1 Background

Since the early 1970s, regulation of lead sources in the United States has dramatically lowered population blood lead levels (BLLs). In particular, two milestone achievements, the gradual elimination of lead from automotive gasoline and the ban on lead-based paint sales, were followed by a more than 93% reduction in mean BLLs between 1970 and 2017 [1, 2]. Although lead was banned from automotive gasoline, it is still permitted as an additive in aviation gasoline (avgas) used to power piston engine aircraft [3]. Leaded avgas emissions from airports have been an area of interest for regulators at both the state and federal levels. In January 2022, after a study found that children living within a half-mile (0.8 km) of the Reid-Hillview Airport (RHV) had increased levels of lead in their blood, the County of Santa Clara in California banned the sale of leaded avgas [4]. The ultimate authority to regulate lead emissions from piston engine aircraft resides with federal authorities, and Sect. 233 of the Clean Air Act prohibits states and local communities from setting their own standards [5]. In October 2023, the U.S. Environmental Protection Agency (EPA) announced a final endangerment finding that leaded avgas contributes to air pollution that “may reasonably be anticipated to endanger public health and welfare.” [6]. The final endangerment finding obligates the agency to propose new regulatory standards for leaded avgas emissions.

The effects of lead poisoning are serious, especially for young children. The most serious health effects include damage to the brain and nervous system, slowed growth and development, learning and behavioral problems, as well as problems with hearing and speech [7]. Due to its physical properties, lead tends to accumulate in the environment, making it difficult and costly to remove completely. Lead emissions from decades ago combine with more recent emissions to create a persistent risk [8]. A recent study estimates that more than half the current U.S. population was exposed to elevated lead levels during early childhood, and significant proportions of U.S. children will experience adverse lead exposures for the next several decades [9]. Meanwhile, evidence continues to emerge that even low levels of lead exposure can result in significant health effects, including reduced intellectual function in children and increased cardiovascular disease mortality in adults [10,11,12,13]. As a result, the U.S. Centers for Disease Control and Prevention (CDC) has repeatedly lowered the blood lead “reference value” to identify children with higher levels of lead in their blood compared to most children, most recently in 2022 from 5 to 3.5 µg per deciliter (µg/dL) [14]. These trends have increased public concern about sources of lead that were once viewed as less important compared to leaded automotive gasoline and lead paint.

In response to public concerns, this study evaluates the available evidence for a link between living in proximity to airports and children’s blood lead levels in Colorado. Colorado’s blood lead testing rates are low, in part because it is a targeted testing state that does not require universal testing of all children and has a younger median housing age compared to the rest of the U.S. [15]. While federal law requires testing for children enrolled in Medicaid and the Children’s Health Insurance Plan (CHP +), there are recommendations for testing children at increased risk for lead poisoning [16], but no other state requirements. Under Colorado Board of Health regulation (6 CCR 1009-7) [17], all providers and laboratories performing blood lead tests are required to report test results to the Colorado Department of Public Health & Environment (CDPHE) for individuals 18 years and under, regardless of the test result. Since the CDPHE policy effectively treats individuals aged 18 years and under consistently concerning reporting requirements and case management recommendations, we selected this age group as our main study population. There are 76 non-military and four military airports in Colorado, but the state’s childhood lead testing rates are too low to analyze all airports in the state. To ensure an adequate sample size surrounding the airports under study, this analysis examines BLL data from individuals ages 18 years and younger over a 10-year study period and focuses on 12 airports that have adequate sample sizes to obtain meaningful results. Our primary exposure variable is the continuous distance from an airport measured in miles, but we also control for the number of days a tested individual’s residence was downwind of an airport in the 60 days prior to testing. After controlling for all covariates and applying a variety of sensitivity analyses, we find that living in proximity to airports has a small but statistically significant effect on the BLLs of children in Colorado.

2 Methods

2.1 Data

We extracted all blood lead test results from the CDPHE blood lead surveillance database for Colorado residents ages 18 years and under who were tested between Jan. 1, 2011, and Dec. 31, 2020. We specifically obtained the numeric BLL test result measured in µg/dL, age, sex, residential address, collection date, the limit of detection, and the sample type reported for each test (i.e., whether the sample was obtained from a venous or capillary blood draw). For children with multiple tests, we assigned a sample order by sorting the data by child and collection date, eliminating exact duplicates, and assigning an integer value for the sample order. We also used the reported collection date to create an indicator variable for the season (spring, summer, autumn, winter) and a numeric vector specifying the year and quarter of each blood draw beginning at 1 for Q1 2011.

