Prenatal Exposure to Ambient PM2.5 and Early Childhood Growth Impairment Risk in East Africa

Height for age is an important and widely used population-level indicator of children’s health. Morbidity trends show that stunting in young children is a significant public health concern. Recent studies point to environmental factors as an understudied area of child growth failure in Africa. Data on child measurements of height-for-age and confounders were obtained from fifteen waves of the Demographic and Health Surveys (DHS) for six countries in East Africa. Monthly ambient PM2.5 concentration data was retrieved from the Atmospheric Composition Analysis Group (ACAG) global surface PM2.5 estimates and spatially integrated with DHS data. Generalized additive models with linear and logistic regression were used to estimate the exposure-response relationship between prenatal PM2.5 and height-for-age and stunting among children under five in East Africa (EA). Fully adjusted models showed that for each 10 µg/m3 increase in PM2.5 concentration there is a 0.069 (CI: 0.097, 0.041) standard deviation decrease in height-for-age and 9% higher odds of being stunted. Our study identified ambient PM2.5 as an environmental risk factor for lower height-for-age among young children in EA. This underscores the need to address emissions of harmful air pollutants in EA as adverse health effects are attributable to ambient PM2.5 air pollution.


Introduction
Ambient air pollution is higher in low-and middle-income countries (LMICs) compared with higher-income countries. Consequently, childhood morbidity and mortality related to prenatal ambient air pollution exposure are assumed to be disproportionately higher in LMICs [1]. However, there is minimal epidemiological data to substantiate these assumptions, particularly for the African continent.
The East African countries of Uganda, Rwanda, Ethiopia, Burundi, Kenya, and Tanzania are low-income countries [2,3] with varying ambient air pollution levels. East Africa shares a considerable burden of childhood mortality, accounting for more than half of all under-five children mortality in Sub-Saharan Africa [4]. Air pollution has been considered a major public health concern as a top risk factor for mortality in these countries and accounts for almost 1 million deaths annually. Cooking with solid fuels in households is also a significant and persistent source of air pollution in sub-Saharan Africa [5,6].
The ambient air pollutant of interest in this study is fine particulate matter with an aerodynamic diameter of 2.5 micrometers or less (PM 2.5 ). Ambient PM 2.5 is ubiquitous and can be found increasingly in all environments, especially in growing urban cities [7,8]. It is harmful to human health, with young children and elderly populations particularly vulnerable to the effects of PM 2.5 exposure [9][10][11]. PM 2.5 is small enough to lodge in the alveolar region of the lower respiratory system, enabling PM 2.5 to evade the body's innate and adaptive immune defenses [12]. This characteristic of PM 2.5 exposure affords some particles access to the bloodstream, and recent evidence shows that it can reach the placenta during gestation [13]. Once PM 2.5 enters, it can lead to cardiovascular effects but may also translocate to the placenta, which can affect the fetus's health [14,15]. Several mechanisms

DHS Survey Data
This project used nationally-representative health survey data from fifteen waves of the DHS for six countries between 2006 and 2019. Table 1 shows the waves and years of DHS data that were included for each country and the sample size of each wave based on the study's inclusion criteria. The DHS is a two-stage cluster sample, where randomly sampled clusters of households from enumeration areas are surveyed. A detailed methodology of DHS is available elsewhere [49]. Therefore, we obtained a representative cross-section of East African children under five. The inclusion criteria for children under 5 in this study included (1) less than 60 months old at the time of the survey, (2) a valid heightfor-age standard deviation value calculated, (3) singleton births, and (4) a matched PM 2.5 estimate (see Figure 1). Other variables from the DHS were obtained for use as covariates in statistical models (see Section 2.4 Statistical Analyses).  The PM2.5 concentration data in this research comes from the Atmospheric Composition Analysis Group (ACAG) global PM2.5 estimates from 1998 to 2020 [50]. ACAG provides data on ground-level prediction surfaces of average monthly PM2.5 concentration available at a spatial resolution of 0.1° × 0.1° (1 km × 1 km grid). ACAG derived these satellite-based global PM2.5 prediction surfaces by combining aerosol optical depth (AOD) data from the satellite instruments MODIS (Moderate Resolution Imaging Spectroradiometer) and MISR (Multiangle Imaging Spectroradiometer) with chemical transport mod-  The PM 2.5 concentration data in this research comes from the Atmospheric Composition Analysis Group (ACAG) global PM 2.5 estimates from 1998 to 2020 [50]. ACAG provides data on ground-level prediction surfaces of average monthly PM 2.5 concentration available at a spatial resolution of 0.1 • × 0.1 • (1 km × 1 km grid). ACAG derived these satellite-based global PM 2.5 prediction surfaces by combining aerosol optical depth (AOD) data from the satellite instruments MODIS (Moderate Resolution Imaging Spectroradiometer) and MISR (Multiangle Imaging Spectroradiometer) with chemical transport models and ground-based air monitoring data [51].

