GPS-DERIVED ENVIRONMENTAL EXPOSURES DURING PREGNANCY AND EARLY POSTPARTUM – EVIDENCE FROM THE MADRES COHORT

Abstract

The built and natural environment factors (e.g., greenspace, walkability) are associated with maternal and infant health during and after pregnancy. Most pregnancy studies assess exposures to environmental factors via static methods (i.e., residential location at a single point in time, usually 3^rd trimester). These do not capture dynamic exposures encountered in activity spaces (e.g., locations one visits and paths one travels) and their changes over time. In this study, we aimed to compare daily environmental exposure estimates using residential and global positioning systems (GPS)-measured activity space approaches and evaluated potential for exposure measurement error in the former. To do this, we collected four days of continuous geolocation monitoring during the 1^st and 3^rd trimesters of pregnancy and at 4–6 months postpartum in sixty-two pregnant Hispanic women enrolled in the MADRES cohort. We applied residential and GPS-based methods to assess daily exposures to greenspace, access to parks and transit, and walkability, respectively. We assessed potential for exposure measurement error in residential vs GPS-based estimates using Pearson correlations for each measure overall and by study period. We found residential and GPS-based estimates of daily exposure to total areas of parks and open spaces were weakly positively correlated (r=.31, P<.001) across pregnancy and postpartum periods. Residential estimates of %greenspace (r=.52, P<.001) and tree cover (r=.55, P<.001) along walkable roads were moderately correlated with GPS-based estimates. Residential and GPS-based estimates of public transit proximity, pedestrian-oriented intersection density, and walkability index score were all highly positively correlated (r>.70, P<.001). We also found associations between residential and GPS-based estimates decreased among participants with greater daily mobility. Our findings suggest the popular approach that assessing the built and natural environment exposures using residential methods at one time point may introduce exposure measurement error in pregnancy studies. GPS-based methods, to the extent feasible, are recommended for future studies.

1. INTRODUCTION

The impact of the built and natural environment factors such as greenspace, walkability on maternal and child health has attracted notable attention over the past decade. Past research has associated exposure to neighborhood built environment characteristics such as shorter distance to public transit stops and higher neighborhood walkability with increased physical activity in pregnant women (Kershaw et al., 2021; Porter et al., 2019; Richardsen et al., 2016; Thomson et al., 2019). Additionally, studies have shown that greater access to parks and open space (e.g., visual access, proximity, coverage) decreases pregnant women’s stress levels and depressive symptoms (Boll et al., 2020; McEachan et al., 2016). Furthermore, research has linked higher levels of satellite imagery-based exposure to greenspace to decreased maternal glucose levels, attenuated risks of gestational diabetes mellitus and weight gain, and higher infant birthweight (Anabitarte et al., 2020; Torres Toda et al., 2020).

However, most studies assess prenatal exposure to the built and natural environment factors at the residential neighborhood level at one time point (i.e., in a circular buffer around the home location as determined by an address obtained at time of birth). This approach has several major limitations. Spatially, relying on residential address at a single point late in the pregnancy ignores residential mobility (or moving) across pregnancy and postpartum (R. F. Banay et al., 2017; Bell and Belanger, 2012; Chen et al., 2010; Hodgson et al., 2015). Also, the method does not capture women’s daily environmental exposures occurring outside their residential neighborhood (e.g., walking trips, errands, other visited locations) which may also influence health outcomes (Kanning et al., 2023; Xu et a., 2022; Matthews and Yang, 2013; Perez et al., 2019). Moreover, they do not consider changes in activities and behaviors as pregnancy progresses over time and into the postpartum period. These activities and behaviors could dramatically vary due to preparations for childbirth, difficulty of physically moving around and increased fatigue later in pregnancy, and childcare responsibilities after birth (Varshavsky et al., 2020). Further, within a particular pregnancy period such as 3^rd trimester, exposures to environmental factors could result from household, occupational, and recreational activities (e.g., grocery shopping, commuting to work), which may differ from day to day. As a result, assessing exposures using residential-based approaches may fail to capture the “true causally relevant” environmental characteristics that exert contextual influence on pregnant women’s and infants’ health and could introduce exposure measurement error and potential bias (Robertson and Feick, 2018; Yi et al., 2019).

Increasingly, studies are integrating human mobility into exposure assessment approaches (i.e., matching dynamic human movement with environmental features) to derive environmental exposures (Jankowska et al., 2021; Liu et al., 2020; Ragettli et al., 2015; Wei et al., 2023a; Zenk et al., 2018, 2011; Yi et al., 2023). In these studies, highly resolved Global Positioning System (GPS) monitoring data is collected (usually every few seconds), which is used as the basis for constructing daily activity spaces containing locations one visits and paths one travels. These activity spaces are integrated with environmental layers (e.g., parks, walkability, air pollution, fast food outlets) in GIS software to measure an individual’s dynamic exposure at high spatiotemporal resolutions (Yi et al., 2019). Consequently, these GPS-based approaches may improve exposure assessment in pregnancy studies compared to residential-based approaches since they incorporate information about where and when individuals spend their time. However, very few studies have been able to assess GPS-based environmental exposures during pregnancy and postpartum given the potentially higher burden of collecting personal, highly resolved geolocation data. Furthermore, to understand whether quantified environmental exposures and observed health relationships are sensitive to the choice of measurement method, it is also important to systematically evaluate the measurement errors associated with residential-based compared to GPS-based estimates, as well as to understand different scenarios in which such errors tend to be higher or lower (different pregnancy periods, high vs. low daily mobility). To date, only a few studies have addressed this issue (Liu et al., 2020; Ragettli et al., 2015; Wei et al., 2023a), and to our knowledge, none has focused on pregnant women.

