Developing an Exposure Burden Score for Chemical Mixtures Using Item Response Theory, with Applications to PFAS Mixtures

Abstract

Background:

There are few existing methods to quantify total exposure burden to chemical mixtures, independent of a health outcome. A summary metric could be advantageous for use in biomonitoring, risk assessment, health risk calculators, and mediation models.

Objective:

We developed a novel exposure burden score method for chemical mixtures, applied it to estimate exposure burden to per- and polyfluoroalkyl substances (PFAS) mixtures, and estimated associations of PFAS burden scores with cardio-metabolic outcomes in the general U.S. population.

Methods:

We applied item response theory (IRT) to biomonitoring data from 1,915 children and adults 12–80 years of age in the 2017–2018 National Health and Examination Survey to quantify a latent PFAS burden score, using serum concentrations of eight measured PFAS biomarkers, each considered an “item.” The premise of IRT is that through using both information about a participant’s concentration of an individual PFAS biomarker, as well as their exposure patterns for the PFAS mixture, we can estimate the participant’s latent PFAS exposure burden, independent of a health outcome. We used linear regression to estimate associations of the PFAS burden score with cardio-metabolic outcomes and compared our findings to results using summed PFAS concentrations as the exposure metric.

Results:

PFAS burden scores and summed PFAS concentrations had moderate-high correlation ($\mathrm{\rho }=0.75$). Isomers of PFOS [$n$-perfluorooctane sulfonic acid ($n$-PFOS) and perfluoromethylheptane sulfonic acid isomers (Sm-PFOS)] were the most informative to the PFAS burden scores. PFAS burden scores and summed PFAS concentrations were both significantly associated with cardio-metabolic outcomes, but associations were generally closer to the null for summed PFAS concentrations vs. the PFAS burden score. Adjusted associations (95% CIs) with total cholesterol (in milligrams per deciliter) were 8.6 (95% CI: 5.2, 11.9) and 2.4 (95% CI: 0.5, 4.2) per interquartile range increase in the PFAS burden score and summed concentrations, respectively. Sensitivity analyses showed similar associations with cardio-metabolic outcomes when only a subset of PFAS biomarkers was used to estimate PFAS burden. In a validation study, associations between PFAS burden scores and cholesterol were consistent with primary analyses but null when using summed PFAS concentrations.

Discussion:

IRT offers a straightforward way to include exposure biomarkers with low detection frequencies and can reduce exposure measurement error. Further, IRT enables comparisons of exposure burden to chemical mixtures across studies even if they did not measure the exact same set of chemicals, which supports harmonization across studies and consortia. We provide an accompanying PFAS burden calculator (https://pfasburden.shinyapps.io/app_pfas_burden/), enabling researchers to calculate PFAS burden scores based on U.S. population exposure reference ranges. https://doi.org/10.1289/EHP10125

Introduction

Most existing statistical methods to elucidate health effects of environmental mixtures are supervised,^1–5 meaning that they study associations between individual exposures or chemical mixtures in the context of a prespecified health outcome. However, few statistical tools are available to summarize exposure burden to a chemical class, independent of a health outcome. A single summary measure to represent total (or underlying) exposure burden is commonplace in other fields, such as polygenic risk scores in genetics, which provide a summary metric for dozens of genes^6,7; depression severity scores in psychology, which are based on several reported symptoms^8,9; and allostatic load scores in medicine, which provide a summary metric for physiological dysregulation via a multitude of biomarkers.^10,11 Here, we create a summary metric that represents total exposure burden to chemical mixtures, which may be advantageous for use in biomonitoring, risk assessment, disease risk calculators, or mediation models. Specifically, we conceptualize an exposure burden score as a latent variable that captures the totality of similar exposures within a chemical class.

We first introduce item response theory (IRT) as a novel method to develop exposure burden scores for chemical mixtures. IRT is a set of well-established latent variable methods developed in the educational testing literature for test scoring (e.g., scoring college entrance exams) and have more recently been applied to patient reported outcomes measurement,^12–22 clinical assessment²³ and epigenetic data.²⁴

One major contribution of IRT is that it enables comparisons of exposure burden to chemical mixtures across studies even if they did not measure the exact same set of chemicals. It is common for studies to measure slightly different sets of chemicals within a particular chemical class, with only a subset of overlapping chemicals. Using conventional approaches to summarize exposure burden, such as summing the concentrations of individual exposure biomarkers, we would be limited to analyzing only the common set of chemicals assayed in all studies. Using IRT, we can instead standardize measures of exposure burden across studies, so that the burden scores are on the same scale and can be compared across studies even if the studies did not measure exactly the same set of chemicals within that chemical class. This feature supports harmonization across studies, consortia, and meta-analyses by enabling larger sample sizes to detect smaller effect sizes or to make inferences across a more heterogenous population.

Second, IRT offers a straightforward way to include exposure biomarkers with low detection frequencies, by allowing for different numbers of categories for different biomarkers. For example, deciles could be used for frequently detected biomarkers, and quartiles could be used for less frequently detected biomarkers. This is known as mixed item types in IRT because of the different numbers of categories for different items (“biomarkers”).

Last, IRT can reduce exposure measurement error by considering both a participant’s biomarker concentrations and their exposure patterns to estimate exposure burden to mixtures. Thus, the underlying burden can be estimated with less measurement error than by using the individual components of the mixture.^25,26 Because IRT takes the pattern of exposure concentrations into account when estimating burden scores, not just the individual chemical concentrations, this can reveal subtle differences in exposure burden not seen with summed concentrations. Summed concentrations can also be sensitive to outliers.

For illustration, we used recent data from the National Health Nutrition and Examination Survey (NHANES) to create an exposure burden score of per- and polyfluoroalkyl substances (PFAS) mixtures, with an item being an individual PFAS biomarker. PFAS, defined as chemicals with at least one aliphatic perfluorocarbon moiety, are a class of $>\mathrm{4,000}$ persistent chemicals that bioaccumulate in the environment and human tissues.²⁷ Their wide use as surfactants and water repellents in industrial applications and consumer products lead to human exposure through contaminated food, water, air, consumer products, and house dust.²⁸ PFAS have been linked to a variety of health outcomes, such as elevated cholesterol,²⁹ and changes in hormone^30,31 and immune system functioning.³² Importantly, the general approach for PFAS management has been to treat them as a single chemical class.^33,34

Here, we applied IRT to estimate PFAS exposure burden using recent, nationally representative U.S. biomonitoring data. We then estimated associations of the IRT-derived PFAS burden score with several cardio-metabolic outcomes: total cholesterol, high-density lipoprotein (HDL), non-HDL, low-density lipoprotein (LDL), triglycerides, C-reactive protein, and systolic and diastolic blood pressure. Serum lipids were chosen because serum PFAS concentrations have consistently been associated with higher lipid levels.^35–37 We included blood pressure because there have been mixed findings of associations between PFAS concentrations and blood pressure.^38,39 Last, studies have found inverse associations of PFAS concentrations with C-reactive protein and that PFAS exposure can suppress immune responses.^40,41 Further, these biomarker outcomes also reduce the potential for reverse causality as compared with using clinical diagnosis of metabolic disease, for which the onset may have occurred many years in the past. Because previous work has used summed concentrations of individual PFAS biomarkers⁴² to represent a summary metric of exposure burden to PFAS mixtures, we also included associations of the summed PFAS concentrations with cardio-metabolic outcomes for comparison.

