Predicted effect of regionalised delivery care on neonatal mortality, utilisation, financial risk, and patient utility in Malawi: an agent-based modelling analysis

Background: Health-care regionalisation, in which selected services are concentrated in higher-level facilities, has successfully improved the quality of complex medical care. However, the effectiveness of this strategy in routine maternal care is unknown. Malawi has established a national goal of halving its neonatal mortality by 2030. In this study, we aimed to assess the effect of obstetric service regionalisation in pregnant women and their newborn babies in Malawi. Methods: In this analysis, we assessed regionalisation through the use of an agent-based simulation model. We used a previously estimated utilisation function, incorporating both patient-specific and health-facility-specific characteristics, to inform patient choice. The model was validated against known utilisation patterns in Malawi. Four regionalisation scenarios were compared with the status quo: scenario 1 restricted deliveries to facilities currently capable of providing caesarean sections; scenario 2 had the same restrictions as scenario 1, but with selected facilities upgraded to provide caesarean sections; scenario 3 restricted delivery to facilities that provided five or more basic emergency obstetric and neonatal care services in the preceding 3 months; and scenario 4 had the same restrictions as scenario 3, but with selected facilities upgraded to provide at least five basic emergency obstetric and neonatal care services. We assessed neonatal mortality, utilisation, travel distance, median out-of-pocket expenditure, and proportion of women facing catastrophic expenditure. The effects of upgrading the obstetric readiness of all facilities, of removing all user fees, and of upgrading without restriction were considered in scenario analyses. Heterogeneity and parameter uncertainty were incorporated to create 95% posterior credible intervals (PCIs). Findings: Scenarios restricting women to give birth in facilities with caesarean section capabilities reduced neonatal mortality by 11·4 deaths per 1000 livebirths (scenario 1; 95% PCI 9·8–13·1) and 11·6 deaths per 1000 livebirths (scenario 2; 10·2–13·1), whereas scenarios restricting women to facilities that provided five or more basic emergency obstetric and neonatal care services did not affect neonatal mortality. Similarly, the caesarean section rate in Malawi, which is 4·6% under the status quo, was predicted to rise significantly in scenario 1 (14·7%, 95% PCI 14·5–14·9; p<0·0001) and scenario 2 (10·4%, 10·2–10·6; p<0·0001), but not in scenarios 3 and 4. Women were required to travel longer distances in scenario 1 (increase of 7·2 km, 95% PCI 4·5–9·9) and in scenario 2 (4·4 km, 1·5–7·2) than in the status quo (p<0·0001). Out-of-pocket costs tripled (p<0·0001; status quo vs scenario 1 and scenario 2), and the risk of catastrophic expenditure significantly increased from a baseline of 6·4% (95% PCI 6·1–6·6) to 14·7% (14·5–14·9) in scenario 1 and 11·3% (11·0–11·5) in scenario 2. This increase was especially pronounced among the poor (p<0·0001; status quo vs scenario 1 and scenario 2). Interpretation: Policies restricting women to give birth in facilities with caesarean section capabilities is likely to result in significant decreases in neonatal mortality and might allow Malawi to meet its goal of halving its neonatal mortality by 2030. However, this improvement comes at the cost of increased distances to care and worsening financial risks among women. Funding: Bill & Melinda Gates Foundation, Damon Runyon Cancer Research Foundation. Malawi is a country with 18·6 million people, located in southeastern Africa. Its total fertility rate of 5·5 translates to 639 000 births per year. Malawi has a life expectancy at birth of 54·8 years, an adult literacy prevalence of 61·3%, and a neonatal mortality rate of 23 deaths per 1000 livebirths. 61·6% of the Malawian population lives below the international extreme poverty line of US$1·25 per day.13 We constructed an agent-based model on the basis of a combined dataset of births in 2013–14 and delivery facilities; this dataset drew from the 2013–14 Malawi Millennium Development Goal Endline Survey and the 2013 Service Provision Assessment and is described in full elsewhere.14 Facilities where women gave birth in the dataset were modelled, including details such as global positioning system (GPS) location, facility type, facility management (private, non-governmental organisation [NGO], or public), whether they charged fees for delivery, and their basic obstetric readiness score. Readiness scores reflect the availability of equipment and resources required for obstetric care, as defined by WHO (appendix).15 Under the status quo, women in Malawi give birth at all levels of the health system: in central hospitals, district hospitals, rural and community hospitals, and other hospitals (primarily consisting of private and NGO hospitals), as well as in health centres, clinics, and maternities. We assessed four regionalisation strategies in comparison with the status quo. In scenario 1, deliveries were restricted to facilities capable of providing caesarean sections. Scenario 2 is scenario 1 plus the upgrade of selected facilities without caesarean section capability to provide caesarean sections; delivery was restricted to facilities from scenario 1 plus the newly-upgraded facilities, with upgraded facilities chosen to maximise population coverage (appendix). In scenario 3, deliveries were restricted to facilities that reported doing five or more basic emergency obstetric and neonatal care procedures in the preceding 3 months. Scenario 4 is scenario 3 plus selected facilities that did not report basic emergency obstetric and neonatal care upgraded to provide such care; delivery was restricted to facilities from scenario 3 plus the newly-upgraded facilities, with upgraded facilities chosen to maximise population coverage (appendix). Official Malawian policy was designed to promote delivery in facilities. Traditional birth attendants are barred from practice16 and, in reality, fewer than 10% of deliveries occur at home.1 As a result, home delivery was not considered in this model. Agent-based models are stochastic models in which each individual actor—patients and health facilities, in this case—responds to its own internal rules. Because of their stochastic nature, these models allow the incorporation of both individual-level heterogeneity and parameter uncertainty (eg, surrounding decision-related factors, such as how important distance or quality is to an individual woman). Agent-based models also allow the easy incorporation of data related to geographical information systems (GIS) such as GPS coordinates, population density, and the road network of a country. Although agent-based models were first developed for transportation, shipping, and other operations-research applications, they have been applied to modelling policies in low-income and middle-income countries in the past decade.17 We modelled a synthetic closed cohort of 20 000 women with GPS locations, wealth, parity, age, literacy, education, marital status, urbanicity, number of antenatal visits, multiple gestations, and predicted risky deliveries calibrated to represent a random sample of pregnant Malawian women. Wealth, GPS location, and urbanicity were modelled as a joint distribution, whereas the remaining variables were assumed to be independent of each other and of the three joint variables. Population density and poverty distribution were derived from the WorldPop project.18 Parity, age, literacy, education, marital status, antenatal visits, risky delivery, and multiple gestation probabilities were derived from the 2013 Millennium Development Goal Endline Survey in Malawi (National Statistical Office, Zomba, Malawi).19 Urbanicity was defined by residence in Blantyre or Lilongwe. Travel routes and distances were calculated with use of the A* search algorithm.20 Each woman's choice was modelled as a two-step process, on the basis of the choice function published by Yorlets and colleagues.12 The choice function was based on discrete-choice methodology applied to the women in the Millennium Development Goal Endline Survey and the 540 possible delivery facilities where they could give birth. Wealth, literacy, parity, antenatal care, predicted risk of delivery, and predicted need for caesarean section served as individual characteristics, whereas distance, indicators of facility quality, fees, and facility type served as choice-specific characteristics. Conditional logistic regression determined the relative importance of these characteristics and the latent classes that underlay individual preferences. Each woman in our study was first probabilistically assigned to a latent preference class, on the basis of her wealth, literacy, parity, receipt of antenatal care, predicted risk of delivery, and predicted caesarean section. Conditional on membership in each latent class, the woman was matched to her most likely delivery facility. Average incomes and out-of-pocket costs for delivery were taken from previously published literature.21, 22, 23, 24 When multiple estimates were encountered, the most conservative one was used. An expense was defined as catastrophic when it was greater than 10% of a woman's yearly expenditure.25, 26 We obtained geographical information of all health facilities from the 2013–14 Service Provision Assessment. Facilities offering routine delivery services (n=540) were selected for this analysis. All delivery facilities within 100 km of a woman were included in her choice set. Factors influencing each woman's facility selection were the following: facility type, basic obstetric readiness score, road distance, and whether the facility charged fees for delivery.12 We calculated the probability of delivery at each facility and the woman was then assigned, probabilistically, to deliver at a single facility. The woman then travelled through Malawi's road network to her assigned facility, and neonatal mortality was then probabilistically determined (appendix). Distance to the facility and travel time to reach it, its obstetric readiness score, its type, and whether fees were charged were recorded. The status quo for maternal delivery in Malawi was modelled first to serve as validation. Results for each regionalisation strategy were then modelled and compared with the status quo. The primary outcomes of our study were neonatal mortality, the location of delivery, and distance travelled for delivery. The secondary outcome was the assessment of individual utility. Outputs of the status quo scenario were compared with the known joint distribution of wealth and location in Malawi, the distribution of maternal delivery across facility types, and the estimated preference class breakdown from Yorlets and colleagues.12 Because utility is not quantifiable and because valid cost data were not available, a so-called willingness to travel calculation was undertaken to make utility more concrete. Utility differences between the status quo and each regionalisation strategy were converted to the number of km that a woman would have to travel to overcome this utility difference. Client costs were estimated on the basis of a white paper published by Partnerships or Health Reform.21 A woman was considered to have faced catastrophic expenses if her out-of-pocket costs equalled more than 10% of her yearly expenditures.26 Because active restriction of maternal delivery to any one type of facility might be difficult to implement, we did a scenario analysis in which the selected facilities in scenarios 2 and 4 were still upgraded, but women were free to choose among all available facilities for delivery. Additionally, because obstetric readiness and user fees have previously been identified as drivers of decision making among women in Malawi,12 we assessed the effect on utility decrement of policies that removed direct user fees, equipped all facilities with full obstetric readiness, or did both. The goal was to assess whether these compensatory policies could overcome the predicted utility decrement that would result from requiring women to travel further for their delivery care. To incorporate parameter uncertainty and heterogeneity, the model was run 200 times for each cohort and scenario combination, leading to 1000 runs of the model (200 for the status quo, and 200 each for the regionalisation scenarios), for a total of 20 000 000 individual patients. Modelling was done in AnyLogic (version 8.1), with analysis done in R (version 3.4.0). Results are presented as mean (95% posterior credible interval [PCI]). The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.