In addition, there are several examples of health studies where residential estimates are used to classify individuals into groups based on environmental exposure estimates. For example, two studies represented high- and low-walkable neighborhoods using top and bottom quartiles of the residential walkability index score (Carlson et al., 2015; Van Dyck et al., 2010). Especially for pregnancy studies, comparing GPS-based exposure estimates to the typically used residential approach might shed light on the potential for exposure measurement error and how this varies over pregnancy and postpartum periods and human mobility patterns (i.e., spatial extent of human movement footprints).

To fill these gaps, we leveraged 4-day smartphone GPS location data from a group of Hispanic, predominantly low-income women during the 1^st and 3^rd trimesters of pregnancy and at 4–6 months postpartum. Participants came from a substudy nested in the larger MADRES cohort in Los Angeles, CA. We then applied geographic information system (GIS) to characterize participants’ daily GPS-based and residential environmental exposures using the 4-day GPS location data. Lastly, we examined the potential for exposure measurement error in relying on residential-based environmental exposures (using the 3^rd trimester residential location) typically applied in previous studies compared to GPS-based exposures. Our study has the three following aims:

1. To describe pregnant participants’ daily GPS-based environmental exposure patterns during the 1^st and 3^rd trimesters of pregnancy and at 4–6 months postpartum;

2. To assess correlations between various residential and GPS-based environmental exposure measures; and

3. To evaluate the potential influences of participants’ daily mobility on exposure measurement error in residential vs. GPS-based environmental exposure measures.

2. METHODS

2.1. Design and Overview

Data for this study comes from the Real-Time and Personal Sampling sub-study of the Maternal and Developmental Risks from Environmental and Social Stressors (MADRES) cohort (O’Connor et al., 2019). This study uses an intensive longitudinal, observational panel study design and examines the daily effects of environmental exposures and social stressors on maternal pre- and post-partum obesity-related biobehavioral responses. A total of 65 Hispanic, predominantly lower income mothers were drawn from the larger MADRES prospective cohort study between 2016–2018. The larger MADRES study participants were recruited from prenatal care providers serving predominantly medically-underserved populations in Los Angeles, California, including two non-profit community health clinics, one county hospital prenatal clinic, one private obstetrics and gynecology practice, and through self-referral from community meetings and local advertisements (Bastain et al., 2019). To be eligible for this sub-study, a participant needed to be >18 years old with a singleton pregnancy and be at less than 30 weeks’ gestation at time of recruitment. In addition, participants who were HIV positive, had physical, mental, or cognitive disabilities that prevented participation, or were currently incarcerated were excluded from the study (Bastain et al., 2019). The University of Southern California Institutional Review Board approved all study procedures, and participants signed an informed consent before enrolling into the cohort and the substudy.

2.2. Data Collection

2.2.1. Geolocation using Global Positioning Systems (GPS)

We provided the study participants with a Samsung MotoG phone (Model Moto G, Samsung) with an Android operating system (Google USA, Inc.), which they could use for the duration of the 4-day data collection across three periods. The phone continuously collected GPS data from 65 study participants at 10-second(s) intervals for four days (two weekdays and two weekend days) during the 1^st and 3^rd trimester and at 4–6 months postpartum. To enable collecting highly resolved and encrypted GPS data, MADRES researchers designed a custom smartphone application (madresGPS app) for Android operating systems. The application on dedicated study smartphones (Samsung MotoG phone) was configured by study coordinators to record geographic coordinates and geolocation/motion metadata, which logged instantaneous GPS location and sensor data every 10 s from the smartphone’s multiple built-in location finding features (cell tower triangulation, Wi-Fi networks, and GPS) and motion sensors (O’Connor et al., 2019). Along with the timestamp, the application recorded metadata such as the number of satellites in use/view, geolocation accuracy, source of GPS, velocity (if GPS source), and network connection status (if network source) (O’Connor et al., 2019). Participants returned the phone at the end of the 4-day data collection period. A study assistant then manually downloaded the GPS data from the phone to a secure server for further processing.

2.3. Data Processing

2.3.1. GPS Data Analysis

The 10-s epoch GPS geolocation data was processed using a custom algorithm. In short, we discarded the raw GPS data collected outside the 4-day data collection window over three time periods. Then, we developed logic to select the most accurate source of geolocation data for each 10-s epoch when two data sources (GPS/Network) were available. Finally, we applied a moving median filter to remove outliers in 1-minute time windows to correct for extreme outliers that might occasionally be present in the data. More details of this algorithm are described in Yi et al. (2022). Missing GPS data was then imputed when participants were very likely to be at their home location in the daytime or nighttime. At night, typical sleep and wake time windows were used along with location before and during to determine whether participants were likely home and fill in any gaps in GPS. During the day, confirmed daily survey reports of being at “Home-Indoors” or “Home-Outdoors” all day were used to fill in gaps with the home location. Days with <6 hours of GPS data (post-imputation) were then flagged as invalid and not included in subsequent analyses. The 6h cut-off was chosen to correspond approximately to ≥6h of GPS data during typical waking hours, as most missing data occurred at night, particularly during the sleep times self-reported by study participants in the survey. This choice allows us to preserve the quality of GPS data and maximize data retention for subsequent analyses, such as environmental exposure assessment.