Methods

Data Set

We analyzed publicly available NHANES 2017–2018 data, which are provided by the National Center for Health Statistics (NCHS) in the Centers for Disease Control and Prevention (CDC). The NHANES is a recurring cross-sectional survey of the noninstitutionalized civilian U.S. population living in the 50 states and the District of Columbia.⁴³ During each 2-y NHANES cycle, ∼5,000 (approximately 5,000) nationally representative adults and children are sampled across counties in the United States using a stratified, multistage, probability cluster design. The NHANES provides a large, nationally representative sample of the U.S. population, ensuring diversity in race/ethnicity by oversampling Hispanics, non-Hispanic Blacks, and non-Hispanic Asians. A one-third subsample of participants ($\ge 12\text{\hspace{0.17em}years}$ of age) was additionally selected for measurement of PFAS concentrations in serum ($n=\mathrm{2,133}$). After excluding those with missing PFAS measurements ($n=204$) and women with a positive urine pregnancy test ($n=14$), $n=\mathrm{1,915}$ participants were included in our study sample.

Serum concentrations of nine biomarkers of seven PFAS were quantitated: perfluorodecanoic acid (PFDeA), perfluorohexane sulfonic acid (PFHxS), 2-($N$-methylperfluoroctanesulfonamido) acetic acid (Me-PFOSA-AcOH), perfluorononanoic acid (PFNA), perfluoroundecanoic acid (PFUA), $n$-perfluorooctanoic acid ($n$-PFOA), branch perfluorooctanoic acid isomers (Sb-PFOA), $n$-perfluorooctane sulfonic acid ($n$-PFOS), and perfluoromethylheptane sulfonic acid isomers (Sm-PFOS).⁴⁴ The NHANES provided imputed concentrations for values below the detection limit using the limit of detection divided by the square root of 2.²¹ Detection frequencies were 10.0% for Sb-PFOA, 59.0% for Me-PFOSA-AcOH, 65.9% for PFUA, 88.7% for PFDeA, 92.5% for PFNA, and $>99\%$ for PFHxS, $n$-PFOA, $n$-PFOS, and Sm-PFOS.

Several cardio-metabolic biomarkers were analyzed: serum total cholesterol (in milligrams per deciliter), HDL (in milligrams per deciliter), non-HDL (in milligrams per deciliter), triglycerides (in milligrams per deciliter), systolic and diastolic blood pressure (in millimeters of mercury), and high-sensitivity C-reactive protein (in milligrams per deciliter). Non-HDL was calculated by subtracting HDL cholesterol from total cholesterol. We used triglycerides measured from the standard biochemistry profile.

Covariates were selected based on previous studies as well as on a directed acyclic graph (DAG; Figure S1),^35–37 and included sex, age (in years), race/ethnicity (using categories prespecified by the NHANES,⁴⁵ including Mexican American, other Hispanic, non-Hispanic White, non-Hispanic Black, non-Hispanic Asian, other/multiple race) as a proxy for racism/structural inequity,⁴⁶ ratio of family income to poverty, and body mass index (BMI, in kilograms per meter squared). The family income-to-poverty ratio was calculated by dividing reported family income by poverty guidelines established by the Department of Health and Human Services. These guidelines were specific to the 2017–2018 survey year, family size, and state (the 48 contiguous states and District of Columbia have the same set of guidelines, whereas Alaska and Hawaii have unique guidelines).⁴⁷ Further, the value of family income-to-poverty ratio was not calculated if the respondent only reported income as $<\$\mathrm{20,000}$ or $\ge \$\mathrm{20,000}$ or if income data were missing. Otherwise, the midpoint of the range of reported income was used to calculate the ratio. Last, ratios $\ge 5.0$ were coded as 5.0 because of disclosure sensitivity.⁴⁵

The procedures and protocol in the NHANES were approved by the CDC/NCHS Research Ethics Review Board. This study was exempted from review by the Icahn School of Medicine at Mount Sinai’s institutional review board (STUDY-21-00654).

Quantifying Latent Mixture Exposure Burden Using IRT

Here, we describe an approach to calculating latent exposure burden to an environmental mixture using an illustrative example of PFAS mixtures. We used IRT,^{12–22,48,49} to understand how the set of assayed PFAS biomarkers provides information about the latent (i.e., underlying) exposure burden to PFAS. IRT-derived scores represent estimates of the latent PFAS burden and differentially weight the contributions of the individual biomarkers based on their data-driven parameters, with a nonlinear relationship between levels of individual biomarkers and the latent burden. IRT is a flexible modeling approach with several similarities to confirmatory factor analysis and attempts to describe and elucidate the relations of measured items (e.g., PFAS biomarker levels) as indicators for an unobservable latent trait (e.g., latent PFAS burden). In an IRT framework, the probability of an individual endorsing an item (e.g., having a higher PFOA level) is a function of the latent trait (e.g., PFAS burden) and item-specific parameters. In addition, the probability of endorsing an item increases monotonically with the level of the trait.

