A scoping review of the methods used to estimate health facility catchment populations for child health indicators in sub-Saharan Africa

Abstract

Background

Evidence indicating persistent geographic inequalities in health outcomes signifies a need for routine subnational monitoring of health-related Sustainable Development Goal targets in sub-Saharan Africa. Health facilities may be an appropriate subnational unit for monitoring purposes, but a lack of suitable demographic data complicates the production of baseline facility-level population denominators against which progress can be reliably measured. This scoping review aimed to map the methods and data sources used to estimate health facility catchment areas and translate them to population denominators for child health indicators in the region.

Methods

Peer-reviewed research publications and grey literature reports were identified by searching bibliographic databases and relevant organisational websites. The inclusion criteria required that studies were conducted in sub-Saharan Africa since January 2000, described quantitative method(s) for estimating health facility catchment areas and/or population denominators, and focussed on children as the population of interest. Following title/abstract then full text screening of search results, relevant data were extracted using a standard form. Thematic analysis was undertaken to extract themes and present a narrative synthesis.

Results

Overall, 33 research publications and 3 grey literature reports were included. Of these, only 7 research studies and 1 technical guidance document outlined aims explicitly framed around methods development and/or evaluation. Studies increasingly estimated catchment areas using complex geostatistical or travel time-based modelling approaches rather than simpler proximity metrics, and produced denominators by intersecting catchment boundaries with gridded population surfaces rather than aggregating area-based administrative counts. Few studies used data produced by or describing health facilities to link estimation methods to service utilisation patterns, inter-facility competition or facility characteristics.

Conclusion

There is a need for catchment population estimation methods that can be scaled to national-level facility networks and replicated across the region. This could be achieved by leveraging routinely collected health data and other readily available and nationally consistent data sources. Future methodological development should emphasise modern geostatistical approaches drawing upon the relative strengths of multiple data sources and capturing the range of spatial, supply-side, individual-level and environmental factors with potential to influence catchments’ extent, shape and demographic composition.

Background

Projections from the 2017 Global Burden of Disease study suggest that many countries of sub-Saharan Africa (SSA) are falling short of the progress required to meet any health-related Sustainable Development Goals (SDGs) target by 2030 [1]. The region also faces challenges in relation to child health; despite recent improvement, levels of mortality [2, 3] and infectious disease incidence [4, 5] remain high, amid evidence of growing non-communicable disease burden [6]. Subnational analyses, however, reveal within-country inequalities in the distribution of health outcomes [7,8,9] that would otherwise be hidden by national-level data, signifying a need for routine monitoring at more granular geographies. By revealing and characterising high risk areas or underserved populations such an approach could also help to address spatial inequalities, contributing to targeted resource allocation and the development of locally-relevant interventions and services [1, 2, 7].

Health facilities (HFs) may be an appropriate subnational unit for monitoring progress against targets in SSA: they provide routine, formal care to populations in small geographic areas and, in so doing, collect continuous and near real-time empirical data describing service utilisation, health status, disease incidence and prevalence, and intervention coverage [10, 11]. Intuitively, progress monitoring at this level presupposes clear knowledge of the geographical ‘catchment’ area served by each HF, together with its baseline denominator population and demographic composition [12]. Though traditionally viewed as the principal sources of demographic data in many low- and middle-income countries (LMICs), censuses and household surveys do not provide direct population estimates at the lowest levels of health service delivery [13]. Moreover, as catchments are rarely delineated by unambiguous administrative boundaries in SSA, so-called ‘natural’ catchments predominate, tending to emerge as a product of interacting factors influencing patient choice [14], including HF type [15], service quality [16, 17] or distance decay (meaning the tendency towards waning utilisation with greater travel distance) [18, 19]. Without the benefit of typical demographic data and methods, a range of statistical and geospatial model-based approaches to the estimation of HF catchment areas and population denominators have been developed, many of which account for these, and other, salient factors [20], but seldom incorporate the data collected by HFs themselves as the product of routine patient care.

