Independent and joint effects of neighborhood-level environmental and socioeconomic exposures on body mass index in early childhood: the Environmental influences on Child Health Outcomes (ECHO) Cohort

Abstract

Past studies support the hypothesis that the prenatal period influences childhood growth. However, few studies explore the joint effects of exposures that occur simultaneously during pregnancy. To explore the feasibility of using mixtures methods with neighborhood-level environmental exposures, we assessed the effects of multiple prenatal exposures on body mass index (BMI) from birth to age 24 months. We used data from two cohorts: Healthy Start (n = 977) and Maternal and Developmental Risks from Environmental and Social Stressors (MADRES; n = 303). BMI was measured at delivery and 6, 12, and 24 months and standardized as z-scores. We included variables for air pollutants, built and natural environments, food access, and neighborhood socioeconomic status (SES). We used two complementary statistical approaches: single-exposure linear regression and quantile-based g-computation. Models were fit separately for each cohort and time point and were adjusted for relevant covariates. Single-exposure models identified negative associations between NO₂ and distance to parks and positive associations between low neighborhood SES and BMI z-scores for Healthy Start participants; for MADRES participants, we observed negative associations between O₃ and distance to parks and BMI z-scores. G-computations models produced comparable results for each cohort: higher exposures were generally associated with lower BMI, although results were not significant. Results from the g-computation models, which do not require a priori knowledge of the direction of associations, indicated that the direction of associations between mixture components and BMI varied by cohort and time point. Our study highlights challenges in assessing mixtures effects at the neighborhood level and in harmonizing exposure data across cohorts. For example, geospatial data of neighborhood-level exposures may not fully capture the qualities that might influence health behavior. Studies aiming to harmonize geospatial data from different geographical regions should consider contextual factors when operationalizing exposure variables.

1. INTRODUCTION

Questions regarding the joint effects of social stressors and environmental exposures, particularly during the prenatal period, on health outcomes are of growing interest (Koman et al., 2018; Padula et al., 2020; Vrijheid et al., 2020). National research agencies such as the National Institute for Environmental Health Sciences (US National Institutes of Health) have called for more investigation into the effects of multiple exposures that occur simultaneously to better reflect “real world” exposures experienced by humans (Carlin et al., 2013; National Institute of Environmental Health Sciences, 2018). Such studies are important because many exposures share common pathways for impacting health and may have additive or synergistic effects. Several statistical methods for assessing the effects of multiple exposures at once have been developed in recent years for environmental epidemiology applications (Joubert et al., 2022). These methods aim to address issues of multicollinearity and to make results interpretable for researchers, practitioners, and decision makers. As these methods become more widespread, it will be important to identify when they will be most applicable depending on the specific purpose of the study (Yu et al., 2022) and other data considerations.

One area of investigation where mixtures methods may be helpful is the effect of the broader neighborhood context on children’s health. For children, the prenatal period is particularly sensitive to several environmental and socioeconomic factors. An indicator of health during the early life period (i.e., before age 2) that may be sensitive to prenatal insults is body mass index (BMI), a crude measure of body composition that has previously been associated with the risk of obesity in childhood (Roy et al., 2016; Woo Baidal et al., 2016). Prenatal factors known to influence BMI, body composition, or growth trajectories before age 2 include maternal factors such as gestational weight gain (GWG), maternal stress, maternal diet and physical activity during pregnancy as well as environmental exposures like air pollution, features of the built environment and green space (de Bont et al., 2020; Fox et al., 2022; Ji et al., 2023; Larqué et al., 2019). Importantly, these risk factors can be influenced by the neighborhood context in which a pregnant person lives. For example, maintaining a healthy weight before and during pregnancy requires a supportive environment, including access to healthy foods and recreation spaces (Rasmussen et al., 2009; Vanstone et al., 2017). Higher walkability, greater access to parks, and a higher number of grocery stores in a neighborhood are each associated with lower odds of excessive GWG (Grobman et al., 2023; Kinsey et al., 2022). Environmental exposures like air pollution and social stressors like neighborhood poverty are able to impact health vial pathways such as oxidative stress and inflammation (Erickson and Arbour, 2014; Rakers et al., 2020). Despite evidence that exposures at the neighborhood level across multiple domains (i.e., the chemical, physical, and social environments) share pathways through which they may influence infant body composition, few studies have investigated how multiple simultaneous prenatal exposures are associated with BMI before age 2.

To explore the feasibility of using mixtures methods to assess the effects of multiple exposures as the neighborhood level, our objective in this study was to assess the independent and joint effects of prenatal environmental exposures on BMI in early childhood. We selected BMI at multiple time points before age 2 because it is an indicator of health in early life that is based on routinely collected data that can be abstracted from medical records. We hypothesized that higher prenatal exposures to adverse environmental and social factors would result in increased BMI z-scores in early childhood. We utilized data from two cohorts representing two different geographical areas: the Healthy Start cohort (metro Denver, Colorado) and the Maternal and Developmental Risks from Environmental and Social Stressors (MADRES) cohort (Los Angeles, California). Our motivation for using data from these two cohorts was to (1) explore a broader range of environmental and social exposures and (2) increase the socioeconomic diversity of included participants. This study provided an important opportunity to explore the feasibility of combining harmonized data from multiple cohorts in an environmental epidemiology study and the potential to apply these methods to the broader ECHO cohort. We considered several neighborhood factors previously linked with childhood obesity, including ambient air pollution, features of the built environment, and neighborhood socioeconomic status (SES). A secondary objective of our analysis was to determine the extent to which substituting geographic information system (GIS)-based estimates of traffic-related air pollution (TRAP) based on distance to major roads with site-specific, spatiotemporal (ST) predictions of TRAP results in models that are comparable and capture similar effects.

2. METHODS

2.1. Study Populations

Our study population included mother-child dyads in two longitudinal cohorts: Healthy Start and MADRES. Healthy Start and MADRES both participate in the Environmental influences on Child Health Outcomes (ECHO) Cohort, which combines data from 69 cohorts around the U.S. with the goal of understanding how the environment influences child health and development. Cohorts that are included in the ECHO cohort were originally established for a variety of reasons; a major function of the ECHO consortium is the retrospective harmonization of data collected independently by these cohorts (Blaisdell et al., 2021). Healthy Start is an ongoing longitudinal pre-birth cohort based in metro Denver, Colorado that initially recruited pregnant participants starting in 2009. The original goal of the study was to examine risk factors for childhood obesity and other metabolic health outcomes; the study has since expanded to include environmental determinants of health and additional health outcomes. Details on study recruitment and enrollment are available elsewhere (Harrod et al., 2014). Briefly, patients seeking prenatal care at the obstetrics clinic at the University of Colorado Hospital were eligible to enroll if they were 16 years of age or older, had no prior stillbirths, were expecting singleton births, and were <24 weeks of gestation. MADRES is a pregnancy cohort that enrolled participants from community health centers in Los Angeles, California starting in 2015. One of the primary goals of the MADRES study is to examine how cumulative environmental exposures influence childhood obesity risk. Details on the MADRES cohort are described elsewhere (Bastain et al., 2019). Participants were eligible to enroll in MADRES if they were 18 years of age or older, Spanish- or English-speaking, < 30 weeks of gestation, and expecting singleton births.

Participants in the Healthy Start and MADRES cohorts were eligible for this study if they lived within the geographical areas for which exposure data were available, if they had at least one study visit or medical record abstraction for which weight, length, and height data were collected, and if they had complete covariate data available. For both cohorts, we were interested in outcome data collected between delivery and age 24 months.

