Evaluation of the Integrated Exposure Uptake Biokinetic (IEUBK) Model for Lead in Children

Abstract

Background:

The Integrated Exposure Uptake Biokinetic Model for Lead in Children (IEUBK model) was developed by the U.S. Environmental Protection Agency to support assessments of health risks to children from exposures to lead (Pb).

Objective:

This study evaluated performance of IEUBK model (v2.0) as it would be typically applied at Superfund sites to predict blood Pb levels (BLLs) in populations of children.

Methods:

The model was evaluated by comparing model predictions of BLLs to 1144 observed BLLs in a population of children at the Bunker Hill Superfund Site for which there were paired estimates of environmental Pb concentrations.

Results:

Predicted population geometric mean (GM) BLLs (GM: 3.4 μg/dL, 95% CI: 3.3, 3.5) were within 0.3 μg/dL of observed (GM: 3.6 μg/dL, 95% CI: 3.5, 3.8). The model predicted the observed age trend in GM BLLs and explained approximately 90% of the variance in the observed age-stratified GM BLLs. The mean predicted probability of exceeding 5 μg/dL (P₅) was 27% (95% CI: 24, 29) and observed P₅ was 32% (95% CI: 29, 35), a difference of 5%. Differences between geographic area stratified mean P₅ (predicted minus observed) ranged from −11 to 14% (mean difference: 2.3%).

Significance:

Although the more general applicability of these findings to other populations remains to be determined in future studies, our results support applications of the IEUBK model (v2.0) for informing risk-based decisions regarding remediation of soils and mitigation of exposures at Superfund sites where the majority of the exposure unit GM BLLs are expected to be ≤ 5 μg/dL and where it is desired to limit the predicted probability of exceeding 5 μg/dL to less than 5%.

Introduction

The objective of this study was to evaluate performance of the Integrated Exposure Uptake Biokinetic Model for Lead in Children (IEUBK model v2.0) as it would be typically applied at Superfund sites to predict blood Pb levels (BLLs) in populations of children. The IEUBK model was developed by the U.S. Environmental Protection Agency (EPA) to support assessments of health risks to children from lead (Pb) exposures [1,2,3]. The model is currently used in human health risk assessments at sites where Pb is a chemical of potential concern [1,2]. These sites are regulated under the Comprehensive Environmental Response, Compensation, and Liability Act (CERCLA) and the Resource Conservation and Recovery Act (RCRA).

In the EPA CERCLA program, the IEUBK model is used to support decisions regarding the need for remediation of soils and sediments at sites [1,2]. Typically, sites are subdivided into exposure units (EUs) representing geographic subareas of the site (e.g., a residential property) in which a typical child (i.e., the receptor) has an equal probability of being exposed to soil anywhere within the EU (e.g., the child’s yard). The IEUBK model is used in site risk assessment to predict a quasi-steady state geometric mean (GM) blood Pb level (BLL) and limit the probability to less than a 5% chance of exceeding a “BLL decision point” (also referred to as a “target BLL”) in the child receptor population. A widely applied BLL decision point was 10 μg/dL, which was the previous Center of Disease Control (CDC) level of concern [1]. Lower BLL decision point alternatives in the range of 2–8 μg/dL are under consideration based on associations between BLLs and neurodevelopmental outcomes in children observed in more recent epidemiologic studies [4,5]. Based on the 97.5^th percentile of BLLs in U.S. children aged 1–5 years, estimated from data collected during the period 2007–2010, the current CDC reference level for elevated BLL is 5 μg/dL [6]; however, the reference level may be lowered to 3.5 μg/dL based on more recent U.S. national data collected during the period 2011–2014 [7].

Site clean-up decisions and soil clean-up levels are generally selected to achieve a probability ($P_x$) of exceeding a BLL decision point ($x$) that is no more than 5% (i.e., $P_x\le 5\%$). The IEUBK model predicts the GM BLL and $P_x$ from a simulation of the relationships between multimedia exposures to Pb and quasi-steady state BLL (μg/dL) in children across an age range of typically 0.5 to 7 years (i.e., 6 to 84 months), although other age ranges may be selected [8]. This is achieved with four sub-modules. An exposure module calculates daily intakes of Pb (μg/day) averaged over each year of age. Daily Pb intake is calculated based on the combined inputs for exposure concentrations or rates of intake of Pb in air, diet, indoor dust, soil, drinking water, or other user-defined sources of exposure. An uptake module calculates media-specific time-averaged rates of absorption of Pb to blood (μg/day). A biokinetic module simulates transfer of absorbed Pb between blood and other body tissues, and elimination of Pb from the body. The output of the biokinetic module is a GM BLL for each of the 84 months after birth. A variability module is used to calculate the probability of occurrence of a specified BLL over a user-specified age range in a population of similarly exposed children. This calculation is based on a lognormal distribution with the predicted GM BLL and a geometric standard deviation (GSD). The GSD parameter in the IEUBK model is referred to as the interindividual geometric standard deviation (${GSD}_i$) to distinguish it from a population’s empirical GSD. The ${GSD}_i$ represents variability in BLLs expected for similarly exposed children, attributed to variability in rates of media intakes (e.g., soil and dust ingestion rate), bioavailability and biokinetics [8].

Confidence in risk-based decisions regarding remediation of soil Pb at CERCLA sites depends on confidence in the IEUBK model predictions of the $P_x$, which depends on confidence in the prediction of the GM BLL and the value assigned to the ${GSD}_i$ parameter. Over-prediction of the $P_x$ could lead to expensive and disruptive remediation activities, whereas under-prediction of the $P_x$ could result in leaving contaminated soil in place and underestimating residual post-remediation risk at the site. Numerous model-attributed factors could contribute to uncertainty in the $P_x$ including uncertainties and/or errors in assumptions about model parameter values. Model uncertainty can be broadly grouped into two categories: (1) uncertainties that affect confidence in the prediction of the GM BLL, including uncertainties related to media Pb concentrations, intakes and bioavailability, and Pb biokinetics; and (2) uncertainty in the interindividual variance in BLLs among similarly exposed children, represented in the model by the ${GSD}_i$ parameter. Several studies have evaluated these two major categories of uncertainties in the IEUBK model by comparing predicted GM BLL distributions with observed BLL distributions in populations of exposed children [9,10,11,12,13,14,15], or by quantifying the effects of parameter uncertainty on uncertainty in model predictions [16,17].

The Hogan et al. [10] study is particularly important to regulatory application of the IEUBK model at CERCLA sites for the purpose of predicting the probability of exceeding a BLL of 10 μg/dL (P₁₀; U.S. EPA 1998). Hogan et al. [10] compared observed population GM BLLs and individual child BLL distributions with predictions made by the IEUBK v99d model, in a sample population of 478 children who resided at four CERCLA sites. Hogan et al. [10] reported several important statistical evaluations of model performance. The three most germane to the regulatory applications of the model were to show that: (1) predicted site GM BLLs were not statistically different from observed GM BLLs and were within 1 μg/dL of observations; (2) predicted site mean P₁₀ values were not statistically different from observed site mean P₁₀ values and were within ± 4% of observations; and (3) >80% of the observed individual child BLLs at each site were within the 95% prediction interval of the model (defined by the predicted GM BLL and ${GSD}_i$).