Nonetheless, the view that routinely collected health data (RCHD) could be better leveraged for population health improvement has gained traction in recent years, with renewed efforts to improve their quality [21, 22] and establish them as a source of intelligence to monitor health indicators and inform local, evidence-based decision-making [23,24,25]. Meanwhile, District Health Information Software (DHIS2), a health management information system (HMIS) for the collection, warehousing and reporting of RCHD, has been adopted by more than 70 LMICs covering some 30% of the world’s population, principally in SSA and south/south east Asia [26], thus strengthening and harmonising their data collection and production infrastructure. These developments are emblematic of the rapidly evolving data landscape of SSA and may have precipitated methodological innovation that could be replicated more widely across the region. This scoping review was conducted with the aims of mapping the: (i) methods and data sources that have been used to estimate HF catchment areas and translate them to population denominators for child health indicators in SSA; (ii) approaches used to evaluate these estimation methods.

Methods

The review followed the methodological framework established by Arksey and O’Malley [27], and is reported in accordance with the Preferred Reporting Items for Systematic reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) checklist [28] (Supplementary file 1). A protocol was also registered on Open Science Framework [29].

Identifying and selecting peer-reviewed publications

Publications were identified using a search strategy developed by the study team and reviewed by two research librarians. Database search strings (Supplementary file 2) consisted of MeSH terms (Medline only) and search terms arranged into four broad ‘concepts’ (children, HFs, catchment areas/population denominators and SSA) using Boolean operators. Exploratory scoping searches were used to determine the combination of terms required to capture each ‘concept’. Searches of Medline, Scopus, Web of Science Core Collection, GeoBase and African Index Medicus bibliographic databases were executed on 25th October 2021. Results were imported into an EndNote X9 (Clarivate Analytics, Philadelphia, USA) database and de-duplicated.

Screening against the inclusion criteria (Table 1) was conducted in two stages: (i) title/abstract screening of de-duplicated search results using the R ‘metagear’ package [30]; (ii) full text screening of results passing the first stage. Where publications were unavailable online, the authors were contacted directly. Prior to each stage a random sample of 20% were independently screened by MJ and WAA to calibrate the inclusion criteria and screening approach [28]. Discrepancies between reviewers were discussed and, where necessary, resolved by JW as arbitrator. Agreement was assessed using Cohen’s kappa coefficient; once a minimum value of 0.80 was achieved, all remaining publications were screened by MJ alone [31]. After completion, the list of full text screening decisions and rationale was verified by WAA and JW. Reference lists were searched to identify additional publications meeting the inclusion criteria.

Table 1 Inclusion criteria used to select peer-reviewed publications for the review

Grey literature

Recognising non-academic organisations’ role in the production of methodological and technical guidance, grey literature searches were executed during January, February and June 2022. Google Scholar and relevant organisational websites were searched using a simplified strategy consisting of keyword combinations representing ‘concepts’ used to identify peer-reviewed publications (Supplementary file 3). As grey literature tends to appear more regularly after around 30 pages of Google Scholar search results [36], the first 50 pages (500 results) were screened by title/preview only, as were the results from organisational websites. EBSCO was searched as a general source of grey literature and ProQuest for dissertations and theses. Results from peer-reviewed journals were excluded. Title/abstract screening used the platforms’ web interfaces, with results passing this stage progressing to full text screening. Screening was conducted by MJ alone and, other than relaxing the third criterion (Table 1) to allow inclusion of methodological guidance not linked to specific countries, followed the approach used for peer-reviewed publications.

Data extraction and synthesis

For consistency, a standard form was developed to extract variables including bibliographic information, study setting and population, data sources, software and methods used, results, findings and limitations. Following independent extraction from 15% of peer-reviewed publications by MJ and WAA as a calibration exercise [28], the remainder were reviewed by MJ alone. After completion, the extracted data were verified by WAA. R v4.0.2 and RStudio v1.3.1073 (R Core Team, Vienna, Austria) were used to create quantitative tables summarising the corpus of publications. Thematic analysis was undertaken to extract key themes and present a narrative synthesis.

