How can childbirth care for the rural poor be improved? A contribution from spatial modelling in Rural Tanzania

Introduction Maternal and perinatal mortality remain a challenge in resource-limited countries, particularly among the rural poor. To save lives at birth health facility delivery is recommended. However, increasing coverage of institutional deliveries may not translate into mortality reduction if shortage of qualified staff and lack of enabling working conditions affect quality of services. In Tanzania childbirth care is available in all facilities; yet maternal and newborn mortality are high. The study aimed to assess in a high facility density rural context whether a health system organization with fewer delivery sites is feasible in terms of population access. Methods Data on health facilities’ location, staffing and delivery caseload were examined in Ludewa and Iringa Districts, Southern Tanzania. Geospatial raster and network analysis were performed to estimate access to obstetric services in walking time. The present geographical accessibility was compared to a theoretical scenario with a 40% reduction of delivery sites. Results About half of first-line health facilities had insufficient staff to offer full-time obstetric services (45.7% in Iringa and 78.8% in Ludewa District). Yearly delivery caseload at first-line health facilities was low, with less than 100 deliveries in 48/70 and 43/52 facilities in Iringa and Ludewa District respectively. Wide geographical overlaps of facility catchment areas were observed. In Iringa 54% of the population was within 1-hour walking distance from the nearest facility and 87.8% within 2 hours, in Ludewa, the percentages were 39.9% and 82.3%. With a 40% reduction of delivery sites, approximately 80% of population will still be within 2 hours’ walking time. Conclusions Our findings from spatial modelling in a high facility density context indicate that reducing delivery sites by 40% will decrease population access within 2 hours by 7%. Focused efforts on fewer delivery sites might assist strengthening delivery services in resource-limited settings. This study was conducted in two rural districts, Iringa and Ludewa, formerly both part of Iringa Region, and presently in Iringa and Njombe Regions (Fig 1), characterised by high maternal services uptake. According to the latest National Demographic Health Survey the regional estimates for antenatal care coverage (at least one visit) and institutional deliveries were respectively 97.3% and 80.4% [26]. Iringa District has a habitable surface of 9857 Km2 and is morphologically and climatically divided into three areas: the Highlands (over 2000 m asl) located in the south-west, the Midlands in the north-east and the Lowlands in the north-west, the latter being extensively covered by a national park and mostly uninhabited. According to the national census the population in 2012 was 254,023, with 85% relying on subsistence farming. The total number of villages was 122 and the road network consisted of 1272 km of tarmac and unpaved roads [28]. Health services in 2012 were available in 72 facilities, of which 65 dispensaries, 6 health centres and one diocesan District-designated hospital. The majority of health facilities were public, with only 27% run by private non-profit organizations. Delivery services were provided by all health facilities except one dispensary located in Ruaha National Park. Ludewa District is part of Njombe Region and borders Lake Malawi. It has a population of 133,218 and a habitable surface of 6012 Km2. The district is predominantly rural with 77 villages and a road network of 970 Km. People derive their livelihood from subsistence farming, fishing, mining and small scale trading [29]. Morphologically Ludewa District is represented in the north-west by a mountainous and rainy area (Livingstone Mountains with an average elevation of over 2000 m asl), in the centre by a midland area with high population density and in the south by a lowland area, mostly flat and fertile near the Ruhuhu river. Health services in 2012 were provided by 55 facilities, of which 46 dispensaries, 6 health centres and 3 hospitals. 20% (11/55) of health facilities were run by private non-profit organizations. Both districts are characterised by limited road infrastructure and scarce transport services. Data on 2012 staffing and health facility deliveries were obtained at district level from the Human Resources Information System and Health Management Information System and were validated during site visits by comparing routinely collected data with local registers. The Iringa District data were part of a larger dataset described in a previous study [18]. Clinicians, enrolled and registered nurses were categorized as skilled birth attendants while nursing assistant were not, according to the health system organization in Tanzania [17]. The lists of villages and existing health facilities were obtained from District Land Offices and District Medical Offices. Geographical data on location of health facilities were collected during site visits from April to August 2013. A waypoint was marked at the main entrance of each facility by using hand-held Global Position System units (Garmin Etrex 10, Nokia E5-00SE) or an USB GPS device connected to a laptop. Longitude and Latitude data were recorded with an accuracy of 5–15 metres in WGS 1984 Datum coordinate system and were converted in WGS_1984_UTM_Zone_36 system. Geographical coordinates were recorded for all health facilities (55) of Ludewa District, and for 71 out of 72 of Iringa District. As mentioned before, data was not collected from one dispensary that does not provide delivery services. In Ludewa District more than 650 km of motorable roads and 24 km of footpaths were recorded with the track function of Garmin Etrex GPS respectively by car transport and walking. Topographic sheets with scale 1:50000 of United Republic of Tanzania (completed between 1977 and 1982) were used as source of data for villages, watercourses, lakes, roads, tracks and footpaths. Road network and location of villages were updated using remote sensing by comparison of available topographic data with Bing™ and Google Earth™ satellite images. Free online datasets on district boundaries, main road network, main cities, main river network and land use were downloaded from FAO GeoNetwork [30]. A 90 metres cell Digital Elevation Model (DEM) was acquired by US Geological Survey (DEM courtesy of the U.S. Geological Survey) [31] and was used to create the slope and aspect maps. Geographical and administrative data were processed with ArcGIS 10.2™ software (ESRI, Redlands, CA). The road network of each district was classified into 5 classes according to road size, type of surface and seasonal condition (Table 1). Road classification was carried out by merging information gathered through Google Earth™ and Bing™ satellite images with 1:50000 maps and data collected in the field. Walking speeds were recorded through a sample walk of 24 km across a morphologically representative area. Data were collected using a healthy male volunteer with a portable GPS. The recorded track log was split into 10% slope intervals and walking speeds were tabulated for each category of slope. The findings from our study were compared with previous published data [32–34]. Despite a similar trend in speed reduction by degree of surface inclination, absolute values were different from those estimated by other authors (Table 2). It was therefore decided to adopt the more conventional Naismith-Langmuir Rule to estimate walking speed for the study population composed mainly of pregnant women (Table 2) [34]. The Langmuir Rule (1984) assumes a basal speed of 4 km/h and modifies the value according to the following conditions: increased travel time by 0.1 min per 1 m in ascent; reduced travel time by about 0.03 min per 1 m in descent, in the range of slope from -5°: to-12°; increased travel time by about 0.03 min per 1 m in descent for slope steeper than 12° (Noor A.M. et al., 2006). Vehicular travel time was calculated by applying to each type of road a set of travel speeds as described by other authors (Table 1) [25, 35]. Raster datasets on population distribution and population density were obtained from Worldpop Project [36]. Population data for Ludewa and Iringa Districts were updated by applying an adjustment factor to the Worldpop 2010 database based on the 2012 National Census Data [37] without changing the spatial distribution of the population. Data on population density were matched to each spatial model to estimate the percentage of population living in the two-hours’ catchment area. Data management and analysis were performed with ArcGIS™ 10.2 for Desktop (ESRI™), and Python scripts compiled by the authors. Raster and network analysis were carried out with ArcGIS™ extensions 3D Analyst, Network Analyst and Spatial Analyst. Geographical conversions and transformations from and to WGS_1984_UTM_Zone_36S (WKID: 32736 Authority: EPSG) were executed by the project tool supplied by ArcGIS™. Topological functions were used to check the consistency and coherence of line data sets. Minor calculations and ancillary data were managed with MS Excel™ spreadsheet functions and Python scripts. Two spatial models, based respectively on raster and network analysis, were developed to estimate travel time to reach the nearest health facility. Facility catchment area within 2 hours’ walking time was defined with both methods and respective findings were compared for validation. Scenarios with reduced number of delivery sites were described and assessed for population coverage. In addition exploratory network analysis was carried out to assess multimodal transport (pedestrian and vehicular). In the raster method the study area was split into unit cells of 90 square meters, the same resolution used by the Digital Elevation Model. Cost raster maps were created by intersecting basic raster maps with slope and land use surface datasets. Time to cross on foot each cell was adjusted for surface inclination (ascendant or descendant), land use, topographic features and seasonal variation by applying friction coefficients to walking speeds (Table 2 and Table 3). Lakes and swamps were classified as non-passing areas. Friction coefficients were applied to each raster cell to estimate the time needed to cross the cells on foot according to surface characteristics. Walking time towards health facilities was estimated with the Path Distance ArcGISTM Tool by identifying the shortest distance to the nearest health facility. Total walking time was estimated by adding up the time needed to cross contiguous cells and by taking into consideration geographical obstacles and slow crossing areas. Cost distance maps were built for two scenarios: one with all existing health facilities and one with a reduced number of facilities that could be accessible within two hours’ travel time by approximately 80% of the population. Selection criteria for including health facilities in the second scenario were based on spatial aspects and population density. The Path Distance Output Raster was matched with population data (AfriPop) to estimate the cumulative percentage of population living within subsequent 20 minutes’ intervals of walking time. The 20 minutes’ time interval was arbitrarily chosen to map consecutive catchment areas around each health facility to provide sufficient details at local level. To compare population coverage by different scenarios three levels of access to health facilities were considered: less than one hour, less than two hours and greater than 2 hours’ walking time. These thresholds were based on previous studies [38, 39] and on general guidelines [21]. The analysis was performed with the function Zonal Statistic as table tool of ArcGIS™. In rural settings such as in Iringa and Ludewa Districts, where road infrastructure and vehicular transport is limited, the application of network-based methods is hampered by lack of sufficient digital data in vector form. To overcome this constraint we constructed a network based on a combination of real and virtual data. Existing road network was merged with a virtual hexagonal /triangular mesh of lines. The surface of Iringa and Ludewa Districts was divided into 342000 and 539000 hexagonal / triangular areas with side lengths up to 223 metres. Each side is characterized by an attribute value of estimated travel time towards the health facilities, according to type of slope (ascendant or descendent) and mode of transport (pedestrian or vehicular). A similar technique was previously applied in urban areas of Dar es Salaam [40]. As validation, findings from network analysis were compared with data recorded during the sample walk. Route analysis methods were applied to the virtual network to draw the shortest track between the starting and ending point of the sample walk. For each health facility catchment areas based on consecutive 20 minutes’ walking time were built. Cumulative population coverage within the three levels of access was estimated with a Zonal Static Tool matched with the Afripop Raster Database. A scenario with a reduced number of delivery sites was produced for pedestrian transport and was compared with the raster methods outputs for validation. An additional scenario for both pedestrian and vehicular transport (multimodal transport) was produced as an attempt to simulate travel patterns with motorised transportation. Data used in this study were either available as unrestricted sources in the public domain or provided with permission from local health authorities. Health information was extrapolated from routinely collected data in an aggregated form. No data was collected at individual level.