All participants provided informed consent at the time of their enrollment in their original cohort. Healthy Start protocols were approved by the Colorado Multiple Institution Review Board (protocol no. 09–0563). MADRES protocols were approved by the University of Southern California Institutional Review Board (protocol no. HS-15–00498).

2.2. Outcome

Our outcome of interest was standardized BMI scores in childhood. Due to the limited availability of data across both cohorts for all of our timepoints of interest, we focused on BMI in the first two years of life (at or before 24 months). BMI was calculated as weight/length² for participants and standardized as z-scores against the World Health Organization (WHO) reference population (World Health Organization, 2006). Because there is no accepted definition for overweight or obesity in children under the age of 2 years (Tiwari et al., 2024), we only included our outcome as a continuous variable and did not attempt to assess associations with overweight or obesity in these cohorts. We calculated BMI z-scores at four time points: birth, 6 months, 12 months, and 24 months. At each time point, we prioritized measurements collected at study visits over measurements obtained from medical record abstraction. The first Healthy Start study visit for infants occurred at delivery, with more than half of the participants completing a follow-up visit at age 6 months. Because the 12-month and 24-month time points were between Healthy Start study visits, we based outcomes at these time points on medical record abstraction when available. For MADRES participants, study visits occurred at one month and 12 months (Bastain et al., 2019); at the time of data collection for this study, few MADRES participants had completed a 24-month visit. For participants with multiple measurements around our timeframe of interest (4–6 months, 10–12 months, and 22–24 months), we used the measurement closest to the index age.

2.3. Exposure Assessment

We were interested in prenatal exposures that have been hypothesized to be associated with childhood BMI or obesity risk in previous studies (Aris et al., 2022; Bloemsma et al., 2019; Cáceres et al., 2023; Chiu et al., 2017; de Bont et al., 2020; Figaroa et al., 2023; Fleisch et al., 2015; Fossati et al., 2020; McConnell et al., 2015; Patterson et al., 2021; Rautava et al., 2021; Seo et al., 2020; Starling et al., 2020; Vazquez and Cubbin, 2020; Zhou et al., 2022). Our goal was to characterize long-term environmental and social conditions within the neighborhoods where participants lived during their pregnancies. We focused on the chemical environment (i.e., ambient and traffic-related air pollutants) that may directly impact a developing fetus and features of the built, natural, or social environment (e.g., park space, food access) that impact the ability to have a healthy pregnancy via pathways such as physical activity, healthy eating, contact with nature, or stress. We included variables for which publicly available data with national coverage were available for both cohorts. A table summarizing all of our exposure variables is included in the Supplement (Table S1).

Exposures were assigned to each study participant based on their residence during pregnancy. Exposures were based either on the point location of the residence or the census tract in which the residence was located. Healthy Start exposures were based on the address at the time of enrollment. Because daily residential history data during the prenatal period were available for MADRES participants, exposures for these participants were time-weighted based on the time spent at each location.

Exposures were summarized across each pregnancy using the estimated date of conception (delivery date minus gestational age in days) and date of delivery. For more than half of our exposures, only data averaged at the annual (or longer) period were available (Table S1). Therefore, our primary focus was on exposure data that characterized long-term exposures during pregnancy. This was done to allow for the greatest possible consistency between exposures across both cohorts. Our goal was to capture the overall environmental quality of the neighborhood by describing environmental conditions across multiple domains; long-term averages for our exposure data therefore primarily reflect spatial differences in neighborhood quality.

For a select few exposures, data were assessed at shorter timeframes and included in supplemental analyses. For Healthy Start, trimester-specific exposure estimates were calculated for air pollutant exposures derived from monitoring data and TRAP exposures derived from spatiotemporal models. For MADRES, trimester-specific exposure estimates were calculated for air pollutant exposures derived from monitoring data and TRAP exposures derived from spatiotemporal models and for participants who moved during pregnancy (due to spatial differences in exposures where only long-term data are available).

For area-based variables, we selected a 500-m (0.3 mile) circular buffer to represent a reasonable walking distance during pregnancy. Although people who walk for transportation often walk farther than 0.25 miles regularly (Yang and Diez-Roux, 2012), physical activity research has found that pregnant people tend to walk less than recommended amounts (Connolly et al., 2019; Kim and Chung, Eunhee, 2015). Therefore, we opted to use the typical buffer distance to characterize the environment around the homes of participants during their pregnancies.

2.3.1. Ambient Air Pollutants and Meteorology

We obtained daily ambient concentrations of fine particulate matter (aerodynamic diameter <2.5 μm, PM_2.5), ozone (O₃), and nitrogen dioxide (NO₂) from local monitoring networks (US Environmental Protection Agency, 2019). PM_2.5 and NO₂ data were summarized as daily mean concentrations, and O₃ data were summarized as daily 8-hour max concentrations. We estimated daily PM_2.5, O₃, and NO₂ concentrations at each residential location using inverse-distance weighting (IDW) with a power of 2. Data from all monitors within 50 km of each residential location were included in the calculations.

Daily temperature and relative humidity data were obtained from the Climatology Lab gridMET data set (Abatzoglou, 2013). Residential exposures were based on the gridded (4 km x 4 km) estimates of daily minimum and maximum temperature and relative humidity. We averaged the daily minimum and maximum values to estimate the mean temperature and relative humidity for each location each day.

2.3.2. Traffic-Related Air Pollutants

We used two different approaches to estimate exposure to TRAP. First, we used distance to major roadways as a proximity-based proxy measure of TRAP exposure. For Healthy Start, we calculated the minimum distance to a major road using the U.S. Census Bureau TIGER/Line shapefile for primary and secondary roads (US Census Bureau, 2022). For MADRES, we calculated the minimum distance to a class 1 or class 2 road (generally corresponding to freeways and highways in Los Angeles, California) using the Streetlytics data from Citilabs.

Second, we estimated TRAP exposures from spatiotemporal prediction models developed for each cohort. Our use of these modeled estimates of TRAP pollutants was motivated by work highlighting the limitations of proximity-based, spatial interpolation, and land-use regression approaches for assessing TRAP exposures, including that these methods often miss important intraurban gradients in exposure (Jerrett et al., 2005; Khreis and Nieuwenhuijsen, 2017; Xie et al., 2017). These spatiotemporal prediction models estimate pollutant concentrations at the residential (point) level. For Healthy Start, we based TRAP exposures on a prediction model for black carbon (BC) described elsewhere (Martenies et al., 2021). Briefly, we collected weekly PM_2.5 samples using low-cost monitors and measured BC mass using transmissometry (Ahmed et al., 2009; Presler-Jur et al., 2017). Our model was developed using the ‘SpatioTemporal’ package in R (Lindstrom et al., 2019) and featured 13 spatial covariates, two spatiotemporal covariates, and two temporal basis functions derived from regional NO₂, PM_2.5, and BC monitoring data. The weekly BC prediction models performed well, with a 10-fold cross-validation R² value of 0.83 and a root mean square error (RMSE) of 0.10 μg/m³. The prediction model generated weekly BC concentrations at Healthy Start residential locations.

For MADRES, we based TRAP exposures on a prediction model for oxides of nitrogen (NO_x) described elsewhere (Li et al., 2019, 2017). Briefly, the model used a flexible, three-stage hierarchical framework with spatiotemporally referenced covariates and data from both long-term routine monitoring stations and short-term, sporadic measurement campaigns. Temporal basis functions were fit to capture seasonality and longer-term temporal variation (Szpiro et al., 2010). The second stage of the model utilized ensemble learning (bootstrap aggregation) to increase the precision of estimates and output their standard deviation (SD) while the third constrained optimization stage adjusted the parameter estimates based on real-life physical and chemical properties and trends. The model performance according to ensemble-based cross-validation was high, with R² and RMSE values of 0.88 and 9.6 ppb, respectively. The prediction model generated weekly NO_x concentrations at MADRES residential locations.

2.3.2. Green Space and Built Environment

We used the normalized difference vegetation index (NDVI) to characterize green space around each residential location. NDVI data at a 250-m resolution are available every 16 days (Didan et al., 2015). Raster files were averaged by year to generate an annual NDVI estimate. Then, the average NDVI value within a 500-m buffer were calculated for each residential location. For Healthy Start participants, NDVI values were based on the year of birth. For MADRES participants, NDVI values were based on data from 2017.

We calculated three variables related to park space: distance to the closest park; total park area within a buffer, and number of parks within a buffer distance. These three variables were used to represent different characteristics that may influence park usage (e.g., proximity to parks, park area, and park count) (Kaczynski et al., 2020) as well as the ability for green space like parks to mitigate outdoor particulate matter concentrations (e.g., park area) (Chen et al., 2019; Lei et al., 2021). We opted to use complementary measures of park access because we cannot directly measure personal exposures to parks in this study and the most appropriate proxy was not known a priori. For each cohort, we used TIGER/Line landmark area shapefiles from the U.S. Census Bureau to identify park spaces (US Census Bureau, 2022). Distance to the closest park (meters) was defined as the minimum distance between the residential location (point) and the boundary of any nearby park. Total park area (meters squared) was defined as the total park area that fell within a 500-m buffer around the residence. The number of parks was defined as the sum of park polygons that (at least partially) intersected with the buffer.

To further characterize the built and natural environment, we used two variables from the National Land Cover Database (NLCD): the percentage of impervious surface and the percentage of tree canopy (Multi-Resolution Land Characteristics Consortium, 2017). NLCD data are available at a 50-m resolution for the conterminous U.S. The mean percentage of impervious surface and the mean percentage of tree canopy in a 500-m buffer were calculated for Healthy Start and MADRES participants using NLCD data from 2011 and 2016, respectively.

2.3.5. Neighborhood Socioeconomic Status

Three variables were used as measures of neighborhood SES. First, we used the Food Access Research Atlas (FARA) to determine whether participants lived in a low-income and low–food access census tract (US Department of Agriculture, 2020). Data from the 2015 and 2019 releases were used to assign values to Healthy Start and MADRES participants, respectively. We used the FARA variable that defined low income and low access using the half-mile designation for urban areas and 10 miles for nonurban areas. Then, we derived two socioeconomic indicators using data from the American Community Survey (US Census Bureau, 2014): the percentage of the population aged 25 years or older without at least a high school diploma or equivalent (low educational attainment) and the percentage of households with a past year income below the federal poverty level (households in poverty). ACS 5-year estimates for 2010–2014 were used for Healthy Start and for 2015–2019 for MADRES.

2.4. Covariates

Covariates were selected a priori and were informed by a directed acyclic graph (Figure S1). We included maternal age at delivery (continuous), maternal educational attainment (categorical: less than high school, high school graduate, some college, bachelor’s degree, master’s degree or higher), living with a partner during pregnancy (binary), maternal pre-pregnancy BMI (categorical: underweight, normal weight, overweight, class I obesity, class II obesity, class III obesity), prenatal tobacco use (binary) and secondhand smoke (SHS) exposure (binary), infant sex, and season of conception (categorical: spring, summer, fall, winter).

2.5. Statistical Analysis

Prior to fitting our models, continuous exposures were scaled to a mean of 0 and an standard deviation (SD) of 1. We examined correlations between exposure variables and between the exposure variables and BMI z-scores at each time points. We used scatter plots to visualize the potential relationships between individual exposures and BMI z-scores.

We then used two complementary statistical approaches to examine the independent and joint effects of our exposures of interest on BMI z-scores. First, we fit single-exposure models using linear regression. The purpose of the single-exposure models was to determine if relationships between our exposures of interested and health outcomes were generally consistent across cohorts and time points. We used the single-exposure models to determine the feasibility of combining data across cohorts. Second, we estimated the effect of the exposure mixture using quantile-based g-computation (Keil et al., 2020). Quantile-based g-computation is a generalization of weighted quantile sum regression (WQS), which estimates the marginal effect of an index developed using quantized exposure variables on an outcome of interest. Importantly, quantile-based g-computation differs from WQS regression in that it does not assume all exposures within the mixture have the same directional effect on the outcome (Keil et al., 2020). This was an important feature for our study as we did not have a priori hypotheses about the direction of association for all components of our mixture. Quantile-based q-computation generates an overall effect estimate ψ for the mixture and a set of weights for each individual exposure within the mixture. The overall coefficient ψ is interpreted as the marginal effect on the outcome of interest when increasing every component of the exposure mixture by one quantile. In our implementation of quantile-based g-computation regression, the exposures were quantized as quartiles.

Because of the limited number of exposures assessed by trimester, models stratified by trimester were considered exploratory. We present the results of these models in the Supplemental Materials

We fit our mixtures models using two different sets of exposures. We first used the set of exposure variables that included TRAP exposures based on distance to major roadways. Then, we replaced the proximity-based TRAP indicator in the mixture with cohort-specific estimates of BC for Healthy Start participants and NO_x for MADRES participants. The purpose of these second models was to see if increasing the specificity of TRAP exposure estimates improved the precision of our results and if these highly spatially and temporally resolved estimates of TRAP provided comparable results across cohorts. Due to the exploratory nature of our work, we did not adjust p-values for multiple comparisons (Althouse, 2016; Bender and Lange, 2001); we considered our results as preliminary to avoid overinterpretation of our findings.

All statistical analyses were conducted using R version 4.3.0 (R Core Team, 2023). We implemented spatial methods using the “sf” package (Pebsema, 2018). Plots were generated using the “ggplot2,” “ggcorrplot,” and “patchwork” packages (Kassambara, 2019; Pedersen, 2020; Wickham, 2016).

2.5.1. Single-Exposure Models

For each exposure, we fit single-exposure linear regression models to estimate associations with BMI z-scores separately for each cohort. Our single-exposure models were adjusted for all relevant covariates (see section 2.4). Results were reported for an SD increase in continuous exposure variables and a one unit increase in discrete or binomial exposure variables. We fit models for pregnancy-wide exposures and trimester-specific exposures separately.

2.5.2. Quantile-Based G-computation Models

Quantile g-computation models were fit using the “qgcomp” package in R (Keil, 2022). Because we observed differences in which exposures were associated with BMI by cohort in the single-exposure models, we fit quantile g-computation models for each cohort separately. Quantile g-computation models for continuous BMI z-score were adjusted for the same individual-level covariates as the single-exposure models. Pregnancy-wide and trimester-specific exposures were modeled separately. We examined both the weights assigned to each exposure in the index during the first stage of the g-computation algorithm and the overall effect estimate ψ. We drew our estimates of the weights of each mixture component to models fit without bootstrapping. For the overall effect estimate, we estimated the variance around ψ using nonparametric bootstrapping with n=1000 iterations and calculated a 95% confidence interval (CI).

