The Effect of Lead Exposure on Children’s Learning in the Developing World: A Meta-Analysis Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Lee Crawfurd , Rory Todd, Susannah Hares, Justin Sandefur, and Rachel Silverman Bonnifield Around half of children in low-income countries have elevated blood-lead levels. What role does lead play in explaining low educational outcomes in these settings? We conduct a new systematic review and meta-analysis of observational studies on the relationship between lead exposure and learning outcomes. Adjusting for observable confounds and publication bias yields a benchmark estimate of a −0.12 standard-deviation reduction in learning per natural log unit of blood lead. As all estimates are non-experimental, we present evidence on the likely magnitude of unobserved confounding, and summarize re- sults from a smaller set of natural experiments. Our benchmark estimate accounts for over a fifth of the gap in learning outcomes between rich and poor countries, and implies mod- erate learning gains from targeted interventions for highly exposed groups (≈ 0.1 standard deviations) and modest learning gains (< 0.05 standard deviations) from broader public health campaigns. JEL Codes: P46, I18, K32 Keywords: lead poisoning, child education, developing countries. Introduction Over 600 million children in low- and middle-income countries have elevated blood- lead levels (Rees and Fuller 2020). This includes just 3 percent of children in high- income countries, and more than half of children in low-income countries (fig. S1.1).1 Children’s scores on standardized tests of reading and mathematics in these same countries typically fall between one and three standard deviations below the level of performance of children in high-income settings (Angrist et al. 2021). The World Bank The World Bank Research Observer © The Author(s) 2024. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Devel- opment / THE WORLD BANK. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. https://doi.org/10.1093/wbro/lkae010 40:229–260 and UNESCO’s modeled estimates suggest 91 percent of children in low-income coun- tries cannot read and comprehend a simple text by age 10, compared to just 8 percent in high-income countries (Azevedo et al. 2022). Are these two broad patterns connected? In this paper we explore the evidence on the link between lead exposure and children’s learning outcomes. Since the relation- ship between lead and cognition varies with level of exposure, we place particular em- phasis on the likely impacts at blood-lead levels observed among children in the devel- Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 oping world. We then ask what potential gains there are to learning from feasible in- terventions to reduce lead exposure in low- and lower-middle-income countries. What proportion, if any, of the large learning gaps between countries might be explained by lead exposure? Further, is lead abatement a viable and cost-effective strategy to im- proving education performance? We approach these questions by revisiting and extending existing meta-analyses of studies on the relationship between measures of blood lead and child learning out- comes. Much of the literature on the effects of lead on cognitive development has fo- cused on impacts on IQ. A key contribution of our study is in additionally gathering effects on standardized test scores for reading and mathematics. We show that the magnitude of the association between blood-lead levels is fairly consistent across these alternative outcomes. On average, we estimate that a one (natural) log unit increase in blood lead is associated with a −0.23σ change in reading and mathematics test scores. Ninety-five percent of estimates are between −0.28σ and −0.18σ . We then review how inclusion of controls for parental characteristics and socioeco- nomic status changes the association between lead and cognitive outcomes in associ- ational studies. Standard controls typically, but not uniformly, lead to smaller associ- ations with magnitudes that remain relevant for public health policy. We also assess the importance of publication bias in the literature. Standard tests based on funnel plots suggest publication bias may exaggerate the magnitude of the relationship be- tween lead and cognitive outcomes. Lower-powered studies report larger effects, and an anomalous share of p-values fall just below conventional significance levels. How- ever, standard approaches to correcting for this bias only moderately reduce our esti- mated average effect. Accounting for study characteristics and correcting for publication bias, we find that a (natural) log unit increase in blood lead is associated with a −0.12σ decrease in read- ing and mathematics scores in the developing world. Magnitudes are similar for effects on IQ, reading, and mathematics scores. While all of these estimates are based on observational studies, we also review the smaller number of quasi-experimental studies, all of which show larger effects for IV estimates than OLS estimates. By contrast, bounding procedures based on coefficient stability imply that unobserved heterogeneity could explain roughly one-third of the OLS association. 230 The World Bank Research Observer, vol. 40, no. 2 (2025) In order to quantify the overall potential learning gains from eliminating lead expo- sure, we combine our estimates of the association between lead and learning, with es- timates of the prevalence of lead poisoning, and estimates of the effect of interventions targeting lead exposure. If given a causal interpretation, the overall results of this meta- analysis imply that reducing children’s mean blood-lead levels from average existing levels in LMICs (5.3 μg/dL) to high-income country levels (0.5 μg/dL in the United States according to the EPA) would close 21 percent of the learning gap between de- Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 veloping and developed countries for which we have data. Projects designed to reduce acute lead exposure at specific polluted sites have reduced mean blood-lead levels by 1.35 to 2.3 log units, implying a 0.16σ to 0.26σ increase in learning for affected children. Projects targeting chronic low-level lead exposure in wider populations have reduced blood-lead levels by 0.34 to 0.42 log units, imply learning gains of 0.04σ to 0.05σ . Even these smaller effect sizes based on conservative assumptions about the effect of broad public health initiatives offer a potentially cost-effective means to improve learning outcomes (ignoring health benefits), if these programs can be implemented effectively and cheaply at scale in low- and middle-income countries. Meta-Analytic Methods We review the literature on the relationship between lead poisoning, measured through blood testing, and learning outcomes. We focus on studies that measure some kind of cognitive test outcome, and also have blood lead-level measures from the same indi- viduals, whether contemporaneously or from different points in time. Search Strategy Our search strategy is summarized in the flowchart in fig. 1. We seek to identify all studies measuring the correlation between blood-lead levels and IQ or test scores. We start with three recent systematic reviews. The first finds 27 papers published between 2010 and 2020 that measure child (aged 0–19 years) blood-lead levels and “full-scale” IQ scores (Galiciolli et al. 2022),2 of which 9 had at least one eligible result according to the criteria described in the next section. The second reviews 8 papers published between 2000 and 2020, focused on effects on IQ, with a study population under the age of 12 years (Heidari et al. 2022), of which 7 had at least one eligible result. The third reviews 34 studies in low- and middle-income countries with measures of lead expo- sure and a standardized measurement of neurodevelopment, from which we extract an additional 7 studies (Heng et al. 2022). In addition, we conducted our own systematic search using Google Scholar. We search for articles with the following terms in their title: (lead) AND (exposure OR blood OR level) AND (intelligence OR intellectual OR cognitive OR cognition OR education OR achievement OR IQ OR score OR math OR reading OR school). From this we found 951 potential results. We also found several Crawfurd et al. 231 Figure 1. Systematic Review Process Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: This flowchart shows the process used to identify relevant studies and extract eligible results. studies through citation searching from an initial relevant paper. These two methods identified an additional 24 papers that had at least one result fulfilling the eligibility criteria described below. This led to a combined total of 47 unique studies. These are all listed in tables S1.1 and S1.2. Extracting Estimates from Studies Two researchers independently extracted estimates from studies following the same protocol, with disagreements resolved by a third researcher. We extract any coefficient relating maternal or child blood-lead levels to any of three cognitive outcomes—IQ, mathematics skills, or reading skills—on the full sample. We include estimates from sub-samples where subgroups are defined according to blood-lead level. We focus on full-scale IQ rather than its separate components (performance and verbal IQ) where those are reportedly additionally. We include studies that use tests that are designed 232 The World Bank Research Observer, vol. 40, no. 2 (2025) as general intelligence or IQ tests, such as the General Cognitive Index (Schnaas et al. 2006; Cooney et al. 1989), the British Ability Scales (Fulton et al. 1987), and the Kohs Block Design Test (Vega-Dienstmaier et al. 2006). We focus on general reading and mathematics composite outcomes rather than any individual subcomponents. We ex- clude estimates which are entirely a combination of other extracted estimates. We ex- clude results which include blood measures for multiple ages in a model separately, as this has a different estimand: the effect of exposure at a particular age, relative to expo- Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 sure at another age. We exclude estimates where the outcome is a binary re-coding of a continuous outcome, and a result using this continuous outcome has been extracted. We exclude estimates where the outcome was measured before the age of three. We exclude results of models which interact lead exposure with other covariates. For re- sults to be eligible, studies need to report—or provide the information to calculate—an effect size, a standard error or confidence interval, a measure of blood-lead exposure, and the standard deviation of the outcome. We also discarded results with fewer than 20 observations. After discarding studies that did not meet these criteria, we are left with 286 estimates from 47 unique studies. We also code a number of other characteristics for each study: the country where the study was carried out, the specific outcome type (IQ, reading, or mathematics), the standard deviation of this outcome in the sample, the functional form for the estima- tion of the effect of blood lead on the cognitive outcome, the mean blood-lead level of the sample,3 whether exposure for a result was based on more than one blood mea- surement, the average age at blood sampling, the average age at outcome testing, and which variables were adjusted for in the analysis. As we have multiple estimates from most studies, we account for the resulting de- pendence between results using robust variance estimation (Hedges et al. 2010), im- puting a conservative value of 1 for the correlation between results from the same study, meaning that studies which contributed more results were not overweighted (Fisher and Tipton 2015). Harmonizing Variables across Studies We follow the guidelines provided by the Cochrane Handbook for Systematic Reviews of Interventions (Higgins and Green 2008) to calculate effect sizes and standard errors where these are not given in the required form. Where studies give effects as correlation coefficients, these are converted to raw regression coefficients using the formula σy Effect size = Correlation coefficient × , σx where σy and σx denote the standard deviation of the outcome (IQ or test scores, in our case) and the independent variable (blood-lead levels), respectively. Crawfurd