In this study, we used graded response IRT models (GRMs), which account for ordinal polytomous data, so that we could combine items with different numbers of response categories.⁵⁰ Specifically, we first dichotomized infrequently detected PFAS biomarkers (Me-PFOSA-AcOH, PFUA, Sb-PFOA) as not detected vs. detected, and we used quartiles for the other six frequently detected PFAS biomarkers (PFDeA, PFHxS, PFNA, $n$-PFOA, $n$-PFOS, Sm-PFOS). We constructed our GRM as follows:

Suppose that there are $m=1,\dots ,M$ PFAS biomarkers. Each ordinal PFAS biomarker can take on $K_m$ possible values (categories), allowing for differing numbers of quantiles for different PFAS biomarkers (mixed item types). For example, we can have a $K_m=2$ category ordinal PFAS biomarker, with $k=1$ representing nondetect and $k=2$ representing detect. Or we can have a $K_m=4$ category biomarker, with $k=1$ representing the PFAS biomarker concentration $<25\text{th}$ percentile, $k=2$ representing the 25th–50th percentile, $k=3$ representing the 50th–75th percentile, and $k=4$ representing the $>75\text{th}$ percentile. For a given subject i, $x_{i,m}$ represents their polytomous level on the $m\text{th}$ PFAS biomarker, such that $x_{i,m}$ can take on $K_m$ possible values. ${\mathrm{\theta }}_i$ denotes the latent PFAS burden for individual i. The GRM relates an individual’s observed level of a PFAS biomarker $x_{i,m}$ to an individual’s latent PFAS burden ${\mathrm{\theta }}_i$ via the following formulae.

The probability of scoring at or above the given response option, $k\ge 2$, given the level of theta is

$$\mathrm{Pr}(x_{i,m}\ge k|{\mathrm{\theta }}_i)=\frac{\text{\hspace{0.17em}exp\hspace{0.17em}}\left\{{\mathrm{\alpha }}_m\left({\mathrm{\theta }}_i-{\mathrm{\beta }}_{m,k}\right)\right\}}{1+\text{\hspace{0.17em}exp\hspace{0.17em}}\left\{{\mathrm{\alpha }}_m\left({\mathrm{\theta }}_i-{\mathrm{\beta }}_{m,k}\right)\right\}}\mathrm{where}{\mathrm{\beta }}_2<{\mathrm{\beta }}_3<\dots <{\mathrm{\beta }}_{K_m}.$$

The probability of scoring at the current response option, k, is

$$\mathrm{Pr}(x_{i,m}=k|{\mathrm{\theta }}_i)=\mathrm{Pr}\left(x_{i,m}\ge k\right)-\mathrm{Pr}\left(x_{i,m}\ge k+1\right).$$

In addition, $\mathrm{Pr}(x_{i,m}\ge 1|{\mathrm{\theta }}_i)=1$ and $\mathrm{Pr}(x_{i,m}=K_m|$${\mathrm{\theta }}_i)=\mathrm{Pr}(x_{i,m}\ge K_m|$${\mathrm{\theta }}_i)$, given that the probability of responding in any categories above the largest $\left(K_m\right)$ is null. ${\mathrm{\theta }}_i$ represents the latent PFAS burden and is assumed to follow a standard normal distribution, with a mean of 0 and a standard deviation of 1. ${\propto }_m$ denotes the discrimination parameter for PFAS biomarker m, which indicates how well the biomarker can distinguish between participants with very similar latent PFAS burdens. Here, we fit the unconstrained GRM, which allows for a different ${\propto }_m$ per PFAS biomarker. The discrimination parameter quantifies how well the biomarker differentiates between participants with higher and lower levels of latent PFAS burden. It refers to how informative having a detected PFAS biomarker is with respect to gauging the participant’s latent PFAS burden level. Biomarkers with high discrimination parameters provide more information about latent PFAS burden differences, whereas biomarkers with low discrimination parameters do not provide much information and may not need to be included in the scale. Item discrimination estimates can theoretically range from negative infinity to infinity; however, negative discrimination estimates need to be further examined, given that the probability of having a higher observed level of PFAS biomarker should not decrease as a participant’s latent PFAS burden increases.

${\mathrm{\beta }}_{m,k}$ denotes the extremity parameter corresponding to PFAS biomarker m for the boundary between responses k and $k\text{-}1$, but it is on a cumulative probability scale, so its interpretation is not direct. The extremities are on the same scale as ${\mathrm{\theta }}_i$, that is, a standard normal scale. For a biomarker m, there will be $(K_m-1)$ extremities, which represent the cut points between the $K_m$ biomarker quantiles. In the case of a binary (detect/nondetect) biomarker, the ${\mathrm{\beta }}_{m,2}$ extremity refers to how high an individual’s latent PFAS burden needs to be to have a 0.5 probability of having a particular PFAS biomarker detected. In the case of an ordinal biomarker, the extremity refers to how high an individual’s latent PFAS burden needs to be to have a 0.5 probability of having that particular observed quantile of exposure, or a higher observed quantile of exposure, on that specific biomarker. Specifically, if a biomarker is coded into quartiles, there will be three extremity parameters. The first extremity, ${\mathrm{\beta }}_{m,2}$, is the latent trait score at which people have a 50/50 chance of having exposure 1 vs. 2 or 3 or 4 for that biomarker. The second extremity, ${\mathrm{\beta }}_{m,3}$, is the latent trait score at which people have a 50/50 chance of having exposure 1 or 2 vs. 3 or 4 for that biomarker; the third extremity, ${\mathrm{\beta }}_{m,4}$, is the latent trait score at which people have a 50/50 chance of having exposure 1 or 2 or 3 vs. 4 for that biomarker.

The marginal likelihood for the responses of the ith subject is denoted by

$$l_i\left(\propto ,\beta \right)=\log \hspace{0.17em}p(x_i;\propto ,\beta )=\text{log}{{\displaystyle \int }}^{}p(x_i|{\theta }_i;\hspace{0.17em}\propto ,\beta )p({\theta }_i)d{\theta }_i.$$

The model is fit using marginal maximum likelihood estimation, with numerical integration using the Gauss–Hermite quadrature rule with 21 quadrature points because there is not a closed form solution for the integral. Models were fit using the R package ltm⁵¹ in R (R Development Core Team).

We estimated an expected a posteriori estimate ${\widehat{\mathrm{\theta }}}_m$, which is quantified as the most plausible value of the latent PFAS burden for subject i, given subject i’s levels of PFAS biomarkers ($x_i$); a standard normal prior distribution, $g\left(\mathrm{\theta }\right)$; and the GRM, as follows:

$${\widehat{\mathrm{\theta }}}_i={{\displaystyle \int }}^{}\theta p(\theta \text{\hspace{0.17em}|\hspace{0.17em}}{x}_i;\propto ,\beta \text{)}g\left(\theta \right)d\theta .$$

In IRT, because the measurement precision can differ across levels of the latent trait, PFAS burden, we plot the standard error of measurement across PFAS burden ($\mathrm{\theta }$) values. IRT allows each of the items (i.e., biomarkers) to be examined graphically. One benefit of using an IRT approach to estimate an exposure burden score is that it is possible to not only organize individuals along an exposure burden continuum, but one can easily determine which biomarkers are more able to discern who is likely to be at the high end of the exposure burden and who is not. Using the item information curve (IIC) as an example of this feature, higher curves on the IIC plot mean that the particular biomarker contributes more to the IRT burden score. From this plot, we can also see the range of the burden score where information is provided. Thus, the IIC shows how well and precisely each biomarker measures the latent trait at various levels of the latent trait. Certain biomarkers may provide more information at low levels of the latent trait, whereas others may provide more information at higher levels of the latent trait.

