Introduction

The global maternal mortality ratio has declined by 44% since 1990, though these gains have been geographically disparate and many countries, including Mozambique, did not reach targets set by the Millennium Development Goals.1 To further reduce mortality to the ambitious targets set by the Sustainable Development Goals, recent gains must be expanded through targeted responses that consider not only national contexts but also the complexity of health system capability, and subnational and local contexts.2 3 Providing planners with better information on variations in geographic access to lifesaving services could help countries target appropriate strategies.4

One traditional measure of geographic access, the ratio of facilities per population, is a crude proxy of a population’s access to services. Yet, it remains a core measure at global and national levels.5 Facilities per capita ratios can mask geographic inequities; when measured against a target, a province may have sufficient facilities per population, but their geographic distribution could leave a substantial percentage of the province’s population unable to reach timely care. These ratios can be calculated at lower administrative levels, but targets for adequate coverage of emergency services are difficult to interpret for small populations,5 rendering them inadequate measures for local health planning. Furthermore, facility-to-population ratios do not account for the time needed to reach services, despite substantial evidence that links improved health outcomes with time to health services.6–8

Understanding access beyond facility-to-population ratios requires estimates of distance or travel time to services. The use of geographic information system (GIS) allows one to determine distance and time, either through simple estimation of straight-line distances between two points (ie, Euclidean distance)9 10 or through distance along walking paths and road networks.11 12 Yet, distance measures alone do not consider the impact of additional geographic barriers on travel time to services and can generate misinformed estimates of access.12 Travel time measures can account for variations in rates of travel across varying slopes, ground cover and road surfaces as well as availability of emergency transportation. The application of GIS in maternal and newborn health is growing, as is the sophistication of access modelling.13 14 Some models have validated the predictive accuracy of spatial accessibility12 15; accounted for the availability of public transportation,16 referral between facilities, and the impact of modifications to the distribution of services and infrastructure11; and applied sophisticated models that define catchment areas.17 18

Some of the aforementioned approaches are limited by the assumption that a woman selects the closest health facility offering the level of service she needs.9–11 16 Evidence suggests that a woman’s choice of facility and whether she uses or bypasses the nearest facility is influenced by individual and facility-level characteristics related to perceptions about quality of services.19–23 Evidence indicates that the odds of a facility being bypassed increase if women perceive it to provide poor quality care and decrease if quality is perceived to be high. Availability of drugs and equipment, cleanliness, privacy of the delivery room and health provider attitudes influence women’s perceptions of quality,20–22 as does the number of emergency obstetric care (EmOC) signal functions provided at the facility.20 22 23 A study in Tanzania found that for every additional signal function provided the odds of that facility being bypassed were almost halved.23 When they do bypass, women appear satisfied with their choice. Compared with non-bypassing mothers, bypassing mothers have reported receiving better care, better availability of drugs and equipment, and higher levels of satisfaction with delivery services, despite having paid more and travelled further for services.22

The objective of this paper was twofold. First, we aimed to adapt the existing methodology modelling geographic access to better reflect the complexity of women’s behaviour using facility characteristics known to influence women’s care-seeking choices. Second, we aimed to demonstrate the utility of the adapted methodology by applying it to measure the impact of common improvements on efforts to achieve universal timely access to emergency care, using data from Mozambique.

Methods

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.

Modelling

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).

Journey segment 1: home to first preferred facility

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.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

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.

Journey segment 2: interfacility referral

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.

Total travel time

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:

• Model 0: This is the current situation of access, status quo.

• Model 1: We identified 37 level 4 facilities that were each suitable for upgrade to level 5, based on a high facility score and degree of geographically strategic location.

• Model 2: We located vehicles at the 659 facilities without transportation and phones at the 140 facilities without communication.

• Model 3: We combined the improvements of models 1 and 2.

Results

Geographic distribution of improvements in access

One important utility of modelling access within a GIS is the ability to specify geographic areas that would benefit from modelled improvements, as well as areas that would remain underserved.

Figure 2 shows the precise geographic areas (bright green) where model 3 improvements changed access to a level 5 from >2 hours to <2 hours (A), and from >5 hours to <5 hours (B). All areas of improvement appeared around one of the 37 upgraded level 5 facilities (shown in pink), indicating the extent to which that approach fills gaps in timely access. Looking at the maps in conjunction with online supplementary File 4 and online esupplementary figure 2, one could quantify the population that lives in the bright green areas. Model 3 increased the population with 2-hour access to a level 5 by 4 percentage points (42% reaching a level 5, whether directly or via referral in model 0, compared with 46% in model 3). The maps also indicate areas (dark grey) where access, though improved, would remain inadequate (>2 hours, or >5 hours, to a level 5). Finally, geographic areas where baseline access is poor are indicated (grey stripes). This area represents 21% of the population, described as outside the models, that would remain en route after 5 hours in each of our models.

Supplementary file 6

[bmjgh-2018-000772supp006.jpg]

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Geographic distribution of improved 2-hour and 5-hour access to level 5: model 0 versus model 3.

Discussion

This paper’s first aim was to demonstrate a refined methodology to estimate women’s access to emergency obstetric services incorporating evidence that choice of facility, though influenced by time to care, is also influenced by perceived quality of services. This approach is a fundamental improvement on others that assume women seek care at their closest facility. The approach also leverages the evidence that the number of EmOC signal functions provided, along with other basic facility characteristics, are strong predictors of facility choice.20 22 23 Our second aim was to apply this methodology to model scenarios of common strategies for improving emergency access in Mozambique to estimate the impact on access over the status quo. The models’ results describe how close Mozambique’s health system is to ensuring universal, timely access to emergency services.