For spatial analysis, we obtained geographic coordinates representing the child’s residential address using MapMarker geocoding software version 3.1. Approximately 8% of tests during the study time period contained no address information. However, of the remaining data, 96.8% were geocoded to a “rooftop” address location, the highest level of accuracy. In addition, we obtained a geographic information systems (GIS) point location layer containing geographic coordinates representing the address location of all airports in Colorado and estimated monthly aircraft traffic from the Colorado Department of Transportation (CDOT) [18]. Next, we limited our study population to children residing in homes within five miles (8 km) of an airport point location. However, to ensure an adequate sample size of blood lead tests for each airport under study, we focused our main analysis on the 12 airport point locations that had at least 300 blood lead tests within three miles (4.8 km) during the 10-year study period. This resulted in a total sample size of 56,002 blood lead tests from 46,322 unique individuals aged 18 years and under. For each of these 12 airports, as well as for the 10 additional airports included in a sensitivity analysis, we verified that general aviation comprised at least 50% of all aircraft operations by consulting the website AirNav.com, a privately owned website that publishes airport and aeronautical data published by the Federal Aviation Administration (FAA). Piston engine aircraft are the most common form of engine used in general aviation [19], and no airport in Colorado sold unleaded avgas for piston engine aircraft prior to the summer of 2023 [20]. Thus, each airport in our analysis serviced piston engine aircraft that used leaded avgas during our 10-year study period. Next, we calculated the cardinal direction, or residential near angle, of each child’s geocoded address with respect to their nearest airport point location. Daily data on the predominant wind direction for each airport point location during the study period were obtained from Visual Crossing’s Timeline Weather API [21]. We used the wind data to compute the number of days the child’s residence was downwind of an airport by examining daily predominant wind direction for the 60 days prior to the collection date. The half-life of lead in the bloodstream is approximately 30 days, thus the 60-day exposure window was selected to overlap with the expected duration of detectable lead in the blood following exposure [22, 23]. If the cardinal direction of the child’s residential near angle matched the cardinal direction of the predominant wind direction at the airport point location, then that child’s residence was considered downwind for a given day.

We linked each child to their 2010 census tract of residence and obtained data on the percentage of homes built before 1960, median household income, percentage of people of color (defined as the percentage of residents who do not identify as non-Hispanic white), and population density from the US Census American Community Survey’s 2015–2019 five-year estimates. These factors have been shown in previous studies of childhood lead poisoning to be statistically associated with blood lead levels [24, 25]. In addition, we retained the residential census tract Federal Information Processing System (FIPS) code for each blood lead test for use as a group-level variable in the statistical model below. Finally, we counted the number of reported lead-releasing Toxic Release Inventory (TRI) facilities within two miles (3.2 km) of the child’s geocoded address [26].

2.2 Statistical analysis

We analyzed the association between childhood BLLs and living in proximity to airports using a correlated random effects model with bootstrapped standard errors. This approach addresses assumptions of ordinary least squares regression requiring homoscedasticity and normally distributed residuals, allows for the inclusion of both individual test-level and census tract-level predictor variables while adjusting the coefficients of interest for census tract fixed effects, and addresses the assumption of zero correlation between unobserved heterogeneity at the census tract-level and the observed explanatory variables [27,28,29,30,31,32]. Crucially, adjusting for tract fixed effects allows the model to account for additional unobserved time-invariant (at least during the study period) tract-level variables, such as climate or background soil lead concentrations, that may influence BLLs of children from the same neighborhood, and can lead to more accurate estimates of the distance and downwind effects by reducing omitted variable bias [28, 32].

We estimate the mean blood lead level Yijt with the following linear model after the specification for the correlated random effects model described in Schunck [31]

$$Y_{{ijt}} = \beta _{0} + \beta _{1} D_{{it}} + \beta _{2} W_{{it}} + \sum\limits_{{k = 3}}^{{11}} {\beta _{k} X_{{itk}} } + \sum\limits_{{k = 12}}^{{15}} {\beta _{k} X_{k} } + \sum\limits_{{k = 16}}^{{19}} {\beta _{k} \overline{{X_{k} }} } + \nu _{j} + \varepsilon _{{ijt}}$$

Here, \({Y}_{ijt}\) is the estimated BLL in μg/dL for sampled child i in tract j at time period t measured as a continuous variable and \({\beta }_{0}\) is the overall mean BLL in μg/dL when all other variables are held constant at their respective initial values. The two coefficients of greatest interest are \({\beta }_{1}\) the effect of a one-mile increase in distance \({D}_{it}\) from an airport and \({\beta }_{2}\) the effect of one additional day \({W}_{it}\) of being downwind of an airport. The coefficients \({\beta }_{3}\) through \({\beta }_{11}\) represent the effects of additional test-level covariates including age, sex (ref: female), sample type (ref: venous), sample draw order (ref: first test), whether the test result was above the limit of detection (ref: not detected), a numeric vector specifying the year and quarter the sample was taken, the season the sample was taken (ref: spring), a count of lead-releasing TRI facilities within two miles of the child’s home, and an estimate of the average monthly aircraft traffic (over the entire study period) at the child’s nearest airport. The coefficients \({\beta }_{12}\), \({\beta }_{13}\), \({\beta }_{14}\), and \({\beta }_{15}\) represent the effects of the census tract-level covariates including the percentage of homes built before 1960, the percentage of the population identifying as people of color, median household income, and population density. Next, the coefficients \({\beta }_{16}\) through \({\beta }_{19}\) represent the difference between the between-tract and within-tract effects of the four continuous variables that vary both within and between census tracts (i.e., distance, downwind days, age, and TRI count) and the predictor variables associated with these four coefficients are the census tract means of the distance, downwind days, age, and TRI count variables, respectively, denoted \(\overline{{X }_{k}}\). Inclusion of the tract-level mean terms allows coefficients \({\beta }_{16}\) through \({\beta }_{19}\) to absorb the correlation between the individual test-level distance, downwind days, age, and TRI count variables and the census tract fixed effect term, \({\nu }_{j}\) [31]. As a result, the coefficients for the primary exposure variables of interest, \({\beta }_{1}\) and \({\beta }_{2}\), as well the coefficients for age and TRI count are all adjusted for census tract fixed effects and are estimated using only the within-tract variation [32]. Finally, \({\epsilon }_{ijt}\) is the overall error term associated with the observed \({Y}_{ijt}\).

