Participants in the current analysis were from the SWAN, an ongoing, multi-site, multi-ethnic, community-based longitudinal study designed to investigate the natural history of the menopausal transition and its effect on midlife health including risk factors for age-related chronic diseases (Sowers et al., 2000). Between 1996 and 1997, a total of 3,302 women from seven study sites, including Boston, MA; Chicago, IL; southeast Michigan, MI; Los Angeles, CA; Oakland, CA; Newark, NJ; and Pittsburgh, PA, participated. Each site enrolled White women and women from one minority group (Black women from Boston, Chicago, Southeast Michigan, and Pittsburgh; Chinese women from Oakland; Japanese women from Los Angeles; Hispanic women from Newark). Eligibility criteria for enrollment into the SWAN cohort included: age 42 to 52 years in 1996/97; having an intact uterus and at least one ovary; having at least one menstrual period and not taking hormone therapy in the past 3 months; and having self-identified with the site’s designated race/ethnic groups. Participants returned for regular examinations approximately annually. Institutional Review Board approval was obtained at each study site, and all participants provided signed informed consent at each study visit.

To evaluate associations between urinary metals and longitudinal glucose outcomes, we used data from the SWAN Multi-Pollutant Substudy (MPS), which was initiated to examine the associations of multiple environmental chemicals with metabolic and reproductive health outcomes in midlife women (Ding et al., 2020; Wang et al., 2019a). A subset of 1,400 SWAN participants from the five SWAN sites (Boston, southeast Michigan, Los Angeles, Oakland and Pittsburgh) who provided urine samples to the SWAN Repository at the third SWAN follow-up visit (V3, 1999–2000, SWAN-MPS baseline) were assayed for metal concentrations. Women from Chicago and Newark were excluded because urine samples were not collected in these two sites. This subpopulation, by design, included self-identified White, Black, Chinese, and Japanese but not Hispanic women who were recruited exclusively from Newark. Among these five sites, women for whom urine samples were not available were less educated and more likely to be current smokers or obese than women with available urine. For this analysis, we excluded 39 participants who had no information on key covariates (education, household income, body mass index (BMI), physical activity, total energy intake), 46 participants with missing information on fasting glucose or insulin levels, and 53 participants who were taking antidiabetic medications at SWAN-MPS baseline, yielding 1,262 participants eligible for the present study. We censored 347 observations in subsequent follow-up visits when a participant was taking antidiabetic medications because the true untreated levels of the outcome parameters are unknown. A final sample of 1,262 women representing 9,527 observations (7.5 per person) through 2016 was used for data analysis. 804 women remained in the cohort at the last follow-up visit. When compared to those 804 women, women who lost to follow-up were more likely to be Black and had higher BMI at SWAN-MPS baseline. An overview of our analytic sample is illustrated in Figure A.1.

HOMA-IR and HOMA-β, widely used tools to assess insulin resistance and β-cell dysfunction in clinical practices and epidemiological studies (Wallace et al., 2004), are the primary outcome measures of interest for this analysis. In SWAN, fasting serum glucose and insulin levels were assayed from serum samples obtained at each follow-up visit. HOMA-IR was calculated from fasting glucose and insulin levels according to the following equation: [insulin (μU/mL) × glucose (mmol/L)]/22.5 (Matthews et al., 1985). A higher HOMA-IR indicates greater insulin resistance. HOMA-β was calculated as follows: 20 × insulin/[glucose − 3.5] (Matthews et al., 1985). A lower HOMA-β indicates the worse pancreatic β-cell function. Fasting serum glucose level was determined by hexokinase method (Boehringer Mannheim Diagnostics, Indianapolis, IN, USA). Fasting serum insulin was measured by a solid phase radioimmunoassay (Coat-ACount, Diagnostics Product Corp., Los Angeles, CA).

Urinary metal concentrations were analyzed with high-resolution inductively coupled plasma-mass spectrometry (ICP-MS) (Thermo Scientific iCAP RQ, Waltham, MA) in first morning spontaneously voided urine samples collected at the SWAN-MPS baseline at the Applied Research Center of NSF International (Ann Arbor, Michigan), a part of the Michigan Children’s Health Exposure Analysis Resource (M-CHEAR) Laboratory Hub. The analytic methods and quality control procedures have been described previously (Wang et al., 2019a). Urinary concentrations of the following 15 metals were used in the current analysis, including arsenic, barium, cadmium, cobalt, cesium, copper, mercury, manganese, molybdenum, nickel, lead, antimony, tin, thallium, and zinc. The limits of detection (LOD) and detection rates are presented in Table A.1. Participants with metal concentration below the limit of detection (LOD) were assigned a value equal to the LOD divided by the square root of 2 (Arunajadai and Rauh, 2012). Urinary concentrations of beryllium, chromium, platinum, uranium, vanadium and tungsten were also determined in the SWAN-MPS. However, due to the relatively low detection rates (<40%) as described previously, these metals were excluded from the current analysis (Wang et al., 2019a). Pairwise Spearman correlations among specific gravity adjusted urinary metal concentrations were calculated and presented in a heat map.

Age (continuous), self-reported race/ethnicity (White, Black, Chinese, or Japanese), and education level (≤ high school, some college, or college degree/post-college) were assessed through a self-administered questionnaire at baseline. At each study visit, annual household income ($19,999 or under, $20,000-$49,999, $50,000-$99,999, or $100,000 or above), smoking (never smoked, former smoked only, or current smoking), alcohol drinking (use less than once per month, use once per month, or twice or more times per month), menopausal status (premenopausal, post-menopausal, unknown menopause status due to hormone therapy or hysterectomy), and use of exogenous hormones were self-reported. Physical activity was evaluated at each visit using a modified version of the Kaiser Physical Activity Survey (Sternfeld et al., 2000), and a total score ranged from 3 to 15 was calculated indicating the activity levels during the previous 12 months in 3 distinct domains: active living, household/caregiving, and sports/exercise. BMI was calculated at baseline as weight in kilograms divided by the square of height in meters. Dietary intake was collected at baseline, using a detailed semi-quantitative food frequency questionnaire (FFQ) adopted from Block FFQ (Block et al., 1986). The 103-food item FFQ included 4 seafood items (fried fish/fish sandwich, tuna fish/tuna salad, shellfish, and other fish) and 1 rice item (rice/dishes made with rice). For analysis, weekly seafood intake was computed by summing the frequency of intake for the 4 fish items. Total energy intake was obtained from the FFQ based on each food intake. Urinary specific gravity was determined using a handheld digital refractometer (ATAGO model PAL-10S, Tokyo, Japan) as a marker of urine dilution at baseline. In this analysis, age, race/ethnicity, study sites, education level, BMI, dietary factors, and urinary specific gravity were time-independent, while all other covariates were modeled as time-varying covariates.

A two-stage modeling approach was used to evaluate the associations of metal mixtures with longitudinal HOMA measures (HOMA-IR and HOMA-β) because there is no available analytical approach that handles correlations for both dependent and independent variables. In stage 1, to account for correlations in outcome measurements within each participant, linear mixed effects models were used to capture changes in HOMA measures over the follow-up period. Given the highly skewed distributions of both HOMA-IR and HOMA-β, logarithmic transformations were applied. Time (year) was modeled using a linear term. Random intercepts and random slopes of time were included in the models to allow for the variability of HOMA levels at baseline and their rates of change between each study participant. The participant-specific baseline HOMA levels and participant-specific annualized slopes (rates of changes) were estimated and used as dependent variables in the next stage of analysis. The conditional R² of linear mixed effects models was 0.66 for HOMA-IR and 0.56 for HOMA-β.