Linkage of PM 2.5 Data with DHS Data
We spatially linked the spatial grid points of the modeled PM 2.5 raster data with DHS data by using the latitude and longitude of each DHS cluster for each child in our study. The DHS provides these geographic coordinates with a 2 km uncertainty in urban clusters and a 5 km uncertainty in rural clusters to protect the confidentiality of participants. Therefore, we assigned each cluster a buffer zone of 2 km and 5 km for urban and rural clusters, respectively [16]. We clipped the PM 2.5 prediction surfaces for each cluster buffer zone to assign exposures. Specifically, we used the Zonal Statistics tool in ArcGIS to calculate an average PM 2.5 of grid-point values within each cluster buffer.

Percent Crop Estimates
We retrieved land-use data from the European Space Agency (ESA) Climate Change Initiative (CCI) project. This project provides maps of land cover classifications at 300 m resolution from the years 1992 to 2015 (v2.0.7) and from 2016 to 2019 (v2.1.1) [52,53]. Studies have used this dataset to track land cover changes at global and regional scales [54,55]. The raster data for global land use cover for each DHS wave year was downloaded and analyzed in ArcGIS Pro. Briefly, we first clipped yearly raster data to the East Africa region. We then converted yearly crop cover raster values to polygons values by averaging the crop cover raster values for buffers of urban and rural clusters. Thus, we calculated an annual percent cropland cover for each DHS cluster, enabling a crop-cover exposure assignment for each child based on birth year.

Prenatal and Postnatal Exposure Estimates
Prenatal PM 2.5 exposure was estimated using the months preceding each child's reported date of birth [16]. Specifically, for each child's gestation period, the child's date of birth (given by the DHS) was subtracted by 268 days, providing the starting month of gestation. Next, we calculated the child-specific PM 2.5 prenatal exposures (including the month of birth) by averaging each month within the estimated gestation period. We calculated the child-specific postnatal exposures using the child's date of birth and the date of the child's height measurement (also given by DHS) [16].

Statistical Analyses
We calculated summary statistics for all study variables and performed bivariate analyses (ANOVA) to compare the levels of height-for-age with each covariate selected for the study. We constructed a directed acyclic graph (DAG) (see Figure S1, Supplemental Materials) to choose a candidate for testing bivariate relationships. We used general additive models (GAMs) to quantify the exposure-response relationship between prenatal PM 2.5 exposure estimates and height-for-age, including random effects for the child's mother, primary sampling unit (PSU) cluster, country, and month of birth. Smoothing terms were fit to adjust for the non-linear effects of birth year and percent crop coverage on height for age. Additional tested models included breastfeeding (Model 2) and postnatal PM 2.5 exposure (Model 3) as other covariates. In addition, we used multivariable logistic regression to estimate the adjusted Odds Ratios (OR) between prenatal and postnatal ambient PM2.5 exposures for stunting. Both exposures were fit as quartiles of exposure and were included in the same model to estimate their independent effects on stunting. Since stunting is considered a common health outcome, we can assume the OR approximates the RR. Therefore, we used the adjusted ORs estimated from the model and converted them to Relative Risk (RR) of stunting using the oddsratio_to_riskratio function in the effectsize package (version 0.8.2) in R [56,57]. Using these RR estimates, we also computed a population attributable fraction (PAF) that can be attributed to ambient PM 2.5 among children exposed prenatally and postnatally. The following formula was used to compute the PAF: pd * (RR − 1)/RR where pd is the proportion of cases (stunted) exposed to the risk factor (in this case, the proportion of children in the highest quartile of PM 2.5 exposure [0.25]) and RR is the relative risk of stunting associated with the exposure [58].