et al. 233 Figure 2. Log-Linear Dose-Response Curves Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: This figure reproduces dose-response curves published from the individual studies named. They have been stan- dardized and centered to converge at 5 μg/dL, to allow gradients to be compared. Curves from Edwards, Anthopolos, and Miranda (2013) refer to those for non-Hispanic Black children. Where studies only provide p-values for effect sizes, we convert these to standard errors using the formulas Standard error = Effect size/z, where z is calculated using the inverse CDF of the p-value divided by 2. Where studies report only that a p-value is less than a particular value, we impute half the threshold value (e.g., if a paper were to report p < 0.01, we would impute p = 0.005). For studies which do not report sample standard deviations, we impute population standard deviations where available. One study (Shadbegian et al. 2019) has as its outcome the percentile in which the child placed in mathematics and read- ing tests; we convert this to a z-score, assuming test scores to be normally distributed. Re-expressing Coefficients from log and Linear Models Existing meta-analyses have documented a roughly log-linear relationship between blood-lead levels and IQ, whereby the natural logarithm of blood-lead levels is pro- portional to IQ scores (see for example fig. 2).4 Letting X denote a vector of individ- ual and family control variables, and subscript i denote individuals, this log-linear 234 The World Bank Research Observer, vol. 40, no. 2 (2025) specification is our reference benchmark in what follows: Yi − μy = α ln(BLLi ) + Xi γ + εi , σy where Yi represents an individual’s IQ or test score on mathematics or reading tests. However, individual studies use a range of functional forms to model the relation- ship between blood-lead exposure and the cognitive outcome. Some report a regres- sion coefficient for a linear increase in blood lead, some report the coefficient for a log- Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 unit increase—with bases of 2, e, or 10—and some report the difference between two groups defined by a value or range of values for blood-lead levels. In order to compare the effects of lead exposure across studies, we first need to harmonize these results. Based on Lanphear et al. (2005) and other studies shown in fig. 2, we assume that the true relationship between blood lead and cognitive outcomes is log-linear, although we allow more flexibility in our meta-regression by including a linear term for a study’s mean blood-lead level. In order to compare across studies, we convert all effect sizes into natural log units. We do this using a “re-expression algorithm.” All such algorithms are necessarily imperfect. Linakis et al. (2024) show that re-expressions are likely to be biased where the distribution of the underlying variable has a skewed distribution, which is true in our case. In table S1.3 we use microdata from the US NHANES study and three other studies in our review, to estimate both logarithmic and linear specifica- tions using the same data, and compare the performance of three algorithms: Linakis et al. (2024), Rodríguez-Barranco et al. (2017), and Dzierlenga et al. (2020). The Linakis et al. (2024) method has the lowest root mean squared error, and so we proceed with this algorithm. We first convert studies which compare discrete low- and high-exposure groups to a linear equivalent. Studies define “low”- and “high”-exposure groups using different thresholds. We therefore convert these effects to linear estimates by dividing the esti- mated effect size by the difference in average BLL between the high- and low-exposure groups.5 Following Linakis et al. (2024), we calculate a conversion factor I as a function of the desired logarithmic base α and median blood-lead level (BLL).6 I = α logα (meanBLL)+0.5 − α logα (meanBLL)−0.5 We then multiply coefficients estimated from linear models by this conversion factor to obtain an estimate of what the coefficient would have been under a logarithmic model. Overall, we see larger effects for estimates from linear models than from logarithmic models, suggesting that some error may still result from this approximation. In order to mitigate against this, we adjust for whether a result was reported in a non-logarithmic form in our meta-regression analysis, described below. We also include an interaction with the average blood-lead level of the sample, as the magnitude of error will be a function of this variable. Crawfurd et al. 235 Adjustments for Publication Bias A common concern in meta-analyses, particularly those involving observational stud- ies, is publication or reporting bias. If only statistically significant results are reported, we will produce a biased estimate of the true average effect. In our case, funnel plots do indicate that publication bias may be an issue in this literature, with relatively few statistically insignificant results ( fig. S1.3). Specifically, results are asymmetrically dis- tributed, with “missing” results in the region of null or even positive associations, and Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 the distribution of z-statistics shows a spike just below the significance threshold. The former pattern is confirmed by an Egger et al. (1997) asymmetry test. However, we note that some of this asymmetry may be due to errors arising from the re-expression algorithm and other conversions we apply to make results comparable. Imprecision in these conversions will have the effect of biasing both the effect size and standard error in the same direction. There is reason to think that this is true in our case: fig. S1.4 shows that results which have not been converted in some way show sig- nificantly weaker asymmetry. While this will make our estimate less precise, it will not upwardly bias the estimate. Adjusting for this spurious “publication bias” will therefore underestimate the true effect. While we still adjust for publication bias in our primary specification, we note that our estimate may be correspondingly conservative. There are several approaches to adjusting for publication bias, none of which are perfect. For our main estimate we include a control for the variance of each estimate, following Stanley and Doucouliagos (2014). We also show in supplementary online ap- pendix table S1.4 that adjusting our uncontrolled result using three other approaches, which generally only moderately weaken the effect. The simplest approach is the Egger regression intercept, which includes the standard error as a control variable. The Eg- ger approach in simulations is shown by Stanley and Doucouliagos (2014) to be overly biased towards zero, and that a better approach (“PET-PEESE”) is to use a nonlinear quadratic approximation to the true unknown nonlinear relationship between effect sizes and effect size variance, as we do in our main result. The trim-and-fill approach estimates the number of results “missing” due to publication bias and impute these results. The p-uniform∗ method makes use of the principle that the p-values of esti- mates should be uniformly distributed at the true effect size (van Aert and van Assen 2018). Finally, the Andrews and Kasy method7 categorizes results into those which are statistically significant, which are modeled as being published with a probability of 1, and those which are not, which are published with estimated probability p. We can also show the sensitivity of average effects to different assumptions about the degree of selection on statistical significance (Copas and Shi 2001). Results from this method, shown in table S1.5, indicate that treatment effects remain statistically significant with a range of reasonable assumptions about the potential degree of selection of estimates on statistical significance ( fig. S1.5). 236 The World Bank Research Observer, vol. 40, no. 2 (2025) Meta-Regression Our focus in this review is estimating the effect of lead exposure on learning in devel- oping countries. Our primary estimate of the average effect is therefore the constant term β0 from a meta-regression in which an observed effect size θ ˆi j for result i in study j is related to m result covariates, 1 through M . Estimates are weighted inversely to their variance, and we use the Hedges et al. (2010) estimator which is robust to unknown correlation between multiple estimates from the same study: Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 ˆi j = β0 + β1 X1i j + β2 X2i j + · · · + βm XMi j + θ i j. The independent variables correspond to characteristics of the underlying studies. These are (a) whether the original result modeled the effect of blood lead on the cogni- tive outcomes using a logarithmic specification or not; (b) the average blood-lead level for the result sample, centered by the average of mean childhood BLLs for the 34 coun- tries estimated by Ericson et al. (2021); (c) an interaction between these two effects; (d) an indicator for whether BLL was measured contemporaneously to the outcome measurement, as the highest BLL measurement among several taken for an individ- ual, or prenatally, rather than as an average over several measurements over a lifetime or as a “lagged” measurement during an earlier potentially critical developmental pe- riod; (e) an indicator variable for studies from high-income countries; ( f) an indicator for results using IQ rather than reading or mathematics test scores as the outcome variable; (g) an indicator for results which fail to control for parental education; (h) an indicator for results which fail to control for family income or socioeconomic sta- tus;8 (i) the variance of the estimate, which may vary systematically with effect size in the presence of publication bias. All of these variables are coded such that a value of zero corresponds to our preferred specification: i.e., effects on reading or mathematics in a low- or middle-income country, in a context with average BLL, using a log-linear specification, where exposure is calculated as a lifetime average or from an earlier pe- riod than the time of outcome measurement, with controls for household income and parental education. Note that no single result fulfils all of these criteria. Meta-Analysis Results Study Characteristics From the total of 47 studies included in our meta-analysis, 17 are from the United States, 12 are from other high-income countries (Australia, Canada, Italy, New Zealand, South Korea, Taiwan, and UK), and 18 are from low- or middle-income countries (Brazil, China, Colombia, Ecuador, Egypt, India, Malaysia, Mexico, Nigeria, Peru, Pak- istan, and the Philippines). All except one study are observational (the exception Aizer et al. (2018) employs an instrumental variable design); however, many are longitudinal. Crawfurd et al. 237 Table 1. Result Characteristics Mean Median SD Age at blood test 4.70 4.00 3.00 Age at outcome 7.74 8.00 2.70 Mean blood lead 7.06 6.70 4.31 Sample Size 13,869.04 389.00 40,211.86 Binary controls Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Used logarithmic specification 0.24 0.00 0.43 Controlled for parent IQ or ed 0.69 1.00 0.46 Controlled for family income or wealth 0.61 1.00 0.49 Used average or lagged BLL measure 0.62 1.00 0.49 Outcome was mathematics or reading 0.40 0.00 0.49 From low/middle-income country 0.37 0.00 0.48 Observations 286 Note: This table shows descriptive statistics for the 286 estimates from 47 studies in our meta-analysis. The median sample size in our estimates is 389. The median age at blood-lead test- ing is four years old, and at cognitive testing is eight years old. The average blood-lead level is 7.06 μg/dL. The majority of results have controls for parent IQ or education and family background or income (table 1). Average Effects Overall, we find that a one log unit reduction in BLL is associated with a −0.12 standard-deviation improvement in test scores. Our preferred meta-regression spec- ification adjusts for the choice of original model specification, and control variables in each study, and is shown in table 2. To explore the influence of specific sets of controls systematically, we report a spec- ification curve à la Simonsohn et al. (2014). Figure 3 displays the average effect size from a meta-analytic regression controlling for each of all possible combinations of