Journal of Exposure Science and Environmental Epidemiology (2015) 25, 411–416 & 2015 Nature America, Inc. All rights reserved 1559-0631/15
www.nature.com/jes
ORIGINAL ARTICLE
Effect measure modification of blood lead–air lead slope factors Jennifer Richmond-Bryant1, Qingyu Meng2, Jonathan Cohen3, J. Allen Davis1, David Svendsgaard1, James S. Brown1, Lauren Tuttle4, Heidi Hubbard3, Joann Rice5, Ellen Kirrane1, Lisa Vinikoor-Imler1, Dennis Kotchmar1, Erin Hines1 and Mary Ross1 There is abundant literature finding that susceptibility factors, including race and ethnicity, age, and housing, directly influence blood lead levels. No study has explored how susceptibility factors influence the blood lead–air lead relationship nationally. The objective is to evaluate whether susceptibility factors act as effect measure modifiers on the blood lead–air lead relationship. Participant level blood lead data from the 1999 to 2008 National Health and Nutrition Examination Survey were merged with air lead data from the US Environmental Protection Agency. Linear mixed effects models were run with and without an air lead interaction term for age group, sex, housing age, or race/ethnicity to determine whether these factors are effect measure modifiers for all ages combined and for five age brackets. Age group and race/ethnicity were determined to be effect measure modifiers in the all-age model and for some age groups. Being a child (1–5, 6–11, and 12–19 years) or of Mexican-American ethnicity increased the effect estimate. Living in older housing (built before 1950) decreased the effect estimate for all models except for the 1–5-year group, where older housing was an effect measure modifier. These results are consistent with the peer-reviewed literature of timeactivity patterns, ventilation, and toxicokinetics. Journal of Exposure Science and Environmental Epidemiology (2015) 25, 411–416; doi:10.1038/jes.2014.46; published online 25 June 2014 Keywords: child exposure/health; criteria pollutants; exposure modeling; empirical/statistical models; metals
INTRODUCTION Recently, Richmond-Bryant et al.1 have merged ambient air quality data from the US Environmental Protection Agency’s Air Quality System (AQS) database with the National Health and Nutrition Examination Survey (NHANES) database to evaluate the relationship between blood lead (PbB) and ambient air lead (PbA) for a cross-section of the US population following the phase-out of leaded gasoline, with recent (1999–2008) PbA averaging 0.14 mg/m3 for 365 days and recent PbB averaging 1.9 mg/dl.1 This study, in conjunction with other studies in the literature of the PbB–PbA relationship, indicates that changes in PbB level are more sensitive to the changes in PbA concentrations for smaller values of PbA.2–8 These results suggest that a larger relative decrease in PbB may be derived from PbA reduction efforts when PbA is already low. There is substantial evidence in the literature that PbB levels are directly influenced by susceptibility factors, including race and ethnicity, life stage, and housing conditions. In an analysis of PbB levels among pregnant women, a life stage in which bone turnover increases, Miranda et al.9 found that the odds of having PbB Z1 mg/dl were significantly higher among Hispanic women (median odds ratio (OR) ¼ 4.92, 95% confidence interval: 2.39– 10.09) and non-Hispanic black women (median OR ¼ 2.90, 95% confidence interval: 1.79–4.70) compared with non-Hispanic white women. Jones et al.10 analyzed NHANES data for PbB for race and ethnicity (defined in NHANES as non-Hispanic black, MexicanAmerican, non-Hispanic white, other Hispanic, and other race) among children aged 1–5 years and observed that the nonHispanic black group had higher proportions of children with PbB
45 mg/dl compared with the other two race/ethnicity groups. At the same time, the disparity in PbB between race/ethnicity groups was shown to decline between 1988 and 1991 and between 1999 and 2004 in the analysis of Jones et al.10 In both surveys, age in years produced a statistically significant increase in PbB within the 20–59-year age groups, whereas age produced significant decreases in PbB among the 6–11-year and 12–19-year age groups.10 Jacobs et al.10 concluded from combined analysis of NHANES with the American Housing Survey from 1970 to 2000 that percent older housing (built before 1950) trends with the geometric mean of PbB for study participants of all ages.11 Egeghy et al.12 observed that indoor PbA and indoor Pb in dust were significantly associated with home age (Po0.028); however, PbB was not associated with home age (home age was not included in the final model because it was not a significant predictor of PbB in the initial model). Likewise, Miranda et al.9 did not observe a significant increase in the odds of having PbB Z1 mg/dl for age of housing (median OR ¼ 0.99, 95% confidence interval: 0.98–1.00) among pregnant women aged 18–44 years. Although there is collective evidence of direct effects of susceptibility factors on PbB, no study explores how these factors influence the relationship between PbB and PbA. Effect measure modifiers are terms that interact with a regression model slope to change the relationship between a dependent and independent variable, in this case PbB and PbA, respectively. The objective of this study is to evaluate whether susceptibility factors, including age group, sex, housing age, and race/ethnicity, act as effect measure modifiers on the PbB–PbA relationship. An effect measure modification example is provided in Strand et al.,13 who
1 National Center for Environmental Assessment, US Environmental Protection Agency, Research Triangle Park, North Carolina, USA; 2School of Public Health, Rutgers University, Piscataway, New Jersey, USA; 3ICF International, Fairfax, Virginia, USA; 4School of Architecture, University of Texas at Austin, Austin, Texas, USA and 5Office of Air Quality Planning and Standards, US Environmental Protection Agency, Research Triangle Park, North Carolina, USA. Correspondence: Dr. Qingyu Meng, School of Public Health, Rutgers University, 683 Hoes Lane West, Piscataway, NJ 08854 USA. Tel.: þ 1 732 235 9754. Fax: þ 1 732 235 4004. E-mail: [email protected] Received 21 January 2014; revised 10 April 2014; accepted 23 April 2014; published online 25 June 2014
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
412
20–59 Years 8066 Med IQR 0.0055 0.0064 1.4 1.3 % 38 29 36 35 47 53 19 10 49 22
Z60 Years 3594 Med IQR 0.0055 0.0064 2.0 1.0 % 17 31 41 28 50 50 15 6 61 19
developed a PM2.5 exposure model with a non-zero random interaction on the slope to incorporate the effect of individual susceptibility related to many factors, such as time–activity data, socioeconomic status, race/ethnicity, and sex, on the exposure– response relationship. The present study explores the influence of effect measure modification using the interaction between each susceptibility factor and PbA levels in a linear mixed effects (LME) PbB model adapted from that presented in RichmondBryant et al.1
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
6–11 years 2337 Med IQR 0.0055 0.0063 1.4 1.2 % 11 29 35 36 50 50 28 11 27 33 0.0
0.0 Race/ethnicity
Age strata When home was originally built
PbA (mg/m3) (365-day average) PbB (mg/dl)
Sex
o1950 1950–1977 41977 Male Female Mexican American Other Hispanic or other race White Black
Value
1–5 Years 2204 Med IQR 0.0057 0.0065 1.9 1.6 % 10 30 36 34 51 49 29 13 29 29 — — % Missing — 19 % Missing 17 27 Age strata N (number)
Summary of PbB, PbA, and effect measure modifier data used in the LME, by age strata.
Table 1.