Variables for Covariate Adjustment
Household-level covariates included the type of residence (urban or rural), wealth index, and use of polluting fuel. Maternal and child-level variables, including maternal education and information on the child's breastfeeding history, were also analyzed. The DHS survey reported the wealth index as quintiles, which we fit as covariates in their original form. We reclassified the DHS variable for polluting fuel into a binary variable as a cleaner energy source (0 = electricity, LPG, natural gas, and biogas) versus a polluting energy source (1 = kerosene, coal, lignite, charcoal, wood, straw/shrubs/grass, agricultural crop, animal dung, briquette, bottle gas, and other). Using multivariable logistic regression, we performed multiple imputation on missing values for polluting fuel use (N = 1616). We reclassified the duration of breastfeeding variable into a binary variable indicating "ever breastfed" (≥1 month) or "never breastfed". Missing observations for breastfeeding were also imputed (N = 9387) using multivariable logistic regression. The breastfeeding variable was conceptualized as "ever" and "never" because the analysis was robust enough to identify the effect of early-childhood nutrition on the outcome.

Summary and Descriptive Statistics
Our final study sample size included 87,716 children under five with prenatal PM 2.5 exposure estimates in six East African countries (Table 2). For the complete sample, children were exposed to an entire-pregnancy average of 25.83 µg/m 3 ambient PM 2.5 , five times above the WHO recommended level of 5 µg/m 3 annually. Rwanda had the highest prenatal Generally, children born during the wet season had lower levels of prenatal PM 2.5 , and during the dry season, higher levels of prenatal PM 2.5. We also observed higher PM 2.5 levels in the dry seasons for Burundi, Kenya, Rwanda, Tanzania, and Uganda ( Figure 2) (see Figure S2, Supplemental Materials, for annual trends). Ethiopia and Kenya had the most negligible monthly variability, with no apparent seasonal trend. Regardless of seasonal variations, children in Burundi and Rwanda experienced prenatal PM 2.5 exposures above the mean. breastfeeding variable was conceptualized as "ever" and "never" because the analysis was robust enough to identify the effect of early-childhood nutrition on the outcome.

Summary and Descriptive Statistics
Our final study sample size included 87,716 children under five with prenatal PM2.5 exposure estimates in six East African countries ( Table 2). For the complete sample, children were exposed to an entire-pregnancy average of 25.83 μg/m 3 ambient PM2.5, five times above the WHO recommended level of 5 μg/m 3 annually. Rwanda had the highest prenatal PM2.  Overall, children's height-for-age z-score in the study population was low (mean = −1.51 [95% CI: −1.49, −1.53]). The height for age Z-scores were lowest in Burundi (mean = −2.17), and the highest was in Kenya (mean = −1.19) ( Table 2). Most of the country's heightfor-age distribution was around the mean for the months of birth, with little to no noticeable variation (Figure 3) (see Figure S3, Supplemental Materials, for annual trends). Height-for-age among children in Burundi was below the height-for-age mean of the study population.
The mean child age was 28.69 months, with 49.4% of the study population being comprised of girls. Most of the children in the sample resided in a rural setting (80.8%), ranging from 69.6% rural in Kenya to 91.3% rural in Burundi. The vast majority (97.5%) of the households included in the sample used solid fuel for cooking. Table 2. Summary statistics of study population characteristics overall and stratified by country. Overall, children's height-for-age z-score in the study population was low (mean = −1.51 [95% CI: −1.49, −1.53]). The height for age Z-scores were lowest in Burundi (mean = −2.17), and the highest was in Kenya (mean = −1.19) ( Table 2). Most of the country's height-forage distribution was around the mean for the months of birth, with little to no noticeable variation ( Figure 3) (see Figure S3, Supplemental Materials, for annual trends). Height-forage among children in Burundi was below the height-for-age mean of the study population. Survey package functions in R were used to account for the complex stratified random sampling used by DHS.

Bivariate Analyses
A bivariate analysis for each study covariate (Table 3) shows that all are associated with height-for-age at a significance level of less than 0.001, except breastfeeding (p > 0.05). Although breastfeeding was not significant, we included it in Model 2 because it is a known risk factor for height for age.  The mean child age was 28.69 months, with 49.4% of the study population being comprised of girls. Most of the children in the sample resided in a rural setting (80.8%), ranging from 69.6% rural in Kenya to 91.3% rural in Burundi. The vast majority (97.5%) of the households included in the sample used solid fuel for cooking.

Bivariate Analyses
A bivariate analysis for each study covariate (Table 3) shows that all are associated with height-for-age at a significance level of less than 0.001, except breastfeeding (p > 0.05). Although breastfeeding was not significant, we included it in Model 2 because it is a known risk factor for height for age.

Association between Pprenatal PM 2.5 Exposure and Height forAge
The crude model (Table 4) indicates that a 10 µg/m 3 increase in prenatal PM 2.5 exposure is associated with a lower height for age Z-score of −0.143 (95% CI: −0.154, −0.131). After controlling for rurality, maternal education, wealth index, percent crop cover, and use of polluting fuel (Model 1), a 10 µg/m 3 increase of prenatal PM 2.5 was associated with a lower height-for-age Z-score by −0.08 (−0.98, −0.62). After controlling for the child ever being breastfed (Model 2), a 10 µg/m 3 increase of prenatal PM 2.5 exposure was associated with a lower height-for-age by −0.095 (−0.114, −0.076). After controlling for the independent effect of postnatal PM 2.5 exposure (Model 3), the effect of prenatal PM 2.5 exposure on height for age was slightly attenuated (β = −0.069 [−0.097, −0.041]). The effect of postnatal PM 2.5 exposure on height-for-age (β = −0.05 [−0.08, −0.02]) was slightly smaller than the prenatal effect estimate, but had a similar magnitude in the exposure-response relationship.  Other covariates were independently associated with height-for-age Z-score, and associations were in the expected directions (Table 4). For instance, children in rural communities had significantly lower height-for-age compared with children in urban communities. Higher maternal education, household wealth index, and ever breastfed were all independently associated with higher height-for-age Z-scores after adjustment. While using a polluting fuel was significantly associated with lower height-for-age Z-score in Model 1, this association became null after adjusting for ever breastfed (Model 2).

Association between Prenatal and Postnatal PM 2.5 Exposure and Stunting
Relative to the lowest prenatal PM 2.5 exposure group, the highest quartile of prenatal PM 2.5 exposure is associated with a 12% (7%, 17%) higher risk of stunting (Table 5). Independent of prenatal PM 2.5 exposure and relative to the lowest postnatal PM 2.5 exposure group, the highest quartile of postnatal PM 2.5 exposure is associated with an 11% (6%, 16%) higher risk of stunting. These results suggest that 2.7% and 2.5% of the population attributable risk of stunting in the study area are due to ambient prenatal PM 2.5 and postnatal PM 2.5 exposures, respectively. The RR of stunting for all other risk factors considered in the model are further summarized in Table 5. The adjusted RR was computed using the modeling output of Odds Ratios estimated with multivariable logistic regression, adjusting for year of birth and percent crop as smoothing terms and random effects for primary sampling unit, mother, the month of birth, and country. Full modeling output is provided in the Supplemental Materials (see Table S1). CI: Confidence Interval; RR: Relative Risk; AIC: Akaiki information criterion.

Discussion
Using multiple waves of nationally representative DHS data for children under five in six East African countries, our results show a robust negative exposure-response relationship between prenatal ambient PM 2.5 exposure and height-for-age Z-scores and a positive prenatal PM 2.5 exposure-response relationship for stunting. Additionally, our analysis found that the overall average prenatal ambient PM 2.5 exposure (25.83 µg/m 3 ) far exceeded the WHO annual guideline for ambient PM 2.5 (5 µg/m 3 [59]).
From the fully adjusted models, we observe significant adverse effects on height for age in the study population (β range: −0.095 to −0.069 SD, Table 4). Our findings suggest an adverse shift in the population mean and distribution of height-for-age Z-scores due to prenatal PM 2.5 exposure. Such population-level impacts imply that more children are likely to fall into the clinically relevant classification of stunting due to elevated PM 2.5 levels. Indeed, the fully adjusted logistic regression model (Table 5) showed that the highest quartile of prenatal PM 2.5 exposure is associated with a 12% higher risk of stunting compared to the lowest exposure group. The elevated risk of stunting from elevated prenatal ambient PM 2.5 exposure underscores the clinical significance of our results.
The population attributable fraction (PAF) is a useful metric for evaluating the relative contribution that a specific risk factor has on important population health indicators. In our study, we estimated that 2.8% and 2.5% of stunting in East Africa is attributed to the highest quartiles of exposure to prenatal PM 2.5 and postnatal PM 2.5 , respectively. The estimated PAF for stunting due to exposure to polluting cooking fuel is 7.2% in our study. Together, the highest quartiles of prenatal and postnatal PM 2.5 exposures, along with cooking with polluting fuel, attribute to up to 12.5% of stunting risk in East Africa. This finding further highlights the need for health policy and health promotion to emphasize air pollution mitigation, for both ambient and household sources, and to achieve the global target to reduce childhood stunting.
Our results compare favorably with data from the limited number of studies that have investigated the exposure-response relationship between ambient prenatal PM 2.5 exposure and childhood height-for-age and stunting using DHS data [16,22]. Spears et al. [16,22] found in India that a 10 µg/m 3 increase in PM 2.5 was associated with a 0.005 standard deviation reduction in child height [16,22]. Their effect estimate is an order of magnitude lower than what we observed in our study. This effect size difference may be explained by various study design differences, such as our use of entire-pregnancy exposure estimates or more spatially resolved exposure estimates. For example, the India study used the nearest monitoring data to reflect exposure for an entire city, which can contribute to significant exposure misclassification that would bias effects downward. Differences in population vulnerability to PM 2.5 effects may also explain these differences.
In Bangladesh, Goyal et al. found that children prenatally exposed to levels of entirepregnancy PM 2.5 in the highest quartile were 1.13 times more at risk of stunting than children prenatally exposed to PM 2.5 levels [16,22]. In our study, we observed a similar effect size, where children in the highest quartile were up to 1.12 times more at risk of stunting compared to the lowest quartile of exposure. This Bangladesh study used similar ACAG PM 2.5 estimates. However, the ACAG estimates were only available annually at the time of their study. In contrast, our study used updated model estimates of monthly PM 2.5 concentrations, which likely reduced exposure misclassification for our study. Moreover, we estimated postnatal exposure to control for any possible residual confounding that may be attributed to ambient PM 2.5 exposure after the prenatal period.
Our estimates for PAF also compares favorably with recent country-level PAF estimates for ambient PM 2.5 on risk of low birth weight (LBW) in Burundi, Ethiopia, Kenya, Rwanda, Tanzania, and Uganda [19]. Ghosh et al. (2021) estimated that a median PAF of 2.04% of LBW risk is attributable to prenatal ambient PM 2.5 exposure in these countries, while our study estimates a prenatal ambient PM 2.5 PAF of 2.7% for stunting [19].
Our findings are notably consistent with the existing literature concerning other childgrowth risk factors. The bivariate and adjusted models showed that children residing in rural areas were likelier to have lower height-for-age and higher stunting than their urban counterparts. This is consistent with other studies identifying rurality as a risk factor for childhood growth impairment [60][61][62]. Like other studies, maternal education and wealth index was also significantly associated with better growth outcomes in our study [63][64][65]. The use of polluting fuel was significantly negatively associated with height-for-age in our study, which is also consistent with other studies that have found strong associations between childhood growth and biomass fuel for cooking [35,37].
The results on seasonal trends showed that prenatal PM 2.5 exposure varied based on being born in the wet or dry season. Noticeably, the dry seasons had higher levels of PM 2.5 , and the wet seasons had lower levels of PM 2.5 . Previous studies also show that PM 2.5 depends on seasonally varying factors, including weather conditions. A study conducted in Uganda found that higher monthly precipitation was significantly predictive of lower PM 2.5 levels [66]. This finding suggests that infants born during the dry season face disproportionately higher levels of prenatal ambient PM 2.5 exposures in our study area. However, there was no clear association between seasonality (wet versus dry) and height-for-age.