Since the Hogan et al. [10] evaluation, several changes were made to the IEUBK model prompting the need for our evaluation. Default input parameter values for the IEUBK model were revised to reflect more recent data. These included adjustments to parameters that govern exposures and intakes of Pb in air, diet, drinking water, soil and indoor dust derived largely from soil. The intake (ingestion) of soil and indoor dust is an influential parameter on predicted BLLs. White et al. [8] advised that if “significant changes to influential parameters affecting estimated exposure levels from lead soil and dust” are made, “a recalibration of the model would be indicated.” Collectively, these parameter changes result in a lower predicted GM BLL compared to the IEUBK model version evaluated by Hogan et al. [10] at the same soil Pb concentrations.

Updated guidance from the EPA and CDC also affect the application of the IEUBK model at Superfund sites. Newer EPA guidance [18] for application of the IEUBK model recommends that the model be used to estimate the $P_x$ based on the average GM BLL predicted for the age range of 1 to ≤6 years (i.e., 12 to 72 months), rather than the 0.5 to ≤7 years (i.e., 6 to 84 months) age range that was previously used. Additionally, newer epidemiological findings and national trends in BLLs have prompted consideration of BLL decision points below 10 μg/dL [4,5,6,7]. The lowest action level at which environmental investigations are recommended by eight U.S. states [19] is at a BLL of 5 μg/dL. Five states have no recommendation and the remaining 37 states recommend action levels between 10 to 25 μg/dL. Action levels are not conceptually the same as the CDC blood Pb reference value (BLRV; [7]). The BLRV will change because it is the 97.5^th percentile of the annual average BLL distribution of children 1–5 years of age as derived from the National Health and Nutrition Examination Survey (NHANES), whereas, action levels are decision points for evaluating clean-up options. Also, our analysis focused on BLLs measured in the autumn of the year that will tend to be higher than at other times and, therefore, should not be directly compared to the BLRV derived from NHANES that represents an annual percentile [20].

Foreseeing potential applications of predicting the probability of exceeding lower BLL decision points, we evaluated the probability of exceeding 5 μg/dL (P₅). Lowering the BLL decision point has several implications for IEUBK model applications. Performance of the IEUBK model for predicting BLLs <5 μg/dL in child populations has not been evaluated. In the Hogan et al. [10] evaluation of the IEUBK model, site GM BLLs ranged from 5 to 7 μg/dL and the P₁₀ (i.e., the chance of exceeding 10 μg/dL) ranged from 19 to 29%. Assuming a log-normal distribution with the IEUBK default ${GSD}_i$ of 1.6 and limiting the probability to less than a 5% chance of exceeding 5 μg/dL, the population GM BLL is only 2.3 μg/dL. A complicating factor in evaluating performance of the IEUBK model at this lower BLL is the substantially larger contribution of dietary Pb to total Pb uptake. As soil Pb levels decrease, agreement between observed and predicted BLLs becomes increasingly more dependent on assumptions about dietary Pb intake in the receptor population, a variable that is not usually measured in CERCLA site assessments. Dietary Pb intakes in the receptor population are typically represented by national-level estimates (e.g., [15]).

Changes made to the IEUBK model defaults (especially revision of the soil/dust ingestion rates, which were adopted from [14]) and interest in applying the model to assessing risks at lower BLLs prompted this evaluation of the most recent version of the model, IEUBK v2.0. To accomplish this objective, child BLLs paired with soil and indoor dust (house dust) Pb exposure concentrations were needed in a population of children whose BLL distribution included a substantial number of children with BLL <5 μg/dL. Laboratory methods have been validated for estimating relative bioavailability (RBA) of Pb in soils that can be incorporated into calculations of Pb uptake [21]. It is recommended that RBA be determined for Pb in soils at sites for accurate assessment of potential risk. These requirements were satisfied from a subset of data on child BLLs and Pb concentrations in soil and house dust collected at the Bunker Hill Mining and Metallurgical Complex (Bunker Hill Superfund Site, BHSS).

Performance of the IEUBK model as it typically would be applied in a CERCLA risk assessment was evaluated by comparing observed BLLs to model predictions, with the following types of inputs: (1) Pb concentrations in residential yard soil and house dust (which is an optional data objective for CERCLA site assessments), and drinking water (when available); (2) area-specific soil and dust Pb RBA; (3) model default levels of Pb in air, drinking water (at residencies where drinking water data were not available), and dietary Pb levels; and (4) the default value for BLL ${GSD}_i$. Model performance was evaluated by comparing model predictions to observations made in the study population for the following three comparison metrics: population GM BLLs, P₅, and BLL empirical distribution. These evaluations focused primarily on the IEUBK v2.0 model using its defaults. However, alternatives to selected parameter values (namely, soil/dust ingestion rates, partitioning of soil/dust exposure, backfill soil Pb concentration, and dietary Pb intake) were also explored.

Methods

Our evaluation relied on data collected from the BHSS from 1995 to 2018. These data were made available for this evaluation by the Idaho Department of Environmental Quality and the Panhandle Health District. All storage and handling of these data were compliant with informed participant consent, confidentiality agreements, and non-disclosure agreements. This project, evaluating the IEUBK model using BHSS site-specific BLLs and environmental data, was reviewed according to the requirements of U.S. EPA Order 1000.17A (Policy and Procedures on Protection of Human Research Subjects) and U.S. EPA Regulation 40 CFR (Code of Federal Regulations) 26 (Protection of Human Subjects), and determined as not human subjects research.

The collection and analysis of blood and environmental samples at the BHSS has been described in various reports on the site available on-line [22]. Information relevant to how these data were used in this performance evaluation is described below.

Site description

The BHSS is in the Coeur d’Alene River basin in northern Idaho. The BHSS includes mining-contaminated areas in the Coeur d’Alene River corridor, tributaries, adjacent floodplains, downstream water bodies, and fill areas (referred to as “the Basin”), as well as the 21-square-mile area surrounding the historic smelting operations at the Bunker Hill complex, referred to as the “Box” (Figures S1, S2). The EPA identified three Operable Units (OU) within the BHSS: OU1 is the populated areas of the Box; OU2 is the non-populated areas of the Box; and OU3 (the Basin) is the areas of mining-related contamination outside the Box in the broader river basin. As part of the CERCLA activities at the BHSS, data on child BLLs, paired with residential soil Pb and indoor house dust Pb, were systematically collected. The data used in the current evaluation included 1144 BLLs measured in 853 children (less than 7 years of age) for which matched data on residential yard soil and house dust Pb concentrations and RBA were available. In this sample, the geometric mean BLL was 3.6 μg/dL, the GSD was 1.8, and approximately 70% of the observed blood Pb concentrations were <5 μg/dL.

Blood Pb data

Blood Pb data for this evaluation came from the BHSS Lead Health Intervention Program (LHIP) voluntary BLL screening program. The LHIP is offered annually as a public health service to area children, resulting in one BLL per child for each year they participated. Blood samples were collected in the late summer to capture anticipated peak BLLs in young children, as recommended by U.S. EPA [23]. The BLL data consist of a mix of venous samples analyzed at a detection limit of <1 μg/dL and capillary blood analyzed at detection limits of <1.4 or <1.9 μg/dL, depending on the type of capillary equipment used. Venous BLL screening was employed at the BHSS from 1988 to 2002, when capillary blood Pb testing was adopted. From 2002 through 2011, confirmatory venous samples were collected when a capillary result was ≥8 μg/dL; in 2012, this confirmatory venous threshold was lowered to 5 μg/dL. If a confirmatory venous sample result was available, then the venous sample result was used as the BLL for that child; otherwise, the capillary result was used as the primary BLL. No venous samples were below the detection limit. BLLs were recorded to either the nearest tenth or whole number. For inclusion in this evaluation, BLLs were restricted to: (1) children in the age range 6 months to <7 years (the IEUBK v2.0 model includes this age range, although the recommended age range for application at CERCLA sites is now 12 to 72 months [18]); (2) BLLs that exceed the detection limit; and (3) BLLs that were <30.5 μg/dL. Restricting the analysis to values above the detection limit made it unnecessary to assign surrogate BLL values (e.g., half the detection limit) to those capillary data that were below the detection limit, which would have introduced varying degrees of error into the BLLs for those individual children, depending on the detection limit of the analytical method. Exclusion of BLLs that were below the detection limit would not appreciably affect evaluation of model performance for predicting observed GM BLLs and P₅ of the study population; however, it would likely introduce bias into inferences made about the BHSS population as a whole, which was not the objective of this study. BLLs ≥30.5 μg/dL were excluded because these levels exceed IEUBK model performance limits [23]. The resulting dataset used in the performance analysis consisted of a total of 1144 BLL observations derived from measurements made on 853 children. Twenty-three percent of these children ($\text{N}=197$) provided more than one BLL in different years. Typically, this was 2 BLLs per child ($\text{N}=136$); a few children provided 3 (N=36) or up to 6 BLLs ($\text{N}=25$). Multiple BLL measurements in individual children spanned years in which the soil and dust Pb concentrations in the child’s residential environment changed as a result of remediation. Since the model predicted GM BLLs and P₅ were based on soil and dust Pb concentrations measured near the time of each BLL sampling, these multiple observations were included to increase the size of the evaluation dataset. Therefore, BLLs from individual children that participated in multiple years, and their corresponding matched soil and dust Pb concentrations, were treated as independent observations in comparisons with model predictions. In this approach, intra- and inter-individual variance in BLLs were not distinguished.

Soil and house dust Pb

All measured BLLs were temporally paired with soil and house dust Pb concentration data available for each residence. Soil samples were collected at a depth of 0–1 inch and were passed through a No. 80 mesh (177 μm) sieve prior to analysis. House dust samples were collected from personal vacuum cleaners and were also sieved to <177 μm prior to analysis. Concentrations below the detection limits were assigned half the detection limit (<1% of paired observations included in the evaluation). Detection limits varied during the study period and ranged up to 50 mg/kg. During the period in which the BHSS data were collected, property remediation resulted in the replacement of contaminated soils with clean backfill. Criteria for clean backfill material were established such that “clean” replacement soil was generally less than 100 mg/kg Pb, with no replacement soil exceeding 150 mg/kg. Confirmatory soil sampling was not conducted at properties after placement of clean backfill; therefore, a value of 100 mg/kg was used to represent the soil Pb concentration of a remediated property. Clean-up was extensive in the Box area and 587 of 875 (67%) observations from the Box area used in this evaluation represent values for clean backfill soil.

Drinking water Pb

Drinking water Pb data were available only for 27 individual residences in the Basin area who used a private water source. Water samples were paired with a BLL regardless of the year in which the BLL was measured. For example, water samples collected from a particular home in 2006 were paired with BLLs from a child residing in that home in any year of the study. The hierarchy for selecting Basin drinking water data to pair with BLLs was as follows: (1) samples from a kitchen sink; (2) if no sample from a kitchen sink was available, samples from a different location within the house; (3) if no sample from within the house was available, samples at the well source were used. If a water sample Pb concentration was below the detection limit, then the sample was given a concentration of half the detection limit (<3% of paired observations included in the evaluation). Detection limits varied during the study period and ranged up to 3 μg/L. In total, 55 measured water Pb concentrations in the Basin were paired with children’s BLLs. Water samples, including well samples, were not available for residences in the Box or Basin areas that were connected to a public water source; therefore, these children’s BLLs at these residences were paired with the IEUBK v2.0 model default value, 0.9 μg/L.

Air Pb

Reliable air Pb data representative of the study population were not available; therefore, BLLs were paired with the IEUBK v2.0 model default value, 0.1 μg/m³.

Other data

Soil and dust Pb levels were assigned the area-specific values for bioavailability (IEUBK model parameter Absorption Fraction Percent, AFP) based on measurements of RBA estimated from in vitro bioaccessibility [14]. In vitro bioaccessibility was measured using EPA Method 1340 [24]. The estimated site-wide mean for soil Pb bioavailability was 33% ± 4% (SD, $\text{N}=73$) and mean for house dust Pb bioavailability was 28% ± 6 (SD, $\text{N}=193$). These estimates are close to the IEUBK model default AFP values of 30% for soil and indoor dust. Community-specific values for bioavailability were estimated for areas within the Box and these values were assigned to residences in each community. A single average bioavailability value was assigned to all residences in the Basin as no RBA measurements were available for the Basin. Age-specific values for the combined soil and dust ingestion rate parameter in the IEUBK model (IR_SD) were estimated for the BHSS [14] and are similar to the estimates reported in the U.S. EPA Exposure Factors Handbook (EFH; [25]; Supplemental Materials Table S1). Performance evaluation of the IEUBK v2.0 model used the von Lindern et al. [14] values for IR_SD; however, the EFH values were also explored as alternative parameter values (see Results). The IEUBK v2.0 model assumes a partitioning of house dust Pb and soil Pb ingestion as 55% from house dust and 45% from yard soil. However, previously reported analyses of the BHSS data have found better conformance between IEUBK model predicted GM BLLs and observed BLLs when partitioning included contributions from neighborhood and community-wide soil Pb [13,14]. Alternative approaches to partitioning house dust and soil Pb exposure were explored in this analysis (see Results). All other parameters used in the IEUBK model were assigned v2.0 default values (see Table 2–2 of [3]).

Data analysis methods

IEUBK model runs were executed using the batch file processing feature of v2.0 (build 1.6), with each row of the batch file representing an individual child and their age at time of BLL sampling, with corresponding media Pb concentrations from their individual residence. The BLL predicted by the IEUBK model for each residence and child combination (i.e., each row of batch file) was the GM for the exposure values assigned to that residential property. Observed individual child BLLs were stratified by age (year) categories (<1 year to age 6–7 years) and by geographic location (Box, Basin, Upper Basin, Lower Basin). Statistical analyses were performed using SAS/STAT^® software, Version 9.4 of the SAS System for Windows.

