We began our descriptive analysis by examining patterns in the linked data. For space reasons, we present here only the descriptive statistics for 4th-grade reading results. The 4th-grade mathematics results follow strikingly similar patterns. The multivariate analyses presented below include both 4th-grade reading and mathematics.
shows the distribution of children across blood lead levels and race categories. Of the total linked children for 4th-grade reading, 44.8% are white and 55.2% are black. Compared with black children, white children are overrepresented in the lower blood lead level categories (blood lead level, 1 to 3) and underrepresented in the higher blood lead level categories (blood lead level, 4 to ≥ 10). This blood lead level cut point at 3 holds for the 4th-grade mathematics scores as well.
Distribution of blood lead levels among white and black children.
thus demonstrates a distribution for black children that is shifted to the right and is characterized by higher variance compared with white children. These sample distributions are statistically different from each other. Construction of a dissimilarity index indicates that 25% of the members of one group would need to be reassigned blood lead levels for the two groups to show equivalent blood lead level distributions. The Mantel-Haenszel chi-square test for equality of distribution shows the two sample distributions to be statistically significantly different from each other (p < 0.0001).
shows the mean reading scores by race and blood lead levels for all linked students in the 4th-grade reading data set, disaggregated by race. This graphic shows a clear negative relationship between test scores and blood lead levels: Higher blood lead levels are associated with lower test scores, with some erratic behavior at blood lead levels of 9 μg/dL, likely due to the small sample size at this higher blood lead level.
4th-grade mean Reading EOG test results stratified by blood lead levels.
At the lower end of the achievement scale, also demonstrates a dose–response effect between blood lead levels and failure on the end-of-grade test. Subgroups of children with lower blood lead levels in early childhood have lower failure rates on both the mathematics and reading end-of-grade tests (data shown only for 4th-grade reading data set); subgroups with higher blood lead levels in early childhood have higher failure rates.
Percent of students failing 4th-grade Reading EOG.
Although this descriptive evidence is consistent with claims of a causal relationship between blood lead levels and test performance, alternative interpretations are plausible and can be addressed using multivariate analysis. For instance, given the higher blood lead level for children of lower socioeconomic status (as measured by free/reduced-price lunch and low parental education), perhaps these factors are responsible for the observed association of blood lead levels and test scores. Thus we used multivariate analysis to control for the covariates noted in “Methods.” The referent group is defined as white female students, enrolled in the Wake County School System, who do not participate in the free or reduced-price lunch program, who do not use a computer daily, and whose parents graduated high school.
To explore the functional form of the association between the lead variable and test scores, we compare three alternative specifications. The 6 analyses (3 models × 2 data sets) are presented in and .
Results of multivariate regression models for 4th-grade Reading EOG score data (n = 8,603).
Results of multivariate regression models for 4th-grade Mathematics EOG score data (n = 8,627).
In all models, the coefficients on the covariates are of the expected sign. The coefficient on the age at which the blood screen occurred is negative and highly significant, indicating that a higher blood lead level at a later age has a stronger depressive effect on test performance. This likely results from the fact that children who have high blood lead levels at 4 or 5 years of age typically would have had even higher blood lead levels at 2 or 3 years of age, given that the latter is typically considered the age of peak exposure (Canfield et al. 2003
; CDC 1997
; Dietrich et al. 2001
The first model represents blood lead level as a continuous variable: We constrain the effect of a one-unit increase in blood lead level to be identical over the full range of observed scores. The coefficient on blood lead level is negative and statistically significant for 4th-grade reading and 4th-grade mathematics (both p < 0.0001). This effect and others discussed below are net of all control variables shown in the table.
The second model includes two dummy variables: one that is set equal to 1 if the blood lead level is 5–9 μg/dL; and one that is set equal to 1 if the blood lead level is ≥ 10 μg/dL. The coefficient on the dummy variable for a blood lead level of 5–9 μg/dL is negative and significant in both the reading and mathematics models (both p < 0.0001). In addition, the coefficient on the dummy variable for a blood lead level of 10 μg/dL is negative and significant in both the reading and mathematics models (again, p < 0.0001). In analysis not shown here, we also estimated a model that used a single dummy variable for blood lead level ≥ 5 μg/dL and a separate model with a single dummy variable for blood lead level ≥ 10 μg/dL. The results in and , in comparison with other models not shown here, indicate that if one is going to conceptualize the association by a threshold value, then ≥ 5 μg/dL captures much more of the variation in these data than does the CDC blood lead action level of ≥ 10 μg/dL.
The third model enters a dummy variable for each blood lead level (2, 3, 4, … 9, ≥ 10 μg/dL). The last dummy variable combines all blood lead levels ≥ 10 μg/dL, and the referent group is a blood lead level of 1 μg/dL. This scoring is the most flexible and allows a distinct estimate at each blood lead level score.
For the 4th-grade reading analysis, the coefficient on the dummy variable for a blood lead level of 2 μg/dL is negative and marginally significant at p = 0.05. The coefficients on the dummy variables for blood lead levels of 3–8 and 10 μg/dL are consistently negative and statistically significant, and generally increase in absolute magnitude as the blood lead levels increase (all p < 0.0001). The coefficient on the dummy variable for a blood lead level of 9 μg/dL is also negative but significant only at the p = 0.02 level, likely due to the small sample size in this grouping. The results for the 4th-grade mathematics analysis follow a very similar pattern to those of the reading analysis, although the coefficient on the dummy variable for a blood lead level of 2 μg/dL is significant at the p = 0.03 level, and the coefficient on the dummy variable for a blood lead level of 9 μg/dL is significant at the p < 0.0001 level.
Model 3 results demonstrate a strong dose–response effect between early childhood lead exposure and performance on elementary school achievement tests. These results indicate clearly that early childhood lead exposure has a statistically significant and negative impact on school performance at levels well below the current CDC blood lead action level. These results are consistent with the observed association between blood lead levels and elementary school achievement scores demonstrated in both the descriptive analysis and regression models 1–2. All three models indicate, net of a set of control variables, that higher blood lead levels are associated with lower test scores. The least constrained model (model 3) reveals a general decline in test scores with rising blood lead levels. Model 1 constrains this decline to be uniform across all blood lead levels. With our data, we cannot reject the latter in favor of the former; any divergence from a linear decline could be attributed to sampling variability. Model 2 can be aligned with the following question: Once we take account of high blood lead levels (i.e., ≥ 10 μg/dL) is additional variation in blood lead levels important? Results clearly indicate that blood lead levels of 5–10 μg/dL are consequential for test scores. We conclude from these various representations that early childhood blood lead levels reduce test scores and that this effect is clear even at levels < 10 and even < 5 μg/dL.
Given the statistical measures of model fit provided in and (adjusted R2, AIC, and root MSE), all three models show adequate and substantially similar model fit. and graphically summarize the results of models 1 and 3 for the 4th-grade reading and mathematics analyses graphically. These figures aptly demonstrate that test scores decline as early childhood blood lead levels increase. Because model 3 allows a distinct estimate at each blood lead level score, it is useful to compare it directly with model 1, which constrains the effect of a one-unit increase in blood lead level to be uniform across observed scores. and show that the decline in both reading and mathematics scores is steeper at lower blood lead levels than at higher blood lead levels.
Figure 5 Comparing model results for 4th-grade Reading score. Based on a referent individual who was screened at 2 years of age and is a white female, living in Wake County, parents with a high school education, not enrolled in the school lunch program, and who (more ...)
Figure 6 Comparing model results for 4th-grade Mathematics scores. Based on a referent individual who was screened at 2 years of age and is a white female, living in Wake County, parents with a high school education, not enrolled in the school lunch program, and (more ...)