2.3 Sensitivity analyses

We also analyzed three alternative specifications of our primary regression model with continuous BLL in µg/dL as the dependent variable. Since children under 3 years of age are at especially high risk for lead poisoning [33,34,35] and make up a disproportionate share of our sample (78%), we first assessed the impact of limiting the study population to children under 3 years of age, which decreased the sample size to 43,879 blood lead tests and 37,383 unique individuals. Next, we assessed the impact of increasing the sample size and number of airports in the analysis by including airports that had at least 50 tests within 3 miles (4.8 km) during the 10-year study period. This increased the number of airports in the analysis from 12 to 22, and increased the sample size to 61,403 blood lead test and 52,085 unique individuals aged 18 and under. In addition, the correlated random effects model employed in our primary analysis is only one of several possible model specifications appropriate for estimating model parameters in the presence of clustered or correlated longitudinal responses [36]. Thus, during model development, we verified our results were consistent across alternative fixed effects study designs, and we assessed the impact of including an additional individual-level random effect parameter in our primary model.

In addition, previous studies of lead poisoning near airports have used a wide variety of dependent and exposure variable specifications. To assess the impact of varying these specifications on our overall results, we also evaluated several models in which we vary the dependent variable and distance exposure specification. We computed a series of models replacing the primary dependent variable measured in µg/dL with its natural log. We fit a series of logistic regression models by transforming our continuous dependent variable into two binary outcome variables indicating whether a child’s BLL exceeded the current and previous CDC reference levels (≥ 3.5 µg/dL and ≥ 5 µg/dL). We also assess the impact of transforming our continuous distance measure into a discrete variable in each of the above models. The resulting specification is a comparison of average blood lead levels in “medium” and “far” distance exposure groups to those of a “near” distance exposure group. However, since residential settlement patterns vary considerably from airport to airport in Colorado, it is challenging to select a uniform near-distance cutoff that is appropriate for all airport point locations (e.g., 0.5 miles). Some airports communities have residential areas in very close proximity, while others have considerably larger setback distances between the airport point location and the airport property boundary. Hence, the discrete distance cutoffs in our study were selected to reflect the interquartile range of our continuous distance variable, with those living less than 2.7 miles (4.3 km) (i.e., between 0 and 25th percentile) from the nearest airport point location serving as the near exposure group, those living between 2.7 and less than 4.3 miles (6.9 km) (i.e., between the 25th and 75th percentile) serving as the medium exposure group, and those living greater than 4.3 but less than 5 miles (i.e., between the 75th and 100th percentile) serving as the far exposure group. We also assessed whether the results of our main model were overly influenced by data from a single airport by removing all data from one airport at a time and re-estimating the primary model for each of the 12 airports in our main analysis. Lastly, we assessed the impact of varying the method for handling non-detects by replacing test results below the limit of detection with several commonly used substitutes (e.g., ½ the detection limit) and excluding non-detects altogether.

All analyses were conducted using the R Statistical Software version 4.3.1 [37]. The distance and near angle calculations were performed using the geosphere package version 1.5-18 [38]. The regression models were estimated using the lme4 package version 1.1-34, and model parameters were estimated using restricted maximum likelihood for linear mixed effects models and maximum likelihood with adaptive Gauss-Hermite quadrature for generalized linear mixed effects models [39]. Following Nakagawa and Schielzeth [40] the goodness-of-fit of the regression models is assessed by marginal and conditional R2 and computed using the performance package version 0.10.4 [41]. Predicted mean BLLs and predicted probabilities of exceeding CDC reference values were generated from the estimated marginal means of the fitted regression models using the emmeans package version 1.8.7 [42]. Due to privacy constraints, the BLL data for this study cannot be released; however, R code for replicating the analysis can be made available upon request.