2.3.2. Exposure Assessment

2.3.2.1. Constructing Activity Spaces

Daily activity spaces during the 1^st and 3^rd trimester and at 4–6 months postpartum were constructed for each participant based on their GPS trajectories via the route buffer and kernel density estimation (KDE) methods (Figure 1) using ArcGIS Pro 10.7.1 (Esri, 2021). We selected these two GPS-based methods from the extensive literature on activity spaces because they were complementary, capturing participants’ daily activity spaces along their daily routes (via route buffer) and around the activity locations where they spent most of their time (via KDE), both of which may be highly relevant to health behaviors and outcomes related to the built environment (Smith et al., 2019; Yi et al., 2019; Zenk et al., 2011).

Figure 1.

An illustration of two GPS-based activity space metrics applied in this study.

Notes. GPS = Global Positioning System.

To construct buffers along routes, successive GPS points were connected into lines based on timestamps (i.e., routes) and then buffered with a radius of 250 meters. We chose this distance based on our previous work (Yi et al., 2022), which corresponds to the depth of several blocks in urban Los Angeles, CA, to approximate two hypothesized mechanisms of how environmental factors could potentially influence health behaviors and outcomes, including viewshed (e.g., seeing a park when walking on a street) and perception (e.g., knowing a recreational facility is located nearby) (Boruff et al., 2012; Helbich et al., 2016; Yi et al., 2019). Additionally, the geographical area of these route buffers was calculated to represent the spatial extent of daily mobility and classified into low, medium, and high mobility days based on tertiles. Additionally, the KDE approach (based on the quartic kernel function) was applied to generate time-weighted activity grids with 50 × 50 m cells with a bandwidth of 250 m (corresponding to the radius of the route buffer). Each grid was assigned a normalized time weight (range from 0–1) based on percentages of staying duration within a given day, with total weights adding up to 1. Compared to the route buffer approach, which only considers spatial extent, the KDE incorporates both spatial and temporal aspects (i.e. dwell time spent in a particular location) in the creation of activity spaces.

2.3.2.2. Defining Residential Neighborhoods

We assessed residential neighborhood exposures based on the 3^rd trimester residential location as most similar to the typically used approach in pregnancy studies (i.e., assessing exposure based on residential address provided on birth certificates or questionnaires after delivery and assuming it represents the entire prenatal period). Briefly, we derived 800 m residential network buffers via the Service Area Analysis tool and StreetMap Premium product in ArcGIS Pro 10.7.1 (Esri, 2021). The network buffer better captures the area to which a pregnant woman could realistically travel from the home residence as compared to a circular buffer (James et al., 2014). An 800 m radius, which corresponds to a 5–10-minute walk, is recommended for populations with relatively restricted mobility (Frank et al., 2017; James et al., 2014). Additionally, a sensitivity analysis was performed using a larger radius (e.g., 1,600 m) and the resulting environmental exposure estimates differed only slightly, so we chose to report 800 m results only.

2.3.2.3. Assessing Environmental Exposures

Eight environmental factors were measured using the residential and GPS-based route buffer and KDE methods, respectively. These included greenness (3 metrics), park and public transit access (3 metrics), street connectivity (1 metric), and walkability (1 metric) as follows: mean Normalized Difference Vegetation Index (NDVI; ranges from −1 to +1 with higher value represents higher greenness; the NDVI dataset was date-matched to corresponding residential buffers and activity spaces when calculating mean NDVI), while the rest of environmental measures, including %greenspace along walkable roads, and %tree cover along walkable roads, distance to the nearest park entrance, distance to the nearest public transit stop, total area of parks and open spaces, pedestrian-oriented intersection density, and walkability index score (ranges from 1 to 20 with higher score represents more walkable), were derived using the dataset at one single time point that closest matched the data collection period. The environmental measures chosen have been reported by previous studies to be associated with various human activities and health outcomes, particularly among pregnant women (R. Banay et al., 2017; Besser and Dannenberg, 2005; Jiang et al., 2016; Pickard et al., 2015; Porter et al., 2019; Pretty et al., 2005; Saelens et al., 2014; Tsai et al., 2019). All environmental measures, corresponding data sources, resolution, and processing steps and interpretation are shown in Table 1. Residential and GPS-based route buffer exposures were calculated by summarizing values of the respective environmental measures (e.g., mean NDVI, total areas of parks and open spaces) within boundaries of corresponding spatial units (i.e., 800 m residential network buffer, 250 m route buffer). GPS-based KDE exposures were calculated by weighting grid-based values of the respective environmental measures with time spent. For example, we multiplied walkability index score value of the grid by percentage of time spent to derive KDE-based walkability exposure estimates. Exposures were calculated using eq. (1) for the residential or route buffer methods and using eq. (2) for the KDE methods as follows:

$$Exposur{e}_{RNB/RB}=\sum _{n=1}^n\left(\text{\%}Are{a}_n\ast En{v}_n\right)$$ eq. (1)

$$Exposur{e}_{KDE}=\sum _{n=1}^n\left(\text{\%}Are{a}_n\ast En{v}_n\ast T{W}_n\right)$$ eq. (2)

where Env_n is the exposure estimate for a environmnetal measure in the n^th spatial unit (e.g., census block group, grid, line) that intersects the residential network buffer, route buffer, or KDE grid; Area_n is the percentage of the area of n^th spatial unit that falls within the residential network buffer, route buffer, or KDE grid; and TW_n is the normalized time weights (range from 0 to 1) of the n^th grid.

Table 1.