Geometric means and corresponding 95% confidence intervals (CIs) for observed and predicted BLLs were calculated on the log transformed BLLs (SAS PROC MEANS) and outcomes were then exponentiated. We tested the normality of the 1144 BLLs using frequency plots, probability plots, and statistical tests (Shapiro-Wilk, Kolmogorov-Smirnov, Cramer-von Mises, Anderson-Darling). The BLLs formed an empirical distribution that exhibited positive skew and was statistically different from normal ($\text{p}=<0.01$). Probability plots of the BLL data with or without natural log transformation showed that a lognormal distribution described the data better than a normal distribution. The 95% CIs on the observed mean BLL is the confidence interval on the sample mean for N observations (e.g., 1144). For the 1144 observed BLLs in the sample, the model predicted 1144 GM BLLs. We calculated 95% CIs on the mean of the sample of predicted GM BLLs (Hogan et al. [10]). Observed P₅ was calculated as the proportion of individual BLLs ≥5μg/dL. The IEUBK model predicts a P₅ for a hypothetical population of children having a GM BLL and GSD=1.6. As noted above, we predicted GM BLLs for exposures at each property, which was paired with the observed BLL at that property, and calculated the corresponding P₅, assuming a GSD of 1.6. For the 1144 observed BLLs in the sample, we predicted 1144 P₅ values. The confidence limits are on the mean of the 1144 predicted P₅s. We did the same for other BLL strata. Confidence intervals on P₅ were calculated using the Wald method, i.e., the ‘Wald’ Normal interval [26]; Eq. (1).

$$Waldinterval=\widehat{p}\pm z_{\alpha /2}\sqrt{\frac{\widehat{p}(1-\widehat{p})}{N}}$$ (1)

where: $\widehat{p}$ is the observed percentage or predicted probability to exceed 5 μg/dL, $Z_{\alpha /2}$ is the Z-score for the standard normal distribution at percentile $\alpha /2$, α is significance level of 0.05, and $\text{N}$ is the number of observations. When Equation 1 resulted in a lower CI <0% or an upper CI >100%, due to a small sample size or an average exceedance near zero, the Wilson score [27] interval method was used (Eq. 2).

$$Wilsonscoreinterval\left(w^{-},w^{+}\right)\equiv \frac{\left(\widehat{p}+\frac{Z_{\alpha /2}^2}{2N}\pm Z_{\alpha /2}\sqrt{\frac{\widehat{p}\left(1-\widehat{p}\right)}{N}+\frac{Z_{\alpha /2}^2}{4N^2}}\right)}{\left(1+\frac{Z_{\alpha /2}^2}{N}\right)}$$ (2)

Predicted P₅ values were calculated from the IEUBK model for each predicted GM BLL and averaged. The average P₅ was treated as a binomial probability for the purpose of estimating 95% CIs (as in [10]) and calculated from Eqs. 1 or 2. The CIs for the average of the predicted P₅ values were calculated using the number of observations (N) in the stratum. In other words, the same sample sizes were used to estimate the confidence intervals for the predicted and observed P₅ values. Predicted and observed GM BLLs or P₅ were considered to be significantly different if their 95% confidence intervals did not overlap.

Regression modeling of observed and predicted BLLs and P₅ was performed to evaluate performance of the IEUBK model to predict age-stratified BLLs and P₅. Regression modeling was performed using weighted linear regression (SAS PROC NLIN) with uncorrelated weights assigned to each data point ($W_i$) (Eq. 3).

$$W_i=\frac{\text{w}\left({\text{X}}_{\text{i}}\right)·\text{w}\left({\text{Y}}_{\text{i}}\right)}{\text{w(}{\text{X}}_{\text{i}}\text{)+w(}{\text{Y}}_{\text{i}}\text{)}{·(\text{β}}^{\text{2}}\text{)}}$$ (3)

where: $w(X_i)$ and $w(Y_i)$ are the weights (1/SE_i²) for the independent and dependent variables, respectively, and β is the slope of the linear regression line fit by minimizing the weighted sum of squared residuals [28]. Studentized residuals (>2.5 or <−2.5) were used to identify and exclude statistical outliers. Weighting standard errors (WSE) were calculated from the 95% confidence intervals on the mean (Eq. 4).

$$WSE=\frac{CI}{t_{(\alpha ,df)}}$$ (4)

where: $\text{CI}$ is the 95% confidence interval and $t_{(\alpha ,\mathrm{d}\mathrm{f})}$ is the t-value at 0.95 (α) confidence level for degrees of freedom $df$.

Differences between empirical distributions of observed individual child BLLs and predicted GM BLLs were tested using the Kolmogorov-Smirnov two-sample test (SAS PROC NPAR1WAY). Lower and upper 95% prediction limits for any predicted GM BLL were calculated as the 2.5^th and 97.5^th percentiles of the lognormal distribution defined by the predicted GM BLL and the ${GSD}_i$ in Eq. (5) [10].

$${PL}_{95}=e^{(\ln \left(GM\right)+\mathrm{l}\mathrm{n}({GSD}_i)·Z_{\%})}$$ (5)

where: ${PL}_{95}$ is the lower or upper 95% prediction limits and $Z_{\%}$ is the Z-score for the standard normal at percentiles 0.025 or 0.975, respectively.

Results

BLL and environmental Pb exposures of the study population

This analysis included 1144 observations of children’s BLLs paired with their residential house dust and yard soil Pb concentrations, and private-source household water Pb concentrations for a small subset of residences in the Basin area ($\text{N}=55$). Summary statistics for BLLs and environmental Pb levels are presented in Table 1. The median BLL was 3.5 μg/dL (range: 1.0, 22). The median residential yard soil Pb concentration was 100 mg/kg (range: 31, 9180), and the median house dust Pb concentration was 345 mg/kg (range: 11, 15300).

Table 1.

Summary of BLL and Pb exposures for study population

	N	Mean	Min	5th %	50th %	95th %	Max
BLL (μg/dL)	1144	4.4	1.0	1.5	3.5	10.0	22.0
Soil (mg/kg)	1144	413	31.1	100	100	1611	9180
House dust (mg/kg)	1144	598	11.3	82.3	345	1570	15300
Tap water (μg/L)	55	2.3	0.9	1.0	1.5	7.5	8.8

BLL, blood Pb level; Max, maximum; Mean, arithmetic mean; Min, minimum; N, number of BLL samples; %, percentile

Prediction of population GM BLLs

Observed and predicted GM BLLs are presented in Table 2. The GM of the predicted BLLs for the population of children ages <7 years ($\text{N}=1144$; site-wide) was 3.4 μg/dL (95% CI: 3.3, 3.5) and within 0.3 μg/dL of the observed GM, 3.6 μg/dL (95% CI: 3.5, 3.8). Confidence intervals for observed and predicted area-stratified GM BLLs overlapped in the Basin, but not the Box area that comprised approximately 76% ($\text{N}=875$) of the site-wide observations (Fig. 1). Differences (predicted minus observed) between GM BLLs in the Box and Basin areas ranged from −0.4 to 0.4 μg/dL (mean difference: −0.1 μg/dL), indicating the model performed well at predicting the observed area GM BLLs. Observed GM BLLs showed a distinct trend with age, which was captured by the IEUBK model (Fig. 2). The model explained approximately 90% of the variance in the age-stratified GM BLLs (${\text{r}}^2=0.90$). There were no statistical outliers. Differences (predicted minus observed) between age-stratified GM BLLs ranged from −0.5 to 1 μg/dL (mean difference: −0.1 μg/dL). The largest difference (1 μg/dL) was in the age group <1 year. Confidence intervals for observed and predicted age-stratified GM BLLs overlapped for all ages except age 6 to <7 years. Collectively, the results suggest that GM BLLs predicted by the IEUBK v2.0 model agreed well with observed GM BLLs; differences in observed and predicted GM BLLs for site-wide and area-stratified data were <1 μg/dL and the model did not systematically over- or under-predict observed GM BLLs.

Table 2.

Summary of observed and predicted GM BLLs

		Observed BLLs (μg/dL)			Predicted BLLs (μg/dL)
Stratum	N (Nchild)a	GM	LCL	UCL	GM	LCL	UCL
Site-wide	1144 (853)	3.6	3.5	3.8	3.4	3.3	3.5
Box	875 (651)	3.9	3.7	4.0	3.4	3.3	3.6
Basin	269 (223)	3.0	2.8	3.1	3.2	3.0	3.4
Upper Basin	209 (173)	2.9	2.8	3.3	3.3	3.0	3.5
Lower Basin	60 (51)	3.1	2.7	3.6	2.9	2.5	3.4
Age (year)
0.5 - <1	70	3.8	3.4	4.4	4.8	4.2	5.6
1 - <2	138	4.6	4.2	5.1	5.0	4.5	5.5
2 - <3	202	4.0	3.6	4.3	3.5	3.2	3.7
3 - <4	176	3.8	3.4	4.1	3.3	3.0	3.6
4 - <5	187	3.5	3.2	3.8	3.2	3.0	3.5
5 - <6	185	3.3	3.0	3.6	3.0	2.8	3.3
6 - <7	186	3.1	2.8	3.3	2.6	2.4	2.7

Predicted GM BLLs are the GMs of the IEUBK model BLL predictions for all records of the stratum.

BLL, blood lead level; GM, geometric mean; LCL, 95% lower confidence limit; N, number of BLL samples; N_child, number of individual children that result in N; UCL, 95% upper confidence limit.

^aN_child substrata counts will not sum to their grouped strata because children moved between areas; as a result, the same child could have a different BLL (from a different age) in the Box and the Basin and was counted in each strata’s N_child. There are no duplicates per age because a child was tested during the peak summer month only once each year; therefore, N_child is the same as N.

Figure 1.

Comparison of observed (circles) and predicted (triangles) area-stratified GM BLLs. Flags are 95% CLs. Values in parentheses are number of BLLs in each area.

Figure 2.

Weighted linear regression model for site-wide observed and predicted GM BLLs. Circles illustrate the GMs BLLs for specified age ranges. The open circle is the GM BLL for children <7 years and was not included in the regression model since it is an aggregate of the other age strata. Flags are 95% CLs for specified age ranges. The curved dotted lines show the 95% CLs on the weighted regression model. The diagonal dotted line is the line of identity. Parameter values are as follows: intercept =1.32, slope = 0.69, ${\text{r}}^2=0.90$.

Prediction of probability of exceeding a BLL of 5 μg/dL (P₅)

The most important output of the IEUBK model in risk assessment applications is the probability of exceeding the BLL decision point because this metric is used by U.S. EPA to identify EUs for possible risk management actions. We examined how well the model predicted the P₅ for the study dataset. Observed P₅ was calculated as the proportion of individual BLLs ≥5 μg/dL (see Methods). Predicted values for P₅ were obtained from the IEUBK model for each predicted GM BLL and averaged. Observed and predicted P₅ values are presented in Table 3 and Fig. 3. The mean predicted P₅ for the exposures estimated for children ages <7 years ($\text{N}=1144$, site-wide) was 26.6% (95% CI: 24.0, 29.1) and the observed P₅ was 32.1% (95% CI: 29.4, 34.8), a difference of 5.5%. Differences (predicted minus observed) between area-stratified mean P₅ values ranged from −11 to 14% (mean difference: 2.3%). The 95% confidence limits on the observed and predicted mean P₅ did not overlap in most areas, with the exception being the Lower Basin, which comprised 5% of the observations ($\text{N}=60$) and had the widest confidence limits. Differences (predicted minus observed) between age-stratified mean P₅ values ranged from −10 to 15% (mean difference: −3.5%). The largest difference (15%) was in the age group <1 year. The IEUBK model predicted the observed age trend in the P₅ (${\text{r}}^2=0.91$, Supplemental Materials, Fig. S3). Collectively, these results suggest that mean P₅ predicted from the IEUBK v2.0 model agreed relatively well with observations (±6% difference for site-wide) and there did not appear to be any strong bias in the direction of differences between observed and predicted area-stratified P₅.

Table 3.

Summary of observed and predicted P₅

		Observed P5 (%)			Predicted P5 (%)
Stratum	N (Nchild)a	AM	LCL	UCL	AM	LCL	UCL
Site-wide	1144 (853)	32.1	29.4	34.8	26.6	24.0	29.1
Box	875 (651)	37.9	34.7	41.2	27.2	24.2	30.1
Basin	269 (223)	13.0	9.0	17.0	24.6	19.5	29.8
Upper Basin	209 (173)	12.0	7.6	16.4	25.8	19.8	31.7
Lower Basin	60 (51)	16.7	7.2	26.1	20.6	10.4	30.8
Age (year)
0.5 - <1	70	30.0	19.3	40.7	44.6	32.9	56.2
1 - <2	138	45.7	37.3	54.0	47.0	38.7	55.3
2 - <3	202	36.6	30.0	43.3	28.0	21.8	34.2
3 - <4	176	33.5	26.5	40.5	25.0	18.6	31.4
4 - <5	187	30.5	23.9	37.1	23.0	17.0	29.1
5 - <6	185	26.5	20.1	32.8	21.1	15.3	27.0
6 - <7	186	23.7	17.5	29.8	13.4	8.5	18.3

Predicted P₅ values are the arithmetic mean of the IEUBK model P₅ predictions for all records of the stratum.