3 Results

Table 1 presents descriptive statistics on geometric and arithmetic mean BLLs and the proportion of children exceeding the current and previous CDC reference values for all variables in the analysis. Nearly all of the covariates behave as expected or reflect biases known to be present in Colorado’s blood lead testing registry [1, 43]. Both mean BLL and the proportion exceeding CDC reference values increase with increased proximity to airports, increased downwind exposure, and increasing age, as well as for male patients, venous test type, increasing draw order, and during summer and autumn months. Among the census tract-level risk factors, we find mean BLL and the proportion of BLLs exceeding CDC reference values increase with increasing percentage of homes built before 1960, decreasing median household income, and decreasing population density. Notable exceptions to expectations include patterns for the TRI count, estimated aircraft traffic counts, trends for the lowest downwind exposure groups, and the tract percentage of people of color, all of which suggest a negative association with mean BLLs and proportions exceeding CDC reference values in the sample. Finally, the overall temporal trend in the sample reflects the slowly declining mean BLLs observed in Colorado and in the United States over the 10-year study period [1]. We also note that the limit of detection for blood lead level tests varies between < 1.0 µg/dL and < 3.3 µg/dL, and that all summary statistics are calculated by replacing non-detects with their labeled limit of detection. Tables displaying the distribution of tests below and above the limit of detection are provided in the supplementary materials.

Table 1 Descriptive statistics of blood lead levels stratified by independent variables
Full size table

Table 2 reports regression coefficients and standard errors for our primary model predicting arithmetic mean child BLLs measured continuously in µg/dL in column 1, as well as for three additional models predicting the natural log of mean child BLLs and the log odds of child BLLs exceeding 3.5 and 5 µg/dL in columns 2–4. Across all four model specifications, we find evidence of a statistically significant decrease in mean BLL and odds of exceeding CDC reference values with increased distance from airports, controlling for all other factors. Specifically, a one-mile increase in distance away from an airport is associated with a 0.068 (95% CI 0.024–0.111) µg/dL decrease in arithmetic mean child BLLs in our primary model. We find the same increase in distance from an airport is associated with a 2.1% (95% CI 0.8–3.2%) decrease in geometric mean blood lead levels when the dependent variable is measured on the natural log scale. Similarly, by exponentiating the coefficients reported in columns 3 and 4 in Table 2, we find that a one-mile increase in distance away from an airport is associated with a 13.2% (95% CI 4.6–21.2%) and 14.8% (95% CI 2.6–25.4%) decrease in the odds of having an elevated blood lead level ≥ 3.5 and 5 µg/dL, respectively.

Table 2 Regression results for our primary model predicting mean child BLLs (column 1) and additional models predicting the natural log of mean child BLLs (column 2) and the log odds of child BLLs exceeding CDC reference values (columns 3–4)
Full size table

We observe a similarly consistent pattern with respect to our downwind exposure variable. Across all four model specifications, we find a statistically significant increase in mean BLL and odds of exceeding CDC reference values for increased downwind exposure, controlling for all other factors. Specifically, for a one-day increase (out of the 60 days prior to testing) in downwind exposure, our primary model predicts a 0.005 (95% CI 0.003–0.008) µg/dL increase in arithmetic mean blood lead levels. The same increase in downwind exposure was associated with a 0.13% (95% CI 0.06–0.21%) increase in geometric mean blood lead level when the dependent variable is measured on the natural log scale, as well as with a 1.2% (95% CI 0.5–1.8%) and 1.6% (95% CI 0.8–2.5%) increase in the odds of a BLL greater than or equal to the two CDC reference values.

With respect to the additional control variables, the magnitude and direction of the coefficients correspond with expectations in nearly all cases. In our primary model, male sex, increasing draw order, summer and autumn test dates, and the tract percentage of homes built before 1960 all had a statistically significant positive association with mean BLL after controlling for other factors. Across all four model specifications, age, male sex, increasing draw order, summer and autumn test dates, tract percentage of homes built before 1960, and tract percentage of people of color all have positive associations with mean BLL and odds of having an BLL exceeding CDC reference values. The effect of residing in a census tract with a greater percentage of homes built before 1960 is of particular interest due to the higher likelihood of lead exposure from deteriorating lead-based paint in these neighborhoods [1, 44]. Our results suggest that children residing in a census tract in the 75th percentile for the percentage of homes built before 1960 (i.e., a tract with ≥ 13.2% homes built before 1960) have arithmetic mean BLLs that are 0.063 µg/dL higher, geometric means that are 1.6% higher, and have odds of exceeding CDC reference values that are 16.2% and 19.1% higher than their counterparts residing in census tracts in the 25th percentile (i.e., a tract with ≤ 1.26% of homes built prior to 1960). In addition, capillary test types, tract median income, tract population density, and the quarter/year variable capturing a linear effect of time all consistently show a negative association with mean BLL and odds of exceeding CDC reference values, controlling for other factors. TRI count and estimated monthly aircraft traffic both show negative associations with mean BLL and odds of exceeding CDC reference values, with the TRI count variable counterintuitively registering a statistically significant and protective effect across all models.