3. RESULTS

3.1. Participant Characteristics, BMI Z-scores, and Exposures

A total of 1,300 participants from the two cohorts (n=997 (75% of potential participants) from Healthy Start and n=303 (50% of potential participants) from MADRES) had complete exposure, outcome, and covariate data at delivery and were included in this study (Table 1, Figure S2). Combined sample sizes were lower for later time points (n=1174, 865, and 453 at 6 months, 12 months, and 24 months, respectively). Healthy Start participants were mostly non-Hispanic White (52.5%) with high educational attainment (45.5% had a bachelor’s degree or higher). MADRES participants were mostly Hispanic (79.2%) and had lower educational attainment (55.2% had a high school diploma or less formal education).

Table 1. Summary of demographic and outcome variables by time point and cohort.

Variable		Birth		6 months		12 months		24 months
		Healthy Start	MADRES	Healthy Start	MADRES	Healthy Start	MADRES	Healthy Start	MADRES
N		997	303	969	205	712	153	416	37
Parent age at delivery (years)	Mean (SD)	27.9 (6.1)	29.6 (5.9)	27.9 (6.2)	29.8 (5.9)	28.4 (6.2)	30.0 (5.7)	29.1 (6.2)	30.4 (5.2)
Parent living with partner	N (%)	785 (78.7)	296 (97.7)	753 (77.7)	203 (99.0)	568 (79.8)	150 (98.0)	351 (84.4)	35 (94.6)
Race and ethnicity
Hispanic	N (%)	243 (24.4)	240 (79.2)	242 (25.0)	176 (85.9)	163 (22.9)	128 (83.7)	79 (19.0)	33 (89.2)
Non-Hispanic Black	N (%)	163 (16.3)	29 (9.6)	155 (16.0)	13 (6.3)	99 (13.9)	10 (6.5)	56 (13.5)	2 (5.4)
Non-Hispanic White	N (%)	523 (52.5)	21 (6.9)	511 (52.7)	11 (5.4)	401 (56.3)	12 (7.8)	252 (60.6)	2 (5.4)
Other Race or Ethnicity	N (%)	68 (6.8)	13 (4.3)	61 (6.3)	5 (2.4)	49 (6.9)	3 (2.0)	29 (7.0)	0 (0.0)
Parent educational attainment
No high school diploma	N (%)	138 (13.8)	72 (23.8)	135 (13.9)	52 (25.4)	77 (10.8)	40 (26.1)	43 (10.3)	9 (24.3)
High school graduate	N (%)	184 (18.5)	95 (31.4)	174 (18.0)	67 (32.7)	115 (16.2)	48 (31.4)	58 (13.9)	15 (40.5)
Some college	N (%)	222 (22.3)	83 (27.4)	213 (22.0)	47 (22.9)	139 (19.5)	37 (24.2)	63 (15.1)	7 (18.9)
Bachelor’s degree	N (%)	230 (23.1)	32 (10.6)	218 (22.5)	24 (11.7)	186 (26.1)	15 (9.8)	121 (29.1)	3 (8.1)
Master’s degree or higher	N (%)	223 (22.4)	21 (6.9)	229 (23.6)	15 (7.3)	195 (27.4)	13 (8.5)	131 (31.5)	3 (8.1)
Pre-pregnancy BMI
Underweight	N (%)	31 (3.1)	6 (2.0)	31 (3.2)	2 (1.0)	19 (2.7)	2 (1.3)	16 (3.8)	1 (2.7)
Normal	N (%)	513 (51.5)	88 (29.0)	508 (52.4)	49 (23.9)	386 (54.2)	39 (25.5)	237 (57.0)	7 (18.9)
Overweight	N (%)	250 (25.1)	99 (32.7)	238 (24.6)	70 (34.1)	178 (25.0)	61 (39.9)	100 (24.0)	14 (37.8)
Class 1 obesity	N (%)	124 (12.4)	71 (23.4)	116 (12.0)	51 (24.9)	80 (11.2)	32 (20.9)	33 (7.9)	9 (24.3)
Class 2 obesity	N (%)	47 (4.7)	24 (7.9)	43 (4.4)	21 (10.2)	29 (4.1)	11 (7.2)	16 (3.8)	4 (10.8)
Class 3 obesity	N (%)	32 (3.2)	15 (5.0)	33 (3.4)	12 (5.9)	20 (2.8)	8 (5.2)	14 (3.4)	2 (5.4)
Any smoking during gestation	N (%)	87 (8.7)	10 (3.3)	84 (8.7)	5 (2.4)	43 (6.0)	3 (2.0)	20 (4.8)	0 (0.0)
SHS exposure during gestation	N (%)	238 (23.9)	21 (6.9)	234 (24.1)	10 (4.9)	145 (20.4)	7 (4.6)	72 (17.3)	1 (2.7)
Male infant sex	N (%)	508 (51.0)	147 (48.5)	495 (51.1)	100 (48.8)	375 (52.7)	82 (53.6)	227 (54.6)	18 (48.6)
Season of conception	N (%)
Winter	N (%)	231 (23.2)	90 (29.7)	225 (23.2)	57 (27.8)	163 (22.9)	39 (25.5)	98 (23.6)	12 (32.4)
Spring	N (%)	225 (22.6)	82 (27.1)	218 (22.5)	54 (26.3)	163 (22.9)	42 (27.5)	78 (18.8)	8 (21.6)
Summer	N (%)	245 (24.6)	66 (21.8)	234 (24.1)	50 (24.4)	175 (24.6)	35 (22.9)	106 (25.5)	11 (29.7)
Fall	N (%)	296 (29.7)	65 (21.5)	292 (30.1)	44 (21.5)	211 (29.6)	37 (24.2)	134 (32.2)	6 (16.2)
BMI z-score	Mean (SD)	−0.5 (1.0)	−0.4 (1.7)	−0.3 (1.1)	0.3 (1.1)	0.0 (1.1)	0.5 (1.1)	0.0 (1.2)	1.3 (1.3)
BMI assessed via medical record	N (%)	0 (0.0)		379 (39.1)		711 (99.9)		416 (100.0)

BMI, body mass index; SD, standard deviation; SHS, secondhand smoke.

BMI z-scores for both cohorts were similar at the delivery time point but diverged by the time participants were 6 months old (Figure 1). Mean (SD) BMI z-scores for Healthy Start and MADRES participants were lowest at the delivery time point (−0.5 [1.0] and −0.4 [1.7], respectively) (Table 1). For Healthy Start participants, mean (SD) BMI z-scores increased slightly to 0.0 (1.2) by the 24-month time point. For MADRES participants, mean BMI z-scores demonstrated greater increases over time, reaching a mean (SD) value of 0.5 (1.1) by 12 months and 1.3 (1.3) by 24 months (although the value at 24 months was based on a much smaller sample size of n=37).

Figure 1. Box plots for the distribution of body mass index (BMI) z-scores by cohort at each of the study time points.

Exposure patterns across pregnancy differed by cohort (Table 2). MADRES participants had higher mean (SD) PM_2.5 exposures (12.1 [1.2] μg/m³) compared with Healthy Start participants (7.9 [0.6] μg/m³). However, higher ozone exposures were observed for the Healthy Start participants (22.4 [2.1] ppb) compared with MADRES participants (16.8 [3.0] ppb). Although both cohorts had on average the same mean number of parks within a 500-m buffer around their residents (0.4 [0.7] and 0.4 [0.6] for Healthy Start and MADRES, respectively), other park variables differed. MADRES participants lived in neighborhoods with higher poverty (54.5% [15.6%]) and in census tracts with lower educational attainment (40.1% [16.4%]) than Healthy Start participants (14.6% [10.8%] and 15.5% [12.5%], respectively). The proportion of participants in Healthy Start and MADRES living in census tracts with low incomes and low food access was similar (48.4% and 43.6%, respectively).

Table 2. Summary of exposure variables averaged across pregnancy for each cohort.

Variable	Units		Healthy Start (n=997)	MADRES (n=303)
PM2.5	μg/m3	Mean (SD)	7.9 (0.6)	12.1 (1.2)
O3	ppb	Mean (SD)	22.4 (2.1)	16.8 (3.0)
NO2	ppb	Mean (SD)	43.3 (4.1)	42.6 (3.5)
Relative humidity	%	Mean (SD)	44.5 (3.9)	60.9 (4.5)
Temperature	°C	Mean (SD)	10.9 (2.7)	19.4 (1.2)
Black carbon	μg/m3	Mean (SD)	1.7 (0.2)	NA
NOx	ppb	Mean (SD)	NA	18.3 (9.2)
Distance to major road	m	Mean (SD)	1174.0 (988.8)	1269.2 (1075.2)
NDVI (500 m)		Mean (SD)	0.3 (0.0)	0.2 (0.1)
Distance to parks	m	Mean (SD)	1522.8 (1908.6)	769.7 (493.6)
Park area (500 m)	m2	Mean (SD)	15,191.7 (40,933.1)	8,689.0 (17,977.1)
Park count (500 m)	N	Mean (SD)	0.4 (0.7)	0.4 (0.6)
Tree cover (500 m)	%	Mean (SD)	3.5 (3.3)	1.3 (1.6)
Impervious surface (500 m)	%	Mean (SD)	47.2 (11.4)	74.4 (9.6)
LILA census tracta		N (%)	483 (48.4)	132 (43.6)
Households in poverty	%	Mean (SD)	14.6 (10.8)	54.5 (15.6)
Low educational attainmentb	%	Mean (SD)	15.5 (12.7)	40.1 (16.4)

^aLow income and low food access census tract designation.

^bPercentage of the population 25 years of age or older without a high school diploma or equivalent.

LILA, low income and low access; NDVI, normalized difference vegetation index; SD, standard deviation.

Correlations between prenatal exposure variables ranged from weak to strong (Figure S3). As expected, air pollutant and meteorological exposures were moderately to strongly correlated, although patterns differed by cohort. Among Healthy Start participants, PM_2.5 exposures were weakly correlated with NO₂ (r=0.3), whereas the correlation was stronger for MADRES participants (r=0.6). For both cohorts, O₃ was negatively correlated with PM_2.5 and NO₂, although correlations were stronger for Healthy Start participants (r=−0.4 and −0.7, respectively) than for MADRES participants (r=−0.1 and −0.3, respectively). Correlations between the social stressors were as expected. For both cohorts, the correlation between low educational attainment and household poverty was high (r=0.7 for Healthy Start and 0.8 for MADRES). For MADRES, our indicators of neighborhood SES (educational attainment and households in poverty) were negatively correlated with percent tree cover (r = −0.5 and −0.5, respectively) and NDVI (r = −0.5, −0.6, respectively) and positively correlated with the percentage of impervious surface (r = 0.5 and 0.5, respectively).

Despite being conceptualized as indicators of TRAP, correlations between distance to major roads and either BC or NO_x were weak for both cohorts. For Healthy Start, BC was weakly negatively correlated with distance to major roads (r=−0.1). For MADRES, NO_x was also weakly negatively correlated with distance to major roads (r=−0.2).

Patterns of exposure for both cohorts were similar when considering exposures averaged by trimester (Table S2, Figure S4).

3.2. Single-Exposure Linear Regression Models

In our single-exposure models adjusted for covariates, only a few pregnancy-wide exposures were associated with early childhood BMI (Table 3).

Table 3. Change in BMI z-score in early childhood per standard deviation increase (continuous variables) or one unit change (binary/discrete variables) in pregnancy-wide exposure by cohort and time period.

Exposure	Birth	6 Months	12 Months	24 Months
	Healthy Start β (95% confidence interval)a
PM2.5	0.02 (−0.05, 0.08)	−0.05 (−0.12, 0.02)	−0.07 (−0.15, 0.00)	−0.07 (−0.18, 0.04)
O3	−0.01 (−0.10, 0.08)	0.02 (−0.07, 0.12)	0.03 (−0.08, 0.13)	−0.13 (−0.28, 0.03)
NO2	−0.04 (−0.11, 0.03)	−0.09 (−0.18, −0.01)	−0.08 (−0.17, 0.01)	−0.08 (−0.22, 0.05)
Relative humidity	0.03 (−0.04, 0.09)	0.03 (−0.04, 0.10)	−0.04 (−0.12, 0.03)	0.00 (−0.12, 0.11)
Temperature	−0.04 (−0.14, 0.07)	−0.07 (−0.19, 0.05)	0.03 (−0.10, 0.17)	−0.04 (−0.23, 0.16)
Black carbon	−0.03 (−0.10, 0.04)	−0.05 (−0.14, 0.03)	−0.02 (−0.12, 0.07)	−0.06 (−0.20, 0.07)
Distance to major road	−0.05 (−0.11, 0.01)	−0.02 (−0.09, 0.05)	0.01 (−0.07, 0.08)	−0.09 (−0.19, 0.02)
NDVI (500 m)	0.01 (−0.05, 0.06)	0.00 (−0.06, 0.07)	−0.02 (−0.10, 0.05)	−0.10 (−0.20, 0.01)
Distance to parks	0.04 (−0.02, 0.10)	0.01 (−0.05, 0.08)	−0.06 (−0.13, 0.02)	−0.13 (−0.24, −0.02)
Park area (500 m)	0.01 (−0.05, 0.07)	−0.01 (−0.07, 0.06)	−0.01 (−0.08, 0.07)	0.08 (−0.03, 0.18)
Park count (500 m)	−0.01 (−0.10, 0.07)	−0.07 (−0.16, 0.03)	−0.01 (−0.12, 0.10)	0.20 ( 0.04, 0.36)
Tree cover (500 m)	−0.02 (−0.08, 0.04)	0.01 (−0.05, 0.08)	0.00 (−0.07, 0.08)	−0.10 (−0.21, 0.00)
Impervious surface (500 m)	−0.02 (−0.08, 0.04)	0.01 (−0.06, 0.08)	0.04 (−0.04, 0.11)	0.12 ( 0.01, 0.23)
LILA census tract	0.02 (−0.11, 0.14)	0.03 (−0.12, 0.17)	−0.03 (−0.19, 0.14)	0.14 (−0.08, 0.37)
Households in poverty	−0.06 (−0.12, 0.00)	−0.02 (−0.10, 0.05)	−0.02 (−0.10, 0.06)	0.11 ( 0.00, 0.22)
Low educational attainment	−0.04 (−0.11, 0.03)	−0.03 (−0.11, 0.04)	−0.02 (−0.11, 0.07)	0.08 (−0.04, 0.20)
	MADRES β (95% confidence interval)a
PM2.5	−0.16 (−0.35, 0.02)	0.03 (−0.13, 0.20)	0.07 (−0.12, 0.25)	0.25 (−0.31, 0.82)
O3	0.15 (−0.05, 0.36)	−0.23 (−0.41, −0.05)	−0.02 (−0.25, 0.20)	0.10 (−0.58, 0.77)
NO2	−0.04 (−0.24, 0.17)	0.03 (−0.14, 0.19)	0.09 (−0.11, 0.29)	0.58 (−0.21, 1.37)
Relative humidity	0.00 (−0.17, 0.18)	0.13 (−0.02, 0.28)	0.00 (−0.17, 0.17)	−0.13 (−0.56, 0.29)
Temperature	0.12 (−0.14, 0.39)	−0.19 (−0.40, 0.02)	−0.05 (−0.32, 0.22)	0.37 (−0.32, 1.07)
Oxides of nitrogen	−0.12 (−0.33, 0.09)	0.02 (−0.17, 0.21)	0.03 (−0.17, 0.22)	−0.02 (−0.63, 0.58)
Distance to major road	0.07 (−0.09, 0.23)	0.09 (−0.06, 0.24)	−0.06 (−0.22, 0.09)	0.39 (−0.14, 0.93)
NDVI (500 m)	−0.11 (−0.30, 0.07)	−0.11 (−0.28, 0.06)	−0.13 (−0.33, 0.08)	0.00 (−0.61, 0.60)
Distance to parks	−0.03 (−0.19, 0.14)	0.02 (−0.12, 0.16)	−0.20 (−0.35, −0.05)	−0.23 (−0.68, 0.21)
Park area (500 m)	−0.01 (−0.18, 0.16)	0.05 (−0.10, 0.21)	0.05 (−0.15, 0.26)	0.24 (−0.42, 0.90)
Park count (500 m)	0.05 (−0.22, 0.31)	−0.01 (−0.22, 0.21)	0.12 (−0.13, 0.38)	0.14 (−0.42, 0.70)
Tree cover (500 m)	−0.04 (−0.22, 0.14)	−0.07 (−0.24, 0.10)	−0.02 (−0.22, 0.19)	−0.08 (−0.63, 0.46)
Impervious surface (500 m)	0.05 (−0.13, 0.23)	0.11 (−0.06, 0.28)	0.09 (−0.10, 0.29)	−0.01 (−0.76, 0.73)
LILA census tract1	−0.16 (−0.50, 0.17)	−0.17 (−0.45, 0.12)	−0.08 (−0.40, 0.24)	−0.45 (−1.32, 0.43)
Households in poverty	−0.10 (−0.28, 0.09)	0.16 (−0.02, 0.33)	−0.02 (−0.20, 0.17)	−0.30 (−0.87, 0.26)
Low educational attainment2	0.03 (−0.16, 0.22)	0.18 ( 0.00, 0.35)	0.08 (−0.11, 0.27)	0.13 (−0.44, 0.69)

