Geographic information systems (GIS)–based lead poisoning ecologic risk models continue to be useful to public health departments because they are inexpensive and easily translate to location-based intervention strategies such as targeted screening and mitigation.1 However, changes to the US Centers for Disease Control and Prevention (CDC) reference level over the years, combined with diverse modeling strategies employed by different researchers, complicate the comparison of GIS-based risk models conducted in different time periods and regions of the United States.1 Previous lead poisoning ecologic risk models2–15 vary widely with respect to the type of outcome, covariates or exposures they evaluate, as well as the region and geographic scale of analysis. Furthermore, evidence regarding which exposure pathways are most relevant to explaining spatial variation in childhood lead poisoning risk is inconclusive.1 In the majority of studies, the assumed primary source of lead exposure is lead-based interior paint, but a growing body of research points to soil lead contamination as a significant contributor to childhood lead poisoning.16–21 Taken together, these observations suggest a need for states to conduct timely, place-specific spatial surveillance of childhood lead poisoning. Building on collaborative efforts of the National Environmental Public Health Tracking (NEPHT) program to improve childhood lead poisoning data quality, we propose a spatial analysis model suitable for use by any public health program managing blood lead level (BLL) data. This work furthers the objectives of the NEPHT program to enhance surveillance by identifying areas and populations most likely to be affected by environmental lead contamination and assessing the relative importance of spatially varying risk factors.22
Colorado is a targeted screening state with respect to childhood lead poisoning, and all BLL tests of children (18 years or younger) are reportable to the state health department. Current state guidelines recommend testing for lead at ages 12 and 24 months for all low-income children in Colorado, defined as children who are eligible for Medicaid, Children's Health Plan Plus (CHP+), or the Colorado Indigent Care Program. Low-income children between the ages of 3 and 6 years are also recommended for screening if they have not been previously tested. In addition, children younger than 6 years are recommended for screening if they reside in one of 4 low-income Denver zip codes, are refugee children, or regularly visit pre-1978 housing in poor condition. In 2016, the Colorado Department of Public Health & Environment (CDPHE) began confirming cases of elevated blood lead levels (EBLL) of 5 μg/dL or more using an algorithm developed by the NEPHT program. Between 2010 and 2014, Colorado screened an average of 20 000 unique children younger than 6 years per year, with a peak in total children tested occurring in 2012, the year the CDC lowered the reference BLL of concern to 5 μg/dL. During each year, the Denver area comprised 67% of total statewide testing for lead. During this same period, the number of confirmed EBLL cases remained consistent, with an average of 177 and 90 per year identified statewide and in the Denver area, respectively.
The targeted nature of Colorado's screening program makes identifying a denominator population difficult and complicates the calculation of incidence rates for routine epidemiologic surveillance. Spatial analysis methods that involve aggregating cases to geographic units are not well-suited to studying spatial patterns of risk in a targeted screening scenario because they assume a standard/accurate denominator that correctly identifies the population at risk. An alternative approach is to adopt a spatial case-control study design, in which the set of cases is treated as a realization of an inhomogeneous Poisson point process, whose spatial intensity can be evaluated against a set of controls representing the population at risk.23,24
There are many examples in the spatial epidemiology literature applying a spatial case-control study design to spatial point patterns of disease cases and appropriate controls geocoded to residential address locations.23,25–30 From these data, a spatial odds ratio or relative risk function can be calculated, which provides a smooth map describing the spatial distribution of the relative frequency of cases compared with the underlying population at risk.31 The central hypothesis is whether the spatial intensity of cases tends to vary with the density of the population at risk or whether cases exhibit spatial clustering in a manner distinct from a simple map of the population at risk. The latter is suggestive of spatial clustering due to some cluster-inducing factor such as an exposure source.32 A number of different statistical approaches have been proposed to analyze spatial case-control data. Several authors have considered an estimator of spatial disease risk given by the ratio between the spatial intensity of cases and controls using kernel density estimation techniques,23–27,31 However, this method requires stratifying the data and conducting separate analyses to adjust for covariates.32 A related method, employing generalized additive models (GAMs) with bivariate smoothing functions of geographic coordinates, has increasingly been used to identify areas of increased or decreased risk for a binary response variable while adjusting for covariates, including spatial confounders.29,32–41 We analyzed a spatial case-control point pattern data set of incident childhood lead poisoning cases using GAMs, where cases represented the residential locations of all children with confirmed EBLL of 5 μg/dL or more between calendar years 2010 and 2014. Controls were a sample of residential locations of the population at risk drawn from the population of individuals screened for lead in Colorado during the same period. We analyzed the spatial patterns of childhood lead exposure among children younger than 6 years and identified areas of significantly increased risk in the Denver metro area. In addition to modeling the effect of spatial location and adjusting for potential spatial confounding factors, we modeled the effect of additional covariates previously identified in the childhood lead poisoning literature to understand the primary risk factors associated with the observed spatial variation in childhood lead poisoning rates.