In addition, the correlated random effects model generates regression coefficients for the four terms representing the tract-level means of the individual test-level continuous variables that vary within and between census tracts (distance, downwind days, age, and TRI count). The coefficients for these terms assess the statistical significance of the difference between the within-tract and between-tract effects for each of these four variables. The substantive interpretation of these coefficients, which are sometimes referred to as contextual effects, is of limited interest in this context [32]. However, the statistical significance of the mean terms for the downwind and TRI variables indicates that the correlated random effects model reduces biases in the individual test-level coefficients that would otherwise be present in the standard random effects model with a tract random intercept term [32].

We implemented a variety of sensitivity analyses to evaluate the impact of modeling choices. First, Table 3 displays regression coefficients and standard errors for the exposure variables in our primary model (where the dependent variable is BLL measured in µg/dL) with three variations. Column 1 replicates the primary model with the sample restricted to only children less than 3 years of age. Column 2 again replicates the primary model but increases the number of airports in the analysis to 22 by relaxing the inclusion criteria. Lastly, column 3 contains results from our primary model modified to include a random effect parameter for each individual child in the sample. The results are consistent with those reported in our primary analysis across all three model variations. Specifically, the effect of a one-mile increase in distance away from an airport is associated with a statistically significant decrease of 0.070, 0.058, and 0.048 µg/dL in arithmetic mean BLL for each of the three models, respectively. A similar statistically significant effect is observed for the downwind exposure variable, with a one-day increase in exposure associated with a 0.005, 0.004, and 0.003 µg/dL increase in arithmetic mean BLL, respectively.

Table 3 Regression coefficients and standard errors for the exposure variables in our primary model (with the dependent variable measured in µg/dL) with three variations
Full size table

Results of expanded sensitivity analyses for our distance exposure variable are displayed in Tables S1–S5 in the supplementary materials. In Tables S1–S3, regression coefficients for the effect of distance from an airport are displayed in columns 1–8 in order of increasing model saturation, and we present coefficients for continuous and discrete distance specifications. Table S1 reports results for models with the dependent variable measured in µg/dL. The statistical significance of the effect of distance from an airport is robust to all levels of model saturation when it is included in the regression model as a linear effect of a continuous variable, and the coefficient ranged from − 0.048 to − 0.086 µg/dL. The results are similar when distance is measured discretely, with the medium distance exposure group having statistically lower arithmetic mean BLLs across all levels of model saturation; however, in the most saturated models, the differences between the near and far distance groups are no longer statistically significant. Table S2 displays results for models with the dependent variable measured on the natural log scale, and these results are nearly identical to those reported in Table S1. Table S3 displays the results of logistic regression models predicting the odds of having BLL greater than or equal to 3.5 and 5 µg/dL, respectively. The results of models predicting odds of exceeding CDC reference values are less robust to increasing model saturation and replacing the continuous distance exposure variable with a discrete distance specification. Specifically, the coefficient for continuous distance is no longer statistically significant for models predicting odds of exceeding CDC reference values when an individual random effect parameter is added to the model. In addition, when the continuous distance variable is transformed into a discrete variable the direction and statistical significance of the association with the odds of exceeding CDC reference values become inconsistent.

Additional sensitivity analyses are displayed in Tables S4, S5. Table S4 displays the results for our leave-one-out analysis of airports. In each case, regression coefficients for the distance exposure are statistically significant, indicating that a one-mile increase in distance from an airport results in a decrease in arithmetic mean BLLs that ranges from − 0.042 to − 0.091 µg/dL, depending on which airport is excluded from the analysis. Table S5 shows the effect of different methods of handling blood lead test results below the limit of detection. Approximately 55% of all tests included in our analysis were non-detects. In our main analysis, we replace non-detects with the limit of detection and include a dummy variable indicating whether each test result is above or below the limit of detection, which is consistent with the approach of Zahran et al. [45]. Table S5 shows the results if we replace non-detects with half the limit of detection, the limit of detection divided by √2, or 0, and if non-detects are excluded from the analysis. We find that a one-mile increase in distance results in a decrease in arithmetic mean BLL ranging from -0.088 to -0.050 µg/dL. The results are statistically significant for all cases where non-detects are included but were not significant at the conventional alpha level if non-detects were excluded from the analysis.

Lastly, Fig. 1 displays fitted values and 95 percent confidence intervals for arithmetic mean BLLs from our primary model conditioned on distance from airport point locations and the level of downwind exposure. The marginal means are computed assuming a detected venous blood draw in summer for a male child, with the time trend fixed to the end of the study period (Q4/2020), with all other covariates held constant at their sample medians. The red line in Fig. 1 displays predicted mean BLLs at varying distances for a hypothetical population experiencing 21 days (i.e., the 75th percentile) of downwind exposure over the 60 days prior to testing. The predicted means range from 2.36 µg/dL (95% CI 2.19–2.52 µg/dL) at 0 miles from an airport (i.e., at a distance that is coterminous with the geographic point representing an airport) to 2.02 (95% CI 1.91–2.12 µg/dL) at 5 miles from an airport. The blue line in Fig. 1 represents the same trend of predicted mean BLLs for a hypothetical population experiencing 0 days of downwind exposure in the 60 days prior to testing and ranges from 2.24 µg/dL (CI 2.07–2.42 µg/dL) to 1.90 µg/dL (CI 1.79–2.01 µg/dL). For reference, we also graph a horizontal line indicating 2.46 µg/dL, which is the observed sample arithmetic mean BLL, reported in Table 1, of all children residing in census tracts that rank at or above the 75th percentile in the percentage of homes built prior to 1960. In addition, Fig. 2 displays predicted probabilities and 95 percent confidence intervals of a child’s BLL exceeding 5 µg/dL, assuming the same conditions as in Fig. 1. Predicted probabilities of exceeding the 5 µg/dL reference value for the hypothetical population experiencing 21 days of downwind exposure range from 5.2% (CI 3.2–8.3%) at 0 miles from an airport to 2.4% (CI 1.8–3.2%) at 5 miles from an airport, while the same probabilities range from 3.8% (CI 2.3–6.2%) to 1.7% (CI 1.3–2.4%) for the minimum downwind exposure group. Once more for reference, a horizontal line indicating 6.08%, which is the observed sample percentage of tests ≥ 5 µg/dL, reported in Table 1, among children residing in census tracts that rank at or above the 75th percentile in the percentage of homes built prior to 1960.

Fig. 1
figure 1

Predicted mean blood lead levels and 95% confidence intervals conditioned on distance and downwind exposure. The marginal means are computed assuming a detected venous blood draw in summer for a male child, with the time trend fixed to Q4/2020, with all other covariates held constant at their sample medians. The red line shows predicted mean BLLs at varying distances for a hypothetical population experiencing 21 days (i.e., the 75th percentile) of downwind exposure over the 60 days prior to testing. The blue line displays predicted means BLLs over the same distances for a population experiencing 0 days of downwind exposure. The horizontal dotted black line represents the sample mean BLL of 2.46 µg/dL for children residing in a census tract that ranks at or above the 75th percentile of homes built before 1960

Full size image
Fig. 2
figure 2

Predicted probabilities and 95% confidence intervals of a child’s BLL exceeding 5 µg/dL, assuming the same conditions as in Fig. 1. The red line displays predicted probabilities of exceeding the 5 µg/dL threshold for the hypothetical population experiencing 21 days. The blue line displays the same probabilities at each distance for a population experiencing 0 days of downwind exposure. The horizontal dotted line represents 6.08%, the sample percentage of BLLs ≥ 5 µg/dL for children residing in a census tract that ranks at or above the 75th percentile for the percentage of homes built before 1960

Full size image

4 Discussion

Only three previous peer-reviewed studies have explicitly examined the link between proximity to airports and children’s blood lead levels in the United States [45,46,47]. An early study by Miranda et al. [46] analyzed more than 125,000 children’s blood lead levels in relation to airport property boundaries in six counties in North Carolina over a nine-year study period. Using the natural log of children’s BLLs as the dependent variable, the authors reported statistically significant increases in geometric mean BLLs of 4.4%, 3.8%, or 2.1% among children living < 500 m , < 1000 m, or < 1500 m, from the airport property boundaries respectively, compared to other children in the study. Zaharan et al. [47] analyzed more than 1 million blood lead tests from children under 5 years of age living within 10 km of airport point locations in Michigan over a nine-year study period. In their analysis predicting the natural log of BLLs, they found a 1 km increase in continuous distance from an airport point location was associated with a 0.7% decrease in geometric mean BLLs, and that children living < 1 km, 1–2 km, and 2–3 km from an airport point location had geometric mean BLLs that were 5.9%, 2.9%, and 2.4% higher when compared to children living ≥ 4 km from an airport point location. In their logistic regression analysis, the authors found that a 1 km increase in continuous distance from an airport point location was associated with a 2.5% decrease in the odds of having a BLL ≥ 5 µg/dL, and children living < 1 km, 1–2 km, and 2–3 km from an airport point location had odds that were 25.2%, 16.5%, and 9.1% higher than those living ≥ 4 km from an airport point location. Lastly, a more recent study by Zahran et al. [45] analyzed more than 14,000 blood lead tests from children living within 1.5 miles (2.4 km) of a point location representing RHV in Santa Clara County, California. The authors report that a one-mile increase in distance from the RHV point location was associated with a 0.093 µg/dL decrease in arithmetic mean BLLs and a 3.7% decrease in geometric mean BLLs. In addition, children living 0.5 to < 1 mile and 1 to < 1.5 miles from the RHV point location have arithmetic means that are 0.234 µg/dL and 0.235 µd/dL lower (and geometric means that 5.4% and 5.6% lower) than their counterparts living < 0.5 miles from the RHV point location. In addition, in comparison to children living < 0.5 miles from the RHV point location, children living 0.5 to < 1 mile and 1 to < 1.5 miles from the RHV point location have 17.3% and 21.4% lower odds of having a BLL ≥ 4.5 µg/dL.