In stage 2, which was used to avoid multicollinearity among correlated exposure variables, AENET was used to select components of metal mixtures associated with baseline levels of HOMA measures and with their rates of changes, respectively. Ordinary least squares-based variable selection methods are commonly used but prone to over-fitting and do not work well in the presence of potentially high-dimensional predictors, or when predictors are highly correlated (multicollinearity) (Tibshirani, 1996). To combat this issue, elastic-net (ENET), a shrinkage regression method, has been introduced (Zou and Hastie, 2005). ENET executes variable selection by shrinking coefficients of “unimportant” predictors towards exact zeros, and has the ability to handle the complex correlations between predictor variables (Zou and Hastie, 2005). Adaptive elastic net (AENET), as its name would suggest, is an adaptive version of ENET that not only deals with the collinearity problem over ENET but satisfies the asymptotic normality assumption that allows us to conduct statistical inference and hypothesis testing by providing large sample standard errors and p-values (Zou and Zhang, 2009). It should be noted that AENET performs variable selection by shrinking certain coefficients to zero but not based on p-values of coefficients (like forward selection, backward elimination, and stepwise selection). In this study, AENET models were fitted as follows: Yi=β0+∑k=115βkXki+βzTZi+εi, where Y_i represents participant-specific baseline HOMA levels or participant-specific rates of changes in HOMA measures estimated from stage 1, X_ki denotes urinary concentration of k^th metal (we have a total of 15 metals in the initial model to be selected), and Z_i indicates the vector of confounders. Two AENET models were performed to select metals associated with (1) baseline levels of HOMA-IR and HOMA-β, and (2) rates of changes in HOMA-IR and HOMA-β, separately. Given the highly skewed distributions of all metal concentrations, logarithmic transformations were applied. To better compare the associations of different metals with HOMA measures, we further standardized the log-transformed urinary metal concentrations by subtracting the mean of the corresponding log-transformed concentrations divided by its standard deviation (SD). This way the association was interpreted as the percent change in baseline level/rate of change in HOMA in relation to a one SD increase in log-transformed concentrations of selected metals in AENET models. All potential confounders, including age at baseline, race/ethnicity, study site, education level, annual household income, BMI, smoking, alcohol drinking, physical activity score, menopausal status, hormone therapy, dietary intake of seafood and rice, total zinc intake from diets and supplements, total energy intake, and urinary specific gravity were always adjusted for (“forced”) in the models. We decided not to include time-varying BMI in the model because of its role as a potential diabetes risk factor and the fact that it could be affected by metal exposures at baseline (Niehoff et al., 2020; Wang et al., 2018). We adjusted for seafood and rice intake because it may impact glucose homeostasis (Hosomi et al., 2012; Zuñiga et al., 2014) and has been identified as important determinants of metal mixtures in our previous study in the SWAN-MPS (Wang et al., 2019a). We adjusted for total zinc intake to better capture the potential effects of urinary zinc excretion that are independent of dietary zinc intake. For other essential metals, such as copper, no dietary intake was adjusted for due to lack of data. We used a Directed Acyclic Graph to show the hypothesis relations between metals, confounders, and HOMA measures (See Figure A.2). To better visualize the results of two separate AENET models (baseline level and rate of change), we plotted trajectories of HOMA measures from baseline to the end of follow-up using the coefficient estimates of these two AENET models for all metal concentrations fixed at their 25th, 50th, 75th and 90th percentiles, respectively, with all other covariates adjusted. AENET penalized parameters were ascertained based on 10-fold cross-validation for minimal prediction errors. The R package ‘gcdnet’ was used to implement AENET (Yi and Zou, 2017).

To better summarize the combined effects of exposure to metal mixtures, we constructed an ERS, as an integrative index health risk of exposure to multiple chemicals in epidemiological research (Park et al., 2017; Wang et al., 2019b, 2018). The underlying idea behind the ERS is to build a risk score as a weighted sum of the chemical concentrations from the simultaneous assessment of multiple chemicals. In this analysis, weights were determined by the magnitudes of the associations of each metal from the AENET models. In this way, ERS was computed as a weighted sum of non-zero metal predictors selected from AENET by ERSi=∑j=1Pβ^jZij, where Zij(j=1,…,p) is a standardized log-transformed concentration of the j-th metal and β^j is the beta coefficient (weight) of the j-th metal. From this equation, the ERS for participant i can be interpreted as the effects on insulin resistance/β-cell dysfunction corresponding to her urinary concentrations of selected non-zero metal predictors from AENET. We further categorized ERS into quartiles and fit the multiple linear regression models to examine the associations between the ERS and HOMA outcomes. To compare the magnitude of the ERS association with those for selected individual metals, we also categorized those metals into quartiles and computed effect estimates for HOMA outcomes.

For metals that were selected in the AENET models, we also included them together in Bayesian kernel machine regression (BKMR) (Bobb et al., 2015), which enabled us to further evaluate the potential interactions between those metals and non-linear relationships between metals and HOMA measures. Specifically, we examined (1) univariate exposure–response functions of each standardized log-transformed metal concentration with other metals fixed at the median; and (2) bivariate exposure-response functions of each standardized log-transformed metal concentration with the second metal fixed at selected percentiles while other metals fixed at the median, to evaluate potential interactions between metals. Gaussian kernel exposure response machine function was used to fit the model. The same covariates from AENET models were adjusted in BKMR model. The R package ‘bkmr’ was used to implement BKMR (Bobb et al., 2015).

We conducted several sensitivity analyses to evaluate the robustness of our primary findings. First, hyperglycemia has been associated with increased urinary zinc excretion (Chausmer, 1998). Because participants who had relatively high glucose levels at the SWAN-MPS baseline may also have high levels of urinary zinc excretion at that time, reverse causation may account for observed associations between urinary zinc concentration and HOMA measures. To examine the potential impact of reverse causation on our results, we excluded participants with fasting glucose level ≥ 100 mg/dL (impaired fasting glucose) or HOMA-IR ≥ 4.2 (90th percentile) at SWAN-MPS baseline. Second, because adjustment for seafood intake may not sufficiently control for the less toxic organic arsenic in evaluation of the association between arsenic and HOMA measures (Navas-Acien et al., 2011), we evaluated the association in a subpopulation with seafood intake less than 1 time/week. In addition, we adjusted BMI at baseline in the primary analysis. However, it is possible that cumulative exposures to metals could impact BMI at baseline, making it a potential intermediate (Niehoff et al., 2020; Wang et al., 2018). Therefore, as a sensitivity analysis, we excluded BMI at baseline from the model adjustment in case of over-adjustment bias. All analyses were conducted using R, version 3.5.3 (www.R-project.org).

Characteristics of the study population at baseline are summarized in Table 1. The mean (SD) age of the 1,262 participants was 49.7 (2.8) years. Geometric means (geometric standard deviations) of HOMA-IR and HOMA-β were 2.1 (1.7) and 154.5 (1.6), respectively, at baseline. Women with HOMA-IR at baseline greater than the median level (1.8) were more likely to be Black, and to have lower education level, lower family income, higher BMI, and higher alcohol consumption. Women with HOMA-β at baseline less than the median level (148.9) were less likely to be Black, more likely to be Chinese or Japanese, and to have higher family income, higher BMI, and higher alcohol consumption. The distributions and detection rates of all 15 urinary metal concentrations are summarized in Table A.1. Highest mean concentration was observed for zinc (272.44 μg/L), and lowest was observed for thallium (0.08 μg/L). In general, most metals were modestly and positively correlated with each other (Figure A.3). The strongest correlation was between copper and nickel (R=0.54).

Table 2 summarizes the associations of selected individual components of metal mixtures with baseline HOMA-IR and its rate of change in the AENET models. A total of 4 metals including copper, molybdenum, lead, and zinc were associated with baseline HOMA-IR out of 15 candidate predictors. The beta coefficients for all other metals were shrunk to zero. After multiple adjustments, a one SD increase in log-transformed urinary metal concentration was associated with 1.57% (95%CI: −1.09%, 4.29%) higher baseline HOMA-IR level for copper, 0.70% (95%CI: −1.59%, 3.05%) higher level for lead, and 5.76% (95% CI: 3.05%, 8.55%) higher level for zinc. Urinary molybdenum concentration was inversely associated with baseline HOMA-IR (mean percent change in HOMA-IR for a one SD increase in urinary molybdenum concentration = −3.25%, 95%CI: −5.45%, −1.00%). HOMA-IR levels increased by 1.51% (95% CI: 1.41%, 1.61%) annually during the follow-up period. Urinary zinc concentration was associated with faster rate of increase in HOMA-IR, a one SD increase in urinary zinc concentration was associated with a 0.06% (95%CI: −0.03%, 0.15%) increase in the annual rate in HOMA-IR. Beta coefficients of selected non-zero predictors in AENET models are shown in Table A.2. Predicted HOMA-IR levels over the follow-up based on coefficients in these two AENET models are shown in Figure 1.

We constructed ERS of HOMA-IR at baseline using estimated weights for copper, molybdenum, lead, and zinc from the AENET model (Table A.2). The ERS ranged from −0.26 to 0.27 with a mean (SD) equal to 0 (0.08). In accordance with the ERS formula, higher ERS for HOMA-IR at baseline indicates higher HOMA-IR (greater insulin resistance) at baseline attributable to higher concentrations of copper, lead, zinc but lower concentration of molybdenum. After adjusting for confounders, women in the highest quartile of ERS, on average, had 15.70% (95% CI: 9.15%, 22.64%) higher HOMA-IR at baseline, compared to those in the lowest quartile (Table 3). When comparing the effect estimates between quartiles of ERS and quartiles of selected individual metals, stronger effect sizes of ERS were observed than those for individual metals (Table A.3). The ERS of rate of change in HOMA-IR was not calculated because only zinc was selected in the AENET.