^aResults are based on multiple linear regression models adjusted for maternal age at delivery (years), mother living with a partner, maternal educational attainment, maternal pre-pregnancy BMI, smoking and secondhand smoke exposure during pregnancy, infant sex, and season of conception.

BMI, body mass index; LILA, low income and low access; NDVI, normalized difference vegetation index.

For Healthy Start participants, the strongest associations were observed between prenatal exposures and BMI z-scores at 24 months. At 24 months, an SD increase in the minimum distance to parks during the prenatal period and an SD increase in the percentage of tree cover were associated with a 0.13 unit decrease (95% CI: −0.25, −0.02) and a 0.10 unit decrease (95% CI: −0.21, 0.00) in BMI z-scores, respectively. In the MADRES cohort, the strongest associations were observed at 6 and 12 months. An SD increase in prenatal O₃ exposure was associated with a 0.23 unit decrease (95% CI: −0.41, −0.05) in BMI z-scores at 6 months. Conversely, an SD increase in the percentage of residents over the age of 25 without a high school diploma was associated with a 0.18 unit increase (95% CI: 0.00, 0.35) in BMI z-scores at 6 months. We observed additional negative associations between air pollutants and meteorological variables when considering trimester-specific exposures for the Healthy Start cohort (Table S3) and the MADRES cohort (Table S4). Additional details on the models using trimester-specific exposures are included in the Supplemental Material (Text S1).

For both cohorts, using a more cohort-specific spatiotemporal measure of TRAP exposure did not improve associations with BMI z-scores at any of the time points assessed. For both exposures in both cohorts, CIs tended to be wide and included the null value (Table S3, Table S4).

Because associations between exposures and BMI z-scores in the single-exposure models were inconsistent between cohorts, we chose to model the effects of the exposure mixture for each cohort separately.

3.3. Quantile-Based G-computation Models

Overall, the quantile-based g-computation models suggested no significant effect of the prenatal exposure mixture on BMI z-scores at any of the time points assessed (Table 4). The estimates of the overall effect of the pregnancy-wide mixture were in the negative direction for all time points and both cohorts, but CIs were wide and included the null value. The results for each cohort were similar whether we used distance to major roads or the spatiotemporal predictor of TRAP in our exposure mixture (Table 4).

Table 4. Marginal effect of a one unit change (i.e., a one quartile change) in all exposures simultaneously on BMI z-scores estimated by the quantile-based g-computation models by cohort, exposure period, time point, and exposure data set.

Exposure period	TRAP	Birth	6 months	12 months	24 months
		Healthy Start ψ (95% confidence interval)a
Pregnancy	Proximityb	−0.30 (−0.74, 0.14)	−0.07 (−0.59, 0.44)	−0.43 (−1.00, 0.14)	−0.57 (−1.40, 0.26)
First trimester		−0.37 (−0.82, 0.09)	−0.20 (−0.74, 0.34)	−0.36 (−0.98, 0.27)	−0.81 (−1.63, 0.01)
Second trimester		−0.20 (−0.66, 0.26)	0.01 (−0.53, 0.54)	−0.51 (−1.11, 0.10)	−0.62 (−1.52, 0.29)
Third trimester		−0.17 (−0.61, 0.27)	0.03 (−0.54, 0.60)	−0.38 (−1.00, 0.24)	−0.55 (−1.44, 0.34)
Pregnancy	STc	−0.23 (−0.68, 0.21)	−0.10 (−0.61, 0.42)	−0.38 (−0.95, 0.19)	−0.54 (−1.35, 0.27)
First trimester		−0.34 (−0.80, 0.12)	−0.26 (−0.81, 0.30)	−0.38 (−1.00, 0.24)	−0.75 (−1.60, 0.09)
Second trimester		−0.13 (−0.60, 0.33)	−0.05 (−0.59, 0.50)	−0.43 (−1.04, 0.18)	−0.66 (−1.54, 0.21)
Third trimester		−0.15 (−0.60, 0.29)	0.00 (−0.57, 0.58)	−0.29 (−0.91, 0.32)	−0.40 (−1.28, 0.49)
		MADRES ψ (95% confidence interval)a
Pregnancy	Proximityb	−0.97 (−2.53, 0.58)	−0.57 (−1.75, 0.61)	−1.26 (−2.62, 0.09)	−1.39 (−8.05, 5.27)
First trimester		−2.02 (−3.96, −0.07)	−0.54 (−1.79, 0.72)	−1.76 (−3.11, −0.40)	−0.17 (−5.88, 5.54)
Second trimester		−1.24 (−2.75, 0.27)	−0.15 (−1.38, 1.08)	−0.61 (−1.94, 0.73)	0.52 (−3.52, 4.56)
Third trimester		−1.14 (−2.56, 0.27)	−0.25 (−1.43, 0.93)	−0.85 (−2.32, 0.61)	0.00 (−8.47, 8.48)
Pregnancy	STc	−1.03 (−2.68, 0.61)	−0.68 (−1.91, 0.55)	−1.23 (−2.68, 0.23)	0.45 (−7.13, 8.03)
First trimester		−2.15 (−4.18, −0.12)	−0.65 (−1.93, 0.64)	−1.86 (−3.27, −0.45)	−0.79 (−6.40, 4.83)
Second trimester		−1.38 (−2.95, 0.18)	−0.02 (−1.27, 1.22)	−0.50 (−1.94, 0.95)	0.13 (−3.96, 4.22)
Third trimester		−1.12 (−2.57, 0.34)	−0.32 (−1.53, 0.88)	−1.03 (−2.55, 0.49)	0.12 (−7.97, 8.22)

^aConfidence intervals are based on nonparametric bootstrapping using 1,000 iterations.

