Distance to available services for newborns at facilities in Malawi: A secondary analysis of survey and health facility data

Background Malawi has halved the neonatal mortality rate between 1990-2018, however, is not on track to achieve the Sustainable Development Goal 12 per 1,000 live births. Despite a high facility birth rate (91%), mother-newborn dyads may not remain in facilities long enough to receive recommended care and quality of care improvements are needed to reach global targets. Physical access and distance to health facilities remain barriers to quality postnatal care. Methods Using data We used individual data from the 2015-16 Malawi Demographic and Health Survey and facility data from the 2013-14 Malawi Service Provision Assessment, linking households to all health facilities within specified distances and travel times. We calculated service readiness scores for facilities to measure their capacity to provide birth/newborn care services. We fitted multi-level regression models to evaluate the association between the service readiness and appropriate newborn care (receiving at least five of six interventions). Results Households with recent births (n = 6010) linked to a median of two birth facilities within 5-10 km and one facility within a two-hour walk. The maximum service environment scores for linked facilities median was 77.5 for facilities within 5-10 km and 75.5 for facilities within a two-hour walk. While linking to one or more facilities within 5-10km or a two-hour walk was not associated with appropriate newborn care, higher levels of service readiness in nearby facilities was associated with an increased risk of appropriate newborn care. Conclusions Women’s choice of nearby facilities and quality facilities is limited. High quality newborn care is sub-optimal despite high coverage of facility birth and some newborn care interventions. While we did not find proximity to more facilities was associated with increased risk of appropriate care, high levels of service readiness was, showing facility birth and improved access to well-prepared facilities are important for improving newborn care. Data for this study were used under an agreement with the DHS Program. The original survey protocol was reviewed and approved by the National Health Sciences Research Committee in Malawi and the ICF Institutional Review Board. Informed consent and voluntary participation were ensured before each interview and data were kept strictly confidential during the survey implementation and identifying information was destroyed after data processing. The King’s College London College Research Ethics Committee granted approval to conduct these analyses (LRS-17/18-5570) and the project has been registered with the King’s College London Data Protection Registration (DPRF-17/18-8170). We analysed individual and health facility data from two data sources: the 2013–14 Malawi Service Provision Assessment (SPA) survey and the 2015–16 Malawi Demographic and Health Survey (DHS) [2, 19]. SPA surveys collect information on health service availability and the readiness to provide these services [20]. The 2013–14 Malawi SPA was conducted as a census, surveying all facilities in Malawi including all hospitals, health centres, clinics, dispensaries, and health posts. Primary level services include community or rural hospitals, health centres, clinics, dispensaries, health posts, and support for community-based health programs. District hospitals provide inpatient and outpatient care and serve as referral hospitals for primary level facilities. Tertiary services are covered by central hospitals, and serve as referral hospitals for district hospitals, which are situated with secondary level services. Hospitals and health centres are almost exclusively responsible for providing normal birth services [19]. SPA data included information from 977 of the 1060 health facilities in Malawi (other facilities refused participation, were inaccessible, closed, not yet operational, or a respondent was not available). DHS surveys collect data through face-to-face interviews with household representatives and women of reproductive age. Complex multistage sampling with stratification is designed to provide representative national estimates of important demographic and health indicators [21]. For the Malawi DHS, 850 standard enumeration areas (SEA) from the 2015–16 Malawi Population and Housing Census were selected with probability proportional to size, independently from 56 sampling strata in the first stage. SEAs which had more than 250 households were split into segments with one segment selected with probability proportional to size such that survey clusters were either an SEA or a segment of an SEA. Following a household listing operation in each cluster, the second stage included selection of a fixed number of households (30 in urban clusters, 33 in rural clusters) using equal probability systematic selection [2]. We selected for inclusion the most recent birth in the two years preceding the survey from the 2015–16 DHS. A priori, newborns were excluded if they were not two days of age at the time of the survey, or had not survived the first two days of life as the interventions of interest included the content of care