The study area (Figure 1) included the entire contiguous urbanized area surrounding the city and county of Denver, Colorado, including parts of Adams, Arapahoe, Broomfield, Douglas, Jefferson, and Weld counties. Figure 1 shows the study area along with several features of interest, including Denver's major roads, at-risk zip codes, lead-emitting facilities in operation during the study period registered with the US Environmental Protection Agency's Toxic Release Inventory (TRI), and 3 superfund sites previously associated with smelting activity.
Lead testing data were extracted from the Colorado Childhood Lead Poisoning Prevention program's lead surveillance database and confirmed using an algorithm conceptualized and adopted by the NEPHT program in 2016. Briefly, children who are tested for blood lead can sometimes have multiple test results reported in a single year, making it necessary to select a single test result to represent the child's BLL for a given year. The NEPHT algorithm performs this selection, either by selecting a child's confirmed BLL for a given year or by selecting the test result with the maximum BLL reported during the year. The NEPHT algorithm considers an EBLL to be confirmed if (1) a child had a venous type test result of 5 μg/dL or more, or (2) a child had a confirmatory pair, defined as a capillary (or unknown) type test result of 5 μg/dL or more, followed by a subsequent capillary (or unknown) type test result of 5 μg/dL or more between 1 and 84 days of the initial elevated test result. When a confirmatory pair is identified, the most recent elevated test result is retained. If a child has more than 1 confirmatory pair during a given year, the algorithm selects the maximum confirmatory test result. Nonelevated venous test results are also considered confirmed, whereas the remaining capillary (or unknown)-type test results that do not meet the criteria for confirmation are narrowed by selecting the maximum test result per child during a given year. The final test result for each tested individual represents either their confirmed BLL test result or the maximum test result during the year. All test results were geocoded to residential location using MapMarker Desktop software. Approximately 11% of records did not have sufficient address information to assign county of residence. In addition, approximately 0.7% of records were omitted because they did not contain information to calculate age such as date of birth or test date. However, of the remaining records that could be assigned to a county in the study area, approximately 93% were geocoded to the residential address level.
We analyzed all confirmed incident EBLL cases recorded in the study area between 2010 and 2014 for children younger than 6 years. A more consequential choice was our selection of controls. Because of the complexity of the target population in Colorado, it is difficult to obtain an appropriate denominator population from publicly available data sources. As an alternative, we estimated the spatial distribution of the population at risk using residential locations of individuals tested for lead. Control data were randomly selected from the set of 30 475 unique children younger than 6 years tested for lead between 2010 and 2014 within the study area whose test results either showed a BLL of 0 μg/dL or were below detectible levels. Before selection, control children with multiple tests across years were limited to their most recent test result, and their residential address locations were ordered by geographic coordinates to ensure spatial representativeness of the sample. Controls were selected to achieve an approximate 3:1 control to case ratio within in each year during the 5-year study period.
A key assumption was that individual-level socioeconomic characteristics for controls were consistent with case populations because they have been drawn into Colorado's targeted screening program. We sought evidence of this assumption by matching our lead surveillance database against Colorado's Medicaid and CHP+ enrollment records for children younger than 6 years obtained for the same time period. We found 65% of all children screened for lead poisoning statewide, 69% of our cases, and 65% of our controls were enrolled in Medicaid or CHP+ in Colorado for at least 1 month between 2010 and 2014. Next, since Colorado is a targeted screening state, it is important to show that during the study time period, our lead screening program was relatively spatially uniform and consistent with estimates of population density of potentially at-risk children younger than 6 years throughout the study area. We used kernel density estimation to create a smoothed map of the density per square mile of the number of unique children younger than 6 years who were tested during the study period. We found the highest quartile of testing density covered the majority of the study area (see Supplemental Digital Content Figure, available at http://links.lww.com/JPHMP/A373). Furthermore, Pearson correlation coefficients between US Census tract–level population per square mile of children younger than 6 years living below the federal poverty line (FPL) and the number unique children tested and the number of control children per square mile showed strong positive correlations (r = 0.87 and r = 0.76, respectively).