study characteristics listed above. All specifications yield a negative relationship be- tween (log) BLL and learning outcomes, though some are statistically insignificant. We also show our results graphically in the forest plots in figs 4 and 5. These fig- ures show the weighted average of effect sizes across studies for our two main out- comes, IQ and mathematics/reading test scores, respectively. The mean of the pooled effect is −0.22 for results with IQ as the outcome, −0.20 for studies with mathematics scores, and −0.24 for studies with reading scores.9 Unlike our meta-regression esti- mate, however, these averages are unadjusted for study characteristics and potential publication bias. 238 The World Bank Research Observer, vol. 40, no. 2 (2025) Table 2. Meta-regression of effect size on study characteristics (1) (2) (3) (4) ∗∗∗ ∗∗∗ ∗∗∗ Constant −0.227 −0.225 −0.154 −0.115 (0.025) (0.026) (0.021) (0.100) Potential confounds and sources of ref. ref. ref. bias Effect variance ( for pub. bias) – −0.139 −0.143 −0.137 Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 (0.488) (0.484) (0.462) No control for parent ed/IQ – – −0.173∗∗∗ −0.167∗∗ (0.058) (0.072) No control for family income – – 0.002 0.009 (0.061) (0.071) Harmonization of specification and ref. context Not logarithmic – – – 0.000 (0.057) Mean BLL – – – 0.000 (0.007) Not log spec × mean BLL – – – −0.008 (0.011) Exposure: Not average or lag – – – −0.036 (0.047) Outcome (IQ) – – – 0.015 (0.061) High-income country – – – −0.046 (0.063) N (estimates) 286 286 286 286 N (studies) 47 47 47 47 Note: We use the Hedges et al. (2010) estimator to account for dependence between multiple estimates from the same study. The dependent variable in each case is the standardized effect size of one log unit increase in lead exposure. The constant represents the average effect. Column 1 presents the unadjusted average. Column 2 includes the PET-PEESE (Stanley and Doucouliagos 2014) adjustment for publication bias. Column 3 includes controls for potentially confound- ing study-level characteristics—so the constant represents the average for studies that estimate a log-linear relation- ship, in a sample with average blood-lead levels, with controls for parental IQ or education, and with a lagged measure of blood lead. In column 4, we additionally include controls to adjust the constant to studies in low- or middle-income country settings, and focused on reading or mathematics test scores rather than IQ. Standard errors in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Results are robust to dropping each individual study sequentially ( fig. S1.2), and also to dropping studies in which we had to convert or impute different statistics (table S1.6). For studies in which we use the population rather than sample standard devia- tion in outcomes, effects are 0.04 standard deviations larger, but not statistically sig- nificantly different. Crawfurd et al. 239 Figure 3. Specification Curve for Meta-Analytic Regression of Estimates of the Relationship between (log) Blood-Lead Levels and Learning Outcomes Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: Each coefficient corresponds to the mean relationship between log blood lead level (BLL) and standardized learn- ing outcomes from a meta-analytic regression controlling for the study characteristics listed in the bottom panel. Heterogeneity Our main meta-analysis shows substantial heterogeneity between studies. The I 2 statistics (the share of variability in effect sizes not caused by sampling error) from figs 4 and 5 are all well above the 75 percent rule of thumb indicating “significant het- erogeneity” (Higgins et al. 2003). As we saw in table 2 however, just one covariate was statistically significantly cor- related with study effect size—whether that study controlled for parental education or IQ. The control for publication bias (the effect variance) also has a large coefficient, but is imprecisely estimated, and does not create a large shift in the overall mean ef- fect size. We see no statistically significant differences for effects on mathematics or reading compared with IQ, for high-income countries compared with middle-income countries, or between studies according to the timing of blood-lead measurement, i.e., concurrent versus lagged or averaged over measurements at multiple time points. We also show each of these controls individually in bivariate regressions in table S1.5. While controlling for most of these study characteristics individually only alters the overall effect of log BLL on learning outcomes modestly, the combined effect of controlling for all of them is quite dramatic. The overall estimate of the BLL–learning link falls from −0.23σ in column 1 with no controls to −0.13σ with various controls intended to adjust for potential confounding, publication bias, and differences in mea- surement (table 2). Finally, because our ultimate interest is in reading and mathematics scores in developing countries, in column 4 we also control for an indicator of studies 240 The World Bank Research Observer, vol. 40, no. 2 (2025) Figure 4. Effects of Lead on IQ Effect size Weight Study with 95% CI (%) Solon et al, 2008 -2.92 [ -13.10, 7.26] 0.00 Alvarez-Ortega et al, 2017 -2.37 [ -3.78, -0.96] 0.14 Kamel et al, 2003 -1.38 [ -2.20, -0.56] 0.40 Reuben et al, 2017 -1.24 [ -1.92, -0.57] 0.57 Nwobi et al, 2019 -0.66 [ -1.27, -0.05] 0.68 Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Chen et al, 2007 -0.55 [ -1.04, -0.07] 0.99 Huang et al, 2012 -0.55 [ -0.98, -0.12] 1.23 Canfield et al, 2003 -0.51 [ -0.83, -0.20] 1.87 Dantzer et al, 2020 -0.45 [ -0.64, -0.25] 3.11 Rasoul et al, 2012 -0.44 [ -0.70, -0.18] 2.33 Chiodo et al, 2007 -0.39 [ -0.63, -0.16] 2.61 Counter et al, 2005 -0.38 [ -0.57, -0.20] 3.22 Kim, Yu, and Lee, 2010 -0.38 [ -0.74, -0.03] 1.60 Min et al, 2009 -0.34 [ -0.55, -0.13] 2.88 Earl et al, 2016 -0.33 [ -0.58, -0.09] 2.51 Kim et al , 2009 -0.33 [ -0.81, 0.15] 1.02 Jusko et al, 2008 -0.33 [ -0.86, 0.20] 0.87 Fulton et al, 1987 -0.32 [ -0.52, -0.11] 2.98 Rahman et al, 2002 -0.29 [ -0.60, 0.03] 1.85 Roy et al, 2013 -0.28 [ -0.47, -0.08] 3.12 Cai et al, 2021 -0.27 [ -0.56, 0.02] 2.04 Lucchini et al, 2012 -0.23 [ -0.43, -0.04] 3.12 Menezes-Filho et al, 2018 -0.23 [ -0.37, -0.08] 3.82 Baghurst et al, 1992 -0.23 [ -0.51, 0.06] 2.15 Hong et al, 2015 -0.20 [ -0.35, -0.04] 3.63 Dietrich et al, 1993 -0.18 [ -0.44, 0.08] 2.38 Vega-Dienstmaier et al, 2006 -0.16 [ -0.32, -0.01] 3.60 Ruebner et al, 2019 -0.16 [ -0.31, -0.02] 3.75 Liu et al, 2013 -0.16 [ -0.44, 0.12] 2.17 Surkan et al, 2007 -0.16 [ -0.42, 0.10] 2.34 Schnaas et al, 2000 -0.15 [ -0.27, -0.02] 4.09 Braun et al, 2012 -0.10 [ -0.20, 0.00] 4.40 Schnaas et al, 2006 -0.09 [ -0.31, 0.14] 2.71 Bellinger et al, 1992 -0.09 [ -0.28, 0.11] 3.14 Lucchini et al, 2019 -0.07 [ -0.18, 0.04] 4.28 Desrochers-Couture et al, 2018 -0.06 [ -0.12, 0.00] 4.85 Pan et al, 2018 -0.05 [ -0.08, -0.02] 5.12 Zailina et al, 2011 -0.04 [ -0.07, -0.00] 5.08 Taylor et al, 2017 0.04 [ -0.16, 0.24] 3.05 Cooney et al, 1989 1.71 [ 0.71, 2.71] 0.28 Overall -0.22 [ -0.27, -0.17] Heterogeneity: 2 = 0.01, I2 = 77.14%, H2 = 4.38 Test of = 0: z = -7.99, p = 0.00 -1 -.5 0 .5 1 Random-effects REML model Sorted by: effect_sd_ln Note: This figure shows average effects for each study. In cases where more than one effect is reported per study, we show here the mean value. The estimated overall effect size (of −0.22 with no moderators) based on these mean effect sizes is only marginally different to the effect size estimated from all individual estimates, with a robust variance estimator to account for unknown dependence within studies. Crawfurd et al. 241 Figure 5. Effects of Lead on Reading and Mathematics Assessments Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: This figure shows average effects for each study. In cases where more than one effect is reported per study, we show here the mean value. The estimated overall effect size (of −0.20 for mathematics and −0.24 for reading with no moderators) based on these mean effect sizes is only slightly different to the effect size estimated from all individual estimates, with a robust variance estimator to account for dependence within studies. 242 The World Bank Research Observer, vol. 40, no. 2 (2025) in high-income countries and results using IQ as the outcome variable, yielding an es- timate of −0.12σ . Comparing our Results to Other Reviews How do our estimates compare to prior meta-analyses? Lanphear et al. (2005) estimate that an increase from 2.4 μg/dL to 30 μg/dL was associated with a 3.9 decrease in IQ, Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 equivalent to a 0.18 standard-deviation change per log unit increase in exposure. An earlier review found that a doubling of lead ( from 10 to 20 μg/dL blood lead or 5 to 10 μg/g tooth lead) was associated with a 1 to 2 point reduction in IQ. This is equivalent to around a 0.14σ reduction per log unit of lead (Pocock et al. 1994). Heidari et al. (2022) find that the standardized difference in mean IQ scores between high (> 10 μg/dL) and low exposure groups (< 10 μg/dL) is 0.541σ . Similarly, Galiciolli et al. (2022) find a difference between exposed and unexposed groups of 7.37 IQ points (equivalent to approximately 0.49σ ). Heng et al. (2022) provide a narrative review but not a meta- analytic average effect size. Our results for effects on IQ are therefore well within the range of other reviews, while we also extend our review to reading and mathematics test scores. Assessing the Role of Unobserved Confounders Since observational studies may suffer from omitted variable bias, what reason is there to believe the associations we report above reflect a causal relationship? One piece of evidence in support of a causal interpretation is that experimental studies con- ducted with animals show significant impacts of lead exposure on cognitive function (see for example Gilbert and Rice (1987) and Tena et al. (2019)). Furthermore, though the mechanisms by which lead exposure causes cognitive impairment are still not fully understood, lead has been shown to interfere with several processes involved in neu- rological development and functioning, adding credibility to a causal interpretation (Ramírez Ortega et al. 2021). While animal studies and the identification of neurological mechanisms lend cre- dence to the existence of some causal effect, they do not tell us much about the mag- nitude of this effect in humans, or whether observational associations are biased by unmeasured confounders. In this section we present two sources of evidence on po- tential bias which can help us pin down the magnitude of the true causal effect. First, we discuss studies focused on natural experiments in lead exposure. Second, we assess the sensitivity of observational estimates to selection on unobserved variables using the coefficient stability method (Oster 2019). First, a handful of natural experiments with a reasonable claim to causality have been published on the lead–cognition relationship, and these studies allow us to di- rectly compare causal with observational estimates. In table 3 we collect observational Crawfurd et al. 243 Table 3. IV Estimates Produce Larger Effects than OLS Estimates Lead Author Outcome OLS IV measure Instrument Aizer et al. (2018) Reading −0.026 −0.073 Blood (1 mcg/dl) Nearby home (0.002) (0.037) lead remediation Aizer et al. (2018) Mathematics −0.017 −0.030 Blood (1 mcg/dl) Nearby home (0.001) (0.034) lead remediation Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Grönqvist, Nilsson, GPA −0.100 −0.155 Moss (SD) Leaded petrol and Robling (2020) (.018) (.047) ban Clay, Portnykh, and Cognitive 0.0029 0.0414 Soil (ppm) 1944 Interstate Severnini (2019) difficulty (0.004) (0.019) Highway Plan Feigenbaum and Homicides 0.219 1.022 Water Distance to Muller (2016) (0.064) (0.257) pipes (0/1) lead refinery Note: For Aizer et al. (2018) and Grönqvist, Nilsson, and Robling (2020) we present effects on standardized test scores. For Aizer et al. (2018) we do this by dividing the coefficients from their table 8 by the test standard deviation of 13 (reported on their page 319). For Grönqvist, Nilsson, and Robling (2020) we divide the coefficients from their table 6 by the standard deviation of the grade point average (GPA) of 28, and multiply by the standard deviation (SD) of the moss lead measure of 17 (both reported in their table 2). For Clay, Portnykh, and Severnini (2019) the outcome is a binary indicator for children with cognitive difficulty. Results are taken from their table 3. For Feigenbaum and Muller (2016) estimates are from their table 2. ordinary least squares (OLS) and quasi-experimental instrumental variable (IV) esti- mates from five studies.10 In all cases, the IV estimate is larger than the OLS estimate, sometimes substantially so. Aizer et al. (2018) find results two to three times larger in the IV specification than the OLS one. They argue that the IV results are larger due to measurement error attenuating the OLS estimates. An alternative explanation for see- ing larger coefficients for IV estimates than OLS estimates is that they estimate treat- ment effects for slightly different populations. OLS estimates the average treatment effect across the entire population, whereas the IV estimates the local average treat- ment effect only among the subgroup of the population for whom the IV shifts their be- havior. We do not, however, have a compelling reason to believe that treatment effects would necessarily be different in the different sub-populations affected and unaffected by the instruments in these contexts. Aizer et al. (2018) report four estimates from lin- ear specifications (effects per unit change in blood lead), for two outcomes (reading and mathematics) and two different instrumental variable strategies. Applying the re- expression algorithm we describe previously, these results are equivalent to an average effect size of −0.13 standard deviations per natural log unit of blood lead, very close to our central meta-analytic estimate of −0.12 standard deviation. 244 The World Bank Research Observer, vol. 40, no. 2 (2025) Grönqvist et al. (2020) study the phase-out of leaded petrol in Sweden in 1980–81. They show that reductions in lead measured in nationwide moss samples is associated with improved high-school test scores (though no impact on another measure of cog- nitive skills). Their IV estimates for test scores imply that a one standard-deviation in- crease in moss lead levels is associated with a decrease in grade point average of 0.155 standard deviations, which is around 50 percent larger than their OLS estimates. Clay et al. (2019) use an instrumental variable strategy based on the layout of the Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 1944 Interstate Highway System Plan to estimate the relationship between county soil lead levels and cognitive difficulties among students. Counties with a highway recom- mended in the 1944 plan were 17 percent more likely to have above-median lead con- centrations in topsoil. In the IV specification, counties predicted to have above-median soil lead concentrations had 4 percentage points higher levels of children with cogni- tive difficulties (10 times larger than in the OLS specification). A similar pattern of larger effect sizes in quasi-experimental estimates than in ob- servational estimates has also been shown in studies looking at the effect of lead expo- sure on other outcomes besides test scores, including school suspensions and juvenile detention (Aizer and Currie 2019) and homicide rates (Feigenbaum and Muller 2016). Other quasi-experimental studies also show causal effects of lead exposure, though without providing a direct comparison of observational and causal estimates. Hollingsworth et al. (2022) focus on the switch to unleaded petrol in US Nascar mo- tor racing in 2007. They estimate that removing the exposure of a school one mile from a race track would be equivalent to increasing school spending per pupil by $750. Sorensen et al. (2019) evaluate lead hazard control programs, finding that each per- centage point reduction in lead poisoning in early childhood led to 0.04 standard devia- tions higher mathematics scores and 0.08 standard deviations reading scores. Rau et al. (2015) study the opening of a toxic waste dump in Chile in a difference-in-difference framework, showing that attending schools closer to the site after it opened lowered test scores. Higney et al. (2022) shows that water treatment in Scotland lead to im- provements in test scores only in areas with high prevalence of lead pipes. Second, we assess the robustness of observational estimates to selection on both observed covariates and proportional selection on unobserved covariates, based on the stability of estimated coefficients and model fit (R-squared). Specifically, we follow Oster (2019) in estimating the coefficient β that would result were we able to con- trol for unobserved confounders. This equation (1) requires that we make assumptions about the proportional degree of selection δ on unobservables (typically set at equal to selection on observables or 1), and about the maximum plausible value of the model fit Rmax (typically set at 1.3 times the model fit in the fully adjusted model R). Given these assumptions and our knowledge of the coefficients in the adjusted β ˜ and unadjusted models β ˚, and model fits of the adjusted R and unadjusted models R ˚, we can calculate Crawfurd et al. 245 Figure 6. Selection on Observed and Unobserved Variables Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: This figure shows the change in the coefficient in a regression of learning on blood lead before (“unadjusted”) and after adjusting for observed covariates (“adjusted”), and proportional selection on unobserved covariates (“Oster”). We show here estimates from all studies in our meta-analysis that report both an adjusted and unadjusted estimate. We are only able to report the Oster-adjusted estimates for the five studies that report the R-squared of their regressions. the coefficient β : ˚−β (β ˜ )(Rmax − R ) ˜−δ β =β . (1) (R − R˚) We identify 18 estimates ( from 16 studies) in which both an unadjusted and adjusted estimate is provided. Only six of these estimates ( from five studies) also report an R- squared which allows us to calculate the Oster-adjusted estimate. These estimates are reported in fig. 6. Among the six estimates for which we can estimate adjustments for both observable and unobservable confounders, the average reduction in the coeffi- cient adjusted for observables is 14 percent, whereas the average reduction for observ- ables and unobservables is 33 percent. Policy Simulations In order to interpret these effects, we consider two benchmarks: first, what the effect on learning would be of reducing current lead exposure in low- and middle-income 246 The World Bank Research Observer, vol. 40, no. 2 (2025) countries to the level in high-income countries, and second the effect of different in- terventions to reduce blood lead, and what this implies for learning. The Effect of Removing Lead Altogether What would be the effect of reducing current lead exposure in low- and middle-income countries to the level in high-income countries? A recent systematic review shows that Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 mean BLL levels (μbll ) in 34 low- and middle-income countries range from 1.7 to 9.3, with an unweighted average across countries of 5.3 μg/dL (Ericson et al. 2021). Levels for under-18-year-olds in the United States are around 0.5 μg/dL (US EPA 2015). To simulate the effect of reducing BLL levels in each country to the US level, we fol- low Ericson et al. (2021) and assume a log-normal distribution of BLL within each coun- try. Thus, for each country we calculate the mean of the log of BLL as μ2 μ = ln bll . bll + σbll μ2 2 Then, because our preferred specification for the relationship between BLL and cogni- tive scores is log-linear, we can calculate the effect on learning as Improvement in learning = β × (μ − 0.5), where β is the average effect size from the meta-analysis above, e.g., −0.12σ after co- variate adjustments. To measure gaps in learning outcomes around the world, we rely on the harmonized learning outcomes reported by the World Bank (Angrist et al. 2021), which are normed to have a mean of 500 and standard deviation of 100 points. Figure 7 shows the implied impact on the World Bank harmonized learning out- comes for each of the 34 countries in the Ericson et al. (2021) sample. Using the ad- justed estimates of lead’s effect on learning from the meta-analysis (−0.12) the incre- ments are meaningful, with a magnitude ranging from 4 to 31 points depending on the country. The results suggest a major role for lead exposure in explaining learning gaps between rich and poor countries. As an example, the lowest scoring country, the Demo- cratic Republic of Congo, lies 250 points below the global mean of 500, and reducing BLL to 0.5 μg/dL would improve scores by 29 points. On average for these 34 countries, reducing BLL to US levels improves learning by 23 points, equivalent to 21 percent of the 110 point learning gap to the global benchmark of 500 points on the World Bank scale. The Effect of Interventions to Reduce Blood Lead Reducing lead exposure altogether down to US levels may be unrealistic, at least in the short term. What kind of reductions are feasible? The most important sources of ex- Crawfurd et al. 247 Figure 7. Simulated Effect of Eliminating Blood Lead on National Learning Outcomes Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: For each country we simulate the effect of reducing blood-lead levels (BLLs) to 0.5 μg/dL, based on the distribu- tions of BLL reported for 34 countries by Ericson et al. (2021), average learning outcomes reported by the World Bank, and the coefficient from our meta-regression of a −0.12σ in learning per natural logarithm increase in BLL. 248 The World Bank Research Observer, vol. 40, no. 2 (2025) posure to lead in low- and middle-income countries are poorly understood. A recent review notes that “geographic variability and overall distribution of lead in these poten- tial exposure sources have not been adequately characterized, particularly in low- and middle-income countries (LMICs)” (Sargsyan et al. 2024). The most commonly cited sources include informal industrial activity such as battery recycling, and contami- nated consumer products such as cookware and food. Here we review studies on four classes of intervention: regulatory, educational, medical, and targeted environmental Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 interventions. Regulation The most important regulatory intervention has been the banning of leaded gasoline or petrol, which has shown that large reductions in blood lead are feasible at low cost in both developed and developing countries (Angrand et al. 2022). Before-after stud- ies around leaded petrol bans have shown large reductions in blood lead in both high- and low-income countries, including in India (Singh and Singh 2006), Kenya (United Nations Environment Programme (UNEP) 2014), Pakistan (Manser et al. 1990; Rahbar et al. 2002), and South Africa (Mathee et al. 2006). While leaded gasoline has now been banned in every country in the world, there may still be considerable room for in- creased regulatory action in other areas. Other regulatory actions taken by the majority of OECD countries include regulations on lead in water, air, food, batteries, food con- tainers, and paint, with several countries having additional regulations on lead in dust, soil, sewage, waste, and pesticides (Silbergeld 1997). The majority of low- and middle- income countries do not have regulations on the sale of lead in paint (UNEP 2020). The evidence base though remains thin—a systematic review into exposure via consumer products found zero studies on either regulatory, educational, or environmental inter- ventions, and no studies on regulatory interventions targeting exposure via drinking water (Pfadenhauer et al. 2016). Recent regulatory action on removing lead chromate added to spices in Bangladesh and Georgia seems to have been effective at very low cost (Forsyth et al. 2023). Targeted Remediation Second, some environmental interventions targeted at high-risk populations have shown large reductions, though few efforts in developing countries have been well doc- umented. One search for remediation efforts in developing countries (whether evalu- ated or not) found just 13 projects in total (O’Brien et al. 2021). Another systematic review of studies on remediation of lead-contaminated soil found just five studies, all in North America, and with mixed results (Dobrescu et al. 2022). In table 4 we summa- rize results from three studies: Nigeria, Dominican Republic, and Bangladesh. All were areas of acute exposure and remediation was expensive, but did lead to large falls in blood lead. For example, a huge reduction was documented from efforts in Zamfara, Crawfurd et al. 249 Table 4. The implied effect of feasible blood-lead reductions on learning. High Low Effect on Study Change in blood lead BLL BLL learning (SD) design One log unit ( from 0.12 meta-analysis) Reduction from LMIC to HIC 5.3 0.5 0.28 levels Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Observed reduction in United 14.6 2.8 0.20 Before-after States (1976 to 1990) 1. Regulatory intervention Leaded petrol ban India 18.1 12.1 0.05 Before-after Kenya 8 5.6 0.04 Before-after Pakistan 38 15.6 0.11 Before-after South Africa 16 6.4 0.11 Before-after 2. Targeted remediation Soil remediation Zamfara, Nigeria 149 15 0.28 Before-after Dominican Republic 20.6 5.34 0.16 Before-after Bangladesh 22.6 14.8 0.05 Before-after Paint remediation United States (NY) 24.3 12.3 0.08 Diff-in-diff United States (NC) 17.85 9 0.08 Diff-in-diff United States (Missouri) 34 29.58 0.02 Diff-in-diff 3. Education (parental) Georgia 9.6 6.8 0.04 Before-after China 10.1 7.9 0.03 RCT 4. Medical (calcium supplements) Indonesia (Medan) 2.1 0.01 0.64 RCT Indonesia (Bandung) 13.7 4.95 0.12 RCT Nigeria 9.9 8.8 0.01 Case-control Mexico 4.1 3.649 0.01 RCT United States 21.4 21.7 0.00 RCT Note: The high and low blood-lead levels (BLLs) shown are the average arithmetic mean for low- and middle-income countries or LMICs (Ericson et al. 2021), high-income countries or HICs (US EPA 2015), and the study in China (Shen et al. 2004), geometric means for the studies in Nigeria (Tirima et al. 2016), Dominican Republic (Ericson et al. 2018), and Bangladesh (Chowdhury et al. 2021), and medians for Georgia (Ruadze et al. 2021). Geometric means are generally smaller than arithmetic means. The total effect on learning is calculated as the product of the log unit gap, and the estimated coefficient of the effect of one log unit on learning outcomes, taken from our meta-analysis. 250 The World Bank Research Observer, vol. 40, no. 2 (2025) Nigeria, from 149 to 15 (Tirima et al. 2016), implying improvement in learning by 0.28 standard deviations. A soil remediation effort at a former lead smelter in Haina, the Dominican Republic, reduced lead levels from 20.6 μg/dL to 5.34 μg/dL, or 0.16 stan- dard deviations in test scores (Ericson et al. 2018). Both of these large gains came from addressing acutely polluted sites. In Bangladesh, soil capping, household cleaning, and awareness raising reduced lead levels from 22.6 μg/dL to 14.8 μg/dL (Chowdhury et al. 2021). All of these studies rely on before-after designs and so may be biased estimates. Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Several studies from the United States show results of targeted paint remediation, demonstrating modest positive effects on blood-lead levels (Billings and Schnepel 2018; Leighton et al. 2003; Staes et al. 1994). Removing lead pipes has also shown promise (Pfadenhauer et al. 2016). Education Parental education efforts in Georgia and China, where exposure is more diffuse and chronic, have found modest reductions in blood lead. Following striking results from a nationally representative blood-lead testing in Georgia, a government program of ac- tion was implemented to educate parents on the issue. This involved initial BLL test- ing, followed by letters sent to all families of children with elevated BLLs, with advice on how to reduce lead exposure and on dietary habits that can help reduce BLLs (in- creased calcium, iron, and vitamin C). Parents were advised to visit a pediatrician to assess physical and mental development and iron deficiency, and pediatricians were provided with training in early detection and management of lead exposure. This re- duced BLLs from 9.6 to 6.8 μg/dL (Ruadze et al. 2021), equivalent to a 0.04 standard- deviation improvement in test scores. A parental education intervention in China re- duced BLL from 10 to 8μg/dL (Shen et al. 2004). A recent study from Bangladesh has shown promising impacts on behavior change, but does not yet report results on blood lead (Jahir et al. 2021). Medical Intervention Finally, evidence on the provision of calcium supplements has shown mixed results, from no effect in the United States (Markowitz et al. 2004) and Mexico (Ettinger et al. 2009), to small effects in Nigeria (Keating et al. 2011) and large effects in Indonesia (Syofyan et al. 2020; Haryanto et al. 2015). Calcium supplementation may therefore hold promise in contexts in which calcium deficiency is of greater concern. Conclusion Can widespread lead exposure explain low average learning outcomes in the devel- oping world? While data coverage is limited in both cases, (a) mean lead exposure is over 5 μg/dL among children in low- and middle-income countries with reasonably Crawfurd et al. 251 representative samples, or roughly 10 times higher than the United States, and (b) learning levels among primary-school-aged pupils lag more than one full standard de- viation behind OECD levels in the same set of countries. We extend existing meta-analyses of studies linking lead exposure to cognitive out- comes, expanding the traditional focus on IQ to include measures of reading and math- ematics performance among primary-school students. Taken at face value, the associ- ation between lead and learning outcomes in individual-level data across 47 studies Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 suggests that one log unit higher blood-lead levels reduces learning levels by roughly −0.23 standard deviations. In a simple model of global learning gaps, this effect size is sufficient to suggest that observed lead levels explain over half of the gap in learning outcomes between developing and developed countries. Raw correlations likely overstate the true causal impact of lead on learning levels, for at least two reasons. First, we find evidence of publication bias in estimates of the lead-learning link. Funnel plots reveal striking asymmetry in the distribution of find- ings, and clustering of p-values just below conventional significance levels. Second, it is impossible to rule out a large role for unobserved confounding in the lead-learning link. We present various pieces of evidence suggesting the true causal link is much smaller than the published literature. Meta-analytic regressions controlling for these and other potential sources of bias suggest that 1 log unit in blood-lead levels reduces learning levels by roughly −0.12 standard deviations, approximately one-half of the naive estimate. Nevertheless, a true causal effect of this magnitude may remain a viable lever for policy action based only on education outcomes – in addition to myriad other health benefits beyond the scope of this paper. Notably, even with our most conservative estimates, the magnitude of learning gains achievable through lead eradication is comparable to many popular policy initiatives to improve education quality in the developing world (e.g., Evans and Yuan (2022) show that the average effect size on learning from RCTs in global education is 0.1 standard deviations). In cost-benefit terms, a focus on lead exposure is likely justified solely on education grounds if countries are able to achieve significant reductions in lead expo- sure through low-cost, large-scale policy reforms such as improved regulation of the lead paint and lead battery industries. Notes Lee Crawfurd (corresponding author) email: lcrawfurd@cgdev.org ; Rory Todd (email: rtodd@cgdev.org); Susannah Hares (email: shares@cgdev.org); Justin Sandefur (email: jsandefur@cgdev.org); Rachel Silver- man Bonnifield (email: rsilverman@cgdev.org), Center for Global Development, Great Peter House, Abbey Gardens, Great College St, London SW1P 3SE, United Kingdom We are grateful to Bruce Lanphear and David Evans for helpful comments, and to Aisha Ali, Zack Gehan, and Maimouna Konate for research assistance. This work was supported by Givewell and the Effective Altruism Global Health and Development Fund. 252 The World Bank Research Observer, vol. 40, no. 2 (2025) 1. These figures are derived from estimates produced by IHME/UNICEF (Rees and Fuller 2020), who define elevated blood-lead levels as exceeding the US CDC reference value of 5 micrograms per deciliter (the CDC lowered its reference value to 3.5 micrograms in 2021). We combine these exposure rates with population estimates from the United Nations (United Nations, Department of Economic and Social Affairs, Population Division 2019). 2. Full-scale IQ scores are the average of scores on five distinct abilities: verbal comprehension, visual spatial, fluid reasoning, working memory, and processing speed. 3. Studies report averages as arithmetic means, geometric mean, and medians. We convert medi- ans to arithmetic means using the method in Wan et al. (2014). Studies which report both arithmetic Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 means and geometric means show the former are generally slightly larger as a result of the skewed dis- tribution of blood-lead levels. However, differences are small and so we treat both types as equivalent in analysis. 4. In fig. 2 we collect published dose-response curves from 13 different studies. These include Lanphear et al. (2005) who combine data from 7 different longitudinal studies that track a total of 1,333 chil- dren followed from birth or infancy until 5–10 years of age. Rothenberg and Rothenberg (2005) di- rectly compare a linear and log-linear specification, finding a substantially better fit with the log-linear specification. 5. Where results are from models with a wider range of covariates which excludes observations with- out selected variables, the mean blood-lead level and outcome standard deviation may be slightly differ- ent to the values for the sample; however, these differences are likely to be negligible. For studies which express effects as the difference between two groups defined by a range of blood-lead levels, we require mean blood-lead levels for the two groups. Where this is not given, we impute these values where possi- ble. Where the range of a group is only 1 μg/dL, we take the midpoint; for example, for a group with blood lead levels 4 to 5 μg/dL, we impute a mean of 4.5 μg/dL (Shadbegian et al. 2019; Edwards et al. 2013). For 7 studies, we simulate log-normal distributions using the overall mean and standard deviation given for the study sample, and then approximate group means given this simulated distribution (Liu et al. 2013; Surkan et al. 2007). For another we instead simulate a normal distribution as this better fits the reported moments of the data (Rasoul et al. 2012). 6. While this algorithm performs best using the median of the exposure distribution, we find it still performs well using arithmetic means, the most commonly reported average in our studies. 7. We use the implementation of this method by Hong and Reed (2021), specifically the ‘AK1’ estimator which assumes equal probability of publication regardless of the sign of the effect. 8. We define this kind of control broadly: for example, we code results which adjust for the Home Observation for Measurement of the Environment, which measures the physical and social environment of children, as controlling for this. 9. Note that here we show results collapsed to their mean for each study where there are multiple estimates from a study. These results are similar to those obtained using robust variance estimation ac- counting for unknown correlation between estimates within studies (Hedges et al. 2010). 10. We do not include all of these studies in our meta-analysis as they do not all have direct blood-lead measures. We do include the one study that does have blood-lead measures, which uses a lead remediation program in Rhode Island as a source of exogenous variation, showing that instrumental variable estimates are larger than observational estimates (Aizer et al. 2018). References Aizer, A., and J. Currie. 2019. “Lead and Juvenile Delinquency: New Evidence from Linked Birth, School, and Juvenile Detention Records.” Review of Economics and Statistics 101(4): 575–87. Aizer, A., J. Currie, P. Simon, and P. Vivier. 