The test information curve (TIC) aggregates the IICs across all the items. It tells us how well the test measures the latent trait at various levels of the latent trait. Ideally, we would want this curve to peak at about the mean of the sample because that is where the most individuals would be. We would also want this to be high across a range of theta values, so that we have good precision of the burden score across the range of the attribute.

We used the following approach to discretize continuous PFAS biomarker data. Survey-weighted quantile cutoffs were used for biomarkers that were discretized into three or more levels. If a biomarker was analyzed based on whether it was detected or not, then survey-weights were not used. We used deciles for more frequently detected ($>10\%$ detection) PFAS biomarkers to allow for finer resolution of biomarker concentrations and binary (detect vs. nondetect) for infrequently detected PFAS biomarkers ($\le 10\%$ detection). For a biomarker, if the survey-weighted deciles for adjacent deciles were the same, then the adjacent categories were merged. Thus, each biomarker had at most 10 levels (e.g., categories), but there could be different numbers of levels for different biomarkers (e.g., mixed item types).

Because our IRT model assumes unidimensionality, in which there is assumed to be a single latent trait that underlies the observed variables, we used the following established rules to determine whether these biomarkers represent (i.e., load) onto a single factor: The first factor explains $>20\%$ of the variance, the ratio of the first eigenvalue to the second eigenvalue is $>4$, and we used visual inspection of the eigenvalue scree plot.⁵²

We have provided an accompanying PFAS exposure burden calculator, via an R Shiny app, available at https://pfasburden.shinyapps.io/app_pfas_burden/. The R code is available at that link, as well as at https://github.com/shelleyhliu/PFAS-burden-score (see also “R code for burden score calculator” in the Supplemental Material). Researchers can calculate PFAS exposure burden scores on our same scale, which is based on recent U.S. population exposure reference ranges. The NHANES data were extracted using the nhanesA R package.⁵³ We used the ltm⁵¹ package for the IRT analyses because we found it to be very accessible for those new to IRT, particularly for making IRT plots. We also fit these models with the mirt⁵⁴ software for R and found very similar PFAS burden score estimates.

Motivating Example: Associations of PFAS with Cardio-Metabolic Biomarkers

As an applied example of the utility of the IRT approach, we estimated the associations of the PFAS burden score with serum total cholesterol, non-HDL cholesterol, HDL cholesterol, triglycerides, C-reactive protein concentrations, and systolic and diastolic blood pressure, adjusted for covariates using linear regression. We calculated the PFAS burden using two approaches that differ on how they treat isomers of PFOS and PFOA. Our primary PFAS burden score includes concentrations for individual isomers of PFOA and PFOS (hereafter referred to as the PFAS burden score). We also provide an alternative way to calculate PFAS burden scores, using total PFOS and PFOA (i.e., summed isomers of these biomarkers). Some studies may not measure isomers of PFOS and PFOA, so this alternative IRT scoring algorithm was provided so that they can use these PFAS burden scores. These burden scores as presented as sensitivity analyses, henceforth referred to as the PFAS burden scores (summed isomers).

We used deciles to categorize PFAS biomarkers because that allows for finer resolution of biomarker concentrations. Regression models were not survey-weighted, given that the focus was on the PFAS burden score. Regression models for C-reactive protein and triglycerides used the natural log of those outcomes to account for skewness. We qualitatively compared the resulting effect estimates with those obtained via regression models with summed PFAS concentrations. The same set of covariates were used in all analyses and included age, sex, race/ethnicity as a proxy for racism and structural inequity,⁴⁶ BMI (assessed as a continuous variable), and socioeconomic status (SES; assessed via poverty-to-income ratio). Standard linear regression models were used to separately model the adjusted associations of PFAS burden scores, PFAS burden scores (summed isomers), and summed PFAS concentrations with cardio-metabolic biomarkers. Statistical significance was assessed at the $\text{alpha}=0.05$ level. For added interpretability of the beta coefficients of these models, we scaled the effect size in terms of an interquartile range (IQR) increase in PFAS burden (PFAS burden at the 75th percentile minus PFAS burden at the 25th percentile) or IQR increase in summed PFAS concentrations (summed PFAS concentrations at the 75th percentile minus summed PFAS concentrations at the 25th percentile).

Sensitivity Analysis to Demonstrate Utility of IRT Burden Score Calculator When Studies Do Not Measure the Exact Same Set of PFAS

Using 2017–2018 NHANES data, we conducted a sensitivity analysis to determine if associations with health outcomes were similar when the PFAS burden score did not include some of the assessed PFAS biomarkers. We estimated participants’ PFAS exposure burdens using the IRT burden calculator, but we used only their PFAS concentrations from a smaller set of PFAS (i.e., $n$-PFOS, Sm-PFOS, PFOA, PFNA, and PFHxS), which corresponded to setting all participants’ values for Me-PFOSA-AcOH, PFUA, and PFDeA to missing. Thus, this would be akin to a study where only five biomarkers were measured, given that concentrations Me-PFOSA-AcOH, PFUA, and PFDeA were set to missing for all participants. The associations with cardio-metabolic outcomes were compared with those estimated when all PFAS biomarker concentrations were used in the IRT burden calculator.

As an additional, related sensitivity analysis, we calculated PFAS burden scores from a study setting where the two most informative biomarkers with respect to estimating PFAS burden were not measured (corresponding to setting all the participant values for these excluded biomarkers to missing). Then, we compared health outcome associations for PFAS burden scores from the original model (i.e., all biomarkers included) to those that excluded the most informative biomarkers.

Secondary Analysis in a Validation Study

We used NHANES data from 2015–2016 as a validation sample to determine if we could replicate our main findings. In this sample, 2,170 participants ($\ge 12\text{\hspace{0.17em}years}$ of age) were selected for measurement of PFAS concentrations in serum. After excluding those with missing PFAS measurements ($n=177$), and women with a positive urine pregnancy test ($n=17$), 1,976 participants were included in our study sample. IRT burden scores for this sample were calculated using the IRT burden calculator derived from the NHANES 2017–2018 data so that PFAS burden scores for this sample could be placed on the same scale. Participants’ PFAS concentrations were discretized according to the survey-weighted 2017–2018 cutoffs, and they were used as the model inputs to calculate PFAS burden scores. In this secondary analysis, we assessed associations between PFAS burden scores, summed PFAS concentrations, and cardio-metabolic outcomes in 2015–2016 NHANES, adjusting for the same set of covariates as in the main study.

Results

Sample Characteristics

Sociodemographic characteristics of the sample are presented in Table 1. The study sample was predominantly non-Hispanic White (34.6%), with a median age of 46 y (IQR: 26–62), and $\sim 50\%$ female. Among the 1,915 participants, 15.4% were Mexican American, 9.2% were other Hispanic, 34.6% were non-Hispanic White, 22.2% were non-Hispanic Black, 13.3% were non-Hispanic Asian, and 5.3% were other race, including multiracial. Most PFAS biomarker concentrations were correlated with each other (Pearson’s $\mathrm{\rho }=-0.01–0.83$), with PFUA and PFDeA having the highest correlation ($\mathrm{\rho }=0.83$) and Sb-PFOA and $n$-PFOS having the lowest correlation ($\mathrm{\rho }=-0.01$) (Figure S2).

Table 1.

Sociodemographic characteristics of sample, NHANES 2017–2018 ($n=\mathrm{1,915}$).

Variable	% or median (IQR)	Missing ($n$)
Race/ethnicity	—	0
Mexican-American	15.4	—
Other Hispanic	9.2	—
Non-Hispanic White	34.6	—
Non-Hispanic Black	22.2	—
Non-Hispanic Asian	13.3	—
Other race, including multiracial	5.3	—
Sex	—	0
Male	49.2	—
Female	50.8	—
Age	46 (26–62)	0
Family income-to-poverty ratio	1.99 (1.15–3.92)	250
Body mass index (${\mathrm{kg}/\mathrm{m}}^2$)	28.0 (23.8–32.9)	24
Total cholesterol ($\mathrm{mg}/\text{dL}$)	179 (153–209)	4
non-HDL ($\mathrm{mg}/\text{dL}$)	125.0 (100.0–154.5)	4
LDL-C (Friedewald) ($\mathrm{mg}/\text{dL}$)	103.0 (83.0–127.0)	1,011
LDL-C (Martin-Hopkins) ($\mathrm{mg}/\text{dL}$)	104.0 (83.0–127.2)	1,011
LDL-C (NIH) ($\mathrm{mg}/\text{dL}$)	105.0 (84.0–129.0)	1,000
HDL cholesterol ($\mathrm{mg}/\text{dL}$)	50 (42–60)	4
Triglycerides (standard) ($\mathrm{mg}/\text{dL}$)	111.0 (77.0–163.0)	10
Triglycerides (reference) ($\mathrm{mg}/\text{dL}$)	87.0 (59.0–126.0)	999
C-reactive protein ($\mathrm{mg}/\mathrm{L}$)	1.67 (0.74–4.22)	6
Systolic BP (mmHg)	120.7 (110.0–133.7)	108
Diastolic BP (mmHg)	70.67 (63.33–78.67)	108
PFDeA ($\text{ng}/\mathrm{m}\mathrm{L}$)	0.20 (0.10–0.30)	0
PFHxS ($\text{ng}/\mathrm{m}\mathrm{L}$)	1.10 (0.60–1.80)	0
Me-PFOSA-AcOH ($\text{ng}/\mathrm{m}\mathrm{L}$)	0.10 (0.07–0.20)	0
PFNA ($\text{ng}/\mathrm{m}\mathrm{L}$)	0.40 (0.30–0.70)	0
PFUA ($\text{ng}/\mathrm{m}\mathrm{L}$)	0.10 (0.07–0.20)	0
$n$-PFOA ($\text{ng}/\mathrm{m}\mathrm{L}$)	1.30 (0.80–1.90)	0
Sb-PFOA ($\text{ng}/\mathrm{m}\mathrm{L}$)	0.07 (0.07–0.07)	0
$n$-PFOS ($\text{ng}/\mathrm{m}\mathrm{L}$)	2.90 (1.70–5.40)	0
Sm-PFOS ($\text{ng}/\mathrm{m}\mathrm{L}$)	1.20 (0.70–2.20)	0
PFOA ($\text{ng}/\mathrm{m}\mathrm{L}$)	1.37 (0.87–2.07)	0
PFOS ($\text{ng}/\mathrm{m}\mathrm{L}$)	4.20 (2.35–7.80)	0

Note: Percentages are presented for categorical variables; median and IQR are presented for continuous variables; statistics are not survey-weighted. —, not applicable; BP, blood pressure; HDL, high-density lipoprotein; IQR, interquartile range; LDL, low-density lipoprotein; Me-PFOSA-AcOH, 2-($N$-methylperfluoroctanesulfonamido) acetic acid; $n$-PFOA, $n$-perfluorooctanoic acid; $n$-PFOS, $n$-perfluorooctane sulfonic acid; NHANES, National Health Nutrition and Examination Survey; NIH, National Institutes of Health; non-HDL, non-high-density lipoprotein; PFDeA, perfluorodecanoic acid; PFHxS, perfluorohexane sulfonic acid; PFNA, perfluorononanoic acid; PFOA, perfluorooctanoic acid; PFOS, perfluorooctane sulfonic acid; PFUA, perfluoroundecanoic acid; Sb-PFOA, branch perfluorooctanoic acid isomers; Sm-PFOS, perfluoromethylheptane sulfonic acid isomers.

IRT-Derived PFAS Burden Score

We began by creating survey-weighted decile cutoffs for all biomarkers. This resulted in a different number of categories for different PFAS biomarkers because for some biomarkers, the survey-weighted deciles for adjacent deciles were the same, and thus, the adjacent categories were merged. This resulted in a 4-category item for Me-PFOSA-AcOH and PFUA, a 5-category item for PFDeA, a 9-category item for PFNA, and a 10-category item for all others. The survey-weighted quantile cutoffs are presented in Table S1. We then calibrated the IRT model by first running the graded response model for all nine biomarkers. After examination of extremity parameters, we removed an infrequently detected biomarker, Sb-PFOA (detected in only 10% of participants), due to poor item fit, given that it had an extremely low extremity threshold [$-55.8$ standard deviation (SD)], indicating that individuals would need to have a very low latent PFAS burden to have a 50% probability of detection. Further, it had a low and negative discrimination parameter ($-0.039$). Taken together, this suggested it was not useful for estimating latent PFAS burden and could be removed from the IRT model.