We identified 38 distinct covariates, including 7 measured at the individual level, 2 measured at the parcel level, and 29 measured at the census tract level, used in previous lead poisoning ecologic risk models. To create consistent spatial support for all covariate data sets, we spatially smoothed parcel and census tract–level covariates to 0.25-mile square grid cell locations, which served as the basic unit of analysis for the predicted surfaces of the GAMs. Following best practices for integrating area and point-based data in geographic analysis,42–45 we created smoothed grid cell estimates of census tract–level rates using binomial area-to-point kriging and used point-to-point kriging to create smoothed grid cell estimates of average nearby home value and housing age from parcel data. Several GIS-based ecologic risk studies have employed kriging methods to smooth area and point-based estimates to a common spatial support.5,11,12 All kriging calculations were conducted in the ArcGIS 10.2 Geostatistical Analyst package.
Generalized additive models extend the generalized linear modeling framework by adding a component function of 2 or more dimensions that allows nonlinear relationships between the dependent and independent variables to be included in the regression equation. The bivariate spatial smoothing function enables model-based estimates of disease risk at unobserved spatial coordinates in the study area. Regardless of the choice of smoothing function, the general formula for a generalized additive logistic regression model for case-control data is defined as follows:
where yi = 1 for cases, yi = 0 for controls; β is a k × 1 vector of linear regression coefficients representing the effects of a n × k matrix of covariates, zi; and f(x1i, x2i) is a bivariate smoothing function of spatial location (x1i, x2i). The model becomes a crude spatial model when covariates are not specified, a generalized linear model when no smoothing function is specified, and a null logit model when both covariates and spatial smoothing function are omitted.38 The locally weighted scatterplot smoothing (LOESS) function has been used extensively in GAMs to estimate spatial variation in disease risk.32–41 In LOESS, the outcome variable yi is regressed on a function of the data values of locations within a neighborhood of xi, where the size of the neighborhood is controlled by a span parameter representing the proportion of the data points in the total data set used to estimate yi. The LOESS function weights data points in the neighborhood according to their proximity to xi, where nearer data points are given a higher weight than points further away from xi. A larger span size increases the number of data points that receive nonzero weight, which leads to a greater degree of smoothing.46 An optimal span size can be selected by evaluating the GAM over a range of span sizes between 0 and 1.0, selecting the span with the lowest associated Akaike Information Criteria (AIC) value. The GAM framework quantifies the global effect of geographic location on the likelihood of disease by assessing the difference in model deviance between the null model and models that include the bivariate spatial smoothing function. Statistical significance of the difference in model deviance can be evaluated using an approximate χ2 test and Monte Carlo permutation tests.32,46,47 Monte Carlo permutation testing allows identifying local spatial clusters of increased or decreased spatial risk by mapping the contours of data points ranked above or below significance cutoffs in the pointwise permutation distribution.
In our study, we applied spatial GAMs with an LOESS spatial smoothing function to the case-control data set in 3 iterations. The first iteration corresponds to a measure of the effect of geographic location on the odds of having a confirmed EBLL of 5 μg/dL or more without adjusting for additional covariates. The second iteration adjusts the spatial odds ratio for 5 potential spatial confounding factors, including age, sex, year, season, and spatially smoothed screening rate (originally calculated at the census tract level). The final iteration attempts to fully adjust the spatial odds for a parsimonious set of covariates in an attempt to identify a consistent regression model that accounts for the observed spatial variation in EBLL risk. We assessed statistical significance of the difference in model deviance using Monte Carlo permutation tests. Following Bliss and colleagues,46 we selected cutoffs in the pointwise permutation distribution using an adjusted type I error rate of 0.004, which corresponds to a nominal type I error rate of 0.01 (99% confidence level). Locally significant spatial risk areas were identified by the spatial points with predicted log odds outside the 0.4% and 99.6% of the ranked values of the pointwise permutation distribution. All spatial GAMs were estimated using the MapGam package in the R statistical software.
The final data set used for analysis consisted of 394 incident cases and 1277 controls. Table 1 displays the variables used in the analysis stratified by case status, as well as the data source and scale at which each variable was originally measured. The median age of the cases was 2.1 years but was only 2.0 years among the controls. Males were slightly overrepresented among cases compared with controls at approximately 54% and 52%, respectively. Cases on average lived closer to major roads, lead-emitting TRI facilities, and superfund sites associated with lead contamination than the sampled controls. Cases tended to live in areas with older housing of lower estimated value compared with the sampled controls. Table 1 also includes means for each data set on selected covariates of childhood lead poisoning identified from previous ecologic risk studies. In general, differences between the cases and sampled controls occurred in the expected direction, with cases on average tending to live in areas with higher values of ecologic covariates hypothesized to be positively associated with childhood lead poisoning, and vice versa, relative to sampled controls. The greatest differences between cases and controls occurred with measures of area socioeconomic status (SES), in particular, area educational attainment, area percentage of renter-occupied housing, area rates of household poverty and below 200% of FPL, and area percentage households receiving public assistance. Furthermore, the cases tended to live in areas with greater Hispanic and fewer white non-Hispanic populations relative to sampled controls. Cases also tended to live in areas with slightly higher African American and Asian populations.