2018. “Do Low Levels of Blood Lead Reduce Children’s Future Test Scores?” American Economic Journal: Applied Economics 10(1): 307–41. Crawfurd et al. 253 Alvarez-Ortega, N., K. Caballero-Gallardo, and J. Olivero-Verbel. 2017. “Low Blood Lead Levels Impair In- tellectual and Hematological Function in Children from Cartagena, Caribbean Coast of Colombia.” Journal of Trace Elements in Medicine and Biology: Organ of the Society for Minerals and Trace Elements (GMS) 44: 233–40. Angrand, R. C., G. Collins, P. J. Landrigan, and V. M. Thomas. 2022. “Relation of Blood Lead Levels and Lead in Gasoline: An Updated Systematic Review.” Environmental Health 21(1): 138. Angrist, N., S. Djankov, P. K. Goldberg, and H. A. Patrinos. 2021. “Measuring Human Capital Using Global Learning Data.” Nature 592(7854): 403–8. Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Azevedo, J. P., H. Rogers, E. Ahlgren, M. Akmal, M.-H. Cloutier, E. Ding, and A. Raza et al. 2022. “The State of Global Learning Poverty: 2022 Update.” Technical report, World Bank, UNESCO, UNICEF, FCDO, USAID, Bill & Melinda Gates Foundation. Baghurst, P. A., A. J. McMichael, N. R. Wigg, G. V. Vimpani, E. F. Robertson, R. J. Roberts, and S.-L. Tong. 1992. “Environmental Exposure to Lead and Children’s Intelligence at the Age of Seven Years.” New England Journal of Medicine 327(18): 1279–84. Bellinger, D. C., K. M. Stiles, and H. L. Needleman. 1992. “Low-Level Lead Exposure, Intelligence and Aca- demic Achievement: A Long-Term Follow-Up Study.” Pediatrics 90(6): 855–61. Billings, S. B., and K. T. Schnepel. 2018. “Life after Lead: Effects of Early Interventions for Children Exposed to Lead.” American Economic Journal: Applied Economics 10(3): 315–44. Blackowicz, M. J., D. O. Hryhorczuk, K. M. Rankin, D. A. Lewis, D. Haider, B. P. Lanphear, and A. Evens. 2016. “The Impact of Low-Level Lead Toxicity on School Performance among Hispanic Subgroups in the Chicago Public Schools.” International Journal of Environmental Research and Public Health 13(8): 774. Braun, J. M., E. Hoffman, J. Schwartz, B. Sanchez, L. Schnaas, A. Mercado-Garcia, and M. Solano-Gonzalez et al. 2012. “Assessing Windows of Susceptibility to Lead-Induced Cognitive Deficits in Mexican Chil- dren.” Neurotoxicology 33(5): 1040–7. Cai, Q.-L., D.-J. Peng, Lin-Zhao, J.-w. Chen, Yong-Li, H.-L. Luo, S.-y. Ou, M.-L. Huang, and Y.-M. Jiang. 2021. “Impact of Lead Exposure on Thyroid Status and IQ Performance among School-Age Children Living nearby a Lead-Zinc Mine in China.” Neurotoxicology 82: 177–85. Canfield, R. L., C. R. Henderson, D. A. Cory-Slechta, C. Cox, T. A. Jusko, and B. P. Lanphear. 2003. “Intellectual Impairment in Children with Blood Lead Concentrations below 10 g per Deciliter.” New England Journal of Medicine 348(16): 1517–26. Chen, A., B. Cai, K. N. Dietrich, J. Radcliffe, and W. J. Rogan. 2007. “Lead Exposure, IQ, and Behavior in Urban 5–7 Year Olds: Does Lead Affect Behavior Only by Lowering IQ?” Pediatrics 119(3): e650–8. Chiodo, L. M., C. Covington, R. J. Sokol, J. H. Hannigan, J. Jannise, J. Ager, M. Greenwald, and V. Delaney- Black. 2007. “Blood Lead Levels and Specific Attention Effects in Young Children.” Neurotoxicology and Teratology 29(5): 538–46. Chowdhury, A. K. I., S. Nurunnahar, M. L. Kabir, M. T. Islam, M. Baker, M. S. Islam, and M. Rahman et al. 2021. “Child Lead Exposure near Abandoned Lead Acid Battery Recycling Sites in a Residential Com- munity in Bangladesh: Risk Factors and the Impact of Soil Remediation on Blood Lead Levels.” Envi- ronmental Research 194: 110689. Clay, K., M. Portnykh, and E. Severnini. 2019. “The Legacy Lead Deposition in Soils and Its Impact on Cog- nitive Function in Preschool-Aged Children in the United States.” Economics & Human Biology 33: 181– 192. Cooney, G. H., A. Bell, W. McBride, and C. Carter. 1989. “Low-Level Exposures to Lead: The Sydney Lead Study.” Developmental Medicine & Child Neurology 31(5): 640–9. Copas, J. B., and J. Q. Shi. 2001. “A Sensitivity Analysis for Publication Bias in Systematic Reviews.” Statistical Methods in Medical Research 10(4): 251–65. 254 The World Bank Research Observer, vol. 40, no. 2 (2025) Counter, S. A., L. H. Buchanan, and F. Ortega. 2005. “Neurocognitive Impairment in Lead-Exposed Children of Andean Lead-Glazing Workers.” Journal of Occupational and Environmental Medicine 47(3): 306–12. Dantzer, J., P. Ryan, K. Yolton, P. J. Parsons, C. D. Palmer, K. Cecil, and J. M. Unrine. 2020. “A Comparison of Blood and Toenails as Biomarkers of Children’s Exposure to Lead and Their Correlation with Cognitive Function.” Science of the Total Environment 700: 134519. Desrochers-Couture, M., Y. Oulhote, T. E. Arbuckle, W. D. Fraser, J. R. Séguin, E. Ouellet, and N. Forget- Dubois et al. 2018. “Prenatal, Concurrent, and Sex-Specific Associations between Blood Lead Concen- trations and IQ in Preschool Canadian Children.” Environment International 121(Pt 2): 1235–42. Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Dietrich, K. N., O. G. Berger, P. A. Succop, P. B. Hammond, and R. L. Bornschein. 1993. “The Developmental Consequences of Low to Moderate Prenatal and Postnatal Lead Exposure: Intellectual Attainment in the Cincinnati Lead Study Cohort following School Entry.” Neurotoxicology and Teratology 15(1): 37–44. Dobrescu, A.-I., A. Ebenberger, J. Harlfinger, U. Griebler, I. Klerings, B. Nußbaumer-Streit, A. Chapman, L. Affengruber, and G. Gartlehner. 2022. “Effectiveness of Interventions for the Remediation of Lead- Contaminated Soil to Prevent or Reduce Lead Exposure—A Systematic Review.” Science of the Total Environment 806: 150480. Duval, S., and R. Tweedie. 2000. “Trim and Fill: A Simple Funnel-Plot-Based Method of Testing and Adjust- ing for Publication Bias in Meta-Analysis.” Biometrics 56(2): 455–63. Dzierlenga, M. W., L. Crawford, and M. P. Longnecker. 2020. “Birth Weight and Perfluorooctane Sulfonic Acid: A Random-Effects Meta-Regression Analysis.” Environmental Epidemiology 4(3): e095. Earl, R., N. Burns, T. Nettelbeck, and P. Baghurst. 2016. “Low-Level Environmental Lead Exposure Still Negatively Associated with Children’s Cognitive Abilities.” Australian Journal of Psychology 68(2): 98–106. Edwards, S., R. Anthopolos, and M. L. Miranda. 2013. “The Impact of Early Childhood Lead Exposure on Educational Test Performance among Connecticut Schoolchildren: Phase II Report.” Technical report, Children’s Environmental Health Initiative, School of Natural Resources and Environment, University of Michigan, Ann Arbor, MI. Egger, M., G. Davey Smith, M. Schneider, and C. Minder. 1997. “Bias in Meta-Analysis Detected by a Simple, Graphical Test.” BMJ (Clinical Research Ed.) 315(7109): 629–34. Ericson, B., J. Caravanos, C. Depratt, C. Santos, M. G. Cabral, R. Fuller, and M. P. Taylor. 2018. “Cost Ef- fectiveness of Environmental Lead Risk Mitigation in Low- and Middle-Income Countries.” GeoHealth 2(2): 87–101. Ericson, B., H. Hu, E. Nash, G. Ferraro, J. Sinitsky, and M. P. Taylor. 2021. “Blood Lead Levels in Low-Income and Middle-Income Countries: A Systematic Review.” Lancet Planetary Health 5(3): e145–53. Ettinger, A. S., H. Lamadrid-Figueroa, M. M. Téllez-Rojo, A. Mercado-García, K. E. Peterson, J. Schwartz, H. Hu, and M. Hernández-Avila. 2009. “Effect of Calcium Supplementation on Blood Lead Levels in Pregnancy: A Randomized Placebo-Controlled Trial.” Environmental Health Perspectives 117(1): 26–31. Evans, D. K., and F. Yuan. 2022. “How Big Are Effect Sizes in International Education Studies?.” Educational Evaluation and Policy Analysis 44(3): 532–40. Evens, A., D. Hryhorczuk, B. P. Lanphear, K. M. Rankin, D. A. Lewis, L. Forst, and D. Rosenberg. 2015. “The Impact of Low-Level Lead Toxicity on School Performance among Children in the Chicago Public Schools: A Population-Based Retrospective Cohort Study.” Environmental Health 14(1): 21. Feigenbaum, J. J., and C. Muller. 2016. “Lead Exposure and Violent Crime in the Early Twentieth Century.” Explorations in Economic History 62: 51–86. Fisher, Z., and E. Tipton. 2015. “Robumeta: An R-Package for Robust Variance Estimation in Meta- Analysis.” arXiv:1503.02220 [stat]. Crawfurd et al. 255 Forsyth, J. E., M. Baker, S. Nurunnahar, S. Islam, M. S. Islam, T. Islam, and E. Plambeck et al. 2023. “Food Safety Policy Enforcement and Associated Actions Reduce Lead Chromate Adulteration in Turmeric across Bangladesh.” Environmental Research 232: 116328. Fulton, M., G. Thomson, R. Hunter, G. Raab, D. Laxen, and W. Hepburn. 1987. “Influence of Blood Lead on the Ability and Attainment of Children in Edinburgh.” Lancet 329(8544): 1221–6. Galiciolli, M. E. A., L. S. Lima, N. d. S. da Costa, D. P. de Andrade, A. C. Irioda, and C. S. Oliveira. 2022. “IQ Alteration Induced by Lead in Developed and Underdeveloped/Developing Countries: A Systematic Review and a Meta-Analysis.” Environmental Pollution (Barking, Essex: 1987) 292(Pt A): 118316. Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Gilbert, S. G., and D. C. Rice. 1987. “Low-Level Lifetime Lead Exposure Produces Behavioral Toxicity (Spa- tial Discrimination Reversal) in Adult Monkeys.” Toxicology and Applied Pharmacology 91(3): 484–90. Grönqvist, H., J. P. Nilsson, and P.-O. Robling. 2020. “Understanding How Low Levels of Early Lead Exposure Affect Children’s Life Trajectories.” Journal of Political Economy 128(9): 3376–433. Haryanto, B., B. Sutrisna, and I. M. Djaja. 2015. “Effect of Calcium Supplementation on School Children’s Blood Lead Levels in Indonesia.” International Journal of Science and Research 2015(1): 5005. Hedges, L. V., E. Tipton, and M. C. Johnson. 2010. “Robust Variance Estimation in Meta-Regression with Dependent Effect Size Estimates.” Research Synthesis Methods 1(1): 39–65. Heidari, S., S. Mostafaei, N. Razazian, M. Rajati, A. Saeedi, and F. Rajati. 2022. “The Effect of Lead Exposure on IQ Test Scores in Children under 12 Years: A Systematic Review and Meta-Analysis of Case-Control Studies.” Systematic Reviews 11(1): 106. Heng, Y. Y., I. Asad, B. Coleman, L. Menard, S. Benki-Nugent, F. H. Were, C. J. Karr, and M. S. McHenry. 2022. “Heavy Metals and Neurodevelopment of Children in Low and Middle-Income Countries: A Systematic Review.” PLoS ONE 17(3): e0265536. Higgins, J. P. T., and S. Green (Eds.). 2008. Cochrane Handbook for Systematic Reviews of Interventions(1st edition). Chichester, England; Hoboken, NJ: Wiley–Blackwell. Higgins, J. P. T., S. G. Thompson, J. J. Deeks, and D. G. Altman. 2003. “Measuring Inconsistency in Meta- Analyses.” BMJ 327(7414): 557–60. Higney, A., N. Hanley, and M. Moro. 2022. “The Impact of Lead Pollution on Human Capital Formation: Size of Dose Matters.” Technical report. Hollingsworth, A., M. Huang, I. Rudik, and N. Sanders. 2022. “A Thousand Cuts: Cumulative Lead Exposure Reduces Academic Achievement.” Publisher: OSF. Hong, S., and W. R. Reed. 2021. “Using Monte Carlo Experiments to Select Meta-Analytic Estimators.” Re- search Synthesis Methods 12(2): 192–215. Hong, S.-B., M.-H. Im, J.-W. Kim, E.-J. Park, M.-S. Shin, B.-N. Kim, and H.-J. Yoo et al. 2015. “Environ- mental Lead Exposure and Attention Deficit/Hyperactivity Disorder Symptom Domains in a Com- munity Sample of South Korean School-Age Children.” Environmental Health Perspectives 123(3): 271–76. Huang, P.-C., P.-H. Su, H.-Y. Chen, H.-B. Huang, J.-L. Tsai, H.-I. Huang, and S.-L. Wang. 2012. “Childhood Blood Lead Levels and Intellectual Development after Ban of Leaded Gasoline in Taiwan: A 9-Year Prospective Study.” Environment International 40: 88–96. Jahir, T., H. O. Pitchik, M. Rahman, J. Sultana, A. K. M. Shoab, T. M. Nurul Huda, and K. A. Byrd et al. 2021. “Making the Invisible Visible: Developing and Evaluating an Intervention to Raise Awareness and Re- duce Lead Exposure among Children and Their Caregivers in Rural Bangladesh.” Environmental Re- search 199: 111292. Jusko, T. A., C. R. Henderson, B. P. Lanphear, D. A. Cory-Slechta, P. J. Parsons, and R. L. Canfield. 2008. “Blood Lead Concentrations < 10 g/dL and Child Intelligence at 6 Years of Age.” Environmental Health Perspectives 116(2): 243–8. 256 The World Bank Research Observer, vol. 40, no. 2 (2025) Kamel, N. M., A. M. Ramadan, M. I. Kamel, Y. A. E. G. Mostafa, R. M. Abo el Naga, and A. M. Ali. 2003. “Impact of Lead Exposure on Health Status and Scholastic Achievement of School Pupils in Alexandria.” Journal of the Egyptian Public Health Association 78(1-2): 1–28. Keating, E. M., P. R. Fischer, J. M. Pettifor, M. Pfitzner, C. O. Isichei, and T. D. Thacher. 