This analysis is a cross-sectional exposome-wide association study (ExWAS) of data pooled across three cycles of NHANES (Figure 1). Since 1999-2000, NHANES is conducted biannually by the National Center for Health Statistics (NCHS) and is nationally representative of non-institutionalized civilians in the United States.¹¹ Data are collected through questionnaires, clinical measurements, and biomarker measurements, including chemical concentrations. Every two years, data on independent samples of participants are released in cycles. Cognition in adults aged 60 years and older was assessed in four cycles: 1999-2000, 2001-2002, 2011-2012, and 2013-2014. We included three of these cycles (1999-2000, 2011-2012, and 2013-2014; N = 29 896). We excluded the 2001-2002 cycle because a key covariate, fish consumption, was not measured in older adults in that cycle. Raw data are publicly available through the NCHS (https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx). Processed data are publicly available through Kaggle, figshare, and Hugging Face repositories.^12-15 Participants provided informed consent at the time of participation in NHANES. The University of Michigan Institutional Review Board (IRB) approved these secondary data analyses (HUM00116291).

NHANES participants provided urine and blood samples in the Mobile Examination Centers. These samples were analyzed for a variety of biomarkers used to indicate exposures. Information on the laboratory methods used for measuring chemical biomarker concentrations in each cycle year are available through the NCHS.¹⁶ The chemical biomarkers were not all measured in the same cycles nor in the same participants. Across the three cycles of interest, 395 environmental biomarkers were eligible for analysis. The environmental biomarker data were quality controlled similar to our previously described process.^17,18 Urinary cadmium measures that had interference with urinary moldybdenum were removed.^19,20 All measurements below the chemical-specific LOD were imputed as LOD/√2 by NHANES. No outliers were removed from the chemical exposure biomarkers. Consistent with prior analyses, chemical measures were log₂ transformed for analysis.¹⁸ After transformation, the measures were also standardized (mean = 0, standard deviation = 1) within each chemical to enable comparisons between chemicals with different units and distributions.

Cognitive testing via the Digit Symbol Substitution Test (DSST) was performed among participants aged 60 years and older.²¹ The DSST consists of a set of symbols corresponding to the numbers 1 through 9. Participants are given a paper with rows of numbers and are asked to draw the matching symbols beneath each number. The test is scored based on the number of symbols that were drawn correctly in the allotted 120s. In NHANES, the maximum possible score was 133 points.²² Our primary analysis used DSST as a continuous measure. As a sensitivity analysis, consistent with previous studies, we categorized DSST score of less than or equal to the survey-weighted 25th percentile (score ≤ 37 in our sample) as mild cognitive impairment (MCI) and DSST greater than the 25th percentile as normal cognition.²³

Covariates were selected based on prior literature and on conceptual framing as confounders (common causes) of the relationship between environmental chemical exposures and cognitive status, as visualized in a directed acyclic graph (DAG, Figure 1). Variables collected by interview included age (years), sex (male, female), race/ethnicity (Mexican American, Other Hispanic, Non-Hispanic White, Non-Hispanic Black, Other Race/Multiracial), education (less than 9th grade, 9-12th grade without diploma, high school graduate/General Education Development degree, some college or Associate of Arts degree, college graduate or above), history of tobacco use (never, ever smoked at least 100 cigarettes, current use), history of alcohol consumption (consumption in the past 12 months), fish and shellfish consumption (number of times eaten in the past 30 days), and the cycle year of participation. We chose to adjust all models for fish consumption due to its association with levels of certain metals and per- and polyfluoroalkyl substances (PFAS) as well as cognitive function.^24,25

To protect participant anonymity, NHANES top-coded ages greater than 85 years in cycles prior to 2007 as age 85 years. Those who reported an age greater than 80 years in 2007 or later were top-coded as age 80 years. Education was recategorized as a binary variable: less than high school and high school/General Education Development (GED) or above. Never smokers were categorized as smoking fewer than 100 cigarettes, former smokers were categorized as smoking over 100 cigarettes but not currently, and current smokers were categorized as reporting current use. The units for alcohol consumption were scaled to the average number of drinks per month based on the past 12 months. The number of fish and shellfish eaten over 30 days were combined into one value and categorized into three groups: 0, 1-3, and 4+. We considered the following categories as reference groups: male, Non-Hispanic White, less than high school, no fish consumption, never smoker, and the first survey cycle with measurements for each chemical.

Waist circumference (cm) was measured in the Mobile Examination Centers by trained study staff. Urinary creatinine concentrations (mg/dL), serum creatinine concentrations (mg/dL), and serum cotinine concentrations (ng/mL) of the participants were also measured from provided samples. Similar to the exposure measures, urinary creatinine and serum cotinine were log₂ transformed and standardized. For chemicals measured in urine, we used urinary creatinine as a covariate to account for urine dilution.²⁶ Cotinine was included as a covariate for all chemicals other than smoking-related chemicals (eg, hydroxycotinine and total cotinine) as a biomarker of short-term cigarette smoke exposure to account for residual confounding not addressed by self-reported smoking status. Estimated glomerular filtration rate (eGFR) was calculated from serum creatinine measurements using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation, excluding the term for race adjustment.^12,27,28 We categorized normal kidney function as eGFR ≥60mL/min/1.73 m².²⁹

All analyses were performed in R version 4.3.0³⁰ and code files are available (https://github.com/bakulskilab/Cognition_ExWAS). To account for the NHANES complex sampling design, we applied survey weights to all analyses using the R package ‘survey’, unless otherwise noted.³¹

Participants were excluded from cognitive testing by NHANES if they were aged less than 60 years at the time of assessment or by our study team if they had no chemical biomarker measures. To focus on environmental exposures relevant to the general population, biomarkers were excluded if they were annotated as dietary components or phytoestrogens, or if they were measured only in smokers. Next, for any lipid-soluble biomarkers measured in blood, we used the lipid-adjusted measurements and excluded the unadjusted versions. Lastly, we excluded biomarkers if more than 50% of measures were below the reported lower limit of detection (LOD). In total, 147 chemical biomarkers were included as exposures to be screened for potential associations with cognitive function. Participant and chemical biomarker inclusion and exclusion were visualized using flow charts. The distributions of continuous covariates were described by mean and standard deviation, while categorical covariates were described by proportions. We compared distributions of descriptive characteristics for the included and excluded participants. Within the included analytic sample, we compared the distributions of covariates by NHANES cycle and by cognitive impairment status.

We calculated summary statistics for each chemical biomarker (participant count, minimum and maximum concentrations, weighted arithmetic and geometric means and standard deviations, weighted median, weighted 25th and 75th percentiles, and percentage of measurements above the LOD). We visualized distributions of a subset of chemical exposures, log₂ transformed and standardized, with histograms. We used a scatterplot to compare the sample size per chemical versus the percentage of measurements above the LOD, colored by chemical family. Additionally, we calculated Spearman correlations between each chemical biomarker measure and used a heatmap to depict the correlations and missing data, grouped by chemical family.

We assessed the distributions of covariates based on missing outcome status, and determined values were missing at random (MAR). We imputed missing covariate and outcome data using multiple imputation by chained equations (MICE) with the R package “mice,” creating m = 5 complete datasets.³² Due to the variation in sample sizes for the chemical exposure measures, we did not impute exposure data. Additionally, we did not impute eGFR to maintain a consistent division between normal and abnormal kidney function. Characteristics of the imputed variables are summarized in a descriptive table (Supplementary Table 1).

To test the associations between individual chemical exposures and continuous DSST scores, for each chemical we used separate survey-weighted linear regression models adjusted for age (continuous), sex (categorical), race/ethnicity (categorical), education level (categorical), urinary creatinine (continuous), serum cotinine (continuous), reported smoking status (categorical), seafood consumption (categorical), and survey cycle (categorical). Consistent with an ExWAS design, each chemical association was analyzed individually in a separate regression model and we did not consider hypotheses about the direction of effect.¹⁰ Of note, because biomarkers were measured in different subsets of the analytic sample, these regression models had varying sample sizes. Urinary creatinine was included only in models for urinary chemical measurements. Models for smoking-related compounds (serum cotinine, total cotinine, hydroxycotinine, and thiocyanate) were not adjusted for cotinine. To account for the MICE procedure, models were run separately on each of the five unique imputed datasets and pooled to produce one set of coefficient estimates per exposure.³³ The chemical biomarker regression coefficients, 95% confidence intervals (CI), and P-values for each exposure were compiled into a table. Because the chemical concentrations were log₂ transformed and standardized prior to modeling, the regression coefficients are interpreted as the adjusted association with DSST per standard deviation of the log₂-scaled chemical concentration. For discovery analyses, we prioritized chemicals with an unadjusted P-value <.01. To account for multiple comparisons, we controlled the false discovery rate (FDR) using the Benjamini-Hochberg method with the “p.adjust” function in the stats R package, and used a significance threshold of FDR <0.05.^30,34 We visualized the regression coefficients and statistical significance for all chemicals using a volcano plot and a forest plot. For the exposure associations with DSST, the coefficient estimates and confidence intervals for all are presented in tables.

We developed five sensitivity analyses to further explore the associations between chemical biomarkers and DSST. First, decreased kidney function is associated with both chemical biomarker concentrations retained in the body and lower cognitive function.^35-37 We subset the sample to participants with normal kidney function as indicated by eGFR ≥60 mL/min/1.73 m² and re-ran the models for each chemical. Second, we ran models stratified by sex to explore differences between males and females. Third, to account for possible confounding by lifestyle characteristics, in the full sample, we further adjusted models for waist circumference (continuous) and alcohol consumption (categorical). These covariates were addressed in a sensitivity analysis rather than the main analysis to prevent potential overadjustment in the main models. Fourth, to present a more clinically relevant analysis, we used MCI status categories as a binary outcome variable. We conducted modified Poisson regression for each chemical, adjusted for the same covariates as the main analysis and reported the relative risks (RR) and 95% CI. Lastly, we again restricted to participants with normal kidney function and re-ran the Poisson regression models. The results for all sensitivity analyses were visualized in volcano plots and the regression output was compiled in tables.