Parameter estimates from all spatial models are displayed in Table 2. Results from the crude spatial model indicate a high degree of spatial dependence in EBLL risk. Crude odds ratios ranged from 0.22 to 5.3. P values from the approximate χ2 and global permutation statistics were less than .001 in both cases. Figure 2 shows the results from mapping approximate 99% significance contours to identify local areas of increased spatial risk from 1000 Monte Carlo simulations. Results from the spatial confounder-adjusted model are displayed in the second column of Table 2. The spatial risk function was influenced by local screening rates, suggesting areas with higher screening rates are more likely to have EBLL cases. Age at test date and a seasonal indicator variable were both positively and significantly associated with spatial risk. Specifically, children closer to the age of 6 years and summer and fall months (compared with winter months) were associated with an increased spatial risk. There were no statistically significant differences in risk among males and females or between years. In general, the effect of adjusting for potential spatial confounding factors is to narrow the spatial extent of the observed spatial clustering, increasing the precision of the identified spatial cluster. Spatial confounder-adjusted odds ratios ranged from 0.22 to 2.7. Areas of increased spatial confounder-adjusted risk are observed in 2 distinct clusters and are shown in Figure 3.
Having detected persistent statistically significant clustering in the spatial confounder-adjusted model, we then specified a fully adjusted statistical model in an attempt to account for observed spatial variation. We first examined each of the covariates individually, ignoring other factors, using separate GAMs that included each individual covariate along with the bivariate spatial smoothing component. Risk was most strongly associated with indicators of neighborhood SES. Areas with lower estimated educational attainment, greater rates of adult and child poverty/low income, fewer vehicles available, lower rates of owner-occupied housing, greater rates of receiving public assistance, greater rates of vacant housing, greater unemployment, and more foreign-born residents were all statistically significantly associated with an increased risk of having an EBLL, ignoring other factors. Nearby housing value and age were strongly associated with risk in expected directions. Risk was also associated with living closer to major roads; however, no statistically significant relationships were identified for proximity to lead-related TRI facilities or superfund sites, ignoring other factors. In addition, higher concentrations of Asian, African American, and Hispanic populations were all positively and significantly associated with an increased EBLL risk, ignoring other factors, whereas higher concentrations of non-Hispanic white populations were associated with a lower risk.
We began multivariate analysis of ecologic predictors of childhood lead poisoning risk by including each of the covariates identified in backward and forward selection stepwise logistic regression models, which only retained variables if they lowered the model AIC value. We calculated variance inflation factors (VIFs) to assess the potential for multicollinearity among the ecologic-level variables to impact the regression results. Unsurprisingly, given the correlation between different measures of SES at the census tract level, the results indicated a high degree of multicollinearity when multiple measures related to neighborhood SES were included in the model (VIF ≥ 10). On the basis of these results, we used principal components analysis with an orthogonal varimax rotation to reduce the dimensionality of the full covariate data set. The first principal component explained 34% of the total variance in the data. Examining the squared cosine values and approximate correlation coefficients between each variable and the first principal component indicated that SES variables were well represented (cos2 > 0.45) and strongly correlated (r ≥ 0.70) with the first principal coordinate axis. The strongest associations were observed for variables describing the percentage of adult and child poverty and low-income status. Furthermore, the percentage of non-Hispanic white population was strongly negatively correlated with the first principal component, whereas the percentage of Hispanic population was strongly associated in the positive direction. When case status was included in the analysis as a qualitative supplementary variable, the results showed the first principal component was strongly associated with an EBLL risk (r = 0.87).
Reevaluating the stepwise regression models including the principal component in place of the variables measuring area SES and the percentage of Hispanic and non-Hispanic white populations resulted in consistent parsimonious models. The parsimonious models retained the variables for age, screening rate, SES component, housing age, percentage Asian population, percentage African American, and percentage vacant housing. In addition, the variables for sex and test year were retained in the fully adjusted model to be consistent with the spatial confounder-adjusted model. Nearby home value was also retained in the fully adjusted model because of its perceived importance in the Denver metro area. We also tested an interaction effect between the variables measuring average neighborhood housing age and SES/Hispanic ethnicity; however, the effect was not statistically significant when controlling for other factors. Results for the fully adjusted and parsimonious models are displayed in the third and fourth columns of Table 2. VIF measurements indicated no serious problems with multicollinearity with either model. Leverage calculation revealed possible outliers; however, removing outliers caused no substantive changes in the interpretation of model parameters and thus all observations were retained in the final analysis.
