SMRU has been based on the Thailand-Myanmar border for over 30 years. In response to an estimated maternal mortality in refugees of 1000 per 100,000 live births in 1985-86, SMRU established a system of weekly ANCs to offer early detection and treatment of Plasmodium falciparum malaria in pregnancy [19]. As the population of migrants grew in the 1990’s and 2000’s, SMRU opened four facilities on the international border formed by the Moei River. The facilities provided both antenatal care and childbirth services to local populations who had settled outside of the region’s refugee and internally displaced person camps (Fig. 2). The antenatal care and childbirth services are provided by skilled birth attendants, who are comprised of local staff trained in line with the WHO guidelines to provide evidence and human rights based, quality, dignified care, manage the physiological processes of labor and delivery, and facilitate timely management and referral of complications [2]. These skilled birth attendants conduct health promotion, screening, diagnosis, and administer treatments using evidence-based protocols with licensed physicians available for 24-h medical back up via telephone [20].
These four antenatal clinics on the international border served populations living in both Myanmar and Thailand. Mawker Thai (MKT) began providing ANC and childbirth services in 1998, Walley (WAL) and Mu Ru Chai (MRC) began providing ANC care in 2001, and Wang Pha (WPA) began providing ANC and childbirth services in 2004. WAL closed operations in Jul 2010 and MRC in Dec 2012 and services were amalgamated at MKT. All services were free of charge [21] and attendance at ANCs was voluntary. Trained local sonographers determined gestational age using ultrasound offered at the first antenatal visit [22]. Ultrasound becomes increasingly imprecise at estimating gestational age in those presenting after 24 weeks [1], and so clinical staff derived gestational age from the Dubowitz assessment of gestation at birth, last menstrual period, or symphysis fundal height in people presenting after 24 weeks [23].
The maternal health facilities kept antenatal medical records for each pregnancy from 2007 to 2015 which were de-identified and include general demographic information (patient age, gravidity, parity, home village name, and time lived at home village), antenatal care attendance information (estimated gestational age at initial presentation, miscarriages), pregnancy complication information (malaria infection with P. vivax, P. falciparum, or both, multiple pregnancy, very young age), and presence of skilled birth attendants at childbirth (loss to follow-up and normal singleton delivery) [24].
Loss to follow-up was defined as a person who enrolled at ANC but then stopped attending and did not return for childbirth. Most people travel on foot to ANC appointments and a minority hire motorbikes or long-tractors for transportation. SMRU partially subsidizes transportation by car to prenatal visits for those who live on the Thai side of the border to MKT and WPA. All four facilities are built on the Thailand bank of the Moei River. Depending on seasonal variations in rainwater people may use temporary bridges or boats to cross the river.
In 2014, SMRU worked to create and update a geographic information system (GIS) database for Kayin state, Myanmar and Tak province, Thailand. In Kayin State, SMRU collaborated with several local community based organizations and travelled to remote villages by car, boat, and foot to obtain coordinates, which was the first systematic geographic study in the area since before World War II [25]. In Tak province, the Tak Malaria Initiative [26] had previously gathered village coordinates (latitude and longitude), and SMRU performed an updated geographic survey in 2014. We used the Kayin and Tak GISs to link each unique pregnancy with the geographic coordinates of their home village.
Linking geospatial data to de-identified patient data
We used geocoding to convert place names written in patient’s medical records into map coordinates. The geocoders were blinded to all information in the patient record except village names, which they matched with coordinate data in the Kayin and Tak GISs. For the portion of villages not listed in GISs (n = 105/1152), ANC clinic administrators with more than 20 years of experience pinpointed villages using Google Earth software. Patient addresses weren’t consistently recorded at ANC clinics until late 2006 and we therefore limited this analysis to 2007 through 2015. We excluded all records that did not include a village name. We also excluded villages that were greater than 35 km from the ANC, given that our accuracy in correctly identifying villages may be better closer to the clinics (a histogram of the distribution of distances is provided in Supplemental Fig. 1).
Univariate analyses
We estimated the catchment areas for patients attending the respective clinics using standard deviational ellipses (SDE) and calculated the catchment area by converting the SDEs into square kilometers (km2), which we describe in greater detail in Supplemental Text 1 and visualized in Fig. 2 (and Supplemental Fig. 2). Briefly, we generated ellipses for each clinic using the home villages of patients who attended that clinic. The SDE measures two-dimensional spread along an X- and Y-axis from the geometric mean center of a set of points (in this case, home village locations for patients). The Y-axis is rotated until the sum of the squares of the distances between points and axes are minimized. The resulting ellipses provide a visual representation of 63, 98, or 99% of all home villages for 1, 2, or 3 standard deviations, respectively.
In order to investigate potential differences in travel distance we use the straight-line (Euclidian) distance between each patient’s village and the facility where they received care. We calculate distance as the distance between GPS coordinates for the patient’s home village and the clinic they visited, which admittedly does not take into account the ruggedness of terrain or actual road distance. We then calculate univariate and bivariate descriptive statistics for travel distances (minimum distance, maximum distance, first quartile (Q1), third quartile (Q3), and median distance) based on year of childbirth, parity, age, malaria infection status, and pregnancy outcome (i.e., singleton delivery, twins, lost to follow-up, or miscarriage). In our study, we define miscarriage as birth before 28 weeks gestational age. We based this decision on the WHO definition of stillbirth and past studies of the newborn population on the Thailand-Myanmar border [27].
Negative binomial regression
We use negative binomial regression to formally analyze variables associated with distance to a clinic, and selected distance between home village and clinic (in kilometers) as the outcome variable. Variables in the model included: pregnancy outcome (normal singleton birth, twins, miscarriage, or lost to follow up), the trimester in which the person first presented at the clinic (first, second, or third), country of home village (Thailand or Myanmar), whether or not the person had a P. falciparum infection during pregnancy, whether or not the person had a P. vivax infection during pregnancy, antenatal clinic where patient received care (MKT clinic, MLC clinic, WAL clinic, WPA clinic) age group (13-14, 15-19, 20-24, 25-29, 30-24, 35-39, and 40+), year of childbirth, parity (0, 1, 2-3, 4-5, 6-9, 10+), and the duration of time the person had lived in their home village (< 1 year, 1-3 years, 4-9 years, 10 or more years). A more detailed description of all the variables we included in the negative binomial regression can be found in Supplemental Table 1.
The output from the negative binomial regression model is a distance ratio (DR), which can be interpreted as a ratio comparing the distance (in km) travelled of one group to a comparison group. For instance, a DR of 1.4 for patients lost to follow-up, in comparison to those with a normal singleton childbirth, indicates that patients lost to follow-up travelled 40% farther than those with normal singleton childbirths after controlling for the other variables in the model. We will henceforth refer to the distance ratio as DR, and we report it alongside a 95% confidence interval (CI).
To check for model sensitivity to potential errors in geocoding, we stratified our negative binomial regression model by distance from the clinic (Supplemental Table 2). We also assess for potential changes in associations between variables over time by stratifying the model by time period (Supplemental Table 3). To assess for differences in our results between people living in Myanmar and Thailand, we also stratify the model by nation of origin (Myanmar or Thailand) (Supplemental Table 4). We also include a variable for the country of home village in the main model to account for important differences between those living in Myanmar and Thailand, including that some people on the Thailand side receive subsidies for transportation to clinic.
Statistical software
We created maps using QGIS version 3.4.9. We use the Python programming language (version 3.6) to merge geocoded home villages to the patient records, and R statistical software version 3.3.2 for all statistical analyses.