The environmental measures, corresponding data source, resolution, and processing steps

Measures	Data Source	Data Resolution	Processing Steps and Interpretation
Mean NDVI (range from −1 to 1)	NASA MODIS 16-Day L3 Global product (2017–19)1	250 × 250 m raster grids	The mean Normalized Difference Vegetation Index (NDVI) was used to represent overall vegetation levels within respective spatial units (i.e., residential neighborhood, activity space). NDVI values range from −1 to 1, with negative values corresponding to areas with water surfaces, 0–.2 representing barren surfaces, .2–.5 for sparse vegetation, and >.5 for dense vegetation.
% greenspace along walkable roads (%)	EPA EnviroAtlas Community Data (2008–13)	Census block groups	Greenspace areas were derived by combining areas of multiple land cover classes including water, trees and forest, grass and herbaceous cover, shrubs, agriculture, orchards, and woody and emergent wetlands. Sidewalk areas were derived by buffering NAVTEQ roads with a speed limit less than 55 miles per hour (potentially walkable roads) by a width of 25 m on each side. The measure then was calculated by intersecting tree cover and sidewalk areas per each city block.
% tree cover along walkable roads (%)	EPA EnviroAtlas Community Data (2008–13)	Census block groups	Tree cover areas were derived by combining areas of three land cover classes - trees, forests, and woody wetlands. Sidewalk areas were derived by buffering NAVTEQ roads with a speed limit less than 55 miles per hr (potentially walkable roads) by a width of 8.5 m on each side. This measure then was calculated by intersecting tree cover and sidewalk areas per each city block.
Distance to the nearest park entrance (m)	EPA EnviroAtlas Community Data (2020)	Buffer zones	This measure was derived by delineating approximate walking areas from a park entrance to any given location along road networks within the EnviroAtlas community area boundary (i.e., Los Angeles County). Areas with walking distances greater than 5km were omitted.
Total parks and open space coverage (km2)	California Protected Areas Database (2020)	Parks and open space polygons	This measure included 1) National/state/regional parks, forests, preserves, and wildlife areas; 2) large and small urban parks that are mainly open space (as opposed to recreational facility structures); 3) land trust preserves; and 4) Special district open space lands (watershed, recreation, etc.) and other types of open space. The measure was calculated by intersecting parks and open space areas with buffers created.
Distance to the nearest public transit stop (m)	EPA Smart Location Database (2021)	Census block groups	This measure was derived by measuring the minimum walk distance in meters between the 2010 population-weighted census block groups (CBG) centroid (as used by SLD version 2.0) to the nearest transit stop of any route type.
Pedestrian-oriented intersection density (m)	EPA Smart Location Database (2021)	Census block groups	This measure was derived based on an analysis of NAVTEQ 2011 Streets data. All 3-way intersections were weighted by .6667 since they reduced street connectivity compared to intersections with 4 or more legs.
Walkability index score (range from 1 to 20)	EPA Smart Location Database (2021)	Census block groups	A composite index score combining household and employment density, street intersection density, and distance to nearest transit stops. This measure represent different environmental features that are known to be supportive of walking. The index scores range from 1 to 20, with higher value represents better walkability.

¹The MODIS NDVI dataset was date-matched to corresponding activity spaces when calculating mean NDVI.

Notes. NASA = National Aeronautics and Space Administration. MODIS = Moderate Resolution Imaging Spectroradiometer. EPA = Environmental Protection Agency.

2.4. Statistical Analysis

Descriptive statistics were calculated for covariates, and residential and GPS-based route buffer and KDE exposures, respectively. Because this is a lower-income group, we also determined the neighborhood socioeconomic status (NSES) for each participant at baseline based on the Deprivation Index Score (0–10 with higher score indicating most disadvantaged block groups) from the Neighborhood Atlas (Kind and Buckingham, 2018). Additionally, intraclass correlation coefficients (ICC) were calculated to describe the proportion of day-to-day variability in the dynamic exposures that was between-participant (i.e., ICC%) compared to within-participant (i.e., 1-ICC%). Two-level (days nested within participants) linear mixed-effects random intercept only models were used to calculate the ICCs for the 8 dynamic environmental measures. ICC value cut-offs of 0.40 and 0.75 were selected to indicate weak (ICC<0.40), moderate (0.40≤ICC<0.75), or strong (ICC≥0.75) within-person correlations (Zenk et al., 2018). Analyses were conducted in R 4.0.2 (R Foundation for Statistical Computing, 2021) using the lme4 package (version 27.1) (Bates et al., 2015).

Pearson’s correlation coefficients were used to examine similarity of environmental exposures assessed using residential vs. GPS-based methods, given the approximately normal distributions of environmental measures. Moreover, Pearson correlation coefficients were calculated by study period (i.e., 1^st trimester, 3^rd trimester, and 4–6 months postpartum) to investigate how well GPS-based estimates (i.e., the closest proxy for the “true” exposure) correlate with 3^rd trimester residential estimates as an indication of potential for measurement error in relying solely on the latter.

Furthermore, we classified daily environmental exposures into four groups (quartiles) using GPS-based KDE as the most dynamic method that is spatiotemporally matched to human movement (i.e., the true exposure) and calculated the percent of “misclassified” days that were assigned to a different group using the residential method as an indication of potential for exposure measurement error in the latter approach. For example, if the KDE method classifies a day into the highest exposure (4^th quartile) group but the residential method classifies it into lower exposure groups (1^st, 2^nd, or 3^rd quartile), we labelled it as a misclassified day. We used Sankey diagrams to illustrate these relationships, where arrows represent flows from quartiles of KDE exposures (left, assumed “true” daily varying exposure) to quartiles of residential exposure (right, static exposure), and the thickness of the lines flowing from left to right (representing percent of days) indicate the potential for exposure misclassification. In this study, we investigated whether exposure misclassification potential was sensitive to daily mobility by creating the same Sankey diagrams for low, medium, and high mobility days. Finally, to test whether the results of the study depended on the chosen buffer size, we also assessed environmental exposure using 100 m route buffers and time-weighted activity grids. Analyses were conducted in R 4.0.2 (R Foundation for Statistical Computing, 2021) and Sankey diagrams were created using the ggalluvial package (version 0.12.3) (Brunson, 2020).