Results from the final models suggested that ecologic-level variables measuring age of housing, SES/Hispanic ethnicity, and percentage of Asian residents were all independently and significantly associated with an increased EBLL risk, controlling for other factors. Controlling for additional factors had a moderating effect on several variables observed to be associated with an EBLL risk in the individual variable analyses. For example, percentage African American, percentage vacant housing, and home value were no longer significantly associated with an EBLL risk at conventional levels in the fully adjusted model, although percentage African American and vacant housing were retained in the parsimonious model because they lowered the AIC value. Distance to major roads was also no longer statistically significant after adjusting for other factors, and retaining this variable in the model did not lower the AIC value. Results from the permutation test of the parsimonious model are shown in Figure 4. Adjusting the spatial odds for the full model accounted for the observed spatial variation in EBLL risk, as adjusted odds ratios in the area formerly identified as spatial clusters are now mostly less than 1.0, and there were no statistically significant spatial clusters after adjustment. The fully adjusted spatial odds ratios ranged from 0.27 to 3.1; however, the median-adjusted spatial odds ratio was 0.87, and the highest values of the spatial odds ratio are now only found at the edges of the study area, where spatial GAMs are subject to edge-related biases.
In the Denver metro area, between 2010 and 2014, childhood lead poisoning cases among children younger than 6 years exhibited a spatially clustered distribution over and above the spatial distribution of the population at risk. This distribution is suggestive of common factors giving rise to lead poisoning cases within the areas of elevated risk. Clustering persisted even after adjusting the risk function for 5 potential spatial confounding factors, including age, sex, year, season, and spatially smoothed screening rates. The spatial confounder-adjusted results revealed 2 distinct statistically significant spatial clusters of childhood lead poisoning cases.
Adjusting for additional covariates in the final GAM accounted for the observed spatial variation in risk. Consistent with previous risk studies of childhood lead poisoning,2–4,6–8,13,14 we found that ecologic indicators of low SES/Hispanic ethnicity, Asian race, and older housing age were all positively and significantly associated with an increased EBLL risk after controlling for other factors. On a bivariate level, proximity to major roads, lower nearby home values, areas with greater numbers of African American residents, and higher rates of vacant housing were all positively associated with an increased risk; however, these effects were no longer statistically significant at conventional levels when controlling for other factors. The effect of nearby housing age is of particular interest as a potential indicator of the primary exposure pathway, and our results suggested that areas with older nearby housing are strongly associated with an increased EBLL risk even after adjusting for neighborhood socioeconomic conditions. Our results indicated that Hispanic ethnicity and SES are tightly linked in the Denver metro area and both are strongly associated with an EBLL risk. The positive association between EBLL risk and greater area percentages of Asian population reflects the overlap between the eastern spatial cluster and areas of greater Asian population in portions of southeast Denver and western Adams and Arapahoe counties, and these findings help inform outreach and prevention efforts. We attempted to measure the potential for lead exposure from soil contamination with measures of proximity to major roads, lead-emitting TRI facilities, and superfund sites associated with previous lead contamination. However, none of these exposure variables were significantly associated with lead exposure after controlling for factors such as nearby housing age and neighborhood SES. Taken together, these results indicate that the spatial distribution of childhood lead poisoning in the Denver area are most strongly associated with indicators of SES and age/condition of housing, lending support for concluding that lead paint is primary exposure source of environmental lead.
The ecologic-level variables measuring age of housing and SES/Hispanic ethnicity had the strongest observed effect on EBLL risk. To further illustrate, we calculated predicted probabilities from our final regression model stratified by these 2 variables. Figure 5 displays predicted probability curves for 3 levels of the neighborhood SES/Hispanic ethnicity—mean SES, high SES, and low SES (defined as the sample mean and ± 1.5 standard deviations). Holding all other continuous variables at their mean and discrete variables at the reference levels listed in Table 2, we stratified the predicted probability across deciles of the average nearby housing age variable used in our analysis. Overall, the highest predicted probabilities of having an EBLL are found on the low SES curve, where all predicted probabilities are greater than 0.20. Using a predicted probability of 0.50 and more as a classification rule, we would then predict an EBLL for any individual living in a low SES neighborhood with an average nearby housing age of 1946 or less. Applying this classification rule would fail to predict an EBLL case for any individual living in a high SES neighborhood. However, for all 3 curves, the largest between-decile increase in predicted probability occurs between the 20th and 10th percentiles of average nearby housing age, suggesting that, regardless of local SES, risk is greatest when the average nearby year of housing construction decreases below 1946. This observation is consistent with a nationwide US survey48 of housing units that observed 68% of housing units built before 1940 were found to have a lead hazard present in the home compared with 8% for housing units built between 1960 and 1977. Taken together, these findings suggest a need to prioritize older housing units (eg, pre-1960) for lead screening in the Denver metro area in addition to the more general recommendation of pre-1978 housing in poor condition.
This study had several limitations. First, the ecologic scale of the analysis limits the causal interpretation that can be applied to the final model, which should be considered exploratory rather than explanatory. However, in practice, lead intervention and outreach efforts have often been based on ecologic-level analyses. Our use of a subset of the tested population as the control data set represents a potential limitation and requires further discussion. A key assumption we have made about our lead surveillance program is that it has successfully targeted the population at risk, in turn, making our control population representative of the population at risk. This assumption was supported by a matching analysis we performed between our study subjects and Medicaid/CHP+ enrollment data, which showed that roughly two-thirds of our test subjects were enrolled in Medicaid or CHP+ at some point during the study time period. We also noted a strong positive correlation between population density of children younger than 6 years living in poverty at the census tract level and tract-level density of unique children tested for lead and the unique children used as controls in our analysis. Taken together, this information suggests that our lead surveillance data are capable of identifying the relative frequency of cases compared with the underlying population at risk. The role of seasonality also requires further analysis, and more information is needed before drawing any definitive conclusions. The observed seasonality in our data could be suggestive of a source of lead contamination outside the home, since it corresponds to the time during the year (spring and fall months) when children are more likely to be outside or for homes to have their windows open. Thus, one possibility for future research would be to compare the spatial risk of lead poisoning with in situ measurements of soil lead concentrations. Alternatively, the observed seasonality could be artificial and potentially correspond to the times when a disproportionate number of lead tests are being conducted throughout the state. One potential source of this type of confounding could involve the scheduling of Early and Periodic Screening, Diagnostic and Treatment (EPSDT) visits among Medicaid and CHP+ enrolled children. According to EPSDT screening guidelines, states must provide or arrange for screening services at established times and on an as-needed basis, and in Colorado many of these services are offered at School Based Health Clinics. A preliminary analysis of the number of lead tests reported to CDPHE during the study period by month suggests that, on average, the number of tests performed between the months of August and October is 25% greater than other periods throughout the year. Thus, it is possible that an artificial seasonality could be introduced that reflects when these school-based health checkups are scheduled. However, we offer up these possibilities merely as suggestions, and it is possible that neither is true. Finally, since Colorado is a targeted screening state, the potential for unidentified cases exists in areas with relatively low screening rates. An assumption of this analysis is that the spatial distribution of any unidentified cases would follow that of identified cases. In future work, this assumption could be tested by estimating the true number of cases from the set of unconfirmed elevated cases that did not meet the criteria for confirmation set by the NEPHT program algorithm.
Implications for Policy & Practice
* We identified important ecologic-level factors associated with an EBLL risk and highlighted the value of conducting timely, spatially specific lead surveillance.
* Adjusting spatial clusters for screening rates could help other targeted screening states be proactive about spatial surveillance of lead poisoning in their major population centers and help correct for inherent selection bias.
* Housing and socioeconomic factors continue to be the primary ecologic risk factors associated with childhood lead exposure and can be used to predict risk at a fine spatial resolution in the Denver study area.
* Risk estimates at fine spatial resolutions are warranted and could enable more spatially specific targeting of lead abatement activity.
* Understanding the characteristics common to areas identified as having an increased risk of lead poisoning can inform risk communication efforts.
* For example, our model predicts that the greatest risk occurs in local areas with low SES and for all social strata when the average year of construction for nearby housing is 1946 or less. However, Colorado's current lead screening guidelines prioritize screening only for pre-1978 housing in poor condition.
1. Akkus C, Ozdenerol E. Exploring childhood lead exposure through GIS: a review of the recent literature. Int J Environ Res Public Health. 2014;11(6):6314–6334.
2. Bailey AJ, Sargent JD, Goodman DC, Jean F, Brown MJ. Poisoned landscapes: the epidemiology of environmental lead exposure in Massachusetts children 1990-1991. Soc Sci Med. 1994;39(6):757–766.
3. Bailey AJ, Sargent JD, Blake MK. A tale of two counties: childhood lead poisoning, industrialization, and abatement in New England*. Econ Geogr. 1998;74:96–111.