offered during this initial time period. Newborns who had not yet reached two days of age might yet receive the interventions of interest and some interventions of interest may not have been appropriate for newborns who died soon after birth. We also excluded newborns if the woman reported not living in the current community at the time of the birth. One-third of Malawi’s land area is forest area [22] and about 20% is covered by water, primarily Lake Malawi. The Highlands reach an elevation of 1,600–3,000 metres above sea level [23]. According to the 2018 Malawi Population and Housing Census, the population was 17,563,749 with 12% of the population residing in one of four major cities (Blantyre, Lilongwe, Mzuzu, and Zomba) and another 4% residing in other urban areas [24]. While Malawi has an extensive road network, walking is the most used mode of travel in rural and urban areas [23]. Community studies have shown people travel long distances to access health facilities under difficult terrain in some areas of Malawi [25]. The key independent variables reflected proximity to health facilities providing birth services, and readiness to provide birth services of those proximate facilities. We constructed a service readiness score for birth and newborn services with an equal weighting approach similar to that used by Wang et al. [26] and based on the WHO Service Availability Readiness Assessment (SARA) manual [7]. Comparison of measures of quality of care have found this type of weighted additive method to be preferable to simple additive methods or principal components analysis [27]. The score assessed six domains of service readiness comprising: 1) basic emergency obstetric care; 2) newborn signal functions and immediate care; 3) general requirements (e.g. electricity, 24/7 skilled birth attendance); 4) equipment (e.g. neonatal bag and mask); 5) medicines and commodities (e.g. antibiotics); and 6) guidelines (e.g. CEmOC), staff training (e.g. thermal care), and supervision. Each domain included 4–15 dichotomous indicators (‘yes’ representing availability, ‘no’ representing no availability) which were summed and standardised to have a maximum score of 100 (definitions of all included indicators are presented in S1 Table). The score is interpreted as the percentage of readiness the facility has to provide services. A facility with 100% has a positive response for every measured indicator and a facility with 0% has none of the measured equipment, staff training or other indicators. The primary outcome measure was receipt of appropriate newborn care. We created a co-coverage index of newborn care interventions, using a method similar to Victora et al. [28] and Carvajal-Aguirre et al. [29], adding the number of care components women reported their newborns had received from six provider-initiated interventions recommended by WHO [12]. We considered newborns who received at least five out of the six interventions to have received appropriate care. This included newborns who received all six interventions (optimal) and those who received any combination of five interventions (pragmatic). The interventions considered included: weighing at birth, mother counselled on breastfeeding, mother counselled on newborn danger signs, breastfeeding episode observed, umbilical cord examined, newborn’s temperature taken. The survey questions for these interventions are presented in S2 Table. Facility and home births were included in this analysis. A qualitative study of women giving birth outside of facilities in Malawi showed that most in the sample subsequently went to a facility the same day as the birth [30] suggesting that proximity to and quality of facilities should be considered for early newborn care even among home births. While the interventions we included were specifically about delivered by health care providers, we did not distinguish between health care providers delivering interventions in the home/community setting or facility setting nor did we distinguish between pre- or post-discharge for facility births. Additional analysis excluding home births is also presented. DHS surveys collect GPS location points at the centroid of household clusters and SPA surveys collect the GPS location of health facilities. GPS location data for health facilities represent the true location of the facility, however, household cluster data were displaced by the DHS programme prior to release to protect the respondents’ identities (urban clusters up to two kilometres (km), rural clusters up to five km with a further randomly selected 1% displaced up to ten km [31]. We used GPS location data from household clusters to link households to nearby health facilities providing birth/newborn services in Malawi using three methods: distance, travel time with the fastest mode of transport, and walking time. For each linking method, we categorised the number of facilities linking to households within the specified distance/time into three groups: no facility, one facility, or two or more facilities. Similarly, for each linking method we grouped service environment scores into terciles based on the highest score among all facilities linking with a household. Low, middle and high terciles were chosen to improve interpretation