2011. “The Effect of Calcium Supplementation on Blood Lead Levels in Nigerian Children.” Journal of Pediatrics 159(5): 845–850. Kim, D.-S., S.-D. Yu, and E.-H. Lee. 2010. “Effects of Blood Lead Concentration on Intelligence and Person- ality in School Children.” Molecular & Cellular Toxicology 6(1): 19–23. Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Kim, Y., B.-N. Kim, Y.-C. Hong, M.-S. Shin, H.-J. Yoo, J.-W. Kim, S.-Y. Bhang, and S.-C. Cho. 2009. “Co-exposure to Environmental Lead and Manganese Affects the Intelligence of School-Aged Children.” Neurotoxi- cology 30(4): 564–71. Kordas, K., R. L. Canfield, P. López, J. L. Rosado, G. G. Vargas, M. E. Cebrián, J. A. Rico, D. Ronquillo, and R. J. Stoltzfus. 2006. “Deficits in Cognitive Function and Achievement in Mexican First-Graders with Low Blood Lead Concentrations.” Environmental Research 100(3): 371–86. Lanphear, B. P., R. Hornung, J. Khoury, K. Yolton, P. Baghurst, D. C. Bellinger, and R. L. Canfield et al. 2005. “Low-Level Environmental Lead Exposure and Children’s Intellectual Function: An International Pooled Analysis.” Environmental Health Perspectives 113(7): 894–9. Leighton, J., S. Klitzman, S. Sedlar, T. Matte, and N. L. Cohen. 2003. “The Effect of Lead-Based Paint Haz- ard Remediation on Blood Lead Levels of Lead Poisoned Children in New York City.” Environmental Research 92(3): 182–90. Linakis, M. W., C. V. Landingham, A. Gasparini, and M. P. Longnecker. 2024. “Re-expressing Coefficients from Regression Models for Inclusion in a Meta-Analysis.” BMC Medical Research Methodology 24: ar- ticle no. 6. Liu, J., L. Li, Y. Wang, C. Yan, and X. Liu. 2013. “Impact of Low Blood Lead Concentrations on IQ and School Performance in Chinese Children.” PLoS ONE 8(5): e65230. Lucchini, R. G., S. Guazzetti, S. Renzetti, M. Conversano, G. Cagna, C. Fedrighi, and A. Giorgino et al. 2019. “Neurocognitive Impact of Metal Exposure and Social Stressors among Schoolchildren in Taranto, Italy.” Environmental Health: A Global Access Science Source 18(1): 67. Lucchini, R. G., S. Zoni, S. Guazzetti, E. Bontempi, S. Micheletti, K. Broberg, G. Parrinello, and D. R. Smith. 2012. “Inverse Association of Intellectual Function with Very Low Blood Lead but Not with Manganese Exposure in Italian Adolescents.” Environmental Research 118: 65–71. Manser, W. W., R. Lalani, S. Haider, and M. A. Khan. 1990. “Trace Element Studies on Karachi Populations. Part V: Blood Lead Levels in Normal Healthy Adults and Grammar School Children.” JPMA The Journal of the Pakistan Medical Association 40(7): 150–4. Markowitz, M. E., M. Sinnett, and J. F. Rosen. 2004. “A Randomized Trial of Calcium Supplementation for Childhood Lead Poisoning.” Pediatrics 113(1 Pt 1): e34–9. Mathee, A., H. Röllin, Y. von Schirnding, J. Levin, and I. Naik. 2006. “Reductions in Blood Lead Levels among School Children following the Introduction of Unleaded Petrol in South Africa.” Environmental Re- search 100(3): 319–22. McLaine, P., A. Navas-Acien, R. Lee, P. Simon, M. Diener-West, and J. Agnew. 2013. “Elevated Blood Lead Levels and Reading Readiness at the Start of Kindergarten.” Pediatrics 131(6): 1081–9. Menezes-Filho, J. A., C. F. Carvalho, J. L. G. Rodrigues, C. F. S Araújo, N. R. dos Santos, C. S. Lima, and M. J. Bandeira et al. 2018. “Environmental Co-exposure to Lead and Manganese and Intellectual Deficit in School-Aged Children.” International Journal of Environmental Research and Public Health 15(11): 2418. Min, M. O., L. T. Singer, H. L. Kirchner, S. Minnes, E. Short, Z. Hussain, and S. Nelson. 2009. “Cognitive Devel- opment and Low-Level Lead Exposure in Poly-Drug Exposed Children.” Neurotoxicology and Teratology 31(4): 225–31. Crawfurd et al. 257 Nwobi, N. L., S. K. Adedapo, O. Olukolade, O. A. Oyinlade, I. A. Lagunju, N. O. Atulomah, I. A. Nwazuoke, and J. I. Anetor. 2019. “Positive and Inverse Correlation of Blood Lead Level with Erythrocyte Acetylcholinesterase and Intelligence Quotient in Children: Implications for Neurotoxicity.” Interdis- ciplinary Toxicology 12(3): 136–42. O’Brien, R. M., T. J. Phelan, N. M. Smith, and K. M. Smits. 2021. “Remediation in Developing Countries: A Review of Previously Implemented Projects and Analysis of Stakeholder Participation Efforts.” Critical Reviews in Environmental Science and Technology 51(12): 1259–80. Oster, E. 2019. “Unobservable Selection and Coefficient Stability: Theory and Validation”. Journal of Busi- ness & Economic Statistics 37(2), 187–204. Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Pan, S., L. Lin, F. Zeng, J. Zhang, G. Dong, B. Yang, and Y. Jing et al. 2018. “Effects of Lead, Cadmium, Arsenic, and Mercury Co-Exposure on Children’s Intelligence Quotient in an Industrialized Area of Southern China.” Environmental Pollution 235: 47–54. Pfadenhauer, L. M., J. Burns, A. Rohwer, and E. A. Rehfuess. 2016. “Effectiveness of Interventions to Reduce Exposure to Lead through Consumer Products and Drinking Water: A Systematic Review.” Environ- mental Research 147: 525–36. Pocock, S. J., M. Smith, and P. Baghurst. 1994. “Environmental Lead and Children’s Intelligence: A System- atic Review of the Epidemiological Evidence.” BMJ (Clinical Research Ed.) 309(6963): 1189–97. Rahbar, M. H., F. White, M. Agboatwalla, S. Hozhabri, and S. Luby. 2002. “Factors Associated with Elevated Blood Lead Concentrations in Children in Karachi, Pakistan.” Bulletin of the World Health Organization 80: 769–75. Rahman, A., E. Maqbool, and H. S. Zuberi. 2002. “Lead-Associated Deficits in Stature, Mental Ability and Behaviour in Children in Karachi.” Annals of Tropical Paediatrics 22(4): 301–11. Ramírez, Ortega D., D. F. González Esquivel, T. Blanco Ayala, B. Pineda, S. Gómez Manzo, P. Marcial Quino, J. Carrillo Mora, and V. Pérez de la Cruz, 2021. “Cognitive Impairment Induced by Lead Exposure during Lifespan: Mechanisms of Lead Neurotoxicity.” Toxics 9(2): 23. Rasoul, G. A., M. Al-Batanony, O. Mahrous, M. Abo-Salem, and H. Gabr. 2012. “Environmental Lead Ex- posure among Primary School Children in Shebin El-Kom District, Menoufiya Governorate, Egypt.” International Journal of Occupational and Environmental Medicine 3(4): 186–94 Rau, T., S. Urzua, and L. Reyes. 2015. “Early Exposure to Hazardous Waste and Academic Achievement: Evidence from a Case of Environmental Negligence.” Journal of the Association of Environmental and Resource Economists 2(4): 527–63. Rees, N., and R. Fuller. 2020. “The Toxic Truth: Children’s Exposure to Lead Pollution Undermines a Gen- eration of Future Potential.” Technical report, UNICEF, Pure Earth, New York Reuben, A., A. Caspi, D. W. Belsky, J. Broadbent, H. Harrington, K. Sugden, R. M. Houts, S. Ramrakha, R. Poulton, and T. E. Moffitt. 2017. “Association of Childhood Blood Lead Levels with Cognitive Function and Socioeconomic Status at Age 38 Years and with IQ Change and Socioeconomic Mobility between Childhood and Adulthood.” JAMA 317(12): 1244. Rodríguez-Barranco, M., A. Tobías, D. Redondo, E. Molina-Portillo, and M. J. Sánchez. 2017. “Standardizing Effect Size from Linear Regression Models with Log-Transformed Variables for Meta-Analysis.” BMC Medical Research Methodology 17(1): 44. Rothenberg, S. J., and J. C. Rothenberg. 2005. “Testing the Dose–Response Specification in Epidemiology: Public Health and Policy Consequences for Lead.” Environmental Health Perspectives 113(9): 1190–5. Roy, A., A. S. Ettinger, H. Hu, D. Bellinger, J. Schwartz, R. Modali, R. O. Wright, K. Palaniappan, and K. Bal- akrishnan. 2013. “Effect Modification by Transferrin C2 Polymorphism on Lead Exposure, Hemoglobin Levels, and IQ.” Neurotoxicology 38: 17–22. Ruadze, E., G. S. Leonardi, A. Saei, I. Khonelidze, L. Sturua, V. Getia, H. Crabbe, T. Marczylo, P. Lauriola, and A. Gamkrelidze. 2021. “Reduction in Blood Lead Concentration in Children across the Republic 258 The World Bank Research Observer, vol. 40, no. 2 (2025) of Georgia following Interventions to Address Widespread Exceedance of Reference Value in 2019.” International Journal of Environmental Research and Public Health 18(22): 11903. Ruebner, R. L., S. R. Hooper, C. Parrish, S. L. Furth, and J. J. Fadrowski. 2019. “Environmental Lead Exposure Is Associated with Neurocognitive Dysfunction in Children with Chronic Kidney Disease.” Pediatric Nephrology (Berlin, Germany) 34(11): 2371–9. Sargsyan, A., E. Nash, G. Binkhorst, J. E. Forsyth, B. Jones, G. Ibarra Sanchez, S. Berg, A. McCartor, R. Fuller, and S. Bose-O’Reilly. 2024. “Rapid Market Screening to Assess Lead Concentrations in Consumer Prod- ucts across 25 Low- and Middle-Income Countries.” Scientific Reports 14(1): 9713. Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Schnaas, L., S. J. Rothenberg, M.-F. Flores, S. Martinez, C. Hernandez, E. Osorio, S. R. Velasco, and E. Perroni. 2006. “Reduced Intellectual Development in Children with Prenatal Lead Exposure.” Environmental Health Perspectives 114(5): 791–7. Shadbegian, R., D. Guignet, H. Klemick, and L. Bui. 2019. “Early Childhood Lead Exposure and the Persis- tence of Educational Consequences into Adolescence.” Environmental Research 178: 108643. Shen, X.-M., C.-H. Yan, S.-H. Wu, and R. Shi. 2004. “Parental Education to Reduce Blood Lead Levels in Children with Mild and Moderate Lead Poisoning: A Randomized Controlled Study.” Zhonghua Er Ke Za Zhi = Chinese Journal of Pediatrics 42(12): 892–7. Silbergeld, E. K. 1997. “Preventing Lead Poisoning in Children.” Annual Review of Public Health 18(1): 187– 210. Simonsohn, U., L. D. Nelson, and J. P. Simmons. 2014. “n p-Curve and Effect Size: Correcting for Publication Bias Using Only Significant Results.” Perspectives on Psychological Science 9(6): 666–81. Singh, A. K., and M. Singh. 2006. “Lead Decline in the Indian Environment Resulting from the Petrol-Lead Phase-Out Programme.” Science of the Total Environment 368(2-3): 686–94. Solon, O., T. J. Riddell, S. A. Quimbo, E. Butrick, G. P. Aylward, Bacate M. Lou, and J. W. Peabody. 2008. “Associations between Cognitive Function, Blood Lead Concentration, and Nutrition among Children in the Central Philippines.” Journal of Pediatrics 152(2): 237–43. Sorensen, L. C., A. M. Fox, H. Jung, and E. G. Martin. 2019. “Lead Exposure and Academic Achievement: Evidence from Childhood Lead Poisoning Prevention Efforts.” Journal of Population Economics 32(1): 179–218. Staes, C., T. Matte, C. G. Copley, D. Flanders, and S. Binder. 1994. “Retrospective Study of the Impact of Lead- Based Paint Hazard Remediation on Children’s Blood Lead Levels in St. Louis, Missouri.” American Journal of Epidemiology 139(10): 1016–26. Stanley, T. D., and H. Doucouliagos. 2014. “Meta-Regression Approximations to Reduce Publication Selec- tion Bias.” Research Synthesis Methods 5(1): 60–78. Surkan, P. J., A. Zhang, F. Trachtenberg, D. B. Daniel, S. McKinlay, and D. C. Bellinger. 2007. “Neuropsycho- logical Function in Children with Blood Lead Levels <10 microg/dL.” Neurotoxicology 28(6): 1170–7. Syofyan, S. S., A. S. Wahyuni, K. Rusmil, and A. Lelo. 2020. “The Effects of Calcium Supplementation on Blood Lead Levels and Short-term Memory of Chronically Exposed Children: A Clinical Trial Study.” Open Access Macedonian Journal of Medical Sciences 8(B): 1144–51. Taylor, C. M., K. Kordas, J. Golding, and A. M. Emond. 2017. “Effects of Low-Level Prenatal Lead Exposure on Child IQ at 4 and 8 Years in a UK Birth Cohort Study.” Neurotoxicology 62: 162–9. Tena, A., E. Peru, L. E. Martinetti, J. C. Cano, C. D. Loyola Baltazar, A. E. Wagler, R. Skouta, and K. Fenelon. 2019. “Long-Term Consequences of Early Postnatal Lead Exposure on Hippocampal Synaptic Activity in Adult Mice.” Brain and Behavior 9(8): e01307. Tirima, S., C. Bartrem, I. von Lindern, M. von Braun, D. Lind, S. M. Anka, and A. Abdullahi. 2016. “Environ- mental Remediation to Address Childhood Lead Poisoning Epidemic due to Artisanal Gold Mining in Zamfara, Nigeria.” Environmental Health Perspectives 124(9): 1471–8. Crawfurd et al. 259 UNEP. 2020. “Update on the Global Status of Legal Limits on Lead in Paint.” Technical report. Nairobi United Nations, Department of Economic and Social Affairs, Population Division. 2019. “World Population Prospects 2019, Online Edition. Rev. 1.” United Nations Environment Programme (UNEP). 