4. Griffith DA, Doyle PG, Wheeler DC, Johnson DL. A tale of two swaths: urban childhood blood-lead levels across Syracuse, New York. Ann Assoc Am Geogr. 1998;88(4):640–665.
5. Griffith DA, Marco M, Matthew V, Johnson DL, Andrew H. Impacts of positional error on spatial regression analysis: a case study of address locations in Syracuse, New York. Trans GIS. 2007;11(5):655–679.
6. Lanphear BP, Byrd RS, Auinger P, Schaffer SJ. Community characteristics associated with elevated blood lead levels in children. Pediatrics. 1998;101(2):264–271.
7. Litaker D, Kippes CM, Gallagher TE, O'Connor ME. Targeting lead screening: the Ohio Lead Risk Score. Pediatrics. 2000;106(5):E69.
8. Miranda ML, Dolinoy DC, Overstreet MA. Mapping for prevention: GIS models for directing childhood lead poisoning prevention programs. Environ Health Perspect. 2002;110(9):947–953.
9. Mielke HW, Dugas D, Mielke PW, Smith KS, Gonzales CR. Associations between soil lead and childhood blood lead in urban New Orleans and rural Lafourche Parish of Louisiana. Environ Health Perspect. 1997;105(9):950–954.
10. Moody H, Darden JT, Pigozzi BW. The racial gap in childhood blood lead levels related to socioeconomic position of residence in metropolitan Detroit. Soc Race Ethnicity. 2015;2(2):200–218.
11. Oyana TJ, Margai FM. Geographic analysis of health risks of pediatric lead exposure: a golden opportunity to promote healthy neighborhoods. Arch Environ Occup Health. 2007;62(2):93–104.
12. Oyana TJ, Margai FM. Spatial patterns and health disparities in pediatric lead exposure in Chicago: characteristics and profiles of high-risk neighborhoods. Prof Geogr. 2010;62(1):46–65.
13. Sargent JD, Brown MJ, Freeman JL, Bailey A, Goodman D, Freeman DH Jr. Childhood lead poisoning in Massachusetts communities: its association with sociodemographic and housing characteristics. Am J Public Health. 1995;85(4):528–534.
14. Sargent JD, Bailey A, Simon P, Blake M, Dalton MA. Census tract analysis of lead exposure in Rhode Island children. Environ Res. 1997;74(2):159–168.
15. Vivier PM, Hauptman M, Weitzen SH, Bell S, Quilliam DN, Logan JR. The important health impact of where a child lives: neighborhood characteristics and the burden of lead poisoning. Matern Child Health J. 2011;15(8):1195–1202.
16. Laidlaw MAS, Mielke HW, Filippelli GM, Johnson DL, Gonzales CR. Seasonality and children's blood lead levels: developing a predictive model using climatic variables and blood lead data from Indianapolis, Indiana, Syracuse, New York, and New Orleans, Louisiana (USA). Environ Health Perspect. 2005;113(6):793–800.
17. Griffith DA, Johnson DL, Hunt A. The geographic distribution of metals in urban soils: the case of Syracuse, NY. GeoJournal. 2008;74(4):275–291.
18. Mielke HW, Anderson JC, Berry KJ, Mielke PW, Chaney RL, Leech M. Lead concentrations in inner-city soils as a factor in the child lead problem. Am J Public Health. 1983;73(12):1366–1369.
19. Mielke HW, Gonzales CR, Eric P, Morten J, Mielke PW. Nonlinear association between soil lead and blood lead of children in metropolitan New Orleans, Louisiana: 2000-2005. Sci Total Environ. 2007;388(1/3):43–53.
20. Schwarz K, Pickett STA, Lathrop RG, Weathers KC, Pouyat RV, Cadenasso ML. The effects of the urban built environment on the spatial distribution of lead in residential soils. Environ Pollut. 2012;163:32–39.
21. Zahran S, Sammy Z, Mark L, Shawn M, Gabriel F, Mark T. Linking source and effect: resuspended soil lead, air lead, and children's blood lead levels in Detroit, Michigan. Environ Sci Technol. 2013;47(6):2839–2845.
22. McGeehin MA, Qualters JR, Niskar AS. National Environmental Public Health Tracking Program: bridging the information gap. Environ Health Perspect. 2004;112(14):1409–1413.
23. Bailey TC, Gatrell AC. Interactive Spatial Data Analysis. Vol 413. London, England: Longman Scientific & Technical Essex; 1995.