Inspection of the final IRT model, in which we removed Sb-PFOA, revealed that multiple indices supported this unidimensional model: Exploratory factor analysis also supported a one-factor model (first factor explains 53.6% of the variance and the ratio of the first eigenvalue to the second eigenvalue is 5.6; see Figure S3 for the scree plot). Item parameters are given in Table S2. All item discrimination estimates were positive, indicating that the probability of having higher observed levels of a PFAS biomarker increases as latent PFAS burden increases.

Discrimination ranged from 0.57 for Me-PFOSA-AcOH to 4.06 for $n$-PFOS, with higher values indicating that the biomarkers were useful in differentiating between individuals with lower or higher PFAS burden. IICs are presented in Figure 1, where higher information values indicate greater precision in estimation of the latent trait, with the value of the latent trait at the peak of the information curve indicating for whom (with respect to the latent trait) the item is most informative. From the IIC, we observed that isomers of PFOS ($n$-PFOS and Sm-PFOS) provided the most information about the exposure burden score, meaning they had greater ability to determine who was estimated to have high PFAS burden vs. those who were not, followed by PFNA, PFHxS, and $n$-PFOA. In general, most PFAS biomarkers provided their maximal information around the mean burden score, with highly discriminating biomarkers, such as $n$-PFOS, providing the most information between $-2$ and 2 SD of the latent PFAS burden score. Last, the TIC (Figure S4) conveyed the cumulative precision of the PFAS mixture with respect to the underlying exposure burden, with the greatest precision occurring around the mean burden score (i.e., $\text{theta}=0$). The standard error of measurement across the range of the latent trait, presented in Figure S4, shows that the PFAS burden was estimated most precisely at the average population burden ($\text{theta}=0$), with increasing standard error of measurement toward both extremes of latent burden. Based on the TIC, this GRM had good precision across a wide range of theta (Figure S4). The PFAS burden scores estimated using this model are presented as our main findings and are henceforth referred to as PFAS burden scores.

Figure 1.

Item information curves (IICs) of the PFAS IRT model. Each curve corresponds to a biomarker, with higher curves on the IIC plot indicating that the particular biomarker contributes more to the IRT burden score for some range of theta. The IICs show how well and precisely each biomarker measures PFAS burden at various levels of the underlying PFAS burden. Certain biomarkers may provide more information at low levels of PFAS burden, whereas others may provide more information at higher levels of PFAS burden. Isomers of PFOS ($n$-PFOS and Sm-PFOS) provide the most information about the exposure burden score, meaning they are most important in determining who is estimated to have high PFAS burden vs. those who are not, followed by PFNA, PFHxS and $n$-PFOA. Note: IRT, item response theory; Me-PFOSA-AcOH, 2-($N$-methylperfluoroctanesulfonamido) acetic acid; $n$-PFOA, $n$-perfluorooctanoic acid; $n$-PFOS, $n$-perfluorooctane sulfonic acid; PFAS, per- and polyfluoroalkyl substances; PFDeA, perfluorodecanoic acid; PFHxS, perfluorohexane sulfonic acid; PFNA, perfluorononanoic acid; PFUA, perfluoroundecanoic acid; Sm-PFOS, perfluoromethylheptane sulfonic acid isomers.

The decile cutoffs for PFAS burden scores (summed isomers) are presented in Table S3. Item parameters are presented in Table S4; IICs (Figure S5) and TIC (with standard error of measurement) are presented in Figure S6.

Summed Concentrations of PFAS

In these analyses, we use eight PFAS biomarkers. Sb-PFOA was excluded to be consistent with the main analyses using IRT.

Comparison of PFAS Burden Scores, PFAS Burden Scores (Summed Isomers), and Summed PFAS Concentrations

We plotted our primary PFAS burden scores (i.e., scores that included isomers of PFOA and PFOS) against summed PFAS concentrations (Figure 2). Summed PFAS concentrations were moderately correlated with PFAS burden scores ($\mathrm{\rho }=0.75$). Correlations of PFAS burden scores, PFAS burden scores (summed isomers), PFAS summed concentrations, and individual PFAS biomarkers are presented in Figure S7. As expected, PFAS burden scores and PFAS burden scores (summed isomers) were highly correlated with one another ($\mathrm{\rho }=0.95$).

Figure 2.

Scatterplot of PFAS burden score vs. summed PFAS concentrations. Note: PFAS, per- and polyfluoroalkyl substances.

Adjusted Associations of PFAS Burden Scores and Summed PFAS Concentrations with Cardio-Metabolic Biomarkers

Figure 3 and Table S5 show the adjusted effect size and 95% confidence interval (CI) corresponding to an IQR increase in PFAS burden score, an IQR increase in PFAS burden score (summed isomers), and an IQR increase in summed PFAS concentrations—across the seven cardio-metabolic biomarker outcomes. PFAS burden score, as well as summed PFAS concentrations, were significantly associated with higher total cholesterol, higher non-HDL cholesterol, and lower C-reactive protein. In general, however, associations were closer to the null for the summed PFAS concentrations vs. the PFAS burden score.

Figure 3.

Adjusted associations of PFAS indices with cardio-metabolic outcomes in NHANES 2017–2018. Adjusted differences and 95% CIs corresponding to 1 IQR increase in three different PFAS indices (PFAS burden scores, PFAS burden scores using summed isomers of PFOS and PFOA, and summed PFAS concentrations) with cardio-metabolic outcomes. All associations are adjusted for age, sex, race/ethnicity, BMI, and SES (assessed via the poverty-to-income ratio). Triglycerides and C-reactive protein are log-scaled. Regression models are not survey-weighted, and are calculated from $n=\mathrm{1,642}$ respondents for total cholesterol, HDL and non-HDL; $n=\mathrm{1,636}$ for triglycerides, $n=\mathrm{1,640}$ for C-reactive protein, and $n=\mathrm{1,568}$ for systolic BP and diastolic BP. Effect estimates and 95% CIs are reported in Table S5. Note: BMI, body mass index; BP, blood pressure; CI, confidence interval; HDL, high-density lipoprotein; IQR, interquartile range; IRT, item response theory; NHANES, National Health Nutrition and Examination Survey; non-HDL, non-high-density lipoprotein; PFAS, per- and polyfluoroalkyl substances; PFOA, perfluorooctanoic acid; PFOS, perfluorooctane sulfonic acid; SES, socioeconomic status.