This study analyzed the association between childhood blood lead levels and living in proximity to airport point locations in Colorado. Along with distance from an airport point location, we identified the number of days sampled children lived downwind of their nearest airport during the 60 days prior to their test date. We also adjusted the model for a number of additional variables, including age, sex, detection limit, test type, sample order, seasonality, time, proximity to lead-releasing TRI facilities, estimated monthly aircraft traffic, and demographic characteristics of the surrounding neighborhood, including the percentage of homes built before 1960, median income, the percentage of people of color, and population density. In addition, we conducted a variety of sensitivity analyses in which we considered alternative continuous and discrete specifications of the dependent variable and distance exposure variable, decreased the sample size by restricting the population under study to children less than 3 years of age, increased the sample size by including additional airports in the analysis, evaluated the impact of adding an additional individual random effect parameter, evaluated the influence of outliers by conducting a leave-one-out analysis for the 12 airport point locations in our primary model, and assessed the impact of varying our method for handling non-detects in the sample.

Our primary model estimates that a one-mile increase in distance from an airport point location decreases arithmetic mean BLLs by 0.068 µg/dL, and the effect of distance from an airport point location was robust to all sensitivity tests when it was included in the regression model as a continuous variable. When we applied the natural log transformation to our dependent variable, we found a one-mile increase in distance from an airport point location decreases geometric mean blood lead levels by 2.1%, and the effect was similarly robust across sensitivity analyses. Logistic regression models predicting the odds of having an elevated blood lead level greater than or equal to the current and previous CDC reference levels of 3.5 and 5 µg/dL showed that living one mile further from an airport lowered the odds of having an elevated blood lead level by 13.2% and 14.8%, respectively. However, the logistic regression results were less robust to sensitivity analyses, and all models were less consistent when the distance exposure was transformed into a discrete variable.

There are several reasons to prefer our primary model with a continuous distance exposure variable over those considered in the sensitivity analyses. First, it is well-known that dichotomizing continuous random variables can lead to a significant loss of statistical power to detect associations in regression analyses, and there is no obvious theoretical reason to prefer a discrete over a continuous distance exposure specification despite its use in previous studies [48]. Second, consistent with the approach of the three previously published peer-reviewed studies, we have retained all non-duplicate tests of children who were tested multiple times during the study period. We include individual random effects models in our sensitivity analysis to make our results comparable to Zahran et al. [45] who utilized this study design to address the lack of statistical independence between observations from the same individual. However, our primary model excludes the individual random effect because more than 80% of tested individuals in our sample have only one observation and because we also control for a sample order covariate and census tract fixed effects. Other researchers have shown that when individual random effects are estimated in samples where most individuals have only one observation the random effect parameter can become confounded with the residual variance and cannot be uniquely identified [49]. The dramatic decrease in the marginal R2 and concomitant increase in conditional R2 observed in our individual random effect model (see Table 3) indicates that this type of confounding is present and that the use of individual random effects is not advisable in our study. The conclusion that our results are not impacted by retaining children with multiple tests is also supported by a cross-tabulation of the data from our main analysis (N = 56,002) showing no differences in the distribution by distance from airport point locations between singleton and repeated tests. This table is included in the supplementary materials.

Thus, the results of our primary models for all dependent variable specifications are generally consistent in magnitude, direction, and goodness-of-fit with those observed in the three previous peer-reviewed studies that explicitly link proximity to airports to children’s BLLs. This is especially noteworthy given that each of these studies was conducted in a different time period, using different distance specifications, different regression methods, and in three distinct geographic regions of the United States, with considerable differences in climate, terrain, and in their historical patterns of urbanization and economic development, leading to different potential risks of lead exposure through background soil lead concentrations or the concentration of older housing stock with possible lead-based paint [50, 51]. Consistent with previous studies, our analysis finds a small but statistically significant increase in mean BLLs and the odds of exceeding CDC reference levels among children who live near airports servicing general aviation aircraft in Colorado. Given that exposure to lead-based paint and associated dust is commonly believed to be the dominant lead exposure pathway, [1, 44, 52] a natural benchmark for comparison is the estimated risk of living in an older home. Figures 1 and 2 present our findings in comparison to the observed mean BLLs and percent of BLLs ≥ 5 µg/dL for children living in census tracts that rank at or above the 75th percentile of homes built before 1960. These graphs indicate that arithmetic mean BLLs and odds of exceeding CDC reference values measurably increase with proximity to airports, but the impact on BLLs is comparable to this commonly used proxy for the risk of living in older housing only at a distance that is coterminous with a geographic point representing the location of an airport and at the highest levels of observed downwind exposure, conditions that are unlikely to exist in actuality. Thus, our modeled results suggest the potential impact on children’s blood lead levels is greatest at distances less than 2 miles from an airport point location, but as a substantive risk factor, proximity to an airport point location is potentially less important than living in a neighborhood with a relatively high percentage of homes built prior to 1960.

One notable difference between the present study and the three previously published studies is Colorado’s relatively smaller sample size of blood lead tests near airports. Low blood lead testing rates, especially in rural areas, are Colorado’s most significant challenge to lead poisoning prevention [53]. Based on federal requirements from the Centers for Medicare & Medicaid Services, 100% of Medicaid/CHP + -enrolled children should have their blood tested for lead at ages 12 and 24 months. In Colorado, only 11% of Medicaid/CHP + -enrolled children under 3 years of age received a blood lead test in fiscal year 2020 [54]. At the outset, we aimed to look at all blood lead tests reported to the CDPHE between 2011 and 2020, from children ages 18 years and younger who lived within 5 miles of an airport point location. However, due to sample size limitations, we focused our primary analysis on 12 airport point locations that had at least 300 blood lead tests within three miles during the 10-year study period in order to help ensure that the airports included in the study all have a relatively similar and even spatial distribution of blood lead tests at varying distances between 0 and 5 miles. The relatively smaller sample size in Colorado may also help explain our finding that the association between BLLs and distance was less robust when the exposure variable was transformed from a continuous to a discrete variable.

Further, this discussion highlights the importance of maintaining and even increasing blood lead testing rates in light of both the continuing decline in mean BLLs at the population level and the increasingly lower CDC reference value. Lowering exposure to lead lowers the risk of health effects, and new research emphasizes the danger of even low levels of exposure [55]. Testing is critical not only to protect children who might have lead exposure but also because it provides valuable public health data needed to investigate less well-studied exposure pathways and how they affect BLLs. Currently, the U.S. Preventive Services Task Force concludes that the available evidence is insufficient to assess the benefits and harms of screening for elevated blood lead levels in asymptomatic children, and the American Academy of Family Physicians recommends against universal lead testing of asymptomatic children ages 1 to 5 years [56]. However, the results of our study may call these recommendations into question and suggest adequate blood lead testing remains vital to uncovering and eliminating lead exposure pathways, as the measurable impact of proximity to airports on children’s BLLs in our study is not likely to produce obvious symptoms [57, 58].

This study has several limitations. Our study assumes exposure to lead occurs at home and does not account for exposures at school, childcare facilities, or other places where children may spend a significant amount of their time. However, our sensitivity analysis restricting the study population to children < 3 years of age may partially mitigate this effect by modeling the distance-to-airport effect among a population that is not old enough to attend school. Furthermore, we can only account for aggregate housing age at the census tract level. We were also unable to obtain GIS data of the exact property boundaries of each airport and had to rely on distances calculated from a point location representing the geocoded address of each airport. The point locations were verified using aerial imagery; however, we are unable to account for the exact spatial pattern of aircraft takeoff and landing, or the location of the exact property boundary of each airport or how variations in setback distance affect the results. We assume distance to an airport is a reasonable proxy for exposure to leaded avgas emissions, and we are unable to say whether the risk is from current or previous lead emissions. There is also no available data on soil lead concentrations in the affected areas. In addition, our results also suggest that proximity to lead-emitting TRI facilities had a protective effect in our sample. We suspect that this may be a reflection of the known limitations of self-reported regulatory data for capturing exposure rather than a legitimate protective effect [59]. However, this implies that we may have been unable to effectively control for additional industrial sources of lead exposure. Further, we relied on an average monthly traffic estimate because detailed data on avgas consumption were not available for each airport in our study. Detailed information on fuel consumption collected for the RHV airport enabled Zahran et al. [45] to conduct extended analyses of how children’s BLLs responded to fluctuations in avgas usage, but we are unable to do so in the present study. Taken together, all of these factors could influence mean BLLs and odds of exceeding CDC reference values in Colorado and represent potentially important omitted variables. However, our control for census tract fixed effects in regression models is intended to reduce omitted variable bias and our results suggest the associations we observe between proximity to an airport and children’s blood lead levels are robust to a variety of sensitivity analyses and variations in study design. Future work should attempt to measure the effects of these additional variables.

5 Conclusions

This study evaluates the available evidence for a link between living in proximity to airports that service general aviation aircraft and children’s blood lead levels in a targeted blood lead testing state with no requirements for mandatory blood lead testing beyond Medicaid/CHP + -enrolled children. Our results suggest that lead emissions from aircraft using leaded avgas have a small but statistically significant effect on the BLLs of children living near airport point locations in Colorado. While the risk from proximity to airports is likely less than the risks associated with deteriorating lead paint and dust, our findings highlight the value of lead testing in children who may be at risk of exposure from all sources, including those who live in proximity to airports. On its own, the measurable impact on children’s BLLs detected in this study may not be severe enough to cause affected children to exhibit clear symptoms of lead poisoning. However, lower exposure to lead lowers the risk of health effects, and thus testing children living near airports could prove useful in better understanding lead risk throughout Colorado.