If we build it, will they come? Results of a quasi-experimental study assessing the impact of maternity waiting homes on facility-based childbirth and maternity care in Zambia

Introduction Maternity waiting homes (MWHs) aim to increase access to maternity and emergency obstetric care by allowing women to stay near a health centre before delivery. An improved MWH model was developed with community input and included infrastructure, policies and linkages to health centres. We hypothesised this MWH model would increase health facility delivery among remote-living women in Zambia. Methods We conducted a quasi-experimental study at 40 rural health centres (RHC) that offer basic emergency obstetric care and had no recent stockouts of oxytocin or magnesium sulfate, located within 2 hours of a referral hospital. Intervention clusters (n=20) received an improved MWH model. Control clusters (n=20) implemented standard of care. Clusters were assigned to study arm using a matched-pair randomisation procedure (n=20) or non-randomly with matching criteria (n=20). We interviewed repeated cross-sectional random samples of women in villages 10+ kilometres from their RHC. The primary outcome was facility delivery; secondary outcomes included postnatal care utilisation, counselling, services received and expenditures. Intention-to-treat analysis was conducted. Generalised estimating equations were used to estimate ORs. Results We interviewed 2381 women at baseline (March 2016) and 2330 at endline (October 2018). The improved MWH model was associated with increased odds of facility delivery (OR 1.60 (95% CI: 1.13 to 2.27); p<0.001) and MWH utilisation (OR 2.44 (1.62 to 3.67); p<0.001). The intervention was also associated with increased odds of postnatal attendance (OR 1.55 (1.10 to 2.19); p<0.001); counselling for family planning (OR 1.48 (1.15 to 1.91); p=0.002), breast feeding (OR 1.51 (1.20 to 1.90); p<0.001), and kangaroo care (OR 1.44 (1.15, 1.79); p=0.001); and caesarean section (OR 1.71 (1.16 to 2.54); p=0.007). No differences were observed in household expenditures for delivery. Conclusion MWHs near well-equipped RHCs increased access to facility delivery, encouraged use of facilities with emergency care capacity, and improved exposure to counselling. MWHs can be useful in the effort to increase delivery at advanced facilities in areas where substantial numbers of women live remotely. Trial registration number NCT02620436. Between March 2016 (baseline) and October 2018 (endline), we implemented a quasi-experimental intervention study in the catchment areas of 40 primarily rural health centres (RHCs) in Choma, Kalomo and Pemba districts in Southern Province, Nyimba and Lundazi districts in Eastern Province, and Mansa and Chembe districts in Luapula Province, Zambia.22 A cluster design was selected because of the inherent nature of the intervention. Each cluster was comprised an RHC and its government-defined catchment area households. The targeted districts were part of the SMGL intervention.9 We included in the sampling frame all RHCs in the selected districts that met the following eligibility criteria: (1) at least 150 deliveries annually; (2) situated ≤2 hours driving time to the nearest referral hospital capable of providing comprehensive emergency obstetric and neonatal care (CEmONC); and either: (3) able to perform at least five of seven basic emergency obstetric and neonatal care signal functions; or (4) had at least one skilled birth attendant, practiced routine management of third stage labour, and had no stock outs of oxytocin or magnesium sulfate in the previous 12 months. Only referral hospitals offered CEmONC services including parenteral antibiotics, blood transfusions and caesarean sections; the study sites did not. Prior to the opening of study MWHs, all study districts had received the SMGL initiative (2012–2016) which sought to rapidly reduce maternal mortality through a comprehensive set of interventions to address challenges, and improve maternal health services demand, access and quality.23 Among others, interventions included: health facility infrastructure, equipment and medicines stock improvements; training and mentorship of healthcare providers to increase access to EmONC services in the districts; and communication campaigns using community leaders, communication materials, and mass media messaging.23 The study districts are primarily rural (67%–95% of the district populations) with pockets of peri-urban centres.24 The populations of these districts are generally poor; those living in remote areas have limited access to improved sources of water or sanitation, or to electricity.25 At the time of the study described here, the districts had generally similar availability of maternal health services. Facility delivery rates were 56%, 68% and 71% in Southern, Luapula and Eastern Provinces, respectively, according to the Demographic and Health survey conducted prior to the outset of this study.26 The unit of assignment to study arm was the RHC and its catchment area; the unit of analysis was the individual. Of 44 RHCs that met the inclusion criteria, the 40 farthest RHCs were included in the study. Half of the study clusters were assigned prospectively to study arm using a matched-pair randomisation procedure (randomised subgroup). Study clusters in Choma, Kalomo, Pemba and Nyimba districts were randomised prior to baseline enrolment with equal probability to either the intervention or the control group. Prior to randomisation, catchment areas were matched in pairs first based on government-reported transfer time to nearest CEmONC facility, then best fit to average monthly volume of deliveries over the previous year. Within each matched pair, one catchment area was randomly assigned by the study team to the intervention group using the RAND function in Microsoft Excel. Due to the nature of the intervention, it was not possible to blind participants. For the other half of study clusters, political considerations precluded random assignment (non-randomised subgroup). Government officials feared community fatigue from so many projects operating in their areas and therefore preferred to identify intervention sites purposefully. Study clusters in Lundazi, Mansa and Chembe districts were non-randomly assigned prior to baseline enrolment. The Ministry of Health assisted in identifying 10 intervention sites. Comparison sites were then identified from eligible RHCs matched on delivery volume and government-reported transport time the nearest CEmONC facility. Sites with a known formal existing MWH structure were not considered in the sample frame; however, baseline data suggested women in the catchment were often accommodated prior to delivery within available bed space or in informal waiting spaces near the health facility, similar to the randomised subgroup.27 All intervention clusters received the Core MWH Model (described below) and control clusters implemented a various ‘standard of care’ for women awaiting delivery in Zambia, which included use of a community-constructed structure, women staying informally within RHC wards, and no dedicated space to wait.17 Aside from intervention assignment, all study procedures were implemented consistently across the sites. The Core MWH Model was developed based on formative research and community input and refined with input from government stakeholders. It aimed to address common barriers to MWH utilisation and to be culturally acceptable.28 29 The Core MWH Model included three primary domains. First, all sites had similar infrastructure, equipment and supplies which included concrete floors, latrines, bathing areas, intact roof, storage space, covered cooking space, location near a water supply, lockable doors, cupboards and windows, lighting, beds, bedding, mattresses, mosquito nets and cooking utensils. Each site had a main dormitory for pregnant women and a smaller dormitory for women and newborns, who had just been discharged from the health facility or had returned for a postnatal care visit. Formative research identified the ‘mixing’ of pregnant with newborns as culturally inappropriate; community members requested a separate space within the structure for postpartum women and newborns to stay. Each site also had a private bathing and drying area, per results from the formative work.28 Second, each site had a formalised management structure responsible for daily operations and a governance structure with representation from the community, government, traditional leadership and the health centre. Third, all sites were situated within 100 m of, and had formal linkages to, the RHC. Each had health centre staff check in daily on waiting women, and health staff or volunteers offered maternal and child health education courses. Clinical care was conducted at the RHC, not at the MWH. Women learnt about the MWH at ANC visits, through health outreach activities, and through traditional leaders. Implementation began July 2015; the first intervention Core MWH Model opened in study sites immediately following baseline observation. Intervention MWHs operated for a minimum of 13 months before the endline study. A cross-sectional sample of households was identified at baseline (March–May, 2016) and endline (September–October, 2018) using a multi-stage random sampling procedure. In the first stage, 10 villages located at least 10 km from the RHC from each catchment area were randomly selected with probability proportional to size. To identify eligible villages, every village centre was visited and GPS coordinates were taken to determine travel distance to the RHC along the most direct route, calculated using ArcGIS Online (ESRI, Redlands, California, USA). In the second stage, eligibility was restricted to households with at least one woman 15 years or older who delivered a child in the previous 12 months, irrespective of her place of delivery or current vital status. An exhaustive list of eligible households was created with input from RHC staff, community health volunteers and local traditional leaders. Households were then ordered randomly and visited in that order to confirm eligibility and enrol until the target of approximately six households per village was reached. In the third stage, if a household had more than one eligible woman, one woman was selected at random during enrolment. While we did not ask about previous participation in the survey, it is possible that some women who had another delivery between the data collection rounds were selected for endline. There is no reason to believe this would disproportionately affect one study arm more than the other. During each round of data collection, a team of trained enumerators who spoke English and the local language spent 6 weeks conducting surveys. Data enumerators were introduced to the household head or senior woman by community volunteers. On confirming household eligibility, the study team consented the household head, geolocated the household and captured demographic information. On completion of this portion of the survey, the eligible woman was selected, consented separately and responded to the remainder of the survey in a private space. Survey topics included household composition and individual characteristics; experience of the last pregnancy including ANC, labour and delivery, birth outcomes and postnatal care; and MWH use. All responses were self-reported; when available, responses were verified against the mother’s antenatal card or the baby’s under-5 card. Data were captured electronically on encrypted tablets using SurveyCTO Collect software (V.2.212; Dobility). Audits on a random selection of 5% of households the following day found few inconsistencies. The primary outcome was facility delivery, defined as delivery at either an RHC or a hospital. Women were asked where they delivered their most recent child, including the facility name if applicable. For analysis, responses were dichotomised based on whether the delivery occurred at a health facility or other location. Secondary outcomes included: (1) use of an MWH while awaiting delivery; (2) maternity care utilisation measures including referral or transfer to a higher-level facility before, during or after labour and attendance at early postnatal care, asked as ‘approximately 3 days after delivery’; (3) hospital-level services received during labour and delivery (parenteral antibiotics, blood transfusion and caesarean section surgery); (4) exposure to counselling services at the time of delivery (family planning, breastfeeding and kangaroo care (ie, early and continuous skin-to-skin contact)); (5) maternal and neonatal vital status after delivery and (6) health behaviours at the time of the interview including use of modern family planning and infant feeding methods. We also include self-reported delivery expenditures in Zambian Kwacha between study arms. Women reported if and how much they had expended on delivery supplies, baby clothes, transportation for delivery, and accommodation while awaiting delivery.30 The target sample size for each round of data collection was 2400. We assumed a baseline estimate of 63% facility delivery in rural Zambia, about the average across the provinces where we were working at the time26 and an estimated 60 households per cluster. The sample size provided 80% power to detect a 10-percentage point increase in facility delivery due to the intervention and assumed an α of 0.05 and an intracluster correlation coefficient of 0.04.31 We did not have the data necessary at the time of conducting the power calculation to make confident estimates of the correlation between baseline and follow-up outcomes for individuals within a cluster. We therefore conservatively assumed this to be zero. When planning for recruitment, we expected approximately 10% refusal. Loss to follow-up was not considered given the repeated cross-section design. The Stata code used for the power calculation was: ‘clustersampsi, binomial beta(0.80) p1(0.63) p2(0.73) k(20) rho(0.04)’. We constructed a wealth index using household asset information from the broad categories of power source, water source, cooking source, household essentials and luxuries, farming supplies, banking, electronics, and access to internet. The sample was split into quintiles for analysis. We compared potential confounders including characteristics of recently delivered women (age, education, marital status, gravida, parity, antenatal visit, months since delivery and delivery location) and characteristics of the households (wealth quintile, household size, dependency ratio, and distance from village centre to nearest health centre) of the intervention and control groups at both rounds of data collection, including comparisons within randomised and non-randomised subgroups (online supplemental table A1) and characteristics of the sites and MWHs (online supplemental table A2). bmjgh-2021-006385supp001.pdf bmjgh-2021-006385supp002.pdf We determined the impact of the intervention on the primary outcome, facility delivery, for the full sample as well as the randomised and non-randomised subgroups using intention-to-treat analysis. Next, we compared use of MWHs across study groups to better understand intervention uptake. We then estimated the impact of the intervention on the secondary outcomes. We fit a set of generalised estimating equations (GEE) specified as having a binomial distribution for the dependent variable, a logit link function and an exchangeable correlation structure to estimate ORs for all outcomes except for health expenditures. We estimated two separate models to understanding the impact of the intervention on expenditures associated with delivery.32 We first estimated a GEE model with the dependent variable indicating whether there was any expenditure on delivery. We then estimated a second GEE model with the dependent variable ln(total expenditure) using only observations with expenditure >0. For this model, we specified a Gaussian distribution for the dependent variable, an identity link function, and an exchangeable correlation structure. For all GEE models, matched-pair was specified as the group variable and robust SEs were estimated using a degrees-of-freedom corrected sandwich estimator. Except for referral from MWH (which had no baseline value prior to intervention), each model included the cluster-level average of the outcomes measured at baseline. Each model also controlled for the variables used in the matching procedure, average monthly volume of deliveries at nearest RHC and transfer time to nearest CEmONC hospital. No additional covariates were included in the main models. Because half of study clusters were non-randomly assigned, we present adjusted estimates of impact on the primary outcome in online supplemental table A3 using models that included the following covariates: woman’s age (years), education (years), marital status, and primigravida, along with household wealth quintile and distance of the village centre to the nearest government assigned RHC (km). These covariates were selected based on a review of the literature and previous work on where women deliver in Zambia. bmjgh-2021-006385supp003.pdf As a robustness check, we also present estimates of impact on the primary outcomes using a set of mixed-effects models that include random effects for matched-pair, health facility catchment area, and village in online supplemental table A4.33 Finally, we present estimates of impact on the primary outcomes from a set of generalised linear probability models (ie, GEE specified as having a Gaussian distribution for the dependent variable and an identity link function) in online supplemental table A5. All analyses were conducted using Stata statistical software (StataCorp. 2015. Release V.14). All data for this analysis are publicly available.(dataset)34 bmjgh-2021-006385supp004.pdf bmjgh-2021-006385supp005.pdf End-users of the MWHs and other key community-level stakeholders including men, community elders and traditional leadership were involved in conceptualising and designing the intervention during a formative research phase.28 29 The intervention design was refined with input from the Ministry of Health. We continued to engage a variety of key stakeholders, including members of the target population, through a rigorous process evaluation that routinely assessed intervention acceptability, and implementation feasibility and fidelity.19 35 36 We were guided by the Consolidated Standards of Reporting Trials checklist extension for cluster randomised trials in preparing this article.