Participant-level data used for this analysis were obtained from the Continuous NHANES survey (1999–2008). NHANES variables used in the analysis included demographic, socioeconomic status, nutrition, housing, and occupational exposure data that are listed in Supplementary Table S1. The NHANES data were obtained in 2-year cycles, and definitions for some categorical variables changed slightly between cycles. Where appropriate, categorical variables were recoded to make the definitions consistent among cycles. Data were stratified by age group (1–5, 6–11, 12–19, 20–59, and Z60 years) to reflect differences in Pb storage, Pb distribution, Pb metabolism, and hand-to-mouth behavior among young children, older children, adolescents, adults, and older adults. The sample initially comprised 41,961 individual participants. This information is also summarized in Richmond-Bryant et al.1 Characteristics of the study population are provided in Table 1 and Supplementary Table S2. Missing or uncertain NHANES data handling varied based on whether the variable was dependent or independent in the statistical analysis described below. In all, 19% of PbB data were missing from the NHANES survey (including true missing data as well as the responses ‘‘refused to answer’’ and ‘‘don’t know’’), and 0.076% of PbB data had values below the method detection limit (MDL) of 0.2 mg/dl. Because PbB was the dependent variable, the record for an individual was omitted from the file if the PbB datum was missing. When other NHANES variables were missing for an individual, statistical imputation was performed with an inhouse program developed using Matlab R2009a (The Mathworks, Natick, MA, USA). The imputation scheme involved estimation of the data’s statistical distribution and then random assignment of a value within the distribution.14 Statistical distributions of non-missing data were used to estimate the missing data as follows. For categorical variables, cumulative frequency distributions were developed for each categorical variable. A random seed between 0 and 1 was then created, and a categorical assignment was made depending on where the random seed fell within the cumulative frequency distribution. For continuous variables, a missing value imputed as the sum of the mean and standard deviation (SD) multiplied by a random seed between 1 and 1 as applied in the appropriate statistical distribution. Data for many of the variables were lognormally distributed, so that log-transformation was applied to those data. Sixteen of the NHANES covariate data had missing values that were imputed, and visual inspection suggests that the covariate data were missing at random before imputation. Percent imputed data by variable are provided in Supplementary Table S2. AQS data for PM10-PbA were merged to the NHANES participants by geocoded location, based on PM10-PbA monitors’ latitude, longitude, and sampling dates. PM10-PbA data were used, because they were found to have the strongest associations with PbB compared with PbA measured in the PM2.5 and total suspended particulate (TSP) size fractions.15 Data from 155 PM10-PbA monitors sampling for 24 h on 1 in every 3 to 6 days and averaged over 365 days produced 16,222 observations for this analysis. Missing PbA observations were not included in the analyses, although a 75% completeness criterion was observed. However, below-MDL data were not omitted to minimize positive bias on the PbA data distribution; 5.8% of data were below the MDL. AQS data were assigned to an NHANES participant if a PM10-PbA monitor was located within 4 km of a Census block centroid in which an NHANES participant resided to better represent personal exposure to Pb through air-related pathways, and the 24-h PM10PbA concentration measurement fell within 7 days of the NHANES participant examination date. The sample population after the data merge was 21,143 (with the breakdown by age provided in the Supplementary Table S2); in other words, 21,143 participants in the NHANES data set fell within the 4 km buffer around an PbA monitor and hence had data assigned to them from the PbA measurements. In addition, GIS data for total length of street per Census block group normalized by Census block
12–19 Years 4942 Med IQR 0.0060 0.0066 1.0 0.80 % 23 29 37 34 51 49 30 8 28 34
METHODS Data Sets
& 2015 Nature America, Inc.
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
413 group area for the year 2000 were also merged to the NHANES data set based on Census block group identification number. The street per Census block group variable was designed to account for living in a heavily trafficked environment and was adjusted for in the model in the same manner as was the other NHANES covariates.
Statistical Analysis The statistical modeling approach was described in Richmond-Bryant et al.1 and is summarized here. Multi-level LME models adjusted for the variables (e.g., age, race/ethnicity, and housing age) are listed in Supplementary Table S3. The LME analyses were implemented in SAS (SAS Institute, Cary, NC, USA) using PROC MIXED via the National Center for Health Statistics Research Data Center’s remote access system. Multi-level modeling was performed at the individual and Census block group levels, and random effects at the Census block group level were random intercepts for each Census block group. The subject age stratum was treated as a stratification variable except when age was used as an effect measure modifier. The geographical location was treated as a random effect. The LME model without effect measure modification is of the form: lnðPbBi;j Þ ¼ b0 þ bi Zi þ bj þ bPbA lnðPbAj Þ þ ei;j
ð1Þ
where PbBi,j is the PbB for the ith individual living in the jth Census block group, Zi is the vector of covariates for the ith individual, bi is the associated coefficient vector, PbAj is the average PbA concentration obtained at the jth Census block group over a 365-day averaging period, and bPbA is the effect estimate of ln(PbA) on ln(PbB). The log-transform was applied after inspection of quantile–quantile distribution plots that indicated lack of normality for the untransformed data and normality once the log-transform was applied. The covariate and PbA concentration variables are fixed effects. b0 is the overall intercept, bj is a Census block group-level random normal intercept with mean zero and variance t2, and ei,j is an individual-level random normal variable with mean zero and variance s2. The LME models were run a second time with an PbA interaction term for age group, sex, housing age, or race/ethnicity to determine whether those factors acted as effect measure modifiers. When effect measure modification was factored into the model, the LME is of the form: lnðPbBi;j Þ ¼ b0 þ bi Zi þ bj þ bint;j Z lnðPbAj Þ þ ei;j
ð2Þ
where Z* is the variable (i.e., age group, sex, housing age, or race/ethnicity) used to test effect measure modification on the PbA term, and bint,j is the effect estimate for the interaction between Z* and ln(PbAj). Only one effect measure modification variable was included for each model run. Age group was used as a stratification variable for all models except for the case where age group was employed as an interaction term in the all-ages model. Likelihood ratio testing was performed to test the statistical significance of the interaction term by comparing model (1) with model (2). NHANES survey weights were not used in the analysis of this data set, as discussed in Richmond-Bryant et al.,1 based on analyses by Gelman16 and Rabe-Hesketh and Skrondal,17 and blood concentration values below the limit of detection (LOD) were substituted with a value of the LOD*2 0.5. Statistical imputation methods were used to substitute for missing values of the NHANES demographic and socioeconomic variables.1 NHANES records were discarded if they did not contain PbB data. For the air quality
Table 2.
RESULTS AND DISCUSSION The likelihood ratio testing results demonstrated which interaction terms in model (2) were statistically significant compared with model (1) (see Table 2). Statistical significance is tested at the 5% level. Age group was statistically significant as an effect measure modifier in the all-ages model (Po0.001). Housing age was a statistically significant effect measure modifier for 6–11-year (P ¼ 0.01), 12–19-year (P ¼ 0.024), 20–59-year (P ¼ 0.017), and the all-ages models (P ¼ 0.001). Race/ethnicity was statistically significant in the 12–19-year group (P ¼ 0.017) and in the all-ages model (P ¼ 0.004), and it was marginally significant in the 20–59-year age group (P ¼ 0.060). Although using sex as an effect measure modifier never produced a slope that was statistically significantly different from model (1), being male produced an increased slope for the allages model and for all age groups except 1–5-year-old children. The effects of age group, housing age, race, and sex on PbB–PbA associations are illustrated in Figure 1, and model parameters are provided in Supplementary Table S3. Age group was only tested with the all-age model, and the effect measure modification model demonstrated statistically significant increases in effect estimate for the 1–5-year and 12–19-year age groups and an observed but not statistically significant increase for the 6–11-year age group (see Figure 1 and Supplementary Table S3). The effect decreased for the 20–59-year age group. Bierkens et al.2 modeled the PbB–PbA relationship among children o6 years old and adult men and women at least 18 years old using a ln-ln model stratified by age (and sex among adults). They observed large contrasts between the effect estimate for children (b ¼ 0.09) compared with adult men (b ¼ 0.44) or women (b ¼ 0.34), corresponding to only a 16% drop in PbB per unit change in PbA over the period 1990–2007 for young children compared with 57% and 48% in men and women, respectively. The results of Bierkens et al.2 contrast those found here, where effect measure modification caused higher effect estimates among children compared with adults. Time–activity studies provide support for this study’s finding of effect measure modification among children. Wu et al.18 differentiated time– activity patterns among children (mostly o8 years old) from their parents (mostly under age 55 years) and older adults (mostly older than age 54 years) and illustrated that children were more likely than either adult group to partake in vigorous outdoor and indoor activities. Wu et al.18 also found that the pattern of time spent in different microenvironments varied among the age groups. Similarly, Klepeis et al.19 surveyed individuals in Canada and the United States about time–activities and found that children
Likelihood ratio testing of effect measure modification.
Age group
Effect measure modification category Age group
1: 1–5 2: 6–11 3: 12–19 4: 20–59 5: 60 þ All ages
variables, observations with missing and zero values of ambient Pb were not used in the analyses because of the logarithmic transformation that might introduce biases into the analyses.
Sex
Housing age
Race/ethnicity
Likelihood ratio
P-value
Likelihood ratio
P-value
Likelihood ratio
P-value
Likelihood ratio
P-value
— — — — — 38.3
— — — — — o0.001
2.5 0.2 0.9 0.3 0.9 1.4
0.114 0.655 0.343 0.584 0.343 0.237
1.2 9.2 7.5 8.2 3.5 14.3
0.549 0.010 0.024 0.017 0.174 0.001
3.1 5.3 10.2 7.4 3.7 13.4
0.376 0.151 0.017 0.060 0.296 0.004
Results are in bold where there was a statistically significant difference (a ¼ 0.05) between models with and without effect measure modification.
& 2015 Nature America, Inc.
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
414
Figure 1. Estimated deviation from b for the all-ages model and the age stratification model, and for models including interactions for age, sex, housing age, and race/ethnicity (n ¼ 21,143). b Has been shifted so that the term without interaction is zero, for comparison of the relative changes. The error bars represent 95% confidence intervals.
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
& 2015 Nature America, Inc.
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
415 o12 years old spent more time outdoors compared with individuals 412 years. Such differences in activities could result in differences among both exposure and intake of inhaled and ingested ambient PbA. Moreover, age-related ingestion via handto-mouth behaviors and toxicokinetic characteristics could also influence the PbB–PbA relationship.20 For example, studies indicate that 40–50% of ingested water-soluble Pb is absorbed by children o8 years old compared with 3–10% absorption in adult subjects.21–23 There was a trend for the effect estimates to be higher for males than for females in the all-age model and for every age strata except for 1–5-year group, for which the converse was observed (see Figure 1 and Supplementary Table S3). Bierkens et al.2 modeled the PbB–PbA relationship for adult men and women and found that model slope was 0.34 among women and 0.44 among men in a ln-ln model, resulting in declines in PbB with respect to unit declines in PbA of 48% in women and 57% in men from 1990 to 2007. Studies of time spent in physical activity and outdoors suggest that males spend more time performing vigorous activities and exercising outdoors compared with females.24–26 However, sex-based differences were not always statistically significantly different.24 These findings are consistent with our observation of a trend for higher effect estimates in males than females. Older housing (i.e., that built before 1950) elicited a statistically significant increase in effect estimate among 1–5-year-old children (Po0.0001) (see Figure 1 and Supplementary Table S3). However, older housing produced a statistically significant reduction in the effect estimate for 12–19 year olds (P ¼ 0.019) and produced reductions that were not statistically significant in the all-ages model and for the 6–19-year, 20–59-year, and Z60-year age groups. Moderate housing age (i.e., built between 1950 and 1977) and newer housing (i.e., built after 1977) typically produced a null effect or an increase in effect estimate in the all-ages model and among all age groups except 1–5-year-old children. The lack of statistical significance of the housing interaction term suggests that a model with no PbA effects in older housing but possible PbA effects in newer housing is not rejected. Hence, older housing does not appear to influence the relationship between PbB and PbA for people living in older housing, with the exception of young children and teens. It is possible that discrepancies in missingness of housing age among race/ethnicity groups contribute to uncertainties about the influence of housing age on the PbB–PbA relationship. For example, 40% of all MexicanAmerican and 39% of all non-Hispanic black participants did not report housing age, compared with 9% of non-Hispanic white participants of all ages. Among 1–5-year-old study participants, 47% of Mexican-American and 47% of non-Hispanic black participants’ heads of household did not report housing age, compared with 25% of non-Hispanic white participants’ heads of household. Imputation was performed using the data distribution of the entire study population combined, and hence it is possible that there is misspecification in the housing age variable. Such misspecification could also influence the role of housing age as an effect measure modifier of the PbB–PbA relationship. Home ventilation may explain how housing age might act as an effect measure modifier of the relationship between PbB and PbA. Jacobs et al.11 studied data from the American Housing Survey from 1970 to 2000 and observed that the use of central air conditioning increased with newer housing, although they noted a trend in addition of central air to older homes during the 1990s. Because central air conditioning systems employ filters, residents lacking central air would more likely be exposed to PbA in particulate matter infiltrating indoors given that the air would be unfiltered and likely introduced into the home through open windows. The only group where effect measure modification related to older housing increased the effect estimate was 1–5year-old children. Young children are more likely than older children or adults to spend time indoors and to have greater & 2015 Nature America, Inc.
hand-to-mouth exposure to indoor dust that may contain settled PbA particles.20,27 Therefore, ventilation differences related to housing age might only have been observable among the youngest cohort. Effect measure modification by race/ethnicity was fairly consistent for the all-age model and across the age strata (see Figure 1 and Supplementary Table S3). Mexican-American ethnicity had the largest increase in effect estimate for the allages model and for all age strata except for adults Z60 years and 6–11-year-old children, for which it still produced an increase. The effect estimate was roughly the same for those of black race as the model that did not test effect measure modification with the exception of elevated effect estimate for 6–11-year-old children and reduced effect estimate for adults 20–59 years old. The effect estimate was consistently reduced for the all-ages model and across all age strata for being of white race. Results were mixed across age groups for being of other Hispanic or other race. Several studies have examined the racial characteristics of populations living in the vicinity of industrial facilities, often using distance or presence of facilities required to report their emissions to the Toxic Release Inventory (TRI) because they emit a Hazardous Air Pollutant that may but not necessarily will include Pb.28–31 Chakraborty and Zandbergen28 observed that both Hispanic and African-American residents were more likely to live within a 4-mi buffer of a TRI industrial facility compared with white residents, and Wilson et al.31 observed statistically significant increases in the presence of a TRI facility in a Census block for nonWhite residents. Wilson et al.31 found that the odds ratio of living in a Census block containing a TRI facility increased for black residents but was not statistically significant.31 Perlin et al.30 did find that people of African American race were generally more likely to live in close proximity to TRI facilities compared with people of white race regardless of age, and Mohai et al.29 estimated statistically significant increases in the odds ratio for living within a mile of a TRI facility for percent of black race but not for percent of white race. Our effect measure modification results suggest that the PbB–PbA relationship is elevated for MexicanAmerican ethnicity, reduced for white race, and neither elevated nor reduced for black race. The results for black race were surprising given the literature documenting risk of living near a TRI for those of black race. Limitations of the study influence interpretation of the study results. The National Center for Health Statistics Research Data Center’s data restrictions, intended to protect NHANES participant confidentiality, included restrictions on visualizing the data and testing the residuals. These restrictions limited the ability to assess the spatial distribution of missing data among the NHANES participants. Averaging the data over 1 year and requiring 75% completeness of the data reduced the impact of missing data, although imputation to replace missing data related to monitor operationality would potentially introduce bias if the location of missing data was non-random.32,33 Missing PbB data in NHANES were likely to have been randomly distributed, and hence their omission would be less likely to bias the model slope.34,35 Although this study presented an analysis for a population spanning the United States, the omission of sampling weights prohibits national inference from the data. Sampling weights were not used in this analysis, because inclusion of sampling weights in multilevel models can lead to biased estimates for small cluster sizes17 and the process of combining sampling weights over several cycles and subpopulations can lead to inclusion of weights that are not inverse probabilities.16 Given that application of the 4 km buffer resulted in curtailing the sample population from 41,961 to 21,143, it is likely that the weights would not have been representative. Application of the buffers likely also resulted in reduced power. Moreover, application of the 4 km buffer may have added uncertainty to the exposure estimate. The coarse fraction of PM10-PbA is more likely to settle from the atmosphere
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
416 at closer distances to PbA sources compared with the fine fraction of PM10-PbA, so that the mass of PM10-PbA may have been underestimated if the NHANES participant’s location was further from the source than that of the monitor or vice versa. The primary finding of this cross-sectional epidemiologic study is that the effect estimate for the PbB–PbA relationship is modified by age strata and race/ethnicity in the all-age models and for some age strata. Specifically, the effect of PbA concentrations on PbB levels was enhanced by childhood (1–5, 6–11, and 12–19-year age brackets) and by being of Mexican-American ethnicity. Living in older housing (built before 1950) detracted from the effect of PbA on PbB except in the 1–5-year group. Our effect measure modification results are generally consistent with studies in the literature of time–activity patterns, ventilation, and toxicokinetics. However, this study is the first to investigate effect measure modification of the PbB–PbA relationship, and hence data from other studies to substantiate these findings are unavailable. CONFLICT OF INTEREST The authors declare no conflict of interest.
ACKNOWLEDGEMENTS We give special thanks to Dr. Tom Long and Dr. Zach Pekar for their helpful comments in review of this manuscript. We thank Ms. Nataliya Kravets of the National Center for Health Statistics, Centers for Disease Control, for linking the NHANES data with the AQS and GIS data. Authors from ICF International were funded by contract EP-C-09-009. Authors from academia and the US EPA did not receive external funding for this research.
DISCLAIMER The research and this manuscript have been reviewed in accordance with US Environmental Protection Agency policy and approved for publication. Mention of trade names or commercial products does not constitute endorsement or recommendation for use. The views expressed in this article are those of the authors and do not necessarily reflect the views or policies of the US Environmental Protection Agency.
REFERENCES 1 Richmond-Bryant J, Meng Q, Davis JA, Cohen J, Svendsgaard D, Brown JS et al. A multi-level model of blood lead as a function of air lead. Sci Total Environ 2013; 461: 207–213. 2 Bierkens J, Smolders R, Van Holderbeke M, Cornelis C. Predicting blood lead levels from current and past environmental data in Europe. Sci Total Environ 2011; 409: 5101–5110. 3 Brunekreef B. The relationship between air lead and blood lead in children - a critical-review. Sci Total Environ. 1984; 38: 79–123. 4 Hayes EB, McElvaine MD, Orbach HG, Fernandez AM, Lyne S, Matte TD. Long-term trends in blood lead levels among Children in chicago - relationship to air lead levels. Pediatrics 1994; 93: 195–200. 5 Ranft U, Delschen T, Machtolf M, Sugiri D, Wilhelm M. Lead concentration in the mood of children and its association with lead in soil and ambient air - Trends between 1983 and 2000 in Duisburg. J Toxicol Env Health Part A 2008; 71: 710–715. 6 Richmond-Bryant J, Meng QY, Davis AJ, Cohen J, Svendsgaard D, Kravets N et al. The influence of declining air lead levels on blood lead–air lead slope factors in children. Environ Health Perspect 2014, (e-pub ahead of print, doi:10.1289/ ehp.1307072). 7 Schwartz J, Pitcher H. The relationship between gasoline lead and blood lead in the United States. J Official Stat 1989; 5: 421–431. 8 Tripathi RM, Raghunath R, Kumar AV, Sastry VN, Sadasivan S. Atmospheric and children’s blood lead as indicators of vehicular traffic and other emission sources in Mumbai, India. Sci Total Environ 2001; 267: 101–108. 9 Miranda ML, Edwards SE, Swamy GK, Paul CJ, Neelon B. Blood lead levels among pregnant women: historical versus contemporaneous exposures. Int J Environ Res Public Health 2010; 7: 1508–1519.
10 Jones RL, Homa DM, Meyer PA, Brody DJ, Caldwell KL, Pirkle JL et al. Trends in blood lead levels and blood lead testing among US children aged 1 to 5 years, 1988-2004. Pediatrics 2009; 123: e376–e385. 11 Jacobs DE, Wilson J, Dixon SL, Smith J, Evens A. The relationship of housing and population health: a 30-year retrospective analysis. Environ Health Perspect 2009; 117: 597–604. 12 Egeghy PP, Quackenboss JJ, Sandra CC, Ryan PB. Determinants of temporal variability in NHEXAS-Maryland environmental concentrations, exposures, and biomarkers. J Exp Anal Environ Epidemiol 2005; 15: 388–397. 13 Strand M, Hopke PK, Zhao W, Vedal S, Gelfand E, Rabinovitch N. A study of health effect estimates using competing methods to model personal exposures to ambient PM2.5. J Exp Sci Environ Epidemiol 2007; 17: 549–558. 14 Donders ART, van der Heijden GJMG, Stijnen T, Moons KGM. Review: a gentle introduction to imputation of missing values. J Clin Epidemiol 2006; 59: 1087–1091. 15 Meng Q, Richmond-Bryant J, Davis JA, Cohen J, Svendsgaard D, Brown JS et al. Contribution of particle-size-fractionated airborne lead to blood lead during the National Health and Nutrition Examination Survey, 1999-2008. Environ Sci Technol 2014; 48: 1263–1270. 16 Gelman A. Struggles with survey weighting and regression modeling. Stat Sci 2007; 22: 153–164. 17 Rabe-Hesketh S, Skrondal A. Multilevel modelling of complex survey data. J R Stat Soc Ser A Stat Soc 2006; 169: 805–827. 18 Wu XM, Fan ZT, Ohman-Strickland P. Time-location patterns of a population living in an air pollution hotspot. J Environ Public Health 2010; 2010: 625461. 19 Klepeis NE, Nelson WC, Ott WR, Robinson JP, Tsang AM, Switzer P et al. The National Human Activity Pattern Survey (NHAPS): a resource for assessing exposure to environmental pollutants. J Exp Anal Environ Epidemiol 2001; 11: 231– 252. 20 USEPA. Integrated Science Assessment for Lead (Final Report). In: Agency USEP (ed). Washington, DC, 2013. 21 James HM, Hilburn ME, Blair JA. Effects of meals and meal times on uptake of lead from the gastrointestinal-tract in humans. Hum Toxicol 1985; 4: 401–407. 22 Maddaloni M, Lolacono N, Manton W, Blum C, Drexler J, Graziano J. Bioavailability of soilborne lead in adults, by stable isotope dilution. Environ Health Perspect 1998; 106: 1589–1594. 23 Watson WS, Morrison J, Bethel MIF, Baldwin NM, Lyon DTB, Dobson H et al. Food iron and lead absorption in humans. Am J Clin Nutr 1986; 44: 248–256. 24 Brownson RC, Baker EA, Housemann RA, Brennan LK, Bacak SJ. Environmental and policy determinants of physical activity in the United States. Am J Public Health 2001; 91: 1995–2003. 25 Corder K, Ekelund U, Steele RM, Wareham NJ, Brage S. Assessment of physical activity in youth. J Appl Physiol 2008; 105: 977–987. 26 McCurdy T, Graham SE. Using human activity data in exposure models: analysis of discriminating factors. J Exp Anal Environ Epidemiol 2003; 13: 294–317. 27 USEPA. Exposure Factors Handbook 2011 Edition (Final). In: Agency USEP (ed). Washington, DC, 2011. 28 Chakraborty J, Zandbergen PA. Children at risk: measuring racial/ethnic disparities in potential exposure to air pollution at school and home. J Epidemiol Commun Health 2007; 61: 1074–1079. 29 Mohai P, Lantz PM, Morenoff J, House JS, Mero RP. Racial and socioeconomic disparities in residential proximity to polluting industrial facilities: evidence from the Americans’ Changing Lives Study. Am J Public Health 2009; 99: S649–S656. 30 Perlin SA, Wong D, Sexton K. Residential proximity to industrial sources of air pollution: Interrelationships among race, poverty, and age. J Air Waste Manage Assoc 2001; 51: 406–421. 31 Wilson SM, Fraser-Rahim H, Williams E, Zhang H, Rice L, Svendsen E et al. Assessment of the distribution of toxic release inventory facilities in metropolitan Charleston: an Environmental Justice Case Study. Am J Public Health. 2012; 102: 1974–1980. 32 Helsel DR. Fabricating data: how substituting values for nondetects can ruin results, and what can be done about it. Chemosphere 2006; 65: 2434–2439. 33 Katsouyanni K, Schwartz J, Spix C, Touloumi G, Zmirou D, Zanobetti A et al. Short term effects of air pollution on health: A European approach using epidemiologic time series data: The APHEA protocol. J Epidemiol Commun Health 1996; 50: S12–S18. 34 Rubin D. Inference and missing data. Biometrika 1976; 63: 581–592. 35 Sterne JAC, White IR, Carlin JB, Spratt M, Royston P, Kenward MG et al. Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls. Br Med J 2009; 338: 2393.