3. RESULTS

3.1. Descriptive Statistics

Descriptive statistics for participant-level covariates (N=62) are shown in Table 2. The mean age was 29 years at baseline (SD=6.1). All were of Hispanic descent and more than half (53%) were born outside the U.S. The majority had a high school diploma or less (68%), four-fifths (81%) were either married or living with their partner at baseline, and 39% were employed during the 3rd trimester. At baseline, more than one-quarter (26%) were pregnant with their first child, and about three-quarters (76%) were overweight or obese, as measured by their pre-pregnancy body mass index (BMI). Their neighborhood deprivation score at baseline was 6.5 (SD=1.7).

Table 2.

Descriptive statistics of participant characteristics in the analytical sample (N=62 Participants).

Variable	Mean (SD) or n (%)
Age at entry (years)	29 (6.1)
Education
High school or less	42 (68%)
Some college/Graduate	20 (32%)
Marital status
Married/Living together	50 (81%)
Single/Divorced/Separated/Widowed	10 (16%)
Missing	2 (3%)
Acculturation
US-Born Hispanic	29 (47%)
Foreign-Born Hispanic	33 (53%)
Maternal parity
First-born	16 (26%)
Already had child	46 (74%)
Pre-pregnancy BMI category
Normal	16 (26%)
Overweight/Obesity	46 (74%)
Employment status during the 3rd trimester
Unemployed	28 (58%)
Employed	19 (40%)
Missing	1 (2%)
Neighborhood deprivation score (1–10)	6.5 (1.7)

Notes. BMI = Body Mass Index, GPS = Global Positioning System. SD = Standard deviation.

Details of GPS data characteristics and participant compliance are described in an earlier study (Yi et al. 2022). Briefly, our final analytical sample comprised a total of 552 valid person-days of GPS data from 62 participants, with 205 person-days during the 1^st trimester, 180 person-days during the 3^rd trimester, and 167 person-days at 4–6 months postpartum. An average of 8.9 valid GPS days (SD=3.0; Range: 3.0–12.0) were provided by participants across the three periods. On average, 21.7 hours (SD= 5.0; Range: 6.2–24.0) of GPS observations were collected on valid days.

3.2. Daily Environmental Exposure Estimates during Pregnancy and Early Postpartum

Descriptive statistics for daily environmental exposure estimates derived by residential, as well as GPS-based route buffer and KDE methods are shown in Table 3. The median geographic extent of an activity space in a person-day was 2.49 km² (IQR=11.20), which is equivalent in coverage to several superblocks (0.5 to 1 mile long) in urban Los Angeles, CA. Depending upon the environmental measure, exposure estimates varied across the three methods. In addition, the exposure estimates derived by the KDE method had the largest variability (i.e., IQR values) across the three methods.

Table 3.

Descriptive statistics of the daily (N=552) environmental exposure estimates for participants using the route buffer, kernel density estimation (KDE), and residential methods.

	N=552 person-days from 62 participants
Variables	Mean	SD	Median	IQR	Min	Max
Geographic extent (km2)
Route buffer 250 m	7.96	11.94	2.49	11.20	.20	91.91
NDVI (range from −1–1)
Route buffer 250 m	.18	.05	.18	.06	.06	.41
KDE 250 m	.18	.05	.17	.06	.08	.41
Residential 800 m	.18	.05	.18	.07	.10	.38
%greenspace along walkable routes
Route buffer 250 m	23.30	5.87	22.76	6.77	6.61	46.76
KDE 250 m	23.28	8.30	21.87	10.32	1.70	46.27
Residential 800 m	23.04	5.97	22.40	7.10	12.14	45.70
%tree cover along walkable routes
Route buffer 250 m	21.45	5.51	20.75	6.03	6.32	44.78
KDE 250 m	22.58	8.16	21.99	10.77	1.74	50.37
Residential 800 m	21.72	5.75	20.57	6.06	10.68	45.63
Distance to the nearest park entrance (m)
Route buffer 250 m	877.97	397.02	791.82	421.93	294.97	2409.52
KDE 250 m	812.31	455.44	668.94	411.95	51.67	2342.11
Residential 800 m	789.29	414.09	660.04	302.91	400.38	2387.07
Total parks and open space area (km2)
Route buffer 250 m	243.53	608.87	15.50	146.21	.00	4817.38
KDE 250 m	.03	.10	.00	.01	.00	.93
Residential 800 m	9.67	15.65	2.84	11.44	.00	77.68
Distance to the nearest public transit (m)
Route buffer 250 m	280.02	103.27	275.38	144.08	1.94	532.25
KDE 250 m	284.95	123.13	282.97	177.60	.03	1041.17
v 800 m	268.94	105.79	265.49	135.86	81.28	645.76
Pedestrian-oriented intersection density (# per mi2)
Route buffer 250 m	47.90	33.35	42.27	26.55	2.20	190.79
KDE 250 m	51.81	40.78	41.73	38.92	.24	240.01
Residential 800 m	57.60	44.70	46.76	32.24	3.00	222.66
Walkability index score (range from 1–20)
Route buffer 250 m	14.93	1.50	14.96	2.02	10.29	18.43
KDE 250 m	14.84	1.72	14.92	2.59	10.09	18.08
Residential 800 m	14.94	1.43	15.17	1.89	11.59	18.11

Notes. NDVI = Normalized Difference Vegetation Index. KDE = Kernel Density Estimation.