We observed associations (95% CIs) with total cholesterol (in milligrams per deciliter) that were larger in magnitude for an IQR increase in PFAS burden score compared with the summed PFAS concentrations. Specifically, an IQR increase in PFAS burden score was associated with an 8.6 (95% CI: 5.2, 11.9)-$\mathrm{mg}/\text{dL}$ higher total cholesterol concentration, whereas an IQR increase in summed PFAS concentrations was associated with a 2.4 (95% CI: 0.5, 4.2)-$\mathrm{mg}/\text{dL}$ higher total cholesterol concentration. Further, an IQR increase in PFAS burden score was associated with a 7.8 (95% CI: 4.4, 11.1)-$\mathrm{mg}/\text{dL}$ higher non-HDL cholesterol concentration, whereas an IQR increase in PFAS summed concentrations was associated with a 1.8 (95% CI: 0.03, 3.7)-$\mathrm{mg}/\text{dL}$ higher non-HDL cholesterol concentration. There was no association between PFAS burden score or summed PFAS concentrations, with HDL or triglycerides.

In addition, an IQR increase in PFAS burden score was associated with a $-0.15$ (95% CI: $-0.23$, $-0.07$) log-$\mathrm{mg}/\text{dL}$ lower C-reactive protein concentration, whereas an IQR increase in PFAS summed concentrations was associated with a $-0.056$ (95% CI: $-0.099$, $-0.012$) log-$\mathrm{mg}/\text{dL}$ lower C-reactive protein concentration. Although there was a significant association between PFAS burden score and diastolic blood pressure, this association was null using PFAS summed concentrations. An IQR increase in PFAS decile burden score was associated with a 1.09 (95% CI: 0.03, 2.16) mmHg higher diastolic blood pressure. There were no significant associations between PFAS burden score or summed PFAS concentrations with systolic blood pressure. Sensitivity analysis for associations of PFAS burden scores (summed isomers) with cardio-metabolic outcomes revealed similar findings (Table S5), with the exception that PFAS burden scores (summed isomers) were also significantly associated with HDL.

Sensitivity Analysis to Demonstrate Utility of IRT Burden Score Calculator When Studies Do Not Measure the Exact Same Set of PFAS

Measurement of $\mathit{n}$-PFOS, Sm-PFOS, PFOA, PFNA, and PFHxS only.

Using 2017–2018 NHANES data, we conducted a sensitivity analysis to determine if associations with health outcomes were similar when the IRT score did not include the same set of biomarkers originally used to derive the PFAS burden score. We estimated participants’ PFAS burden score using the IRT burden calculator but using only their PFAS concentrations from a smaller set of PFAS ($n$-PFOS, Sm-PFOS, PFOA, PFNA, and PFHxS). Although a study may have measured only a smaller core set of PFAS, exposure concentrations of these biomarkers could still be used to calculate the participants’ PFAS burden on the same, original scale. Associations between exposure burden and cardio-metabolic outcomes were similar whether or not the entire set of PFAS biomarkers was used (Table S6; Figure S8).

Measurement of PFAS biomarkers excluding PFOS isomers.

As a related sensitivity analysis, because $n$-PFOS and Sm-PFOS were the two most informative biomarkers to the PFAS burden score, we estimated participants’ PFAS burden score without those biomarker concentrations (e.g., setting $n$-PFOS and Sm-PFOS values to missing for all participants) to determine whether a PFAS burden score could be calculated from a study setting where the most informative biomarkers for the burden score were not measured. We observed that associations for PFAS burden score without these biomarkers and cardio-metabolic outcomes were comparable to those when these biomarkers were included in the PFAS burden score (Table S6; Figure S9).

Secondary Analysis in a Validation Study

We used 2015–2016 NHANES data as a validation sample to evaluate if the observed stronger association of PFAS burden scores and lipids was inherent to the IRT approach used, which may be more sensitive to detecting associations with health outcomes compared with summed PFAS concentrations, or if this was a spurious observation. Whereas summed concentrations were associated with total cholesterol and non-HDL in 2017–2018 NHANES data, we observed null associations using 2015–2016 NHANES data (Figure 4; Table S7). However, using IRT-derived PFAS burden scores, we observed consistent positive associations with total cholesterol and non-HDL in both NHANES cycles.

Figure 4.

Association of PFAS indices with cardio-metabolic outcomes in NHANES 2015–2016. Adjusted differences and 95% CIs corresponding to 1 IQR increase in three different PFAS indices (PFAS burden scores, PFAS burden scores using summed isomers of PFOS and PFOA, and summed PFAS concentrations) with cardio-metabolic outcomes. All associations are adjusted for age, sex, race/ethnicity, BMI, and SES (assessed via the poverty-to-income ratio). Triglycerides and C-reactive protein are log-scaled. Regression models are not survey-weighted and are calculated from $n=\mathrm{1,763}$ respondents for total cholesterol, HDL and non-HDL; $n=\mathrm{1,761}$ for triglycerides, $n=\mathrm{1,762}$ for C-reactive protein, and $n=\mathrm{1,725}$ for systolic BP and diastolic BP. Effect estimates and 95% CIs are reported in Table S6. Note: BMI, body mass index; BP, blood pressure; CI, confidence interval; HDL, high-density lipoprotein; IQR, interquartile range; IRT, item response theory; non-HDL, non-high-density lipoprotein; PFAS, per- and polyfluoroalkyl substances; PFOA, perfluorooctanoic acid; PFOS, perfluorooctane sulfonic acid; SES, socioeconomic status.

Discussion

We demonstrated the feasibility of applying IRT to create a latent burden score of exposure to PFAS mixtures. This approach relates quantiles of biomarker concentrations to a latent burden score via nonlinear functions, allowing for the inclusion of mixed item types (e.g., using deciles for frequently detected chemicals, and quartiles for less frequently detected chemicals). Using recent NHANES data, we estimated a PFAS burden score as a method to capture the totality of exposure to PFAS mixtures. We conceptualized the burden score to reflect exposure to underlying PFAS exposure sources, then we assessed associations for the PFAS burden scores with seven cardio-metabolic outcomes in the general U.S. population.

In cross-sectional analyses, we found that higher PFAS burden scores were associated with higher total cholesterol, higher non-HDL cholesterol, higher diastolic blood pressure, and lower C-reactive protein. These findings are supported by previous studies reporting associations between single PFAS biomarkers with serum lipids levels.^36,37 PFAS exposure has also been linked to having less inflammation as measured by C-reactive protein.^40,41 Some previous studies have found null associations between individual PFAS and blood pressure.^38,39 In this analysis, we were limited by the cross-sectional design of the study, and by the smaller set of covariates included, given that our focus was to illustrate the IRT approach. In future work, it may be advantageous to account for a broader set of covariates, such as sedentary behavior and medication use, particularly in a prospective study design.

