Discussion
By collecting data from clinical records of pregnant women using CEmOC facilities, we have, for the first time, been able to compare the travel time estimates obtained using popular tools with the reality that the women would have experienced. For our study sample, median travel time was estimated as 5, 11 and 40 min using the cost-friction surface approach, OSRM and Google Maps, respectively. In reality, however, median drive time was 60 min as replicated by two independent private car drivers. Agreement between travel time estimates and real-life driving time was particularly poor for the cost-friction surface approach and OSRM, and for weekday night-time travel and journeys longer than 120 min in all methods.
Modelled approaches such as the cost-friction surface approach and OSRM have been widely used to estimate travel time and to assess the extent of accessibility of healthcare in LMICs in general, not just for obstetric emergencies.6–8 While the limitations of using modelled travel time to indicate real-life travel have been acknowledged,6 30 31 our findings highlight that the errors produced by using cost-friction and OSRM using currently available associated speed data are too large for these tools to be deemed representative of actual travel time in cities. A publication in 2020 by Rudolfson and colleagues showed that even in rural settlements, self-reported times are longer than modelled estimates by a factor of 1.50 for women seeking CEmOC services.32 We show that in a large sub-Saharan African megacity like Lagos, model-based methods are closest to actual travel time only over short distances (especially journeys <10 min), and underestimate travel time by an order of magnitude for longer journeys. This underestimation has significant impact on maternal and perinatal survival. Researchers who have used these two modelling approaches have either assumed that women travelled to the nearest facility or that they were referred from BEmOC to CEmOC facilities.33 However, this is not always the case.19 34 35 While there are personal reasons for some women not to travel directly to the nearest facility,19 our study suggests that for a fifth of our sample a referral to apex referral facilities farther away from women’s residence was needed. In addition, in megacities like Lagos, decisions about whether, where and how to travel are influenced by traffic conditions, perceptions of safety, cost of transport, time of travel and poor roads.19 In Bangladesh, an Asian megacity, congested traffic conditions mean that 37% of the city’s slum population cannot access emergency services within 1 hour.10 These crucial influences on travel time are not accounted for in cost-friction surface approach and OSRM. When both cost-friction surface approach and OSRM were compared previously, larger discrepancies were reported for long travel time estimates.36
On the other hand, Google Maps, which benefits from real-time traffic data, had a median estimate as a percentage of actual travel time of 85%. As per a 2019 systematic review, there has been no use of such web-based platform in estimating travel time in LMICs,6 as has been done in high-income countries.37 This may be because Google Maps queries after a certain number of requests need to be paid for. Such costs could be prohibitive for LMIC researchers.6 There is also the question of scale of analysis. In using Google Maps for our study, we had to individually trace journeys. In a study conducted in rural Mzuzu community, Malawi, the authors using Google Earth combined with global positioning satellite individually found locations of 79 traditional birth attendants and traced from their facilities to formal health facilities in the area.38 Our study was conducted across multiple facilities and on a wider scale. This approach of travel time extraction can be time consuming and is not an efficient process. However, as per our study findings, it was clear that while not perfect, a web-based platform like Google Maps offered estimates that were closer to reality. Strengths of Google Maps have been highlighted in the literature, including its use of relatively up-to-date road data set, capacity to account for traffic and consideration given to peak and off-peak hours.39 These are all important elements altering travel experiences in megacities.16 In our study, there were few instances in which Google Maps time estimates were higher than the median replicated travel time. We attributed this to the application of native intelligence of driver 2 in using short cuts to reach their destination.
Whichever method was used, we found that the smallest differences between replicated travel and estimated travel time were for night-time journeys at weekends, while the largest difference was in night-time journeys on weekdays. This might be because traffic is worst in the Lagos metropolis from evenings onwards during the weekdays when commuters are returning from work.16 We also found that journeys which took less time had the closest estimates using all three methods, and journeys that took longer time had most errors and least agreement. This was expected, as shorter journeys will have fewer encumbrances that can prolong actual travel time.
Comparing our travel time estimates to the global benchmark of 80% of women reaching CEmOC facilities within 2-hour travel time,5 all three methods estimated that this target was exceeded. However, both cost-friction surface and OSRM estimated that all women got to facilities within 2 hours, though in reality 8% did not. This error obscuring the inequality in access to critical services has large implications for advocacy for service provision and service planning for life-saving maternal and perinatal care.