In terms of KDE-measured neighborhood greenness, participants on average were exposed to an NDVI (range from −1 to +1) of 0.18 (SD=0.05) across three time points, only slightly above the Los Angeles County average of 0.14. This value indicates their daily activity locations on average had barren surfaces or very sparse vegetation. In addition, the KDE-measured percent of greenspace and tree cover along walkable roads at their daily activity locations on average was 23.3% (SD=8.3) and 22.6% (SD=8.2), respectively. Both numbers were lower than the Los Angeles County average of 32.4% (%greenspace along all walkable roads of Los Angeles County) and 28.1% (% of tree cover along all walkable roads of Los Angeles County). Mean exposure estimates of all greenness measures varied slightly for the same measure across the residential, route buffer, and KDE methods.

As for park access, participants’ KDE-measured mean daily distance from their daily activity locations to the nearest park entrance was 812 m (SD=455; corresponding to 15–20-minute walk). The estimates differed slightly with the residential or route buffer-measured values. The total estimated exposures to parks and open space area, however, varied tremendously across methods. Participants were exposed to an average of only 0.03 km² parks and open space per day (about the size of a tiny neighborhood park in urban Los Angeles) using KDE, which was on average < 0.5% of exposure to parks and open space areas measured by residential and route buffer methods. Participants’ KDE-measured mean distance to the nearest public transit stop at daily activity locations was 285 m (SD=123; corresponding to <5-min walk).

Lastly, regarding street connectivity and walkability, the KDE-measured mean number of pedestrian-oriented street intersections at daily activity locations participants visited was about 52 intersections per square mile (SD=41), and the mean walkability index score was 14.8 (SD=1.72; range from 1 to 20). These numbers indicated better street connectivity (34.7% higher) and walkability (1.5 scores higher) in personal activity spaces than the Los Angeles County average. Exposure estimates of these two measures differed slightly among the three methods.

3.3. Correlations between Residential and GPS-based Environmental Exposures

Pearson correlation coefficients between exposure estimates of the same environmental measure derived by static vs. dynamic methods are shown in Figure 2. In general, stronger positive correlations were found between residential and GPS-based KDE estimates than between the residential and route buffer estimates. Moreover, the strength of correlations varied substantially by environmental measure. Specifically, residential estimates of total area of parks and open space were weakly positively correlated (r=.31, P<.05) with KDE estimates. The residential and KDE exposure estimates for %greenspace and %tree cover along walkable roads were moderately positively correlated (r=.52, P<.01 and r=.55, P<.01, respectively). Lastly, the exposure estimates derived by residential and KDE for the remainder of environmental measures were strongly positively correlated (r>.7, P<.01). The correlation coefficients between residential and KDE exposure estimates decreased slightly at 4–6 months postpartum and decreased substantially during the 1^st trimester compared to the 3^rd trimester (see Figure 2). This pattern is manifested in the distance to the nearest public transit stop measure, %greenspace along walkable roads, and %tree cover along walkable roads.

Figure 2.

Pearson correlation coefficients between residential vs. day-level GPS-based exposure estimates for the same environmental variable by the 1^st and 3^rd trimesters and 4–6 months postpartum.

*P<.05.

Notes. KDE = Kernel Density Estimation. NDVI = Normalized Difference Vegetation Index. ResB = Residential Buffer. RB = Route Buffer.

3.4. Impact of Daily Mobility on Potential for Exposure Misclassification Introduced by the Residential Method

We further divided days into low, medium, and high mobility days using the tertile values of the geographic extent of daily activity spaces. Low mobility days have an extent of less than 0.24 km² (equivalent to a few city blocks in urban Los Angeles), whereas high mobility days have an extent of more than 7.78 km² (equivalent to a neighborhood such as downtown LA or West Hollywood). Impact of daily mobility (i.e., low, medium, and high mobility days) on potential exposure misclassification using the residential method (at a single 3rd trimester residential location) compared to the GPS-based KDE method (at a daily level based on GPS trajectories) were then examined, and the results are shown in Figures 3–a to 3–c and Supplementary Figures 1–a to 1–e. According to these figures, impact of daily mobility on potential exposure misclassification was observed across all environmental measures when relying on the residential method vs the GPS-based KDE method. The impact was largest for measures such as %greenspace along walkable roads (Figure 3a), NDVI (Supplementary Figure 1a), distance to nearest park entrance (Supplementary Figure 1b) for days when participants were highly mobile. However, this result was less consistent for, pedestrian-oriented intersection density (Supplementary Figure 1e), and walkability index score measures (Figure 3c) and reversed for mean %tree cover along walkable roads measure (Figure 3b), distance to nearest transit stop (Supplementary Figure 1c), and total areas of parks and open space (Supplementary Figure 1d).

Figure 3.

The extent and direction of daily exposure misclassification using third trimester residential exposures for the environmental measures compared to trimester-specific GPS-based KDE estimates.

Notes. The geographical area of route buffers was calculated to represent spatial extent of daily mobility and classified into low, medium, and high mobility days based on tertiles.

Notably, in the highest exposure group (4^th quartile) of high mobility days, 49.6% and 57.9% of person-days were misclassified into different groups by the residential method for % greenspace along walkable roads and % tree cover along walkable roads measures (Figures 3a–b). The percent of misclassification ranged from 25–35% for the highest exposure group for the remainder of environmental measures.