Additionally, there were differences in air pollution between countries. These differences may be partly explained by the economic development differences between countries in our study. According to the World Bank, the three countries below the population mean PM 2.5 levels: Kenya, Tanzania, and Ethiopia, had the highest gross domestic product in 2021 [67,68].
There are several limitations to this study. There is a risk of exposure misclassification in our study attributed to maternal factors during pregnancy. These maternal factors could include moving between homes during pregnancy or gestation periods that did not go to full term. For instance, our PM 2.5 exposure estimation assumed that the mother lived in the same house during the nine months preceding the DHS household survey, which may not be the case for some mothers. In addition, other pollutants potentially affect the outcome of height-for-age, such as sulfur dioxide, nitrogen dioxide, and ozone [69,70]. These co-pollutants were not included in the analysis as covariates because the data was unavailable for the region. Also, our study did not explore the composition of PM25. Ambient PM 2.5 air pollution is composed of different types of chemicals, which our study could not assess.
Furthermore, we used calculated PM 2.5 exposures with primarily satellite-driven estimates, as ground-level data was unavailable for much of the study area to validate predictions. However, the reported R 2 from the ACAG cross-validated model was 0.90-0.92, suggesting reliable estimates [50]. When going through the inclusion criteria, several individuals were excluded (N = 63,284) from the analysis. However, we assume that these unmatched observations are missing randomly, which should not introduce selection bias in the study. Additionally, even though we used the DAG to identify the variables to be included in the models, residual confounding could still be present in the model and should be considered when interpreting the results of the estimated exposure-response relationships between prenatal and postnatal PM 2.5 .
The strengths of this study include analyzing nationally representative data for six East African countries using several waves of the Demographic Health Surveys (DHS). This approach enabled us to spatially link births reported in the DHS with monthly PM 2.5 values and estimate PM 2.5 prenatal exposures and effects at the individual level after controlling for multiple relevant confounding variables. Using GAMs in a hierarchical modeling framework allowed us to further account for non-linear temporal confounders, the hierarchical structure of the DHS survey itself, and height-for-age outcome. Our study also leveraged newly available high spatial and temporal resolution data on PM 2.5 estimates provided on a global scale. While previous studies using DHS data in Bangladesh and India relied on long-term annual global PM 2.5 values and month-of-birth PM 2.5 measurement data, we analyzed PM 2.5 global estimates available at a monthly scale and a prediction model that improved from previous versions. Therefore, we could derive entire-pregnancy exposure estimates incorporating between-season variability of PM 2.5 , which previous studies did not include. Therefore, our approach to exposure estimation should entail lower exposure misclassification and, thus, more refined PM 2.5 effect estimates than the India and Bangladesh studies. Another significant strength of our analysis is the use of multiple models with different risk factors that allowed us to interrogate the robustness of prenatal PM 2.5 effect estimates. For instance, we could control for postnatal exposures that could confound the relationship between prenatal exposures and height for age. Indeed, adjusting for postnatal PM 2.5 resulted in attenuation in the prenatal effect estimate for height-for-age.

Conclusions
We observed a significant negative exposure-response relationship between prenatal ambient PM 2.5 and childhood height-for-age in East Africa. We also found a positive exposure-response relationship between postnatal PM 2.5 exposure estimates and stunting. Currently, stunting remains a high-priority issue in East Africa because it has implications for morbidity and mortality and the overall economy in the long term. These results highlight that exposure to ambient air pollution is a priority for childhood health and development, starting from gestation through early childhood. There is also a need for further expansion of ground air monitoring in these countries and substantial policy changes concerning air pollution reduction in East Africa.
Supplementary Materials: The following supporting information can be downloaded at: https: //www.mdpi.com/article/10.3390/toxics10110705/s1, Figure S1: Directed Acyclic Graph; Figure S2: Trend of prenatal PM 2.5 exposures stratified by child's year of birth and by country; Figure S3: Trend of height-for-age stratified by child's year of birth and by country; Table S1: Crude and adjusted odds ratios for a 10 µg/m 3 increase of prenatal PM 2.5 exposure and stunting in children under 5 years old.

Informed Consent Statement: Not applicable.
Data Availability Statement: The data generated and analyzed in this study is available upon request from the corresponding author.