Our study has some limitations. First, we replicated journeys but cannot confirm that these were the actual times it took the women, since new road constructions or further damage to the roads may have reduced or increased travel time. Second, our data did not capture whether and how long women stopped on the journeys to the destination hospital. Third, though we had data on the months of presentation and could have aggregated these to assess seasonal patterns that may influence travel, we have not replicated the journeys in the months that the women presented. This would have required yearlong replication of travel yet no one of the methods of travel time estimates has capacity to show seasonal variation in travel time. In any case, by driving during the rainy season, we were able to replicate their journeys during the worst-case scenario. However, replicating journeys in the rainy season also gives a worst-case difference between the modelled times and the replicated times, and it may be the case that this difference could have been smaller if journeys were also replicated during the dry season when road infrastructure is better. However, reports from regular road users in Lagos and researchers suggest that many roads are in deplorable conditions and traffic is significant in both dry and rainy seasons.15 40 Fourth, not all women would have travelled by private car, so it is likely that our driver replication would be the most direct and probably fastest way of getting to the hospitals, as public transport would require some waiting time. We also have not accounted for the time it might have taken women to get transport ready after they have decided to go to a facility. This means that the difference in women’s actual journeys and the estimates produced by models may in reality be even greater. Attempts to retrospectively contact the women to capture self-reported travel time raised serious ethical concerns, especially for those who had traumatic birth experiences. In any case, issues of recall bias and subjectivity, which minimises validity of estimates, have been reported with this approach.32 However, we know that in emergencies, some pregnant women in Lagos take extreme measures such as driving in breakdown or oncoming lanes illegally or deserting their motor vehicles to hail motorcycles that would move faster through gridlocks to reach facilities.19 Altogether, re-enacting the exact journeys of the women was not our goal. To account for any variation in the actual travel, we had two drivers independently replicating travel. This is a strength in our study design.
Going forward, models need to take into account the variable traffic conditions, as was done by Ahmed et al.10 While this draws models closer to reality, there will still be a gap in linking populations to actual facilities of care, as we have done in our study. This is particularly important if available tools are going to be truly effective for tackling inequities in geographical coverage in LMICs. On the other hand, web-based platforms should be mainstreamed into efforts to support evidence generation for health service delivery at next to no cost to researchers and service planners. While advocacy to big tech mapping companies like Google and ESRI to provide these ‘life-saving’ data will be a sensible next step, open-source platforms such as the World Bank-supported OpenTraffic (http://opentraffic.io/) should be promoted. Supporting the growth, development and distribution of free transport data for academic and policy use will make models more relevant for urban settings.
For benchmarking, there is a need to review the 2-hour threshold set for travel to access services like EmOC.3 5 Recommending indicators that capture geographical access was a necessary first step.41 However, if the data feeding into these indicators are not reflective of reality, then their validity will be questionable. LMIC urban settings, more so, megacities, are getting too congested to use such wide travel benchmarks for service planning. In any case, pregnancy and childbirth complications can quickly escalate ‘in less than two hours’, and for some women even in minutes.42 43 In our study, we identified women who travelled for longer than 2 hours within the city, and this is without any other personal and structural points of delays that women may have faced. Our finding is particularly concerning when consideration is given to women who live in more periurban and slum areas need to reach these apex referral facilities, which can be remote even where they exist within cities. In a Tanzanian study, excluding time to first facility, modelled median travel time estimates of actual facility referrals from periurban to urban area were 156.4 min (IQR: 7.9–356.6 min).44 As per our findings, real travel time will be longer. Even after getting to these apex hospitals, evidence shows that women could still face an additional 60 min (IQR: 21–215 min) delay between diagnosis and receiving life-saving interventions.45 In many LMIC settings including cities, ambulances are not particularly effective and even if an ambulance is used during referral, due to lack of regard by other road users and emergency road lanes, women are still not guaranteed to reach the facility in good time.19 Another key consideration is that a CEmOC facility being available within 2 hours does not guarantee that the facility can provide the service that the woman needs. Due to such service gaps, service-specific geographical accessibility metrics that reflect service availability (24 hours/not) will be more informative for service planners. Many of these considerations require a ‘local gaze’, minimising the relevance of large-scale model-based studies that assume availability and functionality of facilities. For example, recent large-scale models suggest that geographical accessibility targets have been met and travel can be done to health facilities within 1 hour in the same geographical space in which we conducted our study.7 8 30