We ran an BKMR model regressing HOMA-IR at baseline while including copper, molybdenum, lead, and zinc to explore the potential non-linearity and interaction between those metals. Positive linear associations of log-transformed copper and zinc concentrations with log-transformed HOMA-IR at baseline and an inverse linear association between log-transformed molybdenum concentration and log-transformed HOMA-IR at baseline were confirmed (Figure A.4A). The association between lead and baseline HOMA-IR was slightly diminished in the BKMR model. Further, there was little evidence for interactions between the metals given associations between each metal and HOMA-IR at baseline did not differ by varying quantiles of the other three metals (Figure A.4B).

Table 4 shows the associations of selected components of metal mixtures with baseline HOMA- β and its rate of change in the AENET models. After adjusting for all potential confounders, a one SD increase in urinary metal concentration was associated with 1.59% (95%CI: −3.63%, 0.50%) lower baseline HOMA- β level for arsenic, and 2.66% (95%CI: −5.07%, −0.30%) lower level for zinc, respectively. In contrast, a one SD increase in urinary cobalt concentration was associated with 2.22% (95%CI: −0.10%, 4.60%) higher HOMA-β at baseline. HOMA-β declined during the follow-up (−1.00% annually, 95%CI: −1.02%, −0.90%). Urinary arsenic concentration was associated with faster rate of decline in HOMA-β, such that a one SD increase in urinary arsenic concentration was associated with 0.02% more rapid decline (95%CI: −0.05%, 0%) in HOMA-β annually. Beta coefficients of selected non-zero predictors in AENET models are shown in Table A.4. Predicted HOMA-β levels over the follow-up based on coefficients in these two AENET models are shown in Figure 2.

We constructed ERS of HOMA-β at baseline using estimated weights for arsenic, cobalt, and zinc from the AENET model (Table A.4). The ERS ranged from −0.1 to 0.1 with a mean (SD) equal to 0 (0.03). Lower ERS for HOMA-β at baseline indicates lower HOMA-β (greater β-cell dysfunction) at baseline attributable to higher concentrations of arsenic and zinc but lower concentration of cobalt. After adjusting for confounders, women in the lowest quartile of ERS, on average, had 8.96% (95% CI: −13.89%, −3.77%) lower HOMA-β at baseline, compared to those in the highest quartile (Table 5). The ERS effect size was stronger than those for individual metals (Table A.5). The ERS of rate of change in HOMA-β was not calculated because only arsenic was selected in the AENET.

We ran an BKMR model regressing HOMA-β at baseline with arsenic, cobalt, and zinc included in the model. Results from BKMR showed a positive linear association between log-transformed cobalt and log-transformed HOMA-β at baseline, and an inverse association between log-transformed zinc concentration and log-transformed HOMA-β at baseline (Figure A.5A). There was a potential interaction between arsenic and zinc given stronger inverse associations between arsenic and baseline HOMA-β were observed when urinary zinc concentrations were fixed at lower values (Figure A.5B). No interactions were detected between other metals.

In the sensitivity analysis, 186 women who had a fasting glucose level ≥ 100 mg/dL or a HOMA-IR ≥ 4.2 at the SWAN-MPS baseline were excluded. After exclusion of these participants, we still observed a positive association between urinary zinc concentration and HOMA-IR; a one SD increase in urinary zinc concentration was associated with 3.97% (95% CI: 1.59%, 6.40%) higher HOMA-IR at baseline and 0.08% (95% CI: −0.01%, 0.18%) increase in the annual rate of increase in HOMA-IR, respectively (Table A.6). Similarly, urinary zinc concentration was inversely associated with HOMA-β at baseline, a one SD increase in urinary zinc concentration was associated with 0.94% (95% CI: −3.17%, 1.34%) lower HOMA-β at baseline (Table A.7).

Urinary arsenic was associated with a lower baseline level and a faster rate of decrease in HOMA-β in the primary analysis. In the sensitivity analysis, the association between urinary arsenic and HOMA-β was evaluated in a subpopulation of 371 participants with seafood intake less than 1 time/week. In this subpopulation, similar associations were observed as in the primary analysis, however, the confidence intervals of arsenic’s effects became wider (Table A.8), possibly due to a reduced statistical power in accordance with the smaller sample size in this analysis.

Coefficients of urinary copper and lead in relation to HOMA-IR at baseline were shrunk to zeros in the AENET model without adjustment for BMI at baseline (Table A.9). Similar findings were observed for rate of change in HOMA-IR, HOMA-β at baseline, and rate of change in HOMA-β in AENET models without adjustments for BMI at baseline (Table A.9 and Table A.10).