BLL, blood lead level; AM, arithmetic mean; LCL, 95% lower confidence limit; N, number of BLL samples; N_child, number of individual children that result in N; P₅, probability of exceeding 5 μg/dL; UCL, 95% upper confidence limit

^aN_child substrata counts will not sum to their grouped strata because children moved between areas; as a result, the same child could have a different BLL (from a different age) in the Box and the Basin and was counted in each strata’s N_child. There are no duplicates per age because a child was tested during the peak summer month only once each year; therefore, N_child is the same as N.

Figure 3.

Comparison of observed (circles) and predicted (triangles) area-stratified P₅ (i.e., probability of exceeding 5 μg/dL). Flags are 95% CLs. Values in parentheses are number of BLLs in each area.

Prediction of distribution of individual child BLLs

The intended application of the IEUBK model is to predict the cumulative distribution of BLLs for populations of similarly exposed children (e.g., P₅). The model is not intended to predict BLLs of individual children. Nevertheless, the model performed well at predicting the overall distribution of observed individual child BLLs. Figure 4 shows the cumulative empirical distribution of observed individual child BLLs ($\text{N}=1144$) and the corresponding cumulative distribution of predicted BLLs. The distribution of predicted GM BLLs closely reproduced the lower and upper quartiles of the distribution of observed individual BLLs. In the middle quartiles, the model tended to predict lower GM BLLs than observed individual BLLs. Although the distributions show relatively close agreement, they were significantly different (p≤0.05, Kolmogorov-Smirnov test). The large number of samples in our evaluation contributed to a small critical value for the statistical test, leading to rejection of the null hypothesis of no difference between distributions for differences that may not be meaningful for Pb risk assessment. The number of GM BLLs exceeding 5 μg/dL for the observed and predicted distributions yielded similar estimates; the P₅ based on the observed BLLs was 32% and the P₅ based on the predicted GM BLLs was 20%. The cumulative lognormal distributions for observed BLLs (GM=3.6, GSD=1.8) and predicted GM BLLs (GM=3.4, GSD=1.6) are compared in Fig. S4.

Figure 4.

Comparison of empirical distribution of observed individual child BLLs (circles, $\text{N}=1144$) with the probability distribution predicted by the IEUBK model (solid line) for GM BLLs. Although visually similar, the distributions were significantly different ($\text{p}\le 0.05$, Kolmogorov-Smirnov test).

In addition, most individual BLLs were within the 95% prediction limits of the model. A plot of observed and predicted BLLs for individual children compared to the 95% prediction limits for the model is shown in Fig. 5. Out of 1144 observed BLLs, 86% fell within the prediction limits; 6% were below the lower prediction limit, and 8% were above the upper prediction limit. When stratified by area, the percentage outside of the prediction limits ranged from 10% for Upper Basin (N=209) to 15% for the Box area (N=875) and Lower Basin area (N=60). By comparison, 20% of the data were outside the 95% prediction limits in the model evaluation by Hogan et al. [10].

Figure 5.

Comparison of predicted and observed individual BLLs (circles) for children <7 years ($\text{N}=1144$). Lines are the 95% lower and upper prediction limits for the IEUBK model.

Alternative parameter values for soil/dust ingestion rate

We evaluated the IEUBK v2.0 model, which uses the combined soil and dust ingestion rate (IR_SD, mg/day) estimated for the BHSS [14]. In the von Lindern et al. [14] study, soil and dust ingestion rates for the BHSS were estimated by calibrating the IR_SD parameter of the IEUBK v1.1 model to optimize fit of predicted GM BLLs to a representative subset ($\text{N}=2176$) of longitudinal data on individual child BLLs at the BHSS. von Lindern et al. [14] concluded that the best fit was achieved when intakes (i.e., ingestion) of Pb in soil and house dust were partitioned as: 50% from house dust, 25% from home yard soil, 10% from neighborhood soil (within 200 feet from home), and 15% from community soil (town average). This partitioning (referred to as the 50/25/10/15 dust/soil partition) differs from the IEUBK model default that attributes 55% of Pb intake to house dust and 45% to yard soil. Although this partition and several other parameter values in v1.1 differ from those of v2.0, the IR_SD estimates from von Lindern et al. [14] are similar to the values recommended in the EFH [25]. Consequently, substitution of the von Lindern et al. [14] values with those from the EFH results in small changes to predictions of GM BLL or P₅. Because the EFH estimates are intended to be generalizable to sites where site-specific estimates are not available (i.e., nearly all CERCLA sites), we also evaluated performance metrics for the IEUBK model using the EFH estimates for IR_SD. Figure 6 compares the observed GM BLLs with predictions from the IEUBK v2.0 model when either set of IR_SD parameter values were used in the model. As expected, because of the similarity in the values for IR_SD, performance of the model was essentially unchanged when the EFH values for IR_SD were used. Model performance using the alternative EFH IR_SD values is illustrated in Supplemental Materials Figs. S5–S8.

Figure 6.

Comparison of observed and predicted area GM BLLs when default or alternative parameter values are used in the IEUBK v2.0 model. The point labeled v2.0 represents the model with its default 55% house dust/45% yard soil partitioning and IR_SD from von Lindern et al. [14]. EFH IR_SD is the prediction when the estimates for IR_SD from the U.S. EPA Exposure Factors Handbook [24] were used in place of the default IR_SD from von Lindern et al. [14]. The point labeled 40/30/30 is the prediction when the von Lindern et al. [13] partitioning of 40% house dust, 30% yard soil, 30% community soil was used in place of the default partitioning of 55% house dust, 45% yard soil, but without changing the overall default IR_SD in v2.0 from von Lindern et al. [14]. Points are GM BLLs ($\text{N}=1144$). Flags are 95% CLs.

Alternative partitions of soil and house dust Pb exposure

This performance assessment relied on the IEUBK model default assumption that 55% of soil and dust Pb intake comes from house dust and 45% derives from exposure to yard soil within the residential EU. At the BHSS, a more realistic and complex scenario was used in which exposures for a given child were derived from multiple sources of soil and dust within the community. von Lindern et al. [13,14] showed that, at the BHSS, prediction of observed GM BLLs was successful if the exposure to soil and dust was represented in the IEUBK model as an average intake weighted as follows: 40% from house dust at the residence, 30% from property soil (EU), and 30% from community soil (town average). The 40/30/30 partition was evaluated in this analysis because that partition has been employed in risk assessment and management at the BHSS, and the 50/25/10/15 dust/soil partition requires a level of property sampling that is unlikely to be realized at other sites. Figure 6 shows the GM BLL predicted when inputs to the model were a 40/30/30 partitioning of Pb in house dust, yard soil and community soil as defined in von Lindern et al. [13,14]. The predicted GM BLL for the 40/30/30 partition was higher than the observed GM BLL and the predicted GM BLL for the 55/45 partition (regardless of the values for IR_SD used in the model).

Alternative assumptions for clean backfill soil Pb concentration

Although an assumed soil Pb concentration of 100 mg/kg was used in this analysis to represent post-remediated soils, sampling of the backfill borrow piles during the years remediation occurred indicated that actual replacement soils averaged approximately 55 mg/kg in the Box area throughout the cleanup. Uncertainty in the backfill concentration over this range (50 to 100 mg/kg) had a negligible effect on the predicted GM BLLs. For example, in the Box area where there was ubiquitous application of backfill, the difference was 0.1 μg/dL (3.4 vs 3.3 μg/dL for 100 mg/kg or 50 mg/kg, respectively).

