Menu
Menu
Menu
Menu
Background: Pregnant women in sub-Saharan Africa (SSA) experience the highest levels of maternal mortality and stillbirths due to predominantly avoidable causes. Antenatal care (ANC) can prevent, detect, alleviate, or manage these causes. While eight ANC contacts are now recommended, coverage of the previous minimum of four visits (ANC4+) remains low and inequitable in SSA. Methods: We modelled ANC4+ coverage and likelihood of attaining district-level target coverage of 70% across three equity stratifiers (household wealth, maternal education, and travel time to the nearest health facility) based on data from malaria indicator surveys in Kenya (2020), Uganda (2018/19) and Tanzania (2017). Geostatistical models were fitted to predict ANC4+ coverage and compute exceedance probability for target coverage. The number of pregnant women without ANC4+ were computed. Prediction was at 3 km spatial resolution and aggregated at national and district -level for sub-national planning. Results: About six in ten women reported ANC4+ visits, meaning that approximately 3 million women in the three countries had <ANC4+ visits. The majority of the 366 districts in the three countries had ANC4+ coverage of 50–70%. In Kenya, 13% of districts had < 70% coverage, compared to 10% and 27% of the districts in Uganda and mainland Tanzania, respectively. Only one district in Kenya and ten districts in mainland Tanzania were likely met the target coverage. Six percent, 38%, and 50% of the districts had at most 5000 women with 20,000 women having <ANC4+ visits were 38%, 1% and 1%, respectively. In many districts, ANC4+ coverage and likelihood of attaining the target coverage was lower among the poor, uneducated and those geographically marginalized from healthcare. Conclusions: These findings will be invaluable to policymakers for annual appropriations of resources as part of efforts to reduce maternal deaths and stillbirths.
Kenya, Uganda, and Tanzania are located in East Africa and share national borders (SI Fig. 1). Each country is subdivided into districts that are used for healthcare planning, 47 in Kenya (counties), 135 in Uganda (districts) and 184 in mainland Tanzania(councils) (SI Fig. 1). Population, health, socioeconomic and demographic indicators for each country are presented in SI Table 1. The healthcare system in the three countries is decentralized, running a hierarchical referral system from primary to tertiary level health facilities with both public and private health facilities [21, 22, 24]. These health facilities are expected to serve ANC clients through a recommended package of interventions [8, 9, 11]. The health sector financing in the three countries is mainly dependent on funds from the government, donors, and out-of-pocket payments [35–37]. Over time, these countries have put in place policies to make maternal health services, including ANC, affordable and accessible through subsidies, incentives, partial or full removal of user fees, vouchers, conditional cash transfers and insurance programs [29, 38–42]. ANC guidelines monitored ANC4+ coverage at the time of the survey in the three countries [21–24]. Percentage of pregnant women with at least 4 ANC visits based on the pregnancy preceding their most recent live birth during the 3 years preceding the survey. Empirical observations (A), predicted surfaces at 3 km spatial resolution (B) aggregated at district level (C) and exceedance probability for a 70% target in Kenya, Uganda, and Tanzania mainland We used data from the most recent nationally representative Malaria Indicator Surveys (MIS) in Kenya 2020 [43], Uganda 2018/19 [44], and Tanzania 2017 [45]. MIS are stand-alone cross-sectional household surveys which collects data on key indicators of malaria and population health, including that of pregnant women. The sampling strategy is detailed in supplementary information 1 (SI) section A2. Our study sample included ANC history of 10,237 women of reproductive age (15–49 years) for their most recent live birth in the 3 years preceding the surveys. The women belong to randomly selected households within sampled enumeration areas (EAs)/clusters. Each cluster is represented by a displaced geographical coordinate to protect respondent confidentiality [46]. Urban and rural clusters are displaced by up to 2 and 5 km, respectively while remaining within boundaries of the district or region considered in the survey. Further, 1% of the rural clusters are displaced by up to 10 km [46]. The outcome variable was the percentage of women who reported receiving ANC4+ visits. Women were asked how many visits they received during pregnancy, and during those visits, to list all types of health providers/professionals they saw. We defined doctors, nurses, midwives, medical assistants, clinical officers, assistant clinical officers, assistant nurses, maternal and child health aides as qualified health professionals for the purpose of ANC provision. Women reporting ANC visits but not listing at least one of these providers were categorized as not receiving ANC. Although the study surveys were conducted during the first phase of implementing the new WHO recommendation of at least eight ANC contacts (ANC8+), none of the three countries had transitioned to the ANC8+ model at the time of data collection or had explicit policy targets for its coverage [21–24]. Further, the observed ANC8+ coverage based on the study surveys was very low (3.5% in Kenya, 1.4% in Uganda, and 1.2% in mainland Tanzania) insufficient for robust geostatistical modelling at high spatial resolution. As such, analyses in this study were based on the previous WHO recommendation of ANC4+ and in line with the EPMM targets [11, 19]. Study variables were based on factors known to influence ANC use [47–49] and data available from the three MIS (Table 1). The outcome and covariates based on Malaria Indicator Survey in Kenya (2020), Tanzania (2017) and Uganda (2018/19). Travel time was modelled while nighttime lights were derived from satellite imagery. Geographical coordinates were available at the cluster level, and all data were resolved at this level Two factors not sourced from the MIS were nighttime lights (NTL) and travel time to the nearest health facility. NTL is a proxy for urbanization, gross domestic product, population density and economic activity [52, 53]. Its inclusion alongside other covariates (Table 1) correlated with the urban/rural clusters in geostatistical models for disease mapping accounts for the sampling design implicitly [54, 55]. Annual NTL, temporally matched to survey year, produced using monthly cloud-free radiance averages, made from low light imaging day/night band data collected by the NASA/NOAA Visible Infrared Imaging Radiometer Suite was used [56]. We extracted NTL per cluster within a buffer to minimize the effect of displaced cluster coordinates in ArcMap version 10.5 (ESRI Inc., Redlands, CA, USA). We modelled travel time to the nearest health facility (spatial access) using approaches that combine several modes of transport in a single journey [57, 58] based on a least-cost path algorithm implemented in AccessMod software alpha version 5.7.8 (WHO, Geneva, Switzerland) [59]. We accounted for the road network, land use, topography, and transport barriers except where a road intersected a barrier [57–59]. We leveraged the SSA master health facility list (MHFL) comprising public health facilities managed by the government, local authority, faith-based and non-governmental organizations capable of offering ANC [57, 60, 61]. The SSA MHFL reflects facilities available around 2015–2018. However, the Kenyan list had been updated (2020) by incorporating data from Kenya’s routine data reporting system and Kenya’s MHFL [62]. We extracted the mean travel time for each cluster as done for the NTL gridded surfaces. Equity stratifiers were based on factors known to influence ANC4+ coverage, within EPMM recommendations, based on data availability and in WHO’s list of the main barriers to receiving or seeking care during pregnancy [7, 18, 19, 26, 47–49]. They included maternal education, household wealth and travel time to the nearest healthcare facility and were stratified as shown in Table 1. The stratification followed a pragmatic approach, with a policy interpretation, supported by literature and ensuring each arm had a considerable number of observations to allow for robust inference using MBG. Districts were then used as the unit of aggregation. Data on maternal autonomy (decision to seek ANC services) were only collected on the Kenya MIS, while data on ANC initiation was not reported on the Tanzania MIS. Women who attended ANC but had a “don’t know” response for the number of ANC visits or when they initiated their first visit were recoded as missing (1.4% in Kenya, 0.6% in Uganda, and 2.1% in mainland Tanzania). However, the three variables with missing data did not exceed 2.1% of the total sample size by country and were excluded from the analysis (SI section A2). Exploratory analysis is the first stage of geostatistical analysis [54]. It entails visualizing the spatial distribution of sampled clusters (Fig. 1A), examining the correlation between covariates, assessing the relationship between ANC4+ and covariates, and testing for residual spatial correlation [54]. We undertook these steps as detailed in SI section A3. Briefly, Pearson’s correlation was implemented in corrplot package in R [63] while empirical logit [64] was used to assess the association between ANC4+ coverage and the covariates and visualized with scatter plots. To select a set of parsimonious predictors used as fixed effects during geostatistical modeling, we used a non-spatial generalized linear model relating the covariates with ANC4+ coverage. The selection was done by country and equity stratifier resulting in 21 models. Finally, we assessed the evidence of spatial correlation after accounting for fixed effects (parsimonious predictors) through an empirical variogram (S1 Section A3). Separate Bayesian geostatistical models were used to model ANC4+ coverage for each country and equity strata. Each model contained explained factors (fixed effect) and unexplained factors (random effect). The fixed effect was modelled using the predictors denoted as d'(x)β, where d(x) is the vector of parsimonious predictors with the corresponding coefficient β. The random effect was modelled using two terms, S(x) to account for the spatial residual variation and Z to account for the measurement error or small-scale variation that is not captured in S(x). Specifically, the variation in ANC4+ coverage P(x) at location x was modelled using a binomial geostatistical model (Eq. 1). S(x)was modelled as a zero-mean discretely indexed Gaussian Markov Random Field (GMRF) with Matérn correlation function [65]. All fixed and random effect parameters were estimated using the integrated nested Laplace approximation (INLA) and Stochastic Partial Differential Equation (SPDE) implemented in INLA package [65, 66]. Prediction of ANC4+ coverage was obtained using the simulation from posterior distributions of all the parameters and summarized using the mean, standard error and 95% confidence interval (CI) at 3 × 3 km spatial resolution. The high-resolution surfaces were aggregated by district. Additional details about geostatistical models are provided in SI section A4. We assessed the likelihood (exceedance probability-EP) that each pixel and district had ANC4+ coverage above 70%, the target coverage based on EPMM strategy [19] (SI section A5). An EP value close to 100% indicates that ANC4+ coverage is highly likely to be above the target; if close to 0%, ANC4+ coverage, is highly likely to be below the target; if close to 50%, ANC4+ coverage, is equally likely to be above or below the target. We validated our models by checking if the fitted correlation function was compatible with the data using a variogram-based procedure [67, 68] detailed in SI section A6. It entailed simulating many variograms from the fitted model and then comparing them with the estimated empirical variogram from the data. We concluded that the adopted correlation function is compatible with our data if the estimated empirical variogram lies entirely in the 95% confidence interval of the simulated empirical variograms. We estimated the number of pregnant women with ANC4+ visits by multiplying the 3 km gridded surfaces showing ANC4+ coverage from geostatistical models and population gridded surfaces of pregnant women obtained from the WorldPop portal [69]. The number of pregnant women with fewer than four visits (<ANC4+) was obtained by subtracting those with ANC4+ visits from the total number of pregnant women. The results were aggregated by country and district. Briefly, to construct the population density maps, mid-year population of under 1 year (corrected for mortality and migration) were extrapolated by Worldpop based on United Nations (UN) data on births and WorldPop’s estimates of children under 1 year to estimate total annual births. The births were adjusted to match the UN total births by country. The Guttmacher birth to pregnancy rate was used to compute the number of annual pregnancies. Gridded pregnancy surfaces were available for 2020 in Kenya and 2017 for Uganda and mainland Tanzania at 1 km spatial resolution [69]. STATA (StataCorp. 2015. Stata Statistical Software: Release 14. College Station, TX: StataCorp LP.) was used for descriptive analysis, R statistical software [70] for geostatistical modelling and ArcMap version 10.5 (ESRI Inc., Redlands, CA, USA) for all cartographies.