Supplementary Information accompanies the paper on the Journal of Exposure Science and Environmental Epidemiology website (http:// www.nature.com/jes)
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
& 2015 Nature America, Inc.
www.nature.com/jes
ORIGINAL ARTICLE
Effect measure modification of blood lead–air lead slope factors Jennifer Richmond-Bryant1, Qingyu Meng2, Jonathan Cohen3, J. Allen Davis1, David Svendsgaard1, James S. Brown1, Lauren Tuttle4, Heidi Hubbard3, Joann Rice5, Ellen Kirrane1, Lisa Vinikoor-Imler1, Dennis Kotchmar1, Erin Hines1 and Mary Ross1 There is abundant literature finding that susceptibility factors, including race and ethnicity, age, and housing, directly influence blood lead levels. No study has explored how susceptibility factors influence the blood lead–air lead relationship nationally. The objective is to evaluate whether susceptibility factors act as effect measure modifiers on the blood lead–air lead relationship. Participant level blood lead data from the 1999 to 2008 National Health and Nutrition Examination Survey were merged with air lead data from the US Environmental Protection Agency. Linear mixed effects models were run with and without an air lead interaction term for age group, sex, housing age, or race/ethnicity to determine whether these factors are effect measure modifiers for all ages combined and for five age brackets. Age group and race/ethnicity were determined to be effect measure modifiers in the all-age model and for some age groups. Being a child (1–5, 6–11, and 12–19 years) or of Mexican-American ethnicity increased the effect estimate. Living in older housing (built before 1950) decreased the effect estimate for all models except for the 1–5-year group, where older housing was an effect measure modifier. These results are consistent with the peer-reviewed literature of timeactivity patterns, ventilation, and toxicokinetics. Journal of Exposure Science and Environmental Epidemiology (2015) 25, 411–416; doi:10.1038/jes.2014.46; published online 25 June 2014 Keywords: child exposure/health; criteria pollutants; exposure modeling; empirical/statistical models; metals
INTRODUCTION Recently, Richmond-Bryant et al.1 have merged ambient air quality data from the US Environmental Protection Agency’s Air Quality System (AQS) database with the National Health and Nutrition Examination Survey (NHANES) database to evaluate the relationship between blood lead (PbB) and ambient air lead (PbA) for a cross-section of the US population following the phase-out of leaded gasoline, with recent (1999–2008) PbA averaging 0.14 mg/m3 for 365 days and recent PbB averaging 1.9 mg/dl.1 This study, in conjunction with other studies in the literature of the PbB–PbA relationship, indicates that changes in PbB level are more sensitive to the changes in PbA concentrations for smaller values of PbA.2–8 These results suggest that a larger relative decrease in PbB may be derived from PbA reduction efforts when PbA is already low. There is substantial evidence in the literature that PbB levels are directly influenced by susceptibility factors, including race and ethnicity, life stage, and housing conditions. In an analysis of PbB levels among pregnant women, a life stage in which bone turnover increases, Miranda et al.9 found that the odds of having PbB Z1 mg/dl were significantly higher among Hispanic women (median odds ratio (OR) ¼ 4.92, 95% confidence interval: 2.39– 10.09) and non-Hispanic black women (median OR ¼ 2.90, 95% confidence interval: 1.79–4.70) compared with non-Hispanic white women. Jones et al.10 analyzed NHANES data for PbB for race and ethnicity (defined in NHANES as non-Hispanic black, MexicanAmerican, non-Hispanic white, other Hispanic, and other race) among children aged 1–5 years and observed that the nonHispanic black group had higher proportions of children with PbB
45 mg/dl compared with the other two race/ethnicity groups. At the same time, the disparity in PbB between race/ethnicity groups was shown to decline between 1988 and 1991 and between 1999 and 2004 in the analysis of Jones et al.10 In both surveys, age in years produced a statistically significant increase in PbB within the 20–59-year age groups, whereas age produced significant decreases in PbB among the 6–11-year and 12–19-year age groups.10 Jacobs et al.10 concluded from combined analysis of NHANES with the American Housing Survey from 1970 to 2000 that percent older housing (built before 1950) trends with the geometric mean of PbB for study participants of all ages.11 Egeghy et al.12 observed that indoor PbA and indoor Pb in dust were significantly associated with home age (Po0.028); however, PbB was not associated with home age (home age was not included in the final model because it was not a significant predictor of PbB in the initial model). Likewise, Miranda et al.9 did not observe a significant increase in the odds of having PbB Z1 mg/dl for age of housing (median OR ¼ 0.99, 95% confidence interval: 0.98–1.00) among pregnant women aged 18–44 years. Although there is collective evidence of direct effects of susceptibility factors on PbB, no study explores how these factors influence the relationship between PbB and PbA. Effect measure modifiers are terms that interact with a regression model slope to change the relationship between a dependent and independent variable, in this case PbB and PbA, respectively. The objective of this study is to evaluate whether susceptibility factors, including age group, sex, housing age, and race/ethnicity, act as effect measure modifiers on the PbB–PbA relationship. An effect measure modification example is provided in Strand et al.,13 who
1 National Center for Environmental Assessment, US Environmental Protection Agency, Research Triangle Park, North Carolina, USA; 2School of Public Health, Rutgers University, Piscataway, New Jersey, USA; 3ICF International, Fairfax, Virginia, USA; 4School of Architecture, University of Texas at Austin, Austin, Texas, USA and 5Office of Air Quality Planning and Standards, US Environmental Protection Agency, Research Triangle Park, North Carolina, USA. Correspondence: Dr. Qingyu Meng, School of Public Health, Rutgers University, 683 Hoes Lane West, Piscataway, NJ 08854 USA. Tel.: þ 1 732 235 9754. Fax: þ 1 732 235 4004. E-mail: [email protected] Received 21 January 2014; revised 10 April 2014; accepted 23 April 2014; published online 25 June 2014
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
412
20–59 Years 8066 Med IQR 0.0055 0.0064 1.4 1.3 % 38 29 36 35 47 53 19 10 49 22
Z60 Years 3594 Med IQR 0.0055 0.0064 2.0 1.0 % 17 31 41 28 50 50 15 6 61 19
developed a PM2.5 exposure model with a non-zero random interaction on the slope to incorporate the effect of individual susceptibility related to many factors, such as time–activity data, socioeconomic status, race/ethnicity, and sex, on the exposure– response relationship. The present study explores the influence of effect measure modification using the interaction between each susceptibility factor and PbA levels in a linear mixed effects (LME) PbB model adapted from that presented in RichmondBryant et al.1
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
6–11 years 2337 Med IQR 0.0055 0.0063 1.4 1.2 % 11 29 35 36 50 50 28 11 27 33 0.0
0.0 Race/ethnicity
Age strata When home was originally built
PbA (mg/m3) (365-day average) PbB (mg/dl)
Sex
o1950 1950–1977 41977 Male Female Mexican American Other Hispanic or other race White Black
Value
1–5 Years 2204 Med IQR 0.0057 0.0065 1.9 1.6 % 10 30 36 34 51 49 29 13 29 29 — — % Missing — 19 % Missing 17 27 Age strata N (number)
Summary of PbB, PbA, and effect measure modifier data used in the LME, by age strata.
Table 1.
Participant-level data used for this analysis were obtained from the Continuous NHANES survey (1999–2008). NHANES variables used in the analysis included demographic, socioeconomic status, nutrition, housing, and occupational exposure data that are listed in Supplementary Table S1. The NHANES data were obtained in 2-year cycles, and definitions for some categorical variables changed slightly between cycles. Where appropriate, categorical variables were recoded to make the definitions consistent among cycles. Data were stratified by age group (1–5, 6–11, 12–19, 20–59, and Z60 years) to reflect differences in Pb storage, Pb distribution, Pb metabolism, and hand-to-mouth behavior among young children, older children, adolescents, adults, and older adults. The sample initially comprised 41,961 individual participants. This information is also summarized in Richmond-Bryant et al.1 Characteristics of the study population are provided in Table 1 and Supplementary Table S2. Missing or uncertain NHANES data handling varied based on whether the variable was dependent or independent in the statistical analysis described below. In all, 19% of PbB data were missing from the NHANES survey (including true missing data as well as the responses ‘‘refused to answer’’ and ‘‘don’t know’’), and 0.076% of PbB data had values below the method detection limit (MDL) of 0.2 mg/dl. Because PbB was the dependent variable, the record for an individual was omitted from the file if the PbB datum was missing. When other NHANES variables were missing for an individual, statistical imputation was performed with an inhouse program developed using Matlab R2009a (The Mathworks, Natick, MA, USA). The imputation scheme involved estimation of the data’s statistical distribution and then random assignment of a value within the distribution.14 Statistical distributions of non-missing data were used to estimate the missing data as follows. For categorical variables, cumulative frequency distributions were developed for each categorical variable. A random seed between 0 and 1 was then created, and a categorical assignment was made depending on where the random seed fell within the cumulative frequency distribution. For continuous variables, a missing value imputed as the sum of the mean and standard deviation (SD) multiplied by a random seed between 1 and 1 as applied in the appropriate statistical distribution. Data for many of the variables were lognormally distributed, so that log-transformation was applied to those data. Sixteen of the NHANES covariate data had missing values that were imputed, and visual inspection suggests that the covariate data were missing at random before imputation. Percent imputed data by variable are provided in Supplementary Table S2. AQS data for PM10-PbA were merged to the NHANES participants by geocoded location, based on PM10-PbA monitors’ latitude, longitude, and sampling dates. PM10-PbA data were used, because they were found to have the strongest associations with PbB compared with PbA measured in the PM2.5 and total suspended particulate (TSP) size fractions.15 Data from 155 PM10-PbA monitors sampling for 24 h on 1 in every 3 to 6 days and averaged over 365 days produced 16,222 observations for this analysis. Missing PbA observations were not included in the analyses, although a 75% completeness criterion was observed. However, below-MDL data were not omitted to minimize positive bias on the PbA data distribution; 5.8% of data were below the MDL. AQS data were assigned to an NHANES participant if a PM10-PbA monitor was located within 4 km of a Census block centroid in which an NHANES participant resided to better represent personal exposure to Pb through air-related pathways, and the 24-h PM10PbA concentration measurement fell within 7 days of the NHANES participant examination date. The sample population after the data merge was 21,143 (with the breakdown by age provided in the Supplementary Table S2); in other words, 21,143 participants in the NHANES data set fell within the 4 km buffer around an PbA monitor and hence had data assigned to them from the PbA measurements. In addition, GIS data for total length of street per Census block group normalized by Census block
12–19 Years 4942 Med IQR 0.0060 0.0066 1.0 0.80 % 23 29 37 34 51 49 30 8 28 34
METHODS Data Sets
& 2015 Nature America, Inc.
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
413 group area for the year 2000 were also merged to the NHANES data set based on Census block group identification number. The street per Census block group variable was designed to account for living in a heavily trafficked environment and was adjusted for in the model in the same manner as was the other NHANES covariates.
Statistical Analysis The statistical modeling approach was described in Richmond-Bryant et al.1 and is summarized here. Multi-level LME models adjusted for the variables (e.g., age, race/ethnicity, and housing age) are listed in Supplementary Table S3. The LME analyses were implemented in SAS (SAS Institute, Cary, NC, USA) using PROC MIXED via the National Center for Health Statistics Research Data Center’s remote access system. Multi-level modeling was performed at the individual and Census block group levels, and random effects at the Census block group level were random intercepts for each Census block group. The subject age stratum was treated as a stratification variable except when age was used as an effect measure modifier. The geographical location was treated as a random effect. The LME model without effect measure modification is of the form: lnðPbBi;j Þ ¼ b0 þ bi Zi þ bj þ bPbA lnðPbAj Þ þ ei;j
ð1Þ
where PbBi,j is the PbB for the ith individual living in the jth Census block group, Zi is the vector of covariates for the ith individual, bi is the associated coefficient vector, PbAj is the average PbA concentration obtained at the jth Census block group over a 365-day averaging period, and bPbA is the effect estimate of ln(PbA) on ln(PbB). The log-transform was applied after inspection of quantile–quantile distribution plots that indicated lack of normality for the untransformed data and normality once the log-transform was applied. The covariate and PbA concentration variables are fixed effects. b0 is the overall intercept, bj is a Census block group-level random normal intercept with mean zero and variance t2, and ei,j is an individual-level random normal variable with mean zero and variance s2. The LME models were run a second time with an PbA interaction term for age group, sex, housing age, or race/ethnicity to determine whether those factors acted as effect measure modifiers. When effect measure modification was factored into the model, the LME is of the form: lnðPbBi;j Þ ¼ b0 þ bi Zi þ bj þ bint;j Z lnðPbAj Þ þ ei;j
ð2Þ
where Z* is the variable (i.e., age group, sex, housing age, or race/ethnicity) used to test effect measure modification on the PbA term, and bint,j is the effect estimate for the interaction between Z* and ln(PbAj). Only one effect measure modification variable was included for each model run. Age group was used as a stratification variable for all models except for the case where age group was employed as an interaction term in the all-ages model. Likelihood ratio testing was performed to test the statistical significance of the interaction term by comparing model (1) with model (2). NHANES survey weights were not used in the analysis of this data set, as discussed in Richmond-Bryant et al.,1 based on analyses by Gelman16 and Rabe-Hesketh and Skrondal,17 and blood concentration values below the limit of detection (LOD) were substituted with a value of the LOD*2 0.5. Statistical imputation methods were used to substitute for missing values of the NHANES demographic and socioeconomic variables.1 NHANES records were discarded if they did not contain PbB data. For the air quality
Table 2.
RESULTS AND DISCUSSION The likelihood ratio testing results demonstrated which interaction terms in model (2) were statistically significant compared with model (1) (see Table 2). Statistical significance is tested at the 5% level. Age group was statistically significant as an effect measure modifier in the all-ages model (Po0.001). Housing age was a statistically significant effect measure modifier for 6–11-year (P ¼ 0.01), 12–19-year (P ¼ 0.024), 20–59-year (P ¼ 0.017), and the all-ages models (P ¼ 0.001). Race/ethnicity was statistically significant in the 12–19-year group (P ¼ 0.017) and in the all-ages model (P ¼ 0.004), and it was marginally significant in the 20–59-year age group (P ¼ 0.060). Although using sex as an effect measure modifier never produced a slope that was statistically significantly different from model (1), being male produced an increased slope for the allages model and for all age groups except 1–5-year-old children. The effects of age group, housing age, race, and sex on PbB–PbA associations are illustrated in Figure 1, and model parameters are provided in Supplementary Table S3. Age group was only tested with the all-age model, and the effect measure modification model demonstrated statistically significant increases in effect estimate for the 1–5-year and 12–19-year age groups and an observed but not statistically significant increase for the 6–11-year age group (see Figure 1 and Supplementary Table S3). The effect decreased for the 20–59-year age group. Bierkens et al.2 modeled the PbB–PbA relationship among children o6 years old and adult men and women at least 18 years old using a ln-ln model stratified by age (and sex among adults). They observed large contrasts between the effect estimate for children (b ¼ 0.09) compared with adult men (b ¼ 0.44) or women (b ¼ 0.34), corresponding to only a 16% drop in PbB per unit change in PbA over the period 1990–2007 for young children compared with 57% and 48% in men and women, respectively. The results of Bierkens et al.2 contrast those found here, where effect measure modification caused higher effect estimates among children compared with adults. Time–activity studies provide support for this study’s finding of effect measure modification among children. Wu et al.18 differentiated time– activity patterns among children (mostly o8 years old) from their parents (mostly under age 55 years) and older adults (mostly older than age 54 years) and illustrated that children were more likely than either adult group to partake in vigorous outdoor and indoor activities. Wu et al.18 also found that the pattern of time spent in different microenvironments varied among the age groups. Similarly, Klepeis et al.19 surveyed individuals in Canada and the United States about time–activities and found that children
Likelihood ratio testing of effect measure modification.
Age group
Effect measure modification category Age group
1: 1–5 2: 6–11 3: 12–19 4: 20–59 5: 60 þ All ages
variables, observations with missing and zero values of ambient Pb were not used in the analyses because of the logarithmic transformation that might introduce biases into the analyses.
Sex
Housing age
Race/ethnicity
Likelihood ratio
P-value
Likelihood ratio
P-value
Likelihood ratio
P-value
Likelihood ratio
P-value
— — — — — 38.3
— — — — — o0.001
2.5 0.2 0.9 0.3 0.9 1.4
0.114 0.655 0.343 0.584 0.343 0.237
1.2 9.2 7.5 8.2 3.5 14.3
0.549 0.010 0.024 0.017 0.174 0.001
3.1 5.3 10.2 7.4 3.7 13.4
0.376 0.151 0.017 0.060 0.296 0.004
Results are in bold where there was a statistically significant difference (a ¼ 0.05) between models with and without effect measure modification.
& 2015 Nature America, Inc.
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
414
Figure 1. Estimated deviation from b for the all-ages model and the age stratification model, and for models including interactions for age, sex, housing age, and race/ethnicity (n ¼ 21,143). b Has been shifted so that the term without interaction is zero, for comparison of the relative changes. The error bars represent 95% confidence intervals.
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
& 2015 Nature America, Inc.
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
415 o12 years old spent more time outdoors compared with individuals 412 years. Such differences in activities could result in differences among both exposure and intake of inhaled and ingested ambient PbA. Moreover, age-related ingestion via handto-mouth behaviors and toxicokinetic characteristics could also influence the PbB–PbA relationship.20 For example, studies indicate that 40–50% of ingested water-soluble Pb is absorbed by children o8 years old compared with 3–10% absorption in adult subjects.21–23 There was a trend for the effect estimates to be higher for males than for females in the all-age model and for every age strata except for 1–5-year group, for which the converse was observed (see Figure 1 and Supplementary Table S3). Bierkens et al.2 modeled the PbB–PbA relationship for adult men and women and found that model slope was 0.34 among women and 0.44 among men in a ln-ln model, resulting in declines in PbB with respect to unit declines in PbA of 48% in women and 57% in men from 1990 to 2007. Studies of time spent in physical activity and outdoors suggest that males spend more time performing vigorous activities and exercising outdoors compared with females.24–26 However, sex-based differences were not always statistically significantly different.24 These findings are consistent with our observation of a trend for higher effect estimates in males than females. Older housing (i.e., that built before 1950) elicited a statistically significant increase in effect estimate among 1–5-year-old children (Po0.0001) (see Figure 1 and Supplementary Table S3). However, older housing produced a statistically significant reduction in the effect estimate for 12–19 year olds (P ¼ 0.019) and produced reductions that were not statistically significant in the all-ages model and for the 6–19-year, 20–59-year, and Z60-year age groups. Moderate housing age (i.e., built between 1950 and 1977) and newer housing (i.e., built after 1977) typically produced a null effect or an increase in effect estimate in the all-ages model and among all age groups except 1–5-year-old children. The lack of statistical significance of the housing interaction term suggests that a model with no PbA effects in older housing but possible PbA effects in newer housing is not rejected. Hence, older housing does not appear to influence the relationship between PbB and PbA for people living in older housing, with the exception of young children and teens. It is possible that discrepancies in missingness of housing age among race/ethnicity groups contribute to uncertainties about the influence of housing age on the PbB–PbA relationship. For example, 40% of all MexicanAmerican and 39% of all non-Hispanic black participants did not report housing age, compared with 9% of non-Hispanic white participants of all ages. Among 1–5-year-old study participants, 47% of Mexican-American and 47% of non-Hispanic black participants’ heads of household did not report housing age, compared with 25% of non-Hispanic white participants’ heads of household. Imputation was performed using the data distribution of the entire study population combined, and hence it is possible that there is misspecification in the housing age variable. Such misspecification could also influence the role of housing age as an effect measure modifier of the PbB–PbA relationship. Home ventilation may explain how housing age might act as an effect measure modifier of the relationship between PbB and PbA. Jacobs et al.11 studied data from the American Housing Survey from 1970 to 2000 and observed that the use of central air conditioning increased with newer housing, although they noted a trend in addition of central air to older homes during the 1990s. Because central air conditioning systems employ filters, residents lacking central air would more likely be exposed to PbA in particulate matter infiltrating indoors given that the air would be unfiltered and likely introduced into the home through open windows. The only group where effect measure modification related to older housing increased the effect estimate was 1–5year-old children. Young children are more likely than older children or adults to spend time indoors and to have greater & 2015 Nature America, Inc.
hand-to-mouth exposure to indoor dust that may contain settled PbA particles.20,27 Therefore, ventilation differences related to housing age might only have been observable among the youngest cohort. Effect measure modification by race/ethnicity was fairly consistent for the all-age model and across the age strata (see Figure 1 and Supplementary Table S3). Mexican-American ethnicity had the largest increase in effect estimate for the allages model and for all age strata except for adults Z60 years and 6–11-year-old children, for which it still produced an increase. The effect estimate was roughly the same for those of black race as the model that did not test effect measure modification with the exception of elevated effect estimate for 6–11-year-old children and reduced effect estimate for adults 20–59 years old. The effect estimate was consistently reduced for the all-ages model and across all age strata for being of white race. Results were mixed across age groups for being of other Hispanic or other race. Several studies have examined the racial characteristics of populations living in the vicinity of industrial facilities, often using distance or presence of facilities required to report their emissions to the Toxic Release Inventory (TRI) because they emit a Hazardous Air Pollutant that may but not necessarily will include Pb.28–31 Chakraborty and Zandbergen28 observed that both Hispanic and African-American residents were more likely to live within a 4-mi buffer of a TRI industrial facility compared with white residents, and Wilson et al.31 observed statistically significant increases in the presence of a TRI facility in a Census block for nonWhite residents. Wilson et al.31 found that the odds ratio of living in a Census block containing a TRI facility increased for black residents but was not statistically significant.31 Perlin et al.30 did find that people of African American race were generally more likely to live in close proximity to TRI facilities compared with people of white race regardless of age, and Mohai et al.29 estimated statistically significant increases in the odds ratio for living within a mile of a TRI facility for percent of black race but not for percent of white race. Our effect measure modification results suggest that the PbB–PbA relationship is elevated for MexicanAmerican ethnicity, reduced for white race, and neither elevated nor reduced for black race. The results for black race were surprising given the literature documenting risk of living near a TRI for those of black race. Limitations of the study influence interpretation of the study results. The National Center for Health Statistics Research Data Center’s data restrictions, intended to protect NHANES participant confidentiality, included restrictions on visualizing the data and testing the residuals. These restrictions limited the ability to assess the spatial distribution of missing data among the NHANES participants. Averaging the data over 1 year and requiring 75% completeness of the data reduced the impact of missing data, although imputation to replace missing data related to monitor operationality would potentially introduce bias if the location of missing data was non-random.32,33 Missing PbB data in NHANES were likely to have been randomly distributed, and hence their omission would be less likely to bias the model slope.34,35 Although this study presented an analysis for a population spanning the United States, the omission of sampling weights prohibits national inference from the data. Sampling weights were not used in this analysis, because inclusion of sampling weights in multilevel models can lead to biased estimates for small cluster sizes17 and the process of combining sampling weights over several cycles and subpopulations can lead to inclusion of weights that are not inverse probabilities.16 Given that application of the 4 km buffer resulted in curtailing the sample population from 41,961 to 21,143, it is likely that the weights would not have been representative. Application of the buffers likely also resulted in reduced power. Moreover, application of the 4 km buffer may have added uncertainty to the exposure estimate. The coarse fraction of PM10-PbA is more likely to settle from the atmosphere
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
Effect modification of blood Pb–air Pb model Richmond-Bryant et al
416 at closer distances to PbA sources compared with the fine fraction of PM10-PbA, so that the mass of PM10-PbA may have been underestimated if the NHANES participant’s location was further from the source than that of the monitor or vice versa. The primary finding of this cross-sectional epidemiologic study is that the effect estimate for the PbB–PbA relationship is modified by age strata and race/ethnicity in the all-age models and for some age strata. Specifically, the effect of PbA concentrations on PbB levels was enhanced by childhood (1–5, 6–11, and 12–19-year age brackets) and by being of Mexican-American ethnicity. Living in older housing (built before 1950) detracted from the effect of PbA on PbB except in the 1–5-year group. Our effect measure modification results are generally consistent with studies in the literature of time–activity patterns, ventilation, and toxicokinetics. However, this study is the first to investigate effect measure modification of the PbB–PbA relationship, and hence data from other studies to substantiate these findings are unavailable. CONFLICT OF INTEREST The authors declare no conflict of interest.