24. Bivand RS, Pebesma E, Gómez-Rubio V. Applied Spatial Data Analysis With R. New York, NY: Springer Science & Business Media; 2013.
25. Bithell JF. An application of density estimation to geographical epidemiology. Stat Med. 1990;9(6):691–701.
26. Davies TM, Hazelton ML. Adaptive kernel estimation of spatial relative risk. Stat Med. 2010;29(23):2423–2437.
27. Kelsall JE, Diggle PJ. Kernel estimation of relative risk. Bernoulli. 1995;1(1/2):3–16.
28. Kelsall JE, Diggle PJ. Spatial variation in risk of disease: a nonparametric binary regression approach. J R Stat Soc Ser C Appl Stat. 1998;47(4):559–573.
29. Webster T, Vieira V, Weinberg J, Aschengrau A. Method for mapping population-based case-control studies: an application using generalized additive models. Int J Health Geogr. 2006;5:26.
30. Wheeler DC. A comparison of spatial clustering and cluster detection techniques for childhood leukemia incidence in Ohio, 1996-2003. Int J Health Geogr. 2007;6:13.
31. Sarojinie Fernando WTP, Fernando WTPS, Hazelton ML. Generalizing the spatial relative risk function. Spat Spatiotemporal Epidemiol. 2014;8:1–10.
32. Siangphoe U, Wheeler DC. Evaluation of the performance of smoothing functions in generalized additive models for spatial variation in disease. Cancer Inform. 2015;14(suppl 2):107–116.
33. Bristow RE, Chang J, Ziogas A, Anton-Culver H, Vieira VM. Spatial analysis of adherence to treatment guidelines for advanced-stage ovarian cancer and the impact of race and socioeconomic status. Gynecol Oncol. 2014;134(1):60–67.
34. Bristow RE, Chang J, Ziogas A, Gillen DL, Bai L, Vieira VM. Spatial analysis of advanced-stage ovarian cancer mortality in California. Am J Obstet Gynecol. 2015;213(1):43.e1–43.e8.
35. Di Salvo F, Meneghini E, Vieira V, et al Spatial variation in mortality risk for hematological malignancies near a petrochemical refinery: a population-based case-control study. Environ Res. 2015;140:641–648.
36. Qi X, Wang Y, Li Y, et al The effects of socioeconomic and environmental factors on the incidence of dengue fever in the Pearl River Delta, China, 2013. PLoS Negl Trop Dis. 2015;9(10):e0004159.
37. Vieira VM, Hart JE, Webster TF, et al Association between residences in U.S. northern latitudes and rheumatoid arthritis: a spatial analysis of the Nurses' Health Study. Environ Health Perspect. 2010;118(7):957–961.
38. Vieira VM, Weinberg JM, Webster TF. Individual-level space-time analyses of emergency department data using generalized additive modeling. BMC Public Health. 2012;12:687.
39. Wheeler DC, De Roos AJ, Cerhan JR, et al Spatial-temporal analysis of non-Hodgkin lymphoma in the NCI-SEER NHL case-control study. Environ Health. 2011;10:63.
40. Wheeler DC, Waller LA, Cozen W, Ward MH. Spatial-temporal analysis of non-Hodgkin lymphoma risk using multiple residential locations. Spat Spatiotemporal Epidemiol. 2012;3(2):163–171.
41. You YA, Ali M, Kanungo S, et al Risk map of cholera infection for vaccine deployment: the eastern Kolkata case. PLoS One. 2013;8(8):e71173.
42. Goodchild MF, Anselin L, Deichmann U. A framework for the areal interpolation of socioeconomic data. Environ Plann A. 1993;25(3):383–397.
43. Gotway CA, Young LJ. Combining incompatible spatial data. J Am Stat Assoc. 2002;97(458):632–648.
44. Goovaerts P. Medical geography: a promising field of application for geostatistics. Math Geol. 2009;41:243–264.
45. Goovaerts P. Combining areal and point data in geostatistical interpolation: applications to soil science and medical geography. Math Geosci. 2010;42(5):535–554.
46. Bliss RL, Weinberg J, Vieira VM, Webster TF. Adjusted significance cutoffs for hypothesis tests applied with generalized additive models with bivariate smoothers. Spat Spatiotemporal Epidemiol. 2011;2(4):291–300.
47. Wood S. Generalized Additive Models: An Introduction with R. Boca Raton, FL: CRC Press; 2006.
48. Jacobs DE, Clickner RP, Zhou JY, et al The prevalence of lead-based paint hazards in U.S. housing. Environ Health Perspect. 2002;110(10):A599–A606.
childhood lead poisoning; disease surveillance; generalized additive models; spatial epidemiology