4. DISCUSSION

In this study, we described daily dynamic exposure to the built and natural environment factors during the 1^st and 3^rd trimesters and at 4–6 months postpartum by analyzing highly resolved smartphone location data collected from a sample of 62 Hispanic pregnant women (across 552 observation days) in Los Angeles, CA. Additionally, we answered a critical yet unanswered research question – how similar were pregnant women’s environmental exposure derived by the residential method applied by previous studies, to GPS-based exposure in their daily activity spaces? Our results indicated participants overall had small daily activity spaces and were exposed to daily activity spaces featuring very low to low vegetation levels and minimal parks and open space. Compared with GPS-based estimates, we also found exposure measurement error of residential estimates were potentially larger in general for parks and open space measures, during early pregnancy and postpartum, and in days when participants were more mobile. Our findings have important implications for future studies investigating the association between the built and natural environment and maternal and infant health, especially in lower income and racial/ethnic minority groups. The implications of our findings are discussed below.

4.1. Participants’ Daily Environmental Exposures during Pregnancy and Early Postpartum

In this study, we found that the participants in our sample had a daily activity space equivalent to several city blocks in Los Angeles. In addition, they were generally exposed to daily activity spaces characterized by very low to low vegetation levels and minimal parks and open space. Our results are consistent with previous studies examining green space exposures for low-SES and ethnical minority groups in urban Los Angeles (Giuliano, 2005; Kim and Kwan, 2021; Wolch et al., 2005). Moreover, our participants were exposed to activity spaces that had a high proximity to public transit (typically less than a 5-minute walk to the nearest public transit stop) and had street connectivity and walkability above Los Angeles County average. This finding suggests that, overall, participants may be exposed to environmental resources that facilitate utilitarian walking (e.g., visits to corner groceries) and active transport (e.g., walking to transit stops) than recreational activities (e.g., exercise in a park). However, this result may also indicate that participants chose activity spaces with higher walkability or better access to public transportation due to their personal preferences, occupations, or vehicle ownership – a concept often referred to as selective daily mobility bias (Chaix et al., 2013; Wei et al., 2023b). Therefore, it is possible that the environmental exposure assessed here was not the true “exposure” per se, but rather a mixture of exposure and behavior.

4.2. Similarities of Exposure Estimates between Residential and GPS-based Methods

Overall, we found correlations between the residential and GPS-based KDE estimates were higher than between residential and GPS-based route buffer estimates. This is not surprising given the KDE method weights exposure based on the time duration of stay (i.e., time-weighted spatial averaging) and participants on average spent a significant amount of time at home locations according to an earlier work (Yi et al., 2022), whereas the route buffer method is time-insensitive (i.e., spatial averaging) and weights the home location similarly to any other visited location. Despite the difference in correlations between residential and two GPS-based methods, one method may be better at measuring some environmental factors than the other. For instance, KDE may be better at capturing exposure that is time-sensitive in terms of how it impacts health. For example, KDE will be more nuanced at capturing time spent in a leisurely walk in a park versus simply driving through a park.

Our findings on the low to moderate correlations between residential and GPS-based green space are broadly consistent with a previous study that discovered similar results and therefore recommended GPS-based assessments for greenspace exposure (Wei et al., 2023a), although this previous study used a slightly different activity space method with distance decay function and distinguished between active and passive transportation modes. A growing number of studies (Kwan et al., 2019; Lee and Kwan, 2019; Roberts and Helbich, 2021; Zhang et al., 2021) have reported that GPS-based environmental exposures are associated with health behaviors and mental health outcomes, but residential-based ones are not. Therefore, the choice of methods to assess exposure to green space could influence the study results to investigate similar associations in the population of pregnant women and should be further explored.

Additionally, among all environmental factors measured, we found stronger correlations between the residential and GPS-based KDE estimates for measures of access to neighborhood resources, such as proximity to parks and public transit stops, as well as walkability. For measures with dose-dependent relationships to health (e.g., %tree cover along a street segment), correlations between residential and GPS-based KDE estimates were low to moderate. A previous study (Jankowska et al., 2021) comparing exposure estimates from different GPS-based methods found that similarities in estimates were generally higher for raster environmental exposure layer types than for point, line and area layer types. Thus, the differences in correlations can be explained by the layer types we used to derive the different exposures of the built environment. For example, the variable area of parks and open space (with low correlation) was measured based on park and open space polygons, which are discrete objects in space. Therefore, the correlations could be much lower than for the continuous raster surfaces such as NDVI (with the high correlations).

Across three pregnancy and early postpartum periods, we found correlations of exposure estimates further decreased during the 1^st trimester and at 4–6 months postpartum compared to the 3^rd trimester. This may be partially caused by the residential mobility (10 of 62 participants changed their place of residence during these periods) described in previous work (Yi et al. 2022). Additionally, this decrease may also be explained by changes in time-activity and mobility patterns across the three study periods, since participants in our study were found to spend less time at commercial and service locations and perform more vehicular trips during the 3^rd trimester (Yi et al. 2022).

4.3. Exposure Misclassification Introduced by the Residential Method Due to Daily Mobility

Our results showed exposure misclassification could potentially occur if we were to rely on the residential methods to estimate exposure levels, especially for days when individuals are highly mobile. Exposure measurement error or misclassification can weaken statistical power to detect associations and potentially bias observed risk estimates in health studies (Zeger et al., 2000). This concern is warranted by recent studies which have increasingly reported mixed results on associations between neighborhood greenspace and pregnant women’s activities and health outcomes (Anabitarte et al., 2020; R. Banay et al., 2017; Nichani et al., 2016; Porter et al., 2019).