Alternative assumptions for dietary Pb intake

All children were assigned IEUBK v2.0 model default values for dietary intake of Pb in market basket foods. These values were derived from the What We Eat In America (WWEIA) dietary interview component of the National Health and Nutrition Examination Survey (NHANES) (2003–2006; [29,30]) and the Food and Drug Administration (FDA) Total Diet Study (TDS) food contaminants data (1995–2005; [31]). Zartarian et al. [15] used an alternative set of dietary Pb intakes in IEUBK model simulations of national exposures to Pb and corresponding BLLs (Supplemental Materials Table S1). For the age range <7 years, the Zartarian et al. [15] dietary intake estimates are approximately 60% of the IEUBK v2.0 model default values and, if used in the IEUBK model, predicted GM BLLs would have decreased by approximately 0.4 μg/dL. This is a relatively small magnitude of difference in the predicted GM BLL; however, the effect on the upper tail of the BLL distribution is more pronounced. For example, for a soil Pb concentration of 200 mg/kg and all other parameters set to defaults, the IEUBK v2.0 model predicts a GM BLL of 2.3 μg/dL and P₅ of 5%; if the Zartarian et al. [15] dietary estimates are used in place of default values, the resulting predicted GM BLL is 2.0 μg/dL and P₅ is 2.3%.

Discussion

This study represents the first empirical evaluation of the IEUBK model version 2.0 performance, which includes several revisions to previously released versions of the model (e.g., adjustments to parameters that govern exposures and intakes of Pb in air, diet, drinking water, soil, and soil derived dust). This evaluation focused on three performance metrics: (1) prediction of population GM BLLs; (2) prediction of probability of the BLL exceeding 5 μg/dL (P₅); and (3) prediction of distribution of individual child BLLs. These performance metrics were selected for the following reasons. The GM was selected as the central tendency variable because BLL output from the IEUBK model is interpreted to be the GM of a lognormal distribution. Although U.S. EPA has recommended application of the model in the CERCLA program for predicting the P₁₀ (probability of the BLL exceeding 10 μg/dL) at residential EUs [1,2]; more recent assessments [5] suggest the need to consider lower decision points, including P₅. Predicting the P₅ depends on a prediction of the GM BLL for a child in the EU. Although predicting BLLs of individual children is not a recommended application of the model, in this study and in all previous studies, evaluation of the GM and P₅ predictions has relied on BLL data of individual children. Therefore, it was of secondary interest to understand how well the model predicts distributions of individual BLLs that comprise the performance evaluation dataset.

In general, our study found that the IEUBK v2.0 model performed well at predicting population GM BLLs of the evaluation dataset. For the complete data set of 1144 BLL observations, the model predicted a GM BLL for the population that was within 0.3 μg/dL of the observed GM. An error in predicted GM BLLs <1 μg/dL is tolerable for applications to site risk assessment, given the relatively large uncertainties in estimating exposure and intake of Pb [9,10,13,15,17]. The model also predicted the age trend in observed GM BLLs and explained approximately 90% of the variance in the age-stratified GM BLLs (${\text{r}}^2=0.90$). The model predicted a mean P₅ for the population that was within 6% of the observed mean P₅ and showed relatively low negative and positive biases in the prediction of the P₅ when applied to area-stratified data (mean difference 2.3%). Collectively, these results provide strong support for application of the IEUBK v2.0 model to human health risk assessment.

Prediction of individual child BLLs is beyond the application scope of the IEUBK model; nevertheless, 86% of the observed BLLs were within the 95% prediction limits of the model. Theoretically, if a representative sample from a population of children within an EU was available, and the model correctly assigned environmental exposures to those children, we would expect 95% of the BLLs observed within that EU to fall within the 95% prediction limits of the model (Supplemental Materials Fig. S9). However, several factors could interfere with realizing this ideal outcome, including: (1) misclassification of exposures to individual children within the EU; (2) small sample size for BLLs within the EU (typically 1 child in each EU), which may not represent the child population GM within the EU; and (3) a ${GSD}_i$ that does not represent the actual variance in BLLs within the EU. The prediction intervals for each GM BLL depend completely on the value selected for the ${GSD}_i$ (1.6 assumed in the model). Were the ${GSD}_i$ to be assigned a higher value, the prediction limits would widen and encompass a larger fraction of the observations. For this data set ($\text{N}=1144$), a ${GSD}_i$ value of 1.97 encompasses 95% of the observed BLLs. This GSD was larger than the GSD_i used in the IEUBK model, which is to be expected, given that exposure variability contributes to the BLL variance in the data set (e.g., variability in soil and house dust Pb concentrations between EUs). Given the above considerations, it is not surprising that more than 5% of the observed BLLs were located outside the prediction intervals. Having 86% of the observed BLLs fall within the 95% prediction limits of the IEUBK v2.0 model suggests a modest improvement from the results of an earlier evaluation of v99d of the IEUBK model where 80% were within the 95% prediction limits [10]. This improvement in performance may be, in part, attributed to having better exposure data using sieved soil and dust with estimates of Pb bioavailability at the BHSS for our model evaluation.

The BHSS data provided several important features that contributed to the strengths of this study. These included: (1) a relatively large number of BLLs (N=1144) measured (above detection and below IEUBK performance limits) in children that could be geographically matched to measured soil and house dust Pb concentrations; (2) the distribution of BLLs included a substantial number of BLLs (approximately 70%) that were ≤5 μg/dL; (3) BLLs were measured in late summer as recommended by U.S. EPA [1]; (4) measurements of Pb RBA in BHSS soils and house dusts supported the values for the absorption parameter (AFP) used in the IEUBK v2.0 model; and (5) estimates of soil/dust ingestion rates at the BHSS supported the values for the IR_SD parameter used in the IEUBK v2.0 model [14].