Doubling the number of facilities providing the highest level of service and making substantial investments in transportation and communication would increase the population with 5-hour access to the highest level by almost 1.5 million people (57 000 expected pregnancies and 8550 expected severe complications), and increase access within 2 hours by almost 950 000 people (36 100 expected pregnancies and 5415 expected severe complications). Furthermore, the mean time it would take a woman to move between her first preferred facility and the highest level of care would be reduced by just over 1 hour. Despite these substantial gains, almost one-third of the population, and a relative proportion of the expected pregnancies and complications, would remain beyond 5-hour access to the highest level of care, and the mean total journey time would remain >2 hours across all models. Neither the status quo nor the envisioned improvements are sufficient to ensure universal timely access.

Some model limitations make our estimates of access conservative. By design, our models capture the population with access to a lower-level facility within 1 hour, or to a higher-level facility within 3 or 5 hours. Twenty-one per cent of the population not in this category was excluded from the models. Women in this population have very poor access, yet some would reach a lower-level facility at some point within 5 hours, though not a level 5; thus, the population en route after 2 hours has been overestimated. Roads increase access. For areas where roads exist but are not found in OpenStreetMap data, our models may underestimate access although we feel confident that the augmented road network data used is quite complete. Also, 80 facilities that could not be geographically located were excluded. In all models, a very small proportion of women access a level 1, 2 or 3 facility as first preferred facility. Therefore, the exclusion of the 65 level 3 or lower facilities is unlikely to have substantially affected our results. However, had the 15 level 4 facilities been included, access in all models would have been improved, particularly in Gaza and Zambézia provinces.

On the other hand, these models in many ways represent optimal conditions, and real access would likely be poorer. For example, the models estimate travel during the dry season and daytime, when roads are most traversable. Other model assumptions could be used to estimate travel times under less favourable conditions; however, seasonal variation fluctuates even within the wet and dry seasons, and complex models are required to understand this well.39 The travel times represent only the physical movement across space; they do not account for delays in the decision to seek care, waiting for transportation once reaching a road, time to be seen by a doctor or delay in a provider’s decision to refer. A woman’s true time to care, therefore, would be longer. Little evidence exists on the magnitude of these delays; thus, we did not include them.

Some important factors and methodological approaches were not used. Whether a facility charges for services, and how much, could impact a woman’s choice of facility. We did not include this characteristic in our model because >98% of facilities in Mozambique did not charge for routine or emergency obstetric services.28 Furthermore, the private sector in Mozambique serves a very small proportion of pregnant women, predominantly the wealthy and urban.40 In countries with a more robust private sector, and where poor women seek care despite the costs, the algorithm for classifying facilities should consider the impact of cost on choice. Our methodology could be improved by conducting sensitivity analyses on the parameters used in the models, as well as ground-truthing results. Funds and resources for ground-truthing were not available for our activity, though we are confident that the results remain useful for decision-makers. We used a geographic layer representing total population rather than pregnancies or women of reproductive age. Raster data of pregnancies and women of reproductive age are available, and they are derived from population data by making assumptions about sex and age distribution along with estimates of fertility. Some evidence indicates that the geographic pregnancy data in Mozambique require further adjustment to adequately represent reality; given this uncertainty, we used the population data set.41 Targeted approaches needing to understand the magnitude of women at risk of obstetric complication would benefit from using accurate geographic pregnancy data, and modelling that links geographic location of pregnancies with health facilities is critical for planning and for identifying hard-to-reach communities.42 Despite these limitations and caveats, the described approach provides a more realistic description of the health system and the way women may use it, thus providing useful information for efforts to ensure universal access.

Some efforts towards universal access have placed delivery services at the first level of the health system, sometimes via training of community health workers.23 43 Now, some have called for gradually shifting the provision of obstetric care away from these lowest levels, towards health centres and hospitals,22 in part to align with women’s apparent preferences to deliver in higher-level facilities.23 44 If quality delivery services are concentrated in upper-level facilities, whether by design or default, performance of the emergency referral system becomes increasingly critical. Attempts to shorten interfacility transfer time are paramount to reducing mortality, particularly among women in remote, hard-to-reach areas. Modelling, as demonstrated, could inform a rational reorganisation of the health and emergency referral systems, as well as inform attempts at equitable distribution of services. This analysis indicates that a combination of upgrading facilities and widely distributing transportation can substantially reduce the mean transfer time between facilities. It also indicates that those interventions alone are insufficient to ensure universal timely access. If universal access is a goal, geographic areas of persistently poor access must be targeted with other interventions, such as maternity waiting homes, telemedicine and community midwifery, along with task-sharing treatment of the most time-critical and lethal complications (eg, postpartum haemorrhage and newborn resuscitation).45 Additionally, intersectoral collaboration with road authorities would be critical, as strategic improvements in road surfaces would likely yield the most substantial improvements in access for women currently in hard-to-reach areas.

Conclusion

The models represent more accurate estimates of current geographic access and the impact that specific structural improvements at the health facility level may have on improving that access. Supporting decision-makers with dynamic models like these can provide more realistic expectations of the impact that planned interventions may have and make the case for further innovation and complementary programming to ensure equitable access.