2014. “The Impact of Phasing Out Leaded Petrol in Kenya: Based on Blood Lead Tests Carried Out by Njoroge Kimani before and after Leaded Petrol Phase- Out.” Technical report. Nairobi US EPA 2015. “Data Tables—Biomonitoring—Lead.” van Aert, R. C. M., and M. A. L. M. van Assen. 2018. “Correcting for Publication Bias in a Meta-Analysis with Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 the p-Uniform∗ Method.” MetaArXiv, https://doi.org/10.31222/osf.io/zqjr9. Vega-Dienstmaier, J. M., J. E. Salinas-Piélago, M. d. R. Gutiérrez-Campos, R. D. Mandamiento-Ayquipa, M. d. C. Yara-Hokama, J. Ponce-Canchihuamán, and J. Castro-Morales. 2006. “Lead Levels and Cognitive Abilities in Peruvian Children.” Revista Brasileira De Psiquiatria (Sao Paulo, Brazil: 1999) 28(1): 33–9. Wan, X., W. Wang, J. Liu, and T. Tong. 2014. “Estimating the Sample Mean and Standard Deviation from the Sample Size, Median, Range and/or Interquartile Range.” BMC Medical Research Methodology 14: 135. 260 The World Bank Research Observer, vol. 40, no. 2 (2025) Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Supplementary Online Appendix The Effect of Lead Exposure on Children’s Learning in the Developing World: A Meta-Analysis Lee Crawfurd , Rory Todd, Susannah Hares, Justin Sandefur, and Rachel Silverman Bonnifield S1 Further Figures and Tables Figure S1.1. Estimates of Overall Population Exposure, by Region Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: This figure shows the share of children with an elevated blood-lead level, defined as exceeding the US CDC ref- erence level of 5 micrograms per deciliter. These estimates are based on our analysis of raw numbers produced by IHME/UNICEF (Rees and Fuller 2020), combined with population estimates from the United Nations (United Nations, Department of Economic and Social Affairs, Population Division 2019). Figure S1.2. Sensitivity to Exclusion of Individual Studies Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: This figure shows the robustness of our primary specification result to leaving out individual studies. The study which substantially weakens the effect when excluded is Taylor et al. (2017). Crawfurd et al. 1 Figure S1.3. Funnel Plot for Potential Publication Bias Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: The two top figures show the distribution of study effect sizes and standard errors. An outlier result (Solon et al. 2008) was excluded to allow for visual inspection. Asymmetry around the vertical dashed line indicates that there may be publication bias present. Contours show that “missing” studies are primarily in the region of statistical insignificance—consistent with a statistical significance filter in publishing leading to this bias. The third figure shows the distribution of z-stats. The spike just right of the vertical dashed line separating significant and non-significant results provides further evidence for selection for significant results. Figure S1.4. Funnel Plot by Original Exposure Transformation Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: This figure shows funnel plots for all results, faceted by the original functional form used for the relationship between blood lead and the outcome. An outlier result (Solon et al. 2008) was excluded to allow for visual inspection. The bottom plot also excludes results which required additional conversions or imputations. The error resulting from the re-expression process and other conversions has the potential to generate a spurious correlation between effect sizes and standard errors, as both would be biased in the same direction. The bottom plot suggests this is at least partly true; results which use a natural log functional form, and which were not converted in some other way, show little asymmetry. This means that the existence of asymmetry for the data set as a whole may be an artefact of imprecision in the various conversion processes, rather than due to publication or reporting bias. While this imprecision would make our estimation of the true effect less precise, it would not upwardly bias the estimate; using funnel-plot-based methods like PEESE to adjust for this spurious “publication bias” would therefore underestimate the effect. Thus, while we still implement PEESE in our main specification, we note that our estimate may be correspondingly conservative. Figure S1.5. Copas Plot Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Note: This figure shows Copas method plots where we have collapsed our results, assuming an intra-study correlation of 1. An outlier result (Solon et al. 2008) was excluded to allow for visual inspection. The funnel plot (top left) is equivalent to fig. S1.3. The contour plot (top right) shows results from simulations assuming different selection probabilities for a study with a given precision. The treatment effect plot (bottom left) shows that the estimated treatment effect is lower if we assume these selection probabilities to be lower, but even at low selection probabilities, there is still a substantial negative effect of lead. Table S1.1. Studies on Association of Lead with IQ Authors Year Country code Country income group Alvarez-Ortega et al. 2017 COL M Baghurst et al. 1992 USA H Bellinger et al. 1992 USA H Braun et al. 2012 MEX M Cai et al. 2020 CHN M Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Canfield et al. 2003 USA H Chen et al. 2007 USA H Chiodo et al. 2007 USA H Cooney et al. 1989 AUS H Counter et al. 2005 ECU M Dantzer et al. 2020 USA H Desrochers-Couture et al. 2018 CAN H Dietrich et al. 1993 USA H Earl et al. 2016 AUS H Fulton et al. 1987 GBR H Hong et al. 2015 KOR H Huang et al. 2012 TWN H Jusko et al. 2008 USA H Kamel et al. 2003 EGY M Kim et al. 2009 KOR H Kim, Yu, and Lee 2010 KOR H Liu et al. 2013 CHN M Lucchini et al. 2012 ITA H Lucchini et al. 2019 ITA H Menezes-Filho et al. 2018 BRA M Min et al. 2009 USA H Nwobi et al. 2019 NGA M Pan et al. 2018 CHN M Rahman et al. 2002 PAK L Rasoul et al. 2012 EGY M Reuben et al. 2017 NZL H Roy et al. 2013 IND M Ruebner et al. 2019 USA H Schnaas et al. 2006 MEX M Schnaas et al. 2000 MEX M Solon et al. 2008 PHL M Surkan et al. 2007 USA H Taylor et al. 2017 GBR H Vega-Dienstmaier et al. 2006 PER M Zailina et al. 2011 MYS M Note: We include here some three cognitive assessments that are similar but not identical to IQ tests—Vega-Dienstmaier et al. (2006) use the “Graphic Test of Reasoning” and the “Kohs Block Design Test,” while Cooney et al. (1989) and Braun et al. (2012) use the McCarthy Scales of Children’s Abilities, General Cognitive Index. Table S1.2. Studies on Association of Lead with Reading and Mathematics Scores Authors Year Country code Country Income Group Reading Aizer et al. 2016 USA H Bellinger et al. 1992 USA H Blackowicz et al. 2016 USA H Chiodo et al. 2007 USA H Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Evens et al. 2015 USA H Fulton et al. 1987 GBR H Kamel et al. 2003 EGY M Kim, Yu, and Lee 2010 KOR H Kordas et al. 2006 MEX M Lanphear et al. 2000 USA H Liu et al. 2013 CHN M McLaine et al. 2013 USA H Min et al. 2009 USA H Shadbegian et al. 2019 USA H Surkan et al. 2007 USA H Maths Aizer et al. 2016 USA H Bellinger et al. 1992 USA H Blackowicz et al. 2016 USA H Chiodo et al. 2007 USA H Evens et al. 2015 USA H Fulton et al. 1987 GBR H Kamel et al. 2003 EGY M Kim, Yu, and Lee 2010 KOR H Kordas et al. 2006 MEX M Lanphear et al. 2000 USA H Liu et al. 2013 CHN M Min et al. 2009 USA H Shadbegian et al. 2019 USA H Surkan et al. 2007 USA H Note: This table lists the studies included in our meta-analysis with reading or mathematics as an outcome. Table S1.3. Alternative Re-expression Algorithms Mean BLL SD BLL True effect (log) RB L D NHANES 99–2000 3.060 3.042 −0.261 −0.244 −0.148 −0.105 NHANES 2001–02 2.629 2.205 −0.194 −0.263 −0.159 −0.121 NHANES 2011–12 1.953 1.657 −0.219 −0.247 −0.150 −0.113 Canfield et al. 2003 9.015 5.488 −0.489 −0.845 −0.512 −0.427 Vega-Dienstmaier et al. 10.330 7.360 −0.672 −2.371 −0.750 −0.522 Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 2006 Kordas et al. maths 11.793 6.400 −0.405 −0.573 −0.348 −0.299 Kordas et al. reading 11.787 6.393 −0.463 −0.674 −0.409 −0.351 Crump et al. 2013 14.607 14.230 −0.119 −0.205 −0.125 −0.089 Note: In this table we assess the performance of three different algorithms used to re-express effects estimated with linear models in terms of log units. We first use microdata to estimate directly the effect of a log unit increase in blood lead levels (BLL) on standardized test-score outcomes (column 2). We then estimate the effect of a linear unit increase in BLL, and apply the algorithms to re-express the effect estimated with a linear model in terms of log units (columns 3– 5). RB indicates results from the Rodríguez-Barranco et al. (2017) algorithm, L from the Linakis et al. (2024) algorithm, and D from the Dzierlenga et al. (2020) algorithm. The root mean squared error indicated that the Linakis et al. (2024) algorithm clearly performed best. Table S1.4. Alternative Adjustments for Publication Bias (1) (2) (3) (4) (5) (6) None Egger Non-linear Trim-and-fill p-uniform∗ Andrews and Kasy Effect – −1.470∗∗∗ – – – standard error (0.290) Effect – – −0.252 – variance (0.188) Constant −0.212∗∗∗ −0.085∗∗∗ −0.207∗∗∗ −0.157∗∗∗ −0.178∗∗∗ −0.140∗∗∗ (0.024) (0.029) (0.023) (0.026) (0.031) (0.023) N (studies) 47 47 47 47 47 47 Note: This table shows four alternative standard approaches for adjusting for publication bias, applied to our results when unadjusted and averaged by study (column 1) (we do not use the adjusted result, as the trim-and-fill and p- uniform methods would not be applicable): in column 2, the Egger intercept (adjusting for the study standard error) (Egger et al. 1997); in column 3, the "precision-effect test and precision-effect estimate with standard errors" (PET- PEESE) nonlinear intercept (adjusting for the variance) (Stanley and Doucouliagos 2014), also used in our main spec- ification; in column 4, the trim-and-fill method (attempting to “fill in” hypothetical results given no publication bias) (Duval and Tweedie 2000); and in column 5, the p-uniform∗ method (van Aert and van Assen 2018). Standard errors in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Table S1.5. Bivariate Meta-Regressions (1) (2) (3) (4) (5) (6) (7) (8) Not logarithmic – −0.042 – – – – – – (0.050) Mean blood lead – – −0.008∗∗ – – – – – (0.004) No control for parent ed/IQ – – – −0.173∗∗∗ – – – – (0.036) No control for family income – – – – −0.123∗∗∗ – – – (0.046) Exposure: Not average or lag – – – – – −0.036 – – (0.049) Outcome: Mathematics – – – – – – 0.021 – (0.055) Outcome: Reading – – – – – – −0.015 – (0.050) High-income country – – – – – – – 0.035 (0.057) Constant −0.227∗∗∗ −0.204∗∗∗ −0.175∗∗∗ −0.155∗∗∗ −0.169∗∗∗ −0.212∗∗∗ −0.233∗∗∗ −0.255∗∗∗ (0.025) (0.033) (0.033) (0.021) (0.023) (0.036) (0.032) (0.047) N (estimates) 286 286 286 286 286 286 286 286 N (studies) 47 47 47 47 47 47 47 47 Note: We use the Hedges et al. (2010) estimator to account for dependence between multiple estimates from the same study. The dependent variable in each case is the stan- dardized effect size of a natural log unit increase in lead exposure on the cognitive outcome. Standard errors in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 Table S1.6. Sensitivity (1) (2) (3) Imputed outcome SD – −0.041 – (0.066) Constant −0.115 −0.094 −0.159 (0.100) (0.110) (0.115) N (estimates) 286 286 212 Downloaded from https://academic.oup.com/wbro/article/40/2/229/7734093 by Laura Mowry user on 05 August 2025 N (studies) 47 47 30 Note: We use the Hedges et al. (2010) estimator to account for dependence between multiple estimates from the same study. This table shows our adjusted meta-regression estimate (column 1), the estimate after additionally including a dummy for whether a population rather than sample standard deviation was used to calculate an effect size (column 2), and the estimate in a sub-sample that excludes those in which various imputations had to be made (column 3). These exclusions include results in which we impute mean blood-lead levels, results in which we calculate the standard error using a reported p-value, and results where the effect size is converted from a reported correlation coefficient. Standard errors in parentheses. ∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. The World Bank Research Observer © The Author(s) 2024. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. https://doi.org/10.1093/wbro/lkae010