, in sensitivity analyses, we evaluated effect modification by both sex and BMI at baseline. Neither sex nor BMI at baseline modified the association between baseline PFAS plasma concentrations and diabetes incidence (Tables S6 and S7). All final adjusted Cox models met the proportional hazards assumption (global test p>0.10), and visual inspection of the Schoenfeld residuals also supported appropriate model fit.