To evaluate the relationship between total or bioaccessible Pb and children’s BLLs, soil samples were collected from yards in selected urban residential homes in Philadelphia, Pennsylvania, USA. The majority of the Philadelphia single-family row-style homes included in the study were built before 1900. Soil samples were collected from neighborhoods (Kensington, Olde Richmond, Port Richmond, North Philadelphia East, and Fishtown) with a history of nearby industrial operations dating back to the mid-1800s, several of which are known to have released Pb in air emissions. The study area comprised ZIP codes, 19125, 19134, 19145, 19146, 19147 and 19148. During 2015, based on PDPH child blood lead surveillance data, the percent of children with BLLs 5–9 μg/dL ranged from <1–6% (PDPH, 2015). Of particular interest in this investigation was a large Pb paint and product manufacturer that operated from 1849 to 1996 (ATSDR, 2015) located near map locations 1, 3, 16, 18, 21, 24, 37 (TOC Map). The area adjacent to this former manufacturing facility has been the subject of various environmental and public health investigations since the 1970s. Most industries in all the neighborhoods where soil was collected had closed by the 1970s, corresponding to a nationwide decline in many types of manufacturing operations.

Many of the yards in this row-home section of Philadelphia are quite small (approximately 15–20 yd²), with the majority of yards sampled containing both bare soil and concrete. Surface soil samples were collected primarily from areas of residential yards where parents or guardians said children play. Samples were randomly selected from a subset of yards in the original study and collected from both bare and vegetative soil depending on yard conditions, although bare soil was preferred for sampling. In some cases, samples were collected from a garden area if the small yard was used solely for this purpose. Five discrete soil samples were collected from approximately 0 to 8 cm deep to make one composite sample for each location. Small rocks or pebbles and other foreign materials were removed from the discrete samples before combining the samples in a 3.78 L (1 gallon) resealable plastic bag. An aliquot of soil from this composite sample was sent to U.S. EPA’s Office of Research and Development (Research Triangle Park, North Carolina, USA) under chain of custody for further processing and analysis. Samples were dried (< 40 ºC), sieved to < 250 μm, homogenized, and riffled for mixing and splitting of samples for total and bioaccessible Pb determination. A total of 38 composite samples from independent homes were included in this study.

Total Pb concentration in soil samples was determined using microwave-assisted acid digestion in accordance with U.S. EPA SW-846 Test Method 3051A (U.S. EPA, 2007b), with analysis by inductively coupled plasma–optical emission spectrometry in accordance with U.S. EPA SW-846 Method 6010D (U.S. EPA, 2014). The in vitro bioaccessible (IVBA) fraction of Pb in soil samples was determined using U.S. EPA SW-846 Method 1340 (U.S. EPA, 2013). Briefly, 1 g of soil sample was extracted in 100 mL of 0.4 M glycine solution acidified to a pH of 1.5 using hydrochloric acid and rotated end over end in a 37 ºC water bath for 1 hour. Extracts were filtered and analyzed for Pb concentration by inductively coupled plasma–mass spectrometry in accordance with U.S. EPA SW-846 Method 6020A (U.S. EPA, 1998). Lead bioaccessibility was reported as both a mass fraction (mg bioaccessible Pb per kg of soil) and a percentage as follows: (1)bioaccessiblePb(mgkg)=IVBAextractedPb(mgL)×extractionvolume(L)soilmass(kg) (2)bioaccessiblePb(%)=(massfractionbioaccessiblePbtotalPb)

To protect human participants, institutional review board approval was obtained from both the Centers for Disease Control and Prevention and the Philadelphia Department of Public Health (PDPH). Children’s BLLs were measured within the last two weeks of July 2014 from 49 children aged 1 to 7 years who were associated with the 38 homes where soil samples were collected (some homes had more than one child). For blood collection from participating children, a trained health care provider collected the venous blood samples. The procedure for collecting and processing blood samples and blood collecting materials is described in the Supporting Information (SI). Venous samples were stored in coolers on cold packs and delivered daily to the PDPH laboratory. Blood sample tubes remained at 10–32 ºC and were analyzed for Pb by the PDPH laboratory within 24 hours of collection using an atomic absorption method developed by PDPH on an AAnalyst 600 (PerkinElmer, Waltham, Massachusetts, USA). The PDPH laboratory is certified by the State of Pennsylvania (state laboratory ID 000255A) and the Centers for Medicare and Medicaid Services (CLIA ID 39D0657854). The lower limit of detection for BLLs was 0.1 μg/dL.

Regression models were developed to evaluate the relationship between children’s BLL (the response variable) and soil Pb mass fraction (the predictor variable). Total Pb and bioaccessible Pb in the soil were used as alternate predictor variables in two candidate model formulas. Models were fit using R statistical software (R Core Team, 2013) and were evaluated based on the coefficient of determination (R²), which represents the proportion of the variance in the response variable that is explained by the model (Kutner, 2005). In addition to R², the mean absolute error (MAE) is also reported to characterize the magnitude of model errors.

Before final model development, potential BLL outliers were identified by calculating their Studentized deleted residual (SDR) (Kutner, 2005). A data point with a high SDR exerts a disproportionate influence in a model and could represent an outlier individual in the sample population (e.g., a child who engages in behavior that increases Pb exposure risk). SDR values were calculated based on the residuals of simple linear regression models between BLL and soil Pb (total or bioaccessible). Outliers were identified from the resulting t distributions using the Bonferroni test procedure. The “car” package (Fox and Weisberg, 2011) in the R statistical software environment was used in this portion of the analysis. Identified outliers were excluded from the final model development but are noted in the results and discussion.

The assumption of independence among samples taken across a geographic area can only be made after accounting for inherent correlation between nearby data locations, i.e., spatial autocorrelation (Tobler, 1970). With spatial autocorrelation present, samples taken in close proximity can over represent their shared environmental conditions in a model if they are given equal influence with samples taken from a broader geographic distribution. In these cases, the spatially clustered samples represent the same environmental phenomenon and thus are not independent observations (Waller and Gotway, 2004). The extent of spatial autocorrelation in the BLL data was assessed by mapping the error (i.e., residuals) of the BLL predictions of simple linear regression models based on total and bioaccessible Pb (SI Figure 1). If spatial autocorrelation were present, similar residual values were expected to be clustered at the spatial scale of correlation.

Based on the mapping of residuals, spatial correlation of BLLs was observed to be limited to within individual households, indicating that blood samples collected from children within the same household were not independent. To address this, a hierarchical model structure (Gelman and Hill, 2007) was developed that grouped samples by households to account for the variation among households separately from the variation among individual samples: (3)BLL=βint+αhouse+βPbPb+ε

The hierarchical model has a varying intercept offset for each household, α_house, which follows a normal distribution, while the intercept (β_int), predictor variable (slope) coefficient (β_Pb), and error (ε) are analogous to their simple linear regression counterparts. The coefficients for the intercept and predictor variable are often referred to as “fixed effects” of the hierarchical model, and the varying intercept offsets for households are known as “random effects” (Gelman and Hill, 2007). Hierarchical models for BLL with total Pb or bioaccessible Pb as the predictor variable were fit using the packages lme4 (Bates et al., 2014) and lmerTest (Kuznetsova et al., 2016) in the R statistical software environment. The R² values reported in this paper represent the variance explained by the fixed effects of the model only and do not include the variance explained by the house-level random effects, which would be unknown for households outside the calibration data set. Similarly, the reported MAE values were based on model predictions that only took fixed effects into account.

A bootstrapping technique (Efron, 1979) was used to assess whether BLL prediction performance (R²) improved significantly when bioaccessible Pb rather than total Pb was used as the predictor variable. Bootstrapping involved resampling with replacement from the original data set to generate simulated data sets and calculating the R² of models fit to each new data set (Ohtani, 2000). In this case, 10,000 bootstrapped data sets were generated, and the difference in R² of the total versus bioaccessible Pb fitted models was calculated for each. The distribution of differences in R² was compared to a null hypothesis of zero difference at a 95% confidence level.

To evaluate the influence of child age on the relationship between total or bioaccessible soil Pb levels and BLLs, hierarchical model residuals were plotted against child age to identify any patterns. The residuals, which represent the variation in data unexplained by soil Pb level, would be expected to exhibit a trend when plotted with child age if age were indeed an additional source of variation in BLLs (Kutner, 2005). Additionally, a hierarchical multiple regression model including both bioaccessible Pb and child age as predictors was explored as an alternative to the hierarchical model with only bioaccessible Pb as the predictor variable.

Table 1 summarizes total and bioaccessible Pb levels in soils. Total soil Pb values for the 38 soil samples ranged from 58 to 2,821 mg/kg (mean = 751 mg/kg; median = 722 mg/kg). In 27 of these samples, Pb concentration exceeded 400 mg/kg. Notably, Pb is considered a hazard in children’s play areas when its concentration equals or exceeds 400 mg/kg in bare soil (U.S. EPA, 2001).

The bioaccessible fraction of Pb in soil ranged from 47 to 2,567 mg/kg (mean = 706 mg/kg; median = 668 mg/kg). These bioaccessibility values corresponded to a range of 66% to 100% of total Pb in soil (mean = 93%; median = 98%).

BLLs for 49 children living in homes paired with the 38 soil samples were included in this study (Table 1). Observed BLLs ranged from 0.3 to 9.8 μg/dL (mean = 3 μg/dL; median = 1.8 μg/dL). Analysis for potential outlier BLLs resulted in the removal of one value (BLL = 9.8 μg/dL). This outlier value was singularly high, as it was more than three standard deviations above the mean BLL and more than twice the value of the other child sampled in the same household. The high value could be due to behavioral differences exhibited by the child (e.g., mouthing activity). The Bonferroni p-values for this sample were 0.00041 for the total Pb model SDR and 0.00032 for the bioaccessible Pb model SDR. These results strongly suggested that this sample lay outside of the trends of the remaining data and could not be included in the modeling data without unduly influencing the model (Kutner, 2005).

Visual inspection of the spatial distribution of the residuals from simple linear regression models of BLL against Pb content (SI Figure 1) indicated that BLL values were correlated at the household scale. Correlation at this scale could be expected due to the likelihood of shared behavioral and environmental factors affecting individuals from the same household. Therefore, hierarchical linear regression models were fit to BLL against total and bioaccessible Pb, grouping samples by their household of origin and the corresponding soil sample.

All models indicated a significant positive relationship between BLL and soil Pb level (total or bioaccessible). The hierarchical BLL model based on total soil Pb had an intercept of 1.3 μg/dL BLL and a slope of 1.4 μg/dL BLL per 1000 mg/kg increase in total Pb in soil (Table 2a). The fitted hierarchical model with bioaccessible Pb as the predictor had an intercept of 1.2 μg/dL BLL and a slope of 1.6 μg/dL BLL per 1000 mg/kg increase of bioaccessible Pb in soil (Table 2b and Figure 1). Total Pb was a significant predictor of BLL at the 99% confidence interval (p = 0.0021), and bioaccessible Pb was significant at the 99.9% confidence level (p = 0.00092). These results are consistent with earlier reports of BLL increases ranging from 0.6 to 6.8 μg/dL for every 1000 mg/kg increase in soil Pb mass fraction (Lanphear et al., 1998; U.S. EPA, 1983; Duggan, 1980).

The BLL hierarchical model with total soil Pb as the predictor accounted for 25% of the variability in child BLL, and the model with bioaccessible soil Pb as the predictor accounted for 28% of BLL variability. There was also a small decrease in MAE from 1.22 to 1.19 μg/dL when predicting BLL using bioaccessible Pb rather than total Pb in soil. These results indicate an improvement in prediction accuracy (based on fixed effects) when predicting BLL with bioaccessible Pb relative to total Pb.

The residual variances of the two models (Table 2a–b) were similar ( σε2=1.4) and greater than the variances of the household random effects ( σhouse2=0.92 and 0.84 for models based on total Pb and bioaccessible Pb, respectively). The household random effect variances are consistent with the observation of spatial correlation at the household level (SI Figure 1). Comparing the household random effect variances for the two models (Table 2a–b) suggests that the use of bioaccessible Pb as a predictor reduced the unexplained variability at the household level relative to the model using total Pb.

The small difference in R² between the two models is not surprising given the high average bioaccessibility of Pb in the soils, such that bioaccessible and total Pb measurements were highly correlated (r = 0.996). Nevertheless, the results suggest that accounting for the bioaccessible fraction of Pb in soils can explain significantly more of the variation in children’s BLL than total Pb in soil. Using the total Pb concentration in soil may overestimate risk associated with soil ingestion. Bioaccessibility data are used to improve the accuracy of exposure and risk calculations (U.S. EPA, 2007a). A bootstrapping analysis comparing the total Pb and bioaccessible Pb hierarchical single-predictor models (Figure 2) showed that the model using bioaccessible Pb explained significantly more variance in BLL by a mean R² margin of 2.7%, with a 95% confidence interval of 0.7% to 5.2%. For 99.2% of the sample bootstraps, this model explained more variance than the model using total soil Pb as the predictor.

Ren et al. (2006) reported that correlation coefficients of linear relationships between bioaccessible Pb in soil and children’s BLLs increased as children aged from 3 to 5 years old. In the present study, which includes children between the ages of 1 and 7, child age did not contribute significantly to the prediction of BLL. Figure 3 shows that there is no clear relationship between child age and the single-predictor model residuals (SI Figure 2) for the hierarchical BLL model using bioaccessible soil Pb (Kutner, 2005). The absence of an observable correlation between child age and the unexplained variance in these models suggests that child age did not have a substantial effect on BLL. Furthermore, a hierarchical multiple regression model including both bioaccessible Pb and child age did not assign a significant coefficient (p = 0.56) to child age (Table 2c). Interestingly, the hierarchical multiple regression model including child age explained less of the variance in BLL (R² = 0.27) than the hierarchical regression using bioaccessible Pb only (R² = 0.28). One would normally expect R² to increase with the inclusion of an additional predictor. In hierarchical modeling, however, this is not necessarily the case. For the hierarchical multiple regression developed here (Table 2c), more of the variance is partitioned to the household random effect component of the model ( σhouse2=1.03), whereas the R² values were determined based on the fixed effect model components only. Additional variables (e.g., age of home), which may influence BLL may be further evaluated at a later time.

Our study compared relationships between total versus bioaccessible Pb in soil and children’s BLLs. The observed statistically significant correlations between both total and bioaccessible Pb in soil and children’s BLLs confirm that soil Pb is an exposure route of concern in this study area, accounting for 25% and 28% of the observed variability in BLLs, respectively. Contributions of soil Pb levels to variability in children’s BLLs were within the range of previously published values. While our study focused on soil Pb values at the local yard scale, a recent study evaluating Pb exposure at a Superfund site in Idaho reported that soil Pb at the neighborhood and community scales contribute independently to total Pb uptake by children (von Lindern et al., 2016). Additionally, other potential sources of Pb exposure not considered in this study, including Pb in house dust or paint, would also contribute to Pb exposure.

Unlike findings reported by Ren at al. (2006), however, our findings did not find age to be a factor in the relationship between soil Pb levels and children’s BLLs. Although accounting for child age did not contribute to improved BLL prediction in our study, recent studies that evaluated Pb intake rates from both soil and dust have reported differences in Pb intake rates from house dust based on age with a general trend of higher dust ingestion rates for younger children (von Lindern et al., 2016; Wilson et al., 2013). These findings suggest that differences in Pb exposure by age might be more attributable to Pb levels in house dust than soil.

The significant improvement in model R² values indicate that the bioaccessible portion of Pb in soil might better account for variance in observed children BLLs than total soil Pb level. This finding underscores the importance of considering soil metal bioavailability, either measured directly or estimated using an IVBA assay in exposure and risk assessments of metal-impacted soils.

When available, environmental public health practitioners may improve their child lead-risk assessments by using the bioaccessible measure of soil Pb instead of total soil Pb content. We found the bioaccessible fraction of Pb explained significantly more of the variation in children’s BLL than total Pb in soil. Additionally, using total soil Pb concentration may overestimate risk associated with soil ingestion. Bioaccessibility data, a measure of bioavailability, can improve the accuracy of exposure and risk calculations. Even small changes in bioavailability can result in substantial cost savings in remediation. It may be helpful to follow EPA decision making guidance regarding assessment and incorporation of site-specific bioavailability information into risk-assessment decision-making. (U.S. EPA, 2007c).