^bExposure data set includes the proximity-based indicator of traffic-related air pollution (distance to major roads).

^cExposure data set includes the spatiotemporal predictor of traffic-related air pollution (black carbon for Healthy Start participants and oxides of nitrogen for MADRES participants).

GIS, geographic information system; ST, spatiotemporal; TRAP, traffic-related air pollution.

For the MADRES cohort, we observed an effect of exposures in the first trimester on BMI z-scores (Table 4). When the exposure mixture included distance to roads as the TRAP indicator, the marginal effect of a 1 quartile increase in all first-trimester exposures was a 2.02 unit decrease (95% CI: −3.96, −0.07) in BMI z-scores at birth and a 1.76 unit decrease (95% CI: −3.11, −0.40) in BMI z-scores at 12 months. Similar results were observed when using NO_x as the TRAP indicator in the exposure mixture. No trimester-specific effects were observed for the Healthy Start participants.

The weights estimated by the first stage of the g-computation model–fitting algorithm suggested differences in the direction of associations between our exposures of interest and BMI z-scores across time periods and between the two cohorts (Figure 2, Figure S5, Figure S6). For example, the weights for PM_2.5, NO₂, O₃, distance to major roads, distance to park, park area, % tree cover, and census tract educational attainment were in opposite directions for Healthy Start and MADRES cohorts at a given time point. Similar patterns emerged when considering the spatiotemporal indicator of TRAP (BC for Healthy Start and NO_x for MADRES).

Figure 2. Estimated weights for each exposure in the mixture by cohort and time point based on g-computation models fit using the geographic information system (GIS)-based indicator (distance to major roads) and spatiotemporal indicators (black carbon for Healthy Start and oxides of nitrogen for MADRES participants) traffic-related air pollution (TRAP) exposure.

Weights are representative of the relative contribution of each exposure to the positive or negative component of the marginal effect estimate. The negative and positive weights for each model total −1 and 1, respectively.

4. DISCUSSION

Here we examined whether a mixture of neighborhood-level environmental exposures and social determinants of health experienced during the prenatal period was associated with BMI assessed at birth through age 24 months. In single-exposure models, we observed associations between air pollutants, features of the built environment, and indicators of neighborhood SES and BMI z-scores at each of our time points. These associations tended to be in the negative direction, where higher exposures were associated with lower BMI z-scores, but associations were not consistent across cohorts. The g-computation models we used suggested negative associations between our overall mixture of prenatal exposures and BMI z-scores at or before age 24 months, although confidence intervals were wide, and our results should be interpreted with caution. Our analysis highlights some of the challenges associated with examining the effects of multiple neighborhood-level exposures on early childhood BMI measurements within two differing cohorts.

There remains inconsistency in the literature regarding the impacts of individual exposures on BMI in early childhood. Our findings of an inverse association between air pollutants and BMI z-scores in the single-exposure models differ from previous studies. Prior work has reported associations between increased exposure to PM_2.5 and higher BMI z-scores or increased risk of overweight or obesity (Bloemsma et al., 2019; Mao et al., 2017; Zhou et al., 2023). Longitudinal studies beyond 24 months also support positive associations between NO₂, PM_2.5, and childhood BMI (Bloemsma et al., 2019; de Bont et al., 2021). However, the direction of associations is not consistent across studies. For example, a recent study using data from the MADRES cohort reported mixed findings for weight trajectories across different early life periods (third trimester to 6 months, 6 months to two years) (Ji et al., 2023). Differences in our study results may be due to differences in the constituents of particulate matter or exposure levels (e.g., particulate matter exposures in Denver are well below the National Ambient Air Quality Standard in effect at the time of data collection). Our lack of an association between TRAP pollutants and BMI is consistent with previous studies (Fioravanti et al., 2018; Fleisch et al., 2018; Fossati et al., 2020; Frondelius et al., 2018). However, other studies have found associations between prenatal TRAP exposure and BMI at later time points (ages 5–10 years) (Jerrett et al., 2014; Kim et al., 2018).

Although findings for associations with early-life BMI are not available, some studies investigating the effects of exposures in different domains (i.e., the chemical, physical, and social environments) on childhood obesity risk have been published in the environmental health literature. For example, index-based approaches have reported associations between higher exposure to adverse environmental factors and childhood obesity (Kim et al., 2020; Kinra et al., 2000; Li et al., 2014). Similarly, living in a neighborhood with more favorable environmental and social conditions, particularly during early life periods, is associated with lower obesity risk later in childhood (Aris et al., 2022). Although these approaches incorporate several exposures of interest into a single summary exposure metric, index-based methods do not provide information about how each component of the mixture contributes to overall risk. In contrast, a mixtures approach that considers modifiable risk factors across key domains may provide an opportunity to better understand the complex relationships between these factors and identify policy levers to reduce the prevalence of childhood overweight and obesity.

Despite the potential for these exposures to jointly impact growth in the early childhood period, we did not detect significant effects of the mixture on BMI z-scores using quantile-based g-computation. However, our use of this method for mixtures was a strength when examining our original research question about the feasibility of examining the effect of multiple neighborhood-level exposures on early life BMI. Because we were interested in characterizing as much of the neighborhood context as possible, we included 15 indicators of environmental quality and SES. We were able to include all these variables, several of which were highly correlated and could not otherwise be analyzed at the same time (Keil et al., 2020). G-computation also provides an indication of variable importance. The positive and negative weights provide a measure of an exposure’s relative contribution to the overall effect estimate. This model feature reduces the need to select a priori which indicators of neighborhood quality may be most important for use in a multipollutant framework. This contrasts with other index-based methods which collapse exposures into a single metric without providing information on which components are most important.

An additional advantage our application of quantile-based g-computation is that the exposure variables are not assumed to have the same directional effect on the outcome of interest. Another widely used mixtures approach, WQS regression, forces the weights of each component in the mixture to be in the same direction (Carrico et al., 2015). The quantile-based g-computation models (Figure 2) demonstrated differences across exposures and time points in the direction of associations between mixture components. Previous studies employing statistical models of environmental mixtures have generally focused on the use of biomarkers to characterize exposure (Hu et al., 2021; Lee et al., 2021; Preston et al., 2020; van den Dries et al., 2021; Welch et al., 2022; Yim et al., 2022). The use of biomarkers in these previous studies allowed investigators to hypothesize a priori the direction of associations between components of the mixture and the outcome of interest. Our results suggest this is a more complicated process when assessing exposure at the neighborhood or residential level using secondary data sets. Although it would have aided the interpretation of our results, a lack of consistency across cohorts and time periods in the direction of associations meant we could not reasonably code our variables to have similar directional associations. This lack of consistency also precluded combining the cohorts into a single cohort, which would have increased our sample size.

Similarly, our use of mixtures methods highlights the challenges in interpreting the effects of neighborhood-level exposures on health outcomes. For example, we included two variables related to park space near the homes of participants: park count and park area. We hypothesized that both of these variables would be associated with lower BMI z-scores, potentially through pathways that increased physical activity during pregnancy (Chen et al., 2021; Harrod et al., 2014; Perales et al., 2020; Porter et al., 2019). Park counts and park area were positively correlated in both cohorts and both variables were negatively correlated with distance to parks in both cohorts, as expected (Figure S3). However, weights derived from the g-computation models for these variables on the effect of the overall mixture were often in different directions with similar magnitudes (Figure 2, Figure S5, Figure S6). These discrepancies may reflect differences in how well these data characterize each study area. Our use of park area and park counts as measures of recreation space in a neighborhood may poorly capture the subjective measures of quality that are more likely to influence health behaviors during the prenatal period (Cohen and Leuschner, 2019; Rollings et al., 2015). Park space in Los Angeles neighborhoods likely looks different from park space in Denver neighborhoods as evidenced from the differences in park area and distance to park variables, and many small parks in a neighborhood may not have the same quality as a single large park; secondary data sets, such as the one we used here from the American Community Survey, will likely not capture these differences. Future studies investigating the effects of combined neighborhood exposures will need to consider alternative strategies for measuring the neighborhood features that best represent the possible pathways connecting exposures and health effects.

4.1. Strengths and Limitations

Despite a lack of significant findings for mixtures effects, our study has several strengths. First, we were able to utilize data from two ethnically and socioeconomically diverse cohorts. Healthy Start and MADRES participants were well-characterized, and we were able to develop harmonized exposure and outcome data sets for both cohorts. Although both cohorts are from urban areas, they represent cities with important geographical and social differences, which allowed us to examine how context plays a role in interpreting our results. Second, we used publicly available data sets for most of our exposure data, which allows for other investigators to replicate these methods and generate results that can be directly compared with ours. Third, we observed a wider range of exposures when examining both cohorts together than would have been available for individual cohorts. This allowed us to make comparisons across geographical regions with differing exposure profiles. Lastly, both cohorts had recently developed spatiotemporal models of TRAP exposure (BC for Healthy Start and NO_x for MADRES), which allowed us to explore different measures of TRAP within our exposure mixture. We found that the results for the spatiotemporal indicators of TRAP were generally consistent across cohorts but did not substantially alter the results of our mixtures methods when used in place of a more generic proximity-based indicator of TRAP.

There are some limitations to note as we interpret the results of our study. First, because we did not adjust for multiple comparisons, our results should be interpreted with caution and assumed to be exploratory. Second, because we relied on publicly available data sets, we were limited in the exposures we could include in our mixture and in the geographical span for future work (i.e., cohorts in countries with similar data collection efforts). For example, water quality data at the neighborhood level were not widely available. Third, our reliance on secondary data sets to assign exposures means that most of our exposure data had coarse temporal resolutions. Because we did not have residential history data for the Healthy Start cohort at the time of exposure assessment, we were not able to fully explore critical windows for many of the exposures of interest. Previous work has identified a risk of exposure misclassification when not considering residential mobility during pregnancy that may be differential by participant characteristics (Bell et al., 2018; Bell and Belanger, 2012), but that the effect of this misclassification on effect sizes may be minimal (Madsen et al., 2010; Pereira et al., 2016; Warren et al., 2018). Given the two cohorts collected data and constructed residential histories differently prior to joining the ECHO consortium, this work also serves to highlight the important issues to consider when merging or harmonizing data. For both of these cohorts, we aimed to use the best exposure estimates available, but differences in data collection likely resulted in differences in how noisy our exposure assessments were. Our work highlights the challenges for large consortia like ECHO that aim to harmonize data across time and space. Fourth, we based our criteria air pollutant exposures on estimates interpolated from the local monitoring network. These monitors are designed to capture regional trends in background and regional pollution but they likely do not capture intraurban gradients in exposure to pollutants like PM_2.5, O₃, and NO₂ (Jerrett et al., 2005). There is likely some underestimation of PM_2.5 and NO₂ exposures and overestimation of O₃ exposures for participants living near major sources like roadways. However, because a significant component of regional PM_2.5 is attributable to secondary formation, we elected to use these monitoring data to capture this regional trend. Fifth, we used PM_2.5 mass in our exposure models, which may not fully account for differences in particle constituents across geographical regions or by season. Similarly, our spatiotemporal TRAP models relied on different TRAP constituents (BC and NO_x), which may not have the same etiological links to health outcomes. Lastly, we relied on BMI z-scores as a measure of childhood growth in infancy. BMI z-scores are widely available based on routinely collected anthropometric data, but are limited surrogates for adiposity in infancy (Bell et al., 2018). Because we were assessing outcomes under the age of 2 years, we did not categorize infants into obesity categories based on their BMI z-scores. We also did not access growth trajectories for the infant participants in our study; changes in BMI over time may be a better indicator of infant obesity risk (Regnault and Gillman, 2014).

5. CONCLUSION

Overall, we found no associations between a mixture of neighborhood-level environmental exposures and social determinants of health on BMI z-scores in early childhood. However, our results highlight some important challenges in assessing the effects of environmental mixtures at the neighborhood level, specifically around the use of publicly available geospatial data sets to characterize neighborhood environments. Although we set out to combine data from these two complementary cohorts, we were not able to due to difference in the observed relationships between exposures and outcomes in single-exposure models. Widely available geospatial data on environmental determinants of health may not fully capture the contextual factors that influence health behaviors (e.g., physical activity). Similarly, these datasets are limited in their ability to fully characterize exposures at the time scales relevant to prenatal exposures (i.e. trimesters or gestational weeks). Further, our exposure data harmonization was impacted by differences in how residential history data were collected across cohorts. These differences, which came about because each cohort was originally conceived for different purposes, are likely to impact future studies aiming to combine data into a single analytic cohort. Future studies may be able to leverage the mixtures methods used here to answer questions about the effects of multiple neighborhood-level determinants.

To facilitate these types of studies, we have several recommendations. First, future studies should carefully consider how geospatial data sets may or may not fully reflect contextual factors at the neighborhood level. This will likely require local knowledge or expertise and may limit the number of variables that can be included if local context differs greatly across cohorts. Second, whenever possible, cohorts should collect data on residential history so that exposure assessments can incorporate residential mobility. This may require retrospective data collection from participants when feasible. Lastly, when practical, studies should aim to characterize exposures within geographic boundaries that most closely align with participant conceptions of neighborhood rather than relying on the default areal units typically used for spatial data reporting (e.g., the census tract level). While this may be difficult in multi-site studies, this could be particularly beneficial for intraurban studies to align exposures with the types of administrative boundaries that can influence policy decisions.

HIGHLIGHTS.

• We assessed the effects of a mixture of prenatal environmental exposures on early childhood BMI.

• Exposures were assessed at the neighborhood level during pregnancy.

• We used quantile g-computation to models the joint effect of neighborhood exposures.

• Air pollutants and socioeconomic status were associated with lower BMI in single-pollutant models.

• The overall mixture was not associated with BMI at any time point.

Abbreviations

BC

black carbon

BMI

body mass index

ECHO

Environmental influences on Child Health Outcomes

GIS

geographic information system

GWG

gestational weight gain

IDW

inverse-distance weighting

MADRES

Maternal and Developmental Risks from Environmental and Social Stressors

NDVI

normalized difference vegetation index

NLCD

National Land Cover Database

RMSE

root mean square error

SD

standard deviation

SES

socioeconomic status

SHS

secondhand smoke

TRAP

traffic-related air pollution

WHO

World Health Organization

WQS

weighted quantile sum regression

Data Availability Statement

Select de-identified data from the ECHO Program are available through NICHD’s Data and Specimen Hub (DASH). Information on study data not available on DASH, such as some Indigenous datasets, can be found on the ECHO study DASH webpage.