ACKNOWLEDGEMENTS We give special thanks to Dr. Tom Long and Dr. Zach Pekar for their helpful comments in review of this manuscript. We thank Ms. Nataliya Kravets of the National Center for Health Statistics, Centers for Disease Control, for linking the NHANES data with the AQS and GIS data. Authors from ICF International were funded by contract EP-C-09-009. Authors from academia and the US EPA did not receive external funding for this research.
DISCLAIMER The research and this manuscript have been reviewed in accordance with US Environmental Protection Agency policy and approved for publication. Mention of trade names or commercial products does not constitute endorsement or recommendation for use. The views expressed in this article are those of the authors and do not necessarily reflect the views or policies of the US Environmental Protection Agency.
REFERENCES 1 Richmond-Bryant J, Meng Q, Davis JA, Cohen J, Svendsgaard D, Brown JS et al. A multi-level model of blood lead as a function of air lead. Sci Total Environ 2013; 461: 207–213. 2 Bierkens J, Smolders R, Van Holderbeke M, Cornelis C. Predicting blood lead levels from current and past environmental data in Europe. Sci Total Environ 2011; 409: 5101–5110. 3 Brunekreef B. The relationship between air lead and blood lead in children - a critical-review. Sci Total Environ. 1984; 38: 79–123. 4 Hayes EB, McElvaine MD, Orbach HG, Fernandez AM, Lyne S, Matte TD. Long-term trends in blood lead levels among Children in chicago - relationship to air lead levels. Pediatrics 1994; 93: 195–200. 5 Ranft U, Delschen T, Machtolf M, Sugiri D, Wilhelm M. Lead concentration in the mood of children and its association with lead in soil and ambient air - Trends between 1983 and 2000 in Duisburg. J Toxicol Env Health Part A 2008; 71: 710–715. 6 Richmond-Bryant J, Meng QY, Davis AJ, Cohen J, Svendsgaard D, Kravets N et al. The influence of declining air lead levels on blood lead–air lead slope factors in children. Environ Health Perspect 2014, (e-pub ahead of print, doi:10.1289/ ehp.1307072). 7 Schwartz J, Pitcher H. The relationship between gasoline lead and blood lead in the United States. J Official Stat 1989; 5: 421–431. 8 Tripathi RM, Raghunath R, Kumar AV, Sastry VN, Sadasivan S. Atmospheric and children’s blood lead as indicators of vehicular traffic and other emission sources in Mumbai, India. Sci Total Environ 2001; 267: 101–108. 9 Miranda ML, Edwards SE, Swamy GK, Paul CJ, Neelon B. Blood lead levels among pregnant women: historical versus contemporaneous exposures. Int J Environ Res Public Health 2010; 7: 1508–1519.
10 Jones RL, Homa DM, Meyer PA, Brody DJ, Caldwell KL, Pirkle JL et al. Trends in blood lead levels and blood lead testing among US children aged 1 to 5 years, 1988-2004. Pediatrics 2009; 123: e376–e385. 11 Jacobs DE, Wilson J, Dixon SL, Smith J, Evens A. The relationship of housing and population health: a 30-year retrospective analysis. Environ Health Perspect 2009; 117: 597–604. 12 Egeghy PP, Quackenboss JJ, Sandra CC, Ryan PB. Determinants of temporal variability in NHEXAS-Maryland environmental concentrations, exposures, and biomarkers. J Exp Anal Environ Epidemiol 2005; 15: 388–397. 13 Strand M, Hopke PK, Zhao W, Vedal S, Gelfand E, Rabinovitch N. A study of health effect estimates using competing methods to model personal exposures to ambient PM2.5. J Exp Sci Environ Epidemiol 2007; 17: 549–558. 14 Donders ART, van der Heijden GJMG, Stijnen T, Moons KGM. Review: a gentle introduction to imputation of missing values. J Clin Epidemiol 2006; 59: 1087–1091. 15 Meng Q, Richmond-Bryant J, Davis JA, Cohen J, Svendsgaard D, Brown JS et al. Contribution of particle-size-fractionated airborne lead to blood lead during the National Health and Nutrition Examination Survey, 1999-2008. Environ Sci Technol 2014; 48: 1263–1270. 16 Gelman A. Struggles with survey weighting and regression modeling. Stat Sci 2007; 22: 153–164. 17 Rabe-Hesketh S, Skrondal A. Multilevel modelling of complex survey data. J R Stat Soc Ser A Stat Soc 2006; 169: 805–827. 18 Wu XM, Fan ZT, Ohman-Strickland P. Time-location patterns of a population living in an air pollution hotspot. J Environ Public Health 2010; 2010: 625461. 19 Klepeis NE, Nelson WC, Ott WR, Robinson JP, Tsang AM, Switzer P et al. The National Human Activity Pattern Survey (NHAPS): a resource for assessing exposure to environmental pollutants. J Exp Anal Environ Epidemiol 2001; 11: 231– 252. 20 USEPA. Integrated Science Assessment for Lead (Final Report). In: Agency USEP (ed). Washington, DC, 2013. 21 James HM, Hilburn ME, Blair JA. Effects of meals and meal times on uptake of lead from the gastrointestinal-tract in humans. Hum Toxicol 1985; 4: 401–407. 22 Maddaloni M, Lolacono N, Manton W, Blum C, Drexler J, Graziano J. Bioavailability of soilborne lead in adults, by stable isotope dilution. Environ Health Perspect 1998; 106: 1589–1594. 23 Watson WS, Morrison J, Bethel MIF, Baldwin NM, Lyon DTB, Dobson H et al. Food iron and lead absorption in humans. Am J Clin Nutr 1986; 44: 248–256. 24 Brownson RC, Baker EA, Housemann RA, Brennan LK, Bacak SJ. Environmental and policy determinants of physical activity in the United States. Am J Public Health 2001; 91: 1995–2003. 25 Corder K, Ekelund U, Steele RM, Wareham NJ, Brage S. Assessment of physical activity in youth. J Appl Physiol 2008; 105: 977–987. 26 McCurdy T, Graham SE. Using human activity data in exposure models: analysis of discriminating factors. J Exp Anal Environ Epidemiol 2003; 13: 294–317. 27 USEPA. Exposure Factors Handbook 2011 Edition (Final). In: Agency USEP (ed). Washington, DC, 2011. 28 Chakraborty J, Zandbergen PA. Children at risk: measuring racial/ethnic disparities in potential exposure to air pollution at school and home. J Epidemiol Commun Health 2007; 61: 1074–1079. 29 Mohai P, Lantz PM, Morenoff J, House JS, Mero RP. Racial and socioeconomic disparities in residential proximity to polluting industrial facilities: evidence from the Americans’ Changing Lives Study. Am J Public Health 2009; 99: S649–S656. 30 Perlin SA, Wong D, Sexton K. Residential proximity to industrial sources of air pollution: Interrelationships among race, poverty, and age. J Air Waste Manage Assoc 2001; 51: 406–421. 31 Wilson SM, Fraser-Rahim H, Williams E, Zhang H, Rice L, Svendsen E et al. Assessment of the distribution of toxic release inventory facilities in metropolitan Charleston: an Environmental Justice Case Study. Am J Public Health. 2012; 102: 1974–1980. 32 Helsel DR. Fabricating data: how substituting values for nondetects can ruin results, and what can be done about it. Chemosphere 2006; 65: 2434–2439. 33 Katsouyanni K, Schwartz J, Spix C, Touloumi G, Zmirou D, Zanobetti A et al. Short term effects of air pollution on health: A European approach using epidemiologic time series data: The APHEA protocol. J Epidemiol Commun Health 1996; 50: S12–S18. 34 Rubin D. Inference and missing data. Biometrika 1976; 63: 581–592. 35 Sterne JAC, White IR, Carlin JB, Spratt M, Royston P, Kenward MG et al. Multiple imputation for missing data in epidemiological and clinical research: potential and pitfalls. Br Med J 2009; 338: 2393.
Supplementary Information accompanies the paper on the Journal of Exposure Science and Environmental Epidemiology website (http:// www.nature.com/jes)
Journal of Exposure Science and Environmental Epidemiology (2015), 411 – 416
& 2015 Nature America, Inc.