Our study has several limitations. The concept underlying the use of the IEUBK model to establish soil cleanup levels is that, if a population of children were exposed to the same levels of environmental Pb, that population would have a distribution of BLLs reflecting variability in environmental media intakes, gastrointestinal absorption of Pb, and biokinetics of absorbed Pb. The IEUBK model predicts this BLL distribution and from this distribution, the probability of exceeding 5 μg/dL (P₅). Rigorously evaluating these predictions would require a population of children who are exposed to the same environmental Pb levels (i.e., multiple children residing in the same home) and comparing the observed population GM BLL and P₅ to the values predicted from the model. This ideal evaluation has never been performed because no such data exist to our knowledge. In deriving the IEUBK model default value for the interindividual geometric standard deviation (GSDi) of 1.6, BLL GSDs were calculated in populations that were stratified by soil Pb exposure. The within-strata GSDs were interpreted as “the variability in BLLs among children of similar age and exposed to similar concentrations of environmental lead” [8,23]. This latter interpretation is the basis for idea that the IEUBK model predicts the BLL distribution of a population of “similarly exposed children.” This leads us to the practical problem of how to best evaluate the model for applications to risk assessment, where it is used to predict the P₅ (or some other BLL probability) corresponding to an observed or specified soil Pb level. In the absence of the ideal dataset described above, alternative data sets have been used. In the evaluation of IEUBK v0.99d (Hogan et al. [10]), and in all subsequent reported evaluations of the model, the approach has been similar to the one we used [9,11,12,13,14]. The core of this approach is to quantify how well the model predicts the BLLs in a sample of children for whom we can reliably specify their individual property soil and dust Pb exposures. We have performed such an evaluation based on a relatively large sample of 1144 BLLs, as well as for area and age strata from this sample. We recognize that is not an ideal evaluation of the model; but we also conclude that it is highly informative for evaluating model performance in the context of its use in risk assessment. In risk assessment applications, the model is used to predict GM BLLs at each individual property where children would experience the same/similar exposures. That it does reliably predict the GM BLLs and P₅ in our sample of BLLs indicates good correspondence between model predictions and observations.

In our analysis, individual child BLLs were matched with measured concentrations of Pb in house dust and yard soil at the children’s residences, which were assumed to be the major source of exposures to Pb from soil and dust. We did not attempt to verify this assumption; however, it might be possible to do so from analysis of data collected in home interviews, as was done in the Hogan et al. [10] study. We did consider an alternative 40/30/30 partitioning of exposures that included contributions from house dust, yard soil, and community-wide soil, respectively [13,14]. Site-wide, this alternative partitioning of exposures yielded poorer agreement between observed and predicted GM BLL than the 55/45 (house dust/soil) default in the IEUBK model. When stratified by geographic area, however, the 40/30/30 partition agreed similarly in the Box or slightly better in the Lower Basin than the 55/45 default (data not shown).

Numerous factors other than measured soil and dust Pb concentrations could have affected model performance that were not measured in our study. Concentrations of Pb in drinking water were known from direct measurement of tap water for approximately 5% of children (in the Basin area of the site); for all other children, no drinking water data were available and they were assigned the IEUBK v2.0 model default concentration of 0.9 μg/L. Private sourced drinking water concentrations measured over a 12-year period in the Basin area averaged 2.3 μg/L, ranging from 0.9 to 8.8 μg/L (N=55). However, an underestimate of the average drinking water Pb concentration of 1.4 μg/L (i.e., 2.3 minus 0.9 μg/L) would have lowered the predicted GM BLL by <0.15 μg/dL. This small difference would not have appreciably affected model performance. All children were assigned the IEUBK model default value for Pb in air of 0.1 μg/m³. Annual averages in the Box for the period 1995 to 1998 reportedly ranged from 0.04 to 0.07 μg/m³. Although these values are uncertain, this suggests the model may have overestimated exposures to air Pb and resulting predicted GM BLL by <0.02 μg/dL (a negligible effect). Consequently, using IEUBK model default air values is unlikely to have introduced substantial error into the predictions of GM BLLs. All children were assigned IEUBK v2.0 model default values for dietary intake of Pb in market basket foods, which may not have represented the dietary Pb exposures of individual children. Uncertainty in the dietary Pb exposures becomes increasingly important at lower population GM BLLs because the contribution of food to total Pb uptake and BLL increases [15]. The IEUBK v2.0 model predicts a GM BLL concentration of 2.3 μg/dL (P₅=5%) when the soil Pb level is 200 mg/kg (and air, water and diet exposures are set to national default values). At a soil Pb level of 200 mg/kg, the contribution of soil (and soil-derived house dust) to total Pb uptake is predicted to be approximately 53%, while the contribution of dietary Pb is approximately 42%.

The above limitations regarding information on individual child exposures would be potentially significant if this study were intended to quantify variance in individual child BLL explained by the model. However, the objective of this study focused on evaluation of performance of the model to predict population GM BLLs and P₅ as the model would be typically applied in a human health risk assessment of CERCLA sites. For these types of assessments, data on dietary Pb, air Pb, site-specific data on exposures away from the household, or estimates of soil and dust ingestion rates are rarely available for use in the IEUBK model. The results of our study suggest, that even in the absence of complete information on exposure and reliance on default values from contributing exposures, the model performed well at predicting the population GM BLL and P₅, as well as the overall distribution of individual child BLLs in the selected BHSS population.

Conclusions

In general, our study found that the IEUBK v2.0 model performed well at predicting population GM BLLs of the evaluation dataset. The model also predicted the age trend in observed GM BLLs and explained approximately 90% of the variance in the age-stratified GM BLLs (r² =0.90). Although the more general applicability of the model remains to be determined in future studies of other populations, these results support applications of the IEUBK v2.0 model for informing risk-based discussions regarding remediation of soils and mitigation of exposures at CERCLA sites where the majority of the EU GM BLLs are expected to be ≤ 5 μg/dL. Seventy percent of the BLLs were below 5 μg/dL, which allowed us to evaluate the use of the model for estimating the probability of exceeding this BLL. Given that uncertainty in the dietary Pb exposures becomes increasingly important at lower population GM BLLs, the default dietary intakes may need to be carefully considered if predicting the probability of exceeding BLLs lower than 5 μg/dL. Overall, our evaluation of the IEUBK version 2.0 model performance supports its use as it would be typically applied at CERCLA sites to estimate preliminary remediation goals for Pb in soil and soil-derived dusts that limit the probability to less than a 5% chance of children’s BLLs exceeding 5 μg/dL.

Impact Statement.

The Integrated Exposure Uptake Biokinetic (IEUBK) Model for Lead in Children was developed by the U.S. EPA to assess risks to children from lead exposures. Since the most recent IEUBK performance evaluation in 1998, several important model changes were made to IEUBK v2.0 that prompted the need for this evaluation. This study is the first empirical evaluation of IEUBK v2.0 performance and provides strong support for applications of the IEUBK v2.0 model in lead-related human health risk assessments as typically applied at Superfund sites.

Data Availability

The data used for this evaluation were collected during routine BLL monitoring at the Bunker Hill Superfund (BHSS) site by the local Kellogg Idaho Panhandle Health District under the auspices of the Idaho Department of Environmental Quality (IDEQ). IDEQ contracts with Alta Science and Engineering, Inc. for technical, scientific, and engineering services at the BHSS, including maintaining the blood Pb and environmental databases. Alta maintains these data in a confidential (coded) format that prevents the identification of the participants and their home locations. With permission of IDEQ, Alta staff used paired BHSS BLL and environmental media concentration data to evaluate the IEUBK model v2.0. Transfer of individual or confidential data from Alta to SRC and EPA was expressly prohibited. Additional model evaluation information is available publicly at https://cfpub.epa.gov/ncea/risk/recordisplay.cfm?deid=351563. Additional BHSS details are publicly available at https://cumulis.epa.gov/supercpad/cursites/csitinfo.cfm?id=1000195.