Despite the overall dependence of exposure misclassification on daily mobility, the evidence of such dependence was less pronounced for urban form measures such as street intersection density and neighborhood walkability. It may be that these measures were less variable in the usual daily activity spaces of our relatively homogeneous study population. As a result, even if more activity locations were visited during high mobility days, the exposure estimates around these locations would only slightly differ from static estimate surrounding their place of residence. It could also be due to the reliance on measures (i.e., intersection density and walkability) that summarize conditions within administrative boundaries such as Census Block Groups (CBG) and thereby “artificially” decrease the variability of these measures even though they may vary a lot in short spatial scales or distances. This lack of variability may be further amplified by the fact that women in our sample had relatively small daily spatial footprints (a few urban blocks in Los Angeles, CA) and as a result, the differences in values between environmental data in adjacent spatial units (e.g., walkability index score by census block group) would be further reduced.

4.4. Study Strengths and Limitations

To the best of our knowledge, this is the first study that examines GPS-based dynamic exposures to the built and natural environment factors of pregnant women across pregnancy and early postpartum periods. A major strength is the estimation of daily exposure to environmental factors by repeatedly collecting highly resolved smartphone location data across the 1^st and 3^rd trimesters of pregnancy and at 4–6 months postpartum. Consequently, we overcome recall biases inherent in self-reported environmental exposures and provide insights into longitudinal changes in these exposures. Additionally, the longitudinal design used for this study allowed us to examine both the variations in environmental exposures between the pregnant women and the day-to-day variations for each woman, which previous studies applying the static method could not do. Moreover, the study applies both spatial averaging (i.e., route buffer) and time-weighted spatial averaging (i.e., KDE) methods to capture participants’ actual exposures to environmental factors in daily activity spaces. As a result, our exposure estimates may have less error than those derived by the static approach. Lastly, our study is among the first to examine similarities and differences between environmental exposure derived by static and dynamic methods and, to the best of our knowledge, the first one that focuses on pregnant women (Jankowska et al., 2021; Zhao et al., 2018). Our findings of low correlations between static and KDE neighborhood greenness measures have important implications for future studies examining the association between prenatal neighborhood greenness and maternal and infant health outcomes.

Our study also has a few limitations. First, the GPS data we collected has some missingness. To mitigate its impacts on analyses, we made efforts to impute GPS data using existing information and ruled out the existence of diurnal patterns for the remaining missing segments (it was roughly invariant throughout the day). Despite these efforts, there are other factors that may still potentially bias our exposure measurements. For instance, missingness patterns of GPS data may be correlated with spatial context (e.g., tall buildings, trees) that could obstruct receiver signals. As a result, environmental exposures in these spatial contexts may not be captured. Second, our study is subject to the uncertain geographic context problem (UGCoP), which refers to the uncertainties of environmental contexts that influence health behaviors and outcomes (Kwan, 2012). Although the GPS-based dynamic method as applied in our study tackles UGCoP better than the static method through reducing the spatial mismatch between exposure and human movements and outcomes, it is limited in addressing other potential methodological issues such as the choice of spatial parameters (e.g., buffer sizes). In this study, we chose a 250 m buffer size when constructing activity spaces given this distance considers multiple pathways on how the environmental factors influences maternal activities and health. In addition, we reran the correlation analyses using exposure estimates derived from buffer size of 100 m. The study results were largely unchanged and are therefore not reported. In this context, our choice of a fixed bandwidth for the KDE approach may not distinguish between certain scenarios in which the same exposures may have different effects on health behaviors and outcomes (e.g., exposure to greenness from driving vs. walking). As a result, our exposure estimates may still be subject to error. Furthermore, our results are subject to the selective daily mobility bias (Chaix et al., 2013), which may lead to an over- or underestimation of GPS-based environmental exposure and thus influence our results. Third, we chose to aggregate exposure estimates derived by GPS-based methods to day level and interpret the daily exposures as the averaged time-weighted exposure value within activity spaces. Other temporal units (e.g., trip-level, minute-level) can be chosen if future studies are interested in finer grain within-day relationships between environmental exposures and health behaviors. For example, one can examine the greenness exposures within 30 minutes preceding a walking episode to better understand the within-day effects of the greenness exposure on physical activity behaviors. Fourth, environmental exposure could be homogeneous within this population, and misclassification estimates could be underestimated. In addition, some of the datasets we selected for the built and natural environment have a relatively coarse spatial resolution (e.g., the MODIS NDVI product has a resolution of 250 m), which may also reduce the variability of daily estimates of environmental exposure, affecting our assessment of misclassification errors between GPS- and residence-based approaches. Fifth, we only collected 4-day GPS data on two weekdays and two weekend days in each study period. Therefore, the time-activity and mobility patterns identified from our samples may not capture some infrequent activities that are more likely to occur weekly or on other days of the week, such as grocery shopping. Sixth, we used data from a health disparity group of low-income, Hispanic women, a population originally recruited for a substudy nested into a larger cohort study, to examine the daily effects of environmental and social stressors on maternal pre- and postpartum obesity-related behavioral responses. Thus, our results may not be generalized to pregnant women in other regions or SES groups; nevertheless, they shed light on an important population and may pave the way for future studies to examine participants’ environmental exposures and health outcomes during pregnancy and postpartum especially in urban settings.

5. CONCLUSION

Pregnancy and early postpartum are critical periods of exposure, and we have shown that environmental exposures will likely vary across days over these periods, with potential implications for both short- and long-term maternal and child health. More importantly, we have demonstrated that the residential methods commonly used in prior studies can introduce exposure measurement error, the magnitude of which varies depending on the type of environmental feature studied, the pregnancy and postpartum period, and daily mobility. Therefore, future studies examining the impacts of the environmental on maternal and infant health should consider, when possible, the use of GPS-based methods to incorporate the spatiotemporal movement patterns of pregnant women into exposure measurements as these directly influence affect the ability to detect meaningful relationships. As the field moves closer to the vision of precision environmental health, these methods will become increasingly important to incorporate in epidemiological studies of the built and natural environment.