IRT offers a straightforward way to include biomarkers with low detection frequencies, by allowing for different numbers of categories for different biomarkers. Further, estimates of the overall health effects of a PFAS mixture, as quantified here by a burden score, may reduce coexposure amplification bias^55,56 compared with analyses of individual PFAS biomarkers. Burden scores estimated using IRT can also be used in exposomic research, such as for risk prediction and adjustment as a covariate or mediator. Importantly, this IRT-based approach can also be extended to other chemical classes.

Both summed PFAS concentrations and the PFAS burden score aim to estimate exposure burden to PFAS mixtures. However, they make very different assumptions. A sum-score approach to estimating a burden index (e.g., summed PFAS concentrations), although seemingly straightforward and assumption-free, is actually a highly constrained version of a latent variable model known as a parallel factor model,⁵⁷ in which each biomarker is assumed to contribute equally to the index. Previous work has shown that sum scores and IRT-derived scores are generally positively correlated and monotonic, but a high sum score does not always correspond to a high IRT-derived score because the latter also depends on the particular item (e.g., biomarker) characteristics and a participant’s particular item response pattern (e.g., exposure pattern to all the biomarkers).¹¹ As an example, two individuals may have equal sum scores, but the underlying exposures that yield their individual sum scores could be quite different (e.g., one person has high PFOA concentrations while the other has high PFHxS concentrations). Although a sum-score approach would treat these individuals as exchangeable, an IRT-derived score would better differentiate between these two individuals given that the IRT approach leverages the pattern of biomarker concentrations and not just the magnitudes of observed concentrations.

Although both approaches are valid for calculating a summary metric for exposure burden and that relations with health outcomes could be evaluated, associations with sum scores may generally be attenuated (compared with IRT-derived scores), as we have shown here and in our previous work with allostatic load.¹¹ In general, methods to summarize exposure burden run the risk of attenuating associations with health outcomes because they are not capturing the variability in individual biomarker concentrations but instead focus on the aggregate exposure burden to the chemical class. Our findings suggest that IRT is more sensitive to capturing individual biomarker variability because it accounts for exposure patterns, whereas the summed concentrations approach does not. In our analysis, we generally found associations with cardio-metabolic outcomes were closer to the null for summed PFAS concentrations compared with the IRT-derived PFAS burden score. As such, IRT-derived burden scores may be more sensitive to detecting associations with health outcomes that were missed when using the sum-score approach, although this needs to be replicated in other studies. Further, the summed concentration approach is sensitive to outliers. Furthermore, we showed that IRT-derived burden scores can be directly compared even when different studies measured different sets of PFAS, whereas summed PFAS concentrations cannot be compared unless the sum is restricted to include only biomarkers measured in all studies.

Recent literature calling for PFAS to be managed as a single class³³ provides conceptual support for using a unidimensional IRT approach, as we did in this study. Identifying how violations of unidimensionality practically affect estimates is an active area of psychometrics research, with simulation studies demonstrating that estimates of the latent trait are often unbiased even when unidimensionality is violated.⁵⁸ Extensions of this work to multiple chemical classes may be possible through the use of multidimensional IRT methods.

We view IRT approaches as complementary to other supervised mixtures approaches. Weighted quantile sum regression² and quantile g-computation⁵⁹ are supervised approaches that assume that the mixture index is a linear combination of quantiles, with potentially different weights for different chemicals depending on the prespecified health outcome. Weighted quantile sum regression assumes a linear, additive index, and both methods require the same number of quantiles for all chemicals. Meanwhile, IRT differs on two counts: a) the focus is on quantifying latent exposure burden to mixtures, independent of a prespecified health outcome, and b) it relates quantiles of chemicals to the latent burden score via nonlinear functions, which are often S-shaped.⁶⁰ IRT is not a single model; rather, it encompasses a wide range of models with various specifications, and here in particular, we used a GRM that handles ordinal data of mixed item types. Exposure burden scores calculated using various IRT models should theoretically be highly correlated, but they will not be exactly the same. Although the IRT approach is flexible to different numbers of categories for the biomarkers, in general, using more categories for PFAS biomarkers will yield greater measurement precision of the exposure burden score because more variability in PFAS concentrations will be captured.

In this paper, we have calculated the IRT-derived PFAS burden scores on a reference population (2017–2018 U.S. population) and have made this PFAS exposure burden scale available for researchers to use so that they can input their own study’s biomarker concentrations for each participant and get estimated PFAS burden scores. This scale can be used even if researchers do not have measures of all the PFAS compounds measured in NHANES but did measure a smaller set. These burden scores could then be compared across different studies because they are on the same scale, with this harmonization being potentially beneficial for consortia, such as the Environmental influences on Child Health Outcomes (ECHO)⁶¹ and for meta-analyses.

Our sensitivity analyses demonstrated that we could calculate participants’ PFAS burden on the same scale, even if only a smaller set of PFAS biomarkers was measured. Importantly, we found similar associations between PFAS burden scores and health outcomes, whether or not all PFAS biomarkers were used. Even when Sm-PFOS and $n$-PFOS, which were the most informative biomarkers for the burden score, were not used to estimate a participant’s burden (by setting those levels to missing), we found similar associations between the IRT score and the cardio-metabolic outcomes. This, again, is because IRT takes into account an individual’s exposure patterns in addition to concentrations of individual biomarkers.

IRT is not without limitations, however. It does not allow for assessment of the effects of individual chemicals directly: The focus is on the exposure burden score. Further, different combinations of biomarker levels could yield similar exposure burden scores. However, our application demonstrated that as expected, PFAS concentrations increase as the PFAS burden score increases. Strengths include that the IRT approach creates a simple, composite index of exposure burden that is independent of a health outcome, and researchers can use our PFAS burden calculator to calculate PFAS burden scores using recent U.S. reference ranges.

Quantifying exposure burden using IRT allows researchers to compare the relative contribution of PFAS burden to a range of health outcomes while maintaining a consistent metric for the exposure burden score. This enables one to compare across a large range of outcomes and physiological pathways to determine which are most perturbed by PFAS exposure burden. Given that PFAS are conceptually conceived and managed as a single chemical class and that there is evidence that some PFAS are phased out and substituted for others that exert similar health effects, we may be less concerned with which specific PFAS are driving the health effect and more concerned with the totality of PFAS exposure burden.