The analysis included N = ,982 participants across three NHANES cycles (Figure 2A). Within the included sample, the mean age was 69.8 years, 55% were female, and 78.2% were Non-Hispanic White (Table 1). On average, participants from cycle year 1999-2000 had lower observed DSST scores compared to participants from cycle years 2011-2014 (Table 1). The included participants were on average more likely to be Non-Hispanic White and had a higher DSST score compared to the excluded participants aged 60 years and older (Supplementary Table 2). Participants with MCI were older, less likely to be Non-Hispanic White, and less likely to have completed high school compared to participants without impairment (Supplementary Table 3).

Our analytic sample included 147 chemicals from 14 chemical families (Figure 2B). Descriptive statistics including concentration minimums and maximums, means, standard deviations, medians, interquartile ranges, and the percentage of measurements above the LOD are available for all 147 chemicals (Supplementary Table 4). Serum cotinine was measured in the most participants (n = 4687), followed by blood cadmium and blood lead (n = 3955 each). Benzaldehyde had the fewest number of measurements with n = 377. Of the 147 included chemicals, 90 had detection rates of 90% or higher (Supplementary Figure 1). A heatmap of Spearman correlations is depicted in Figure 3. Polychlorinated biphenyls, polyaromatic hydrocarbons, and smoking-related compounds were highly correlated within their families. Histograms depicting the distributions of nine selected exposures, log₂ transformed and standardized, are available in Supplementary Figure 2.

In our primary adjusted regression analysis, five of 147 exposures were associated (p < 0.01) with DSST score (Figure 4). The identified exposures included chemicals in the classes of metals, personal care compounds, smoking-related compounds, and volatile organic compounds. Increases in the concentrations of serum cotinine and urinary tungsten were associated with lower DSST scores (Table 2). A standard-deviation increase in the log₂-transformed concentration of serum cotinine was associated with 2.71 points lower DSST (95% CI: −3.69, −1.73). Similarly, a standard-deviation increase in the log₂-transformed concentration of urinary tungsten was associated with 1.34 points lower DSST (95% CI: −2.11, −0.56). An increase in three identified exposures (urinary benzophenone-3, blood m-/p-xylene, and urinary 2-methylhippuric acid) were associated with higher DSST scores (Table 2). Two exposures, serum cotinine and blood m-/p-xylene, had FDR-adjusted P-values <.05. Results for all screened exposures, regardless of statistical significance, are available in Supplementary Table 5 and as a forest plot (Supplementary Figure 3).

When restricting the analytic group to participants with normal kidney function (N = 3593), three chemical biomarkers were associated (P < .01) with DSST score (Supplementary Figure 4 and Supplementary Table 6). All identified associations were consistent with the results of the main analysis.

In sex-stratified models, associations differed between males and females. Of the five exposures associated (P < .01) with cognition in the primary analysis, four were also associated in at least one sex group, with serum cotinine being the only shared exposure across groups at P < .01 (Supplementary Figure 5 and Supplementary Table 7).

When adjusting further for waist circumference and alcohol consumption in the full analytic sample, the previously identified associations between blood mercury and 2-methylhippuric acid and DSST were no longer significant (Supplementary Figure 6 and Supplementary Table 8). Notably, an increase in urinary cadmium concentration was associated with a decrease in DSST score by 2.72 points (95% CI: −4.27, −1.18).

After MICE, the weighted 25th percentile of DSST score defined 37 as the cutoff for MCI. A total of five biomarkers were associated (P < .01) with MCI in modified Poisson regression models (Figure 5 and Table 3). Consistent with the continuous results, serum cotinine was associated with higher prevalence of MCI (RR 1.23, 95% CI: 1.14, 1.33). Four biomarkers were associated with lower prevalence of mild cognitive impairment: 2-methylhippuric acid (RR 0.81, 95% CI: 0.71, 0.93), m-/p-xylene (RR 0.80, 95% CI: 0.71, 0.90), perfluorooctanoic acid (PFOA) (RR 0.83, 95% CI: 0.73, 0.94), and zinc (RR 0.84, 95% CI: 0.75, 0.94). Poisson regression results for all exposures are available in Supplementary Table 9.

When restricted to participants with normal kidney function, serum cotinine continued to be associated (P < .01) with MCI (RR 1.24, 95% CI: 1.10, 1.39). Full results are available in Supplementary Figure 7 and Supplementary Table 10.