and understanding over use of a continuous score. We calculated the straight-line distance between every DHS household cluster and every health facility in Malawi. A Euclidean buffer link method [32–34] was used to create a buffer centred around each DHS cluster using a radius of 5 km in urban areas and 10 km in rural areas to account for displacement of household clusters. All health facilities falling in these 5–10 km buffer areas are considered linked to the household cluster, without consideration of sub-national boundaries. Fig 1A shows an illustrative example of household-facility distance linking. We calculated travel time from household clusters to facilities providing birth services using two scenarios: a) fastest possible mode of transportation (i.e. best-case scenario, similar to other studies estimating travel time to hospitals [16, 35, 36]) and b) walking only (i.e. worst-case scenario). A) Fastest mode of transportation. we used the 2015 Malaria Access Project Global Friction Surface [37, 38] which divides the world into one kilometre-square grid cells with each cell value representing the difficulty of crossing the one kilometre cell based on road quality, bodies of water, and sloping terrain. Assuming use of the fastest possible mode of transportation, an algorithm was applied to identify the path requiring the least time to travel between any two points on the friction surface [37, 39]. We employed the algorithm for all possible pairs of DHS household clusters (n = 850) and health facilities with birth services (n = 540), a total of 459,000 pairs. Travel time could not be calculated for 17,360 pairs (3.8%), however, largely for combinations involving one point on Likoma or Chizumulu Island and one point on Malawi mainland as well as for a few health facilities on the national border with Mozambique. All household-facility pairs within a two-hour travel time (fastest mode) were classified as linked. B) Walking only. As many women in Malawi walk to health centres [40], we also calculated travel times for walking as a worst-case scenario. To calculate walking times we used the Open Source Routing Machine (OSRM) API [41] and Google Maps Platform Directions API [42, 43] to plot the optimal route by foot and compute an estimated travel time. We first attempted to calculate walking times for all 459,000 household-cluster-health-facility pairs using OSRM, however, OSRM walking times could not be calculated for 47,381 pairs. For these 47,381 pairs, we calculated the distance and identified pairs <50 km apart (n = 1,775) using the Haversine method which assumes a spherical earth ignoring ellipsoidal effects [44]. Using the Google Maps API, we calculated walking times for 1,742 of the pairs <50 km apart, however walking times could not be calculated for 33 pairs. These totalled 411,619 OSRM walking time estimates (89.7%) and 1,742 Google Maps walking time estimates (0.4%). No walking time estimate was calculated for 45,639 pairs, however all but 33 were determined to be >50km apart. Examination of the coordinate pairs for missing points showed them to be in national parks, forest reserves, or across bodies of water. All household-facility pairs within a two-hour walking time were classified as linked. Fig 1B shows an illustrative example of household-facility travel time linking. Simple weighted descriptive statistics on coverage of appropriate care and proximity and service readiness of facilities were calculated. We fitted generalised linear mixed models to examine the relationship between co-coverage of newborn care and the number of linked facilities or the service environment of linked facilities. To estimate risk ratios for our binary outcome variable, we used a Poisson distribution with a logarithm link function and robust standard errors [45]. We controlled for socio-demographic and birth-related factors (population density, place of birth (home/health facility), wealth quintile (DHS-provided), maternal age at birth, and maternal education). Population density was obtained from The DHS Program’s Spatial Data Repository Geospatial Covariates which uses the average United Nations population density within the surrounding buffer area (2 km for urban clusters or 10 km for rural) [46]. We grouped households into terciles based on population density. As individuals are nested within clusters (SEAs) and our main predictors are cluster-level variables, the multilevel models account for this nesting and simultaneously test the effects of cluster-level and individual-level predictors on the individual outcome. While covariates had fixed effects, intercepts could vary randomly across clusters. Level-1 and level-2 weights were computed using a method described by Elkasabi et al. where level-1 individual weights are denormalised and the level-2 cluster weights are approximated by equally allocating the variation between the individual and cluster levels (α = 0.5) [47]. All geographic linking and descriptive statistical analyses were conducted in R [48], using the survey package [49] to adjust for the complex sampling design. DHS-provided weights were used to account for sampling probability and non-response. Multilevel models were fitted in STATA 16 using the svy: melogit command [50].