Results

A PRISMA flow diagram outlining the selection process is presented in Fig. 1. Twenty-nine of 1087 unique peer-reviewed publications were included, as well as 4 identified via reference list searches. Also included were 3 grey literature reports, consisting of one case study [37], one thesis [38] and one technical guidance document [39], of which the latter is excluded from the forthcoming quantitative tables and synthesis.

Fig. 1

PRISMA flow diagram outlining the process used to identify and select relevant publications and reports

Study setting and health facility locations

Excepting 2 region-wide [40, 41] and 2 multi-country studies [37, 42], most (88.6%) peer-reviewed publications and grey literature reports (hereafter, publications) described single-country studies (Table 2). Overall, 15 countries were represented; while most had 3 publications or fewer, 13 (37.1%) originated from Kenya. Although 7 (20.0%) publications implemented methods across the national HF network [43,44,45,46,47,48,49], most (68.6%) analysed a defined subnational area or subset of specific, purposively selected HFs. Only 3 (8.6%) publications stated that analyses captured public and private sector HFs [42, 46, 50] (Table 3). All studies utilised data describing HF locations (Table 4). Although study staff sometimes (14.3%) conducted onsite geolocation by field survey [15, 42, 44, 51, 52], information held within the national health system or by HFs were the most common data source (60.0%). Recent years have seen the development of open databases geolocating public sector HFs across SSA [53, 54], and there were instances of their use to augment within-country data [46], or as the primary source for region-wide analyses [40, 41].

Table 2 Geographic area and scale of the included publications

Table 3 Detailed summary of the data extracted from the publications included in the review

Table 4 Characteristics of the included publications

Study aims

While only 7 (20.0%) publications were framed around methods development and/or evaluation (Table 4), 12 (34.3%) each utilised similar methods to estimate catchment areas and/or population denominators, or to compute measures of health service accessibility as covariate to models testing associations with outcomes such as stunting [55] or hospital admission [56]. Several different health indicators were analysed (Table 3): indicators related to malaria or fever treatment-seeking were used for 15 (42.9%) publications, but 4 (11.4%) each analysed child mortality, access to/utilisation of surgical services, and immunisation coverage.

Catchment area estimation

Catchment boundaries were sometimes (17.1%) aligned with those of established administrative units (Table 4), usually selected following an algorithmic process informed by patient-level data extracted from the HMIS [37, 57,58,59,60]. To produce catchment areas independent of administrative boundaries, many (31.4%) publications described the use of GIS software to assign locations to HFs based on straight-line distance [38, 47,48,49,50,51,52, 56, 61,62,63]. In cases where the network of roads and/or footpaths was also mapped, this was enhanced by measuring ‘network’ distance [64,65,66]. Similarly, additional spatial datasets describing other topographic features, such as land cover or slope, were often (31.4%) combined as input to cost impedance models measuring the travel time ‘cost’ of health-seekers’ most efficient route between locations [15, 19, 41,42,43,44, 46, 55, 67,68,69]. More complex approaches integrated spatial data with that gathered from nationally representative household surveys [18, 40, 70] or the HMIS [71] in geostatistical models assembling catchments based on location-specific estimates of health-seeking probability. Most publications presented a map of HF locations (65.7%) (Table 4) but, while those describing methods development or catchment estimation typically delineated their boundaries explicitly, others computing an accessibility-based model covariate tended to produce isochrones visualising travel time or distance strata.

Denominator estimation

Although not all publications (68.6%) translated catchments to population denominators (Table 4), the methods for doing so were divided almost equally between intersecting catchment boundaries with fine spatial-scale gridded population surfaces (34.3%) and aggregating nationally- or locally-produced population counts (31.4%) at the level of districts [37], enumeration areas [38, 52], or villages [72], for example. Overall, 7 (20.0%) publications reported denominators for individual HFs and 6 (17.1%) at other levels of spatial aggregation.

Evaluation of methods

Most (60.0%) publications described methods evaluation (Table 4), which commonly entailed sensitivity analyses (28.6%) or comparing multiple methods within a single analysis dataset (11.4%), but in the case of model-based approaches used a validation subset to assess the performance of candidate model specifications (8.6%). Only 4 (11.4%) described comparisons against independent [64, 70, 72] or purposively collected data [52].

Data sources and software

Although most publications utilised open, or widely available, secondary data sources only (Table 4), several accessed data linking service utilisation events with health-seekers’ origin location, which are not routinely available in this setting: 5 (14.3%) were conducted within a Health and Demographic Surveillance System area [61, 65, 66, 68, 69], 7 (20.0%) extracted patient-level data from the HMIS [37, 57,58,59,60, 71, 72] and 2 (5.7%) surveyed health-seekers attending local HFs [15, 51].

Over half were published from 2017 onwards (Table 3). This may be linked to wider adoption of open-source software: while Geographic Information System (GIS) products such as ArcGIS and ArcView (Esri, Redlands, USA) and the AccessMod extension [73] were most common overall, use of R packages for spatial analysis (including ‘geoR’, ‘gstat’ and ‘R-INLA’) was evident from 2017 [19, 64, 71, 72].

Discussion

This review of the literature on HF catchment population estimation for child health indicators in SSA found that few of the 36 included publications took methods development and/or evaluation as the primary focus. Of these, 7 were subnational research studies [15, 38, 51, 52, 64, 71, 72] and the eighth was a technical guidance document concerning single-hospital denominator estimation [39]. Though data inequity has previously been cited as a barrier [15], recent efforts to strengthen health data infrastructure and ongoing advances in the availability, coverage and resolution of spatial and demographic data may now offer the opportunity for development of reproducible methods that can be scaled to national-level networks. This will be essential if the HF is to be taken forward as a credible subnational unit for routine monitoring of health indicators.

A successful catchment estimation method should, without need for empirical data tracing actual health-seeking flows, be able to outline the geographic area from which the users of a given HF are expected to originate [20]. The most basic method aligns catchments with established administrative units. This is problematic, however, in that subnational administrative boundaries do not usually impede population movement and are thus unlikely to accurately represent health-seeking flows. Most publications described methods underpinned by measures of spatial accessibility, which focus upon the space or distance separating health-seekers from services [74]. The chosen measure has implications for catchments’ extent, shape and configuration, however. By conceptualising the catchment network as a complete areal tessellation encapsulating the entire population [38, 47, 49, 51, 52], the simplest straight-line distance methods carry the unrealistic assumption that all health-seekers have access to one, and only one, HF. Recognising that some may, in reality, reside beyond practical reach of any HF, buffers were sometimes used to constrain catchments to a distance threshold provided by policy targets [46, 52, 55], guidance around the health indicator under consideration [41, 44, 50], or the inflection point of a modelled decay curve [18, 19, 70]. As straight-line distance inherently overlooks transport infrastructure and topographic barriers to movement, additional spatial data may be used to produce a more realistic measure [75]. Network distance may have limited utility in SSA, where pedestrian travel is common, and not necessarily restricted to roads and footpaths [76]. Instead, the process of converting the study area to a grid representation, assigning all cells a traversal ‘cost’ based on their aggregate spatial characteristics, then fitting cost impedance models to measure the travel time associated with health-seekers’ most efficient route to the nearest HF is often preferred, despite increased data and computational needs [75]. Aligning catchment boundaries more closely with topographical features that bar or facilitate movement may better represent the real-world travel experience in this setting, but is sensitive to the quality and resolution of spatial data. One publication noted that the coarse resolution required to handle the regional inconsistency of road network and other spatial data may have overestimated accessibility in rural areas or near major roads [41]. Distance measures or cost impedance models were sometimes integrated within a broader geostatistical modelling framework alongside other supply-side or individual-level factors, such that catchments were defined by the combined effect of multiple covariates on location-specific probability of health-seeking and service utilisation [70, 71]. Overall, the included publications depicted a trade-off between catchment estimation methods that are comparatively easy to implement, but oversimplistic and likely to yield unrealistic denominators, and others that more accurately represent reality but entail additional data needs and methodological complexity (Fig. 2).

Fig. 2

Outline of the catchment estimation methods identified by the review, together with their associated inputs and outputs

Increasingly complex modelling approaches present the further challenge that the optimal specification and appropriateness of any assumptions are likely context-specific [75, 77], underlining the need for evaluation. While cost impedance modelling facilitated comparison of alternate travel scenarios based on variable transport modes [44, 69] or seasonal conditions [43, 78] as a means of sensitivity analysis, geostatistical approaches further allowed for the performance of candidate model specifications to be compared using a validation dataset [19, 40, 71]. Nonetheless, empirical ‘origin/destination’ data specifying health-seekers’ origin location and the HF attended, though scarce in SSA, expose actual health-seeking flows and should be seen as the ‘gold standard’ for evaluation. Where available, these data provided new, otherwise unattainable insight: two related studies collecting origin/destination data via onsite surveys showed that travel beyond the nearest HF was common [15, 51]. Catchment boundaries bisecting the space between adjacent HFs typify straight-line distance methods, but the finding that a substantial proportion of health-seekers attended a more distant, but higher-tier, HF (hospitals rather than health centres, for example) suggests that inter-facility competition or other HF characteristics may also influence health-seeking and, as such, have a role in appropriate boundary placement. Indeed, while accessibility measures, in isolation, implicitly assume that all health-seekers attend, and can be served by, the nearest HF, methods adjusting for facilities’ capacity [55] or capabilities in respect of specific conditions [42], or health-seekers’ individual-level characteristics [19, 70], for example, may produce more realistic results. Though simpler methods based on spatial accessibility have arguably been necessary in the absence of data specifying health-seekers’ origin location or capturing the range of aspatial supply-side and individual-level factors known to influence patient choice [20], recent advances may now permit the use of more comprehensive, yet scalable, methods leveraging multiple data sources with national coverage.

Directly linking estimation methods to RCHD could help to narrow the gap between modelling and reality. Two publications [37, 57] followed disease-specific technical guidance issued by the World Health Organisation [39, 79], which proposed algorithmic case detection and geolocation from retrospective hospital records then catchment delineation at the geographic extent of rank-ordered administrative units contributing a cumulative 80–85% of cases. Though a relatively simple and intuitive algorithm, replication is limited by the burden of manual retrieval and review of physical records, which were often difficult to locate, incomplete or illegible [57]. Instead, clinical surveillance databases appear a more practicable foundation for algorithm development [60] or amalgamation of RCHD from multiple HFs [71, 72]. Although the included publications described local, purpose-built databases, they lend credence to the notion that HMIS, bolstered by recent strengthening initiatives, may be a viable platform for scalable estimation methods. Indeed, DHIS2 has been instrumental to the development of a novel approach to district-level denominator and intervention coverage estimation [80], subsequently replicated elsewhere in SSA [81, 82]. Having thus far been applied in established administrative units only, this method did not meet the review inclusion criteria but may have potential at more granular geographies such as HF catchments. Perhaps reflecting the long-standing prominence of malaria within the international health agenda [83], nearly half of the included publications focussed on related indicators, with other pressing concerns such as lower respiratory infections and diarrhoeal disease [84, 85] comparatively underrepresented. The breadth of RCHD could address this imbalance by enabling parallel, indicator-specific estimation using a common methodological approach, a valuable innovation given the propensity for health-seeking and distance decay to vary by type or severity of health event [61, 86]. Realising these aspirations will depend on consistent, complete and high-quality data throughout the health system, however, a concern that has historically led to structