Geographic access to emergency obstetric services: A model incorporating patient bypassing using data from Mozambique

Introduction Targeted approaches to further reduce maternal mortality require thorough understanding of the geographic barriers that women face when seeking care. Common measures of geographic access do not account for the time needed to reach services, despite substantial evidence that links proximity with greater use of facility services. Further, methods for measuring access often ignore the evidence that women frequently bypass close facilities based on perceptions of service quality. This paper aims to adapt existing approaches for measuring geographic access to better reflect women’s bypassing behaviour, using data from Mozambique. Methods Using multiple data sources and modelling within a geographic information system, we calculated two segments of a patient’s time to care: (1) home to the first preferred facility, assuming a woman might travel longer to reach a facility she perceived to be of higher quality; and (2) referral between the first preferred facility and facilities providing the highest level of care (eg, surgery). Combined, these two segments are total travel time to highest care. We then modelled the impact of expanding services and emergency referral infrastructure. Results The combination of upgrading geographically strategic facilities to provide the highest level of care and providing transportation to midlevel facilities modestly increased the percentage of the population with 2-hour access to the highest level of care (from 41% to 45%). The mean transfer time between facilities would be reduced by 39% (from 2.9 to 1.8 hours), and the mean total journey time by 18% (from 2.5 to 2.0 hours). Conclusion This adapted methodology is an effective tool for health planners at all levels of the health system, particularly to identify areas of very poor access. The modelled changes indicate substantial improvements in access and identify populations outside timely access for whom more innovative interventions are needed.

Mozambique, a low-income country in southern East Africa, had a gross national per capita income of US$480 in 2016.24 Divided into 11 administrative units (provinces), including the capital city of Maputo, the country is crossed from west to east by large rivers, isolating some parts during the rainy season. The population in 2012 was 23.6 million.25 The national health system provides the vast majority of healthcare services, while private sector facilities are available primarily in urban areas.26 The health system has four levels of progressively complex care: the first level (health centres and posts) provides primary care including basic maternal and child health services; the secondary level (rural, district or general hospitals) may offer surgical services and serves as a referral level for the first; the tertiary level (provincial hospitals located primarily in capitals) functions as the next referral level; and the quaternary level (central hospitals) serves as the regional referral level.26 However, women’s choice of facility does not necessarily follow this clearly defined pyramidal structure, particularly in emergencies. The maternal mortality ratio in Mozambique has remained high, hovering around 480/100 000 live births, over the past two decades.27 In 2012, the proportion of deliveries occurring in facilities was 67%, yet the caesarean delivery rate was 2.8% of all expected deliveries, indicating that not all women needing this life-saving intervention are able to access it.28 Using multiple data sources and modelling within a GIS environment, we calculated two segments of a patient’s journey to care. The first segment was home to the first preferred facility assuming women, if given a choice, would travel longer to reach a facility perceived to be of higher quality. If the first preferred facility did not offer the highest level of service, the woman may need to be referred upwards to the closest facility that did. This would be the woman’s second journey segment. Together these two segments made up the total travel time to highest care. We then used these models to measure resulting changes in the proportion of the population with access to the highest level of care after upgrading strategically located facilities to provide higher-level services and placing ambulances and communication modes at midlevel to lower-level facilities. Facility data from Mozambique’s 2012 assessment of emergency obstetric and newborn care were used with permission from Mozambique’s Ministry of Health. Data were collected between November and December 2012 through a cross-sectional survey of health facilities. The survey included a census of all health facilities that provided delivery services in the previous year, regardless of volume (946 facilities). Six data collection modules were used in the assessment, though this secondary analysis included data from just four28: facility infrastructure; human resources; essential drugs, equipment and supplies; and facility case statistics. Data collectors were medical students in their final year or recent medical graduates. All were trained with a standardised curriculum over 5 days. The survey received approval from the local ethics committee. Data were double-entered into EpiData and exported to Stata V.13.29 For complete methods, see the final survey report28; for distribution of facilities by type and province, see online supplementary file 1, Table 1.1. bmjgh-2018-000772supp001.pdf A master list of facility geographic coordinates was provided by the Ministry of Health. This list, accurate as of 2012, included 1266 health facilities, some of which did not provide delivery services. Using data management techniques in Excel and multiple rounds of manual matching, we matched health facilities in the assessment to facilities on the master list. For those not on the master list, we triangulated data from facility location information in the assessment (eg, facility name, district, zone, region) with data available through several online sources (Google Maps, Google Earth, GeoNames, OpenStreetMap and The Fuzzy Gazeteer) to identify a settlement that matched the facility location. We then selected coordinates from a central location of that settlement. If a settlement was large, more than one settlement was possible, or we found no likely settlement, that facility was excluded from our analysis (online supplementary file 1, Table 1.2). For each facility, we used assessment data to calculate a score based on facility characteristics known to influence women’s perceptions of quality. We did not attempt to categorise facilities by actual quality of service. We also did not use the formal definition of fully functioning basic EmOC and comprehensive EmOC, as determined by recent performance of the life-saving interventions known as the EmOC signal functions.5 Rather, our score was graduated along five levels that were largely based on the facility’s readiness to provide each of the nine signal functions (online supplementary file 2, Table 2.1). Using a facility’s signal function readiness rather than actual provision allowed inclusion of facilities that might not have provided the function recently, perhaps due to low patient volume, but theoretically could have if a patient needed it. The score had a maximum of 14 points and was based on two dimensions: bmjgh-2018-000772supp002.pdf Based on the facility score and two other items—presence of functioning transport and recent performance of caesarean delivery—we ultimately placed facilities in one of five levels (table 1) (detail in online supplementary File 2 Section 2.2). Definition of facility levels and time-bounded catchment areas The access models were created in stages within ArcGIS V.10.3 software30 using the Spatial Analyst and Network Analyst extensions. To calculate travel time to health facilities, a single cost–distance raster layer estimated the cost in minutes required to cross each cell. The cost was determined using three data sources: land cover, road networks and elevation. Land cover31 was used to represent likely pedestrian travel times, and walking speeds were defined for each type of land cover.32 The road network vector data set was downloaded from OpenStreetMap33 and used to estimate motorised travel speeds.34 We replaced pedestrian values with motorised travel speeds in land cover cells where roads overlapped; some road cells overlapped empty spaces once occupied by river features, indicating bridges that pedestrians and vehicles could move across. Finally, this raster of combined pedestrian and motorised rates was attenuated by slope using a digital elevation model35 and the Van Wagtendonk formula.36 Additional information, including assumed rates of travel, can be found in online supplementary File 3. The population data set used was created by WorldPop with a population projection to the year 2010, in a raster format with approximately 100 m × 100 m cells.37 The 2010 population was used in the modelling to establish proportions of the population with various levels of access. Absolute values of population were calculated based on population projected to 2012 to align with the year of the Mozambique facility assessment. Where estimates of expected pregnancies and severe obstetric complications are reported, they were determined by applying Mozambique’s 2012 crude birth rate to the 2012 population25 and the estimate that 15% of expected pregnancies will result in severe complications.5 bmjgh-2018-000772supp003.pdf We defined catchment areas with time-bound limits to simulate decision points at which women might evaluate a trade-off between perceived quality of services and time to care (table 1). The health system must be organised to ensure timely, universal access for all scenarios pregnant women may face. We framed our model assumptions for measuring timely access against the most time-critical complication, postpartum haemorrhage, thus the selected time bounds include the clinically relevant 2-hour mark and a maximum travel time of 5 hours.38 These time-bound intervals were exploratory, as we found no literature that investigated the travel time parameters around women’s choice of facility. Women’s decision points are likely more fluid; yet, to operationalise the approach we defined finite ranges. Each time-bound catchment area was prioritised beginning with primary catchment area level 5, followed by level 4, and then lower levels (table 1). In ArcGIS, we layered health facilities onto the cost–distance raster described above, and for each facility created polygons of time-bound catchment areas outwards into the surrounding space until the primary and secondary travel time thresholds were reached. This generated multiple overlapping time-bound catchment areas—one for each facility’s primary and secondary catchment areas. For cells where catchment areas overlapped, we maintained the area with the highest priority and deleted the others. This created a complex patchwork of catchment areas modelling how women in these areas might bypass a nearby facility in favour of one further away. Figure 1 helps visualise this approach: if a woman lives within the primary catchment area of a level 3 (defined by 0–1 hour of travel time) and within the primary catchment area of a level 5 (defined as between 0 and 2 hours of travel time), our model prioritises the level 5 catchment area and assumes she would choose to bypass the closer level 3 facility in favour of accessing the perceived better-quality services of the level 5 facility. In this article, that level 5 would be her first preferred facility. Alternately, if a woman lives within the primary catchment area of the level 3 and within the secondary catchment area of the level 5, the model assumes she would travel directly to the level 3, and that would be her first preferred facility. Illustrative diagram of 1-hour catchment area boundaries around two facilities, their relative priority and resulting modelled behaviour. Finally, we overlaid the population layer to calculate the population with access to their first preferred facilities within hourly segments. For women whose first preferred facility is not a level 5, we modelled interfacility referral using the ArcGIS Network Analyst extension to calculate the direct transfer time between each facility and the closest level 5. Transfer time was estimated along the road network based on assumed travel speeds for each road type. We adjusted the direct transfer time per the availability of vehicles and communication at the sending facility: if the facility had a functioning vehicle, the total transfer time was maintained as the direct time; however, if the facility had no vehicle, the transfer time was doubled since we assumed the receiving facility would send a vehicle to retrieve the patient. All but one level 5 had transport; our status quo model would require functional transportation be placed there. If the sending facility did not have communication, we further increased the transfer time by 30 min to allow time to locate a phone.11 To describe changes in journey segment 2, we reported the population-weighted interfacility transfer times. It was necessary to weight by population because the units of analysis were segments of catchment areas of varying population sizes. In Stata V.13, we calculated total travel time to a level 5 by adding travel times of both journey segments. To evaluate changes across models, we reported population-weighted total travel times. We then applied these methods to four models: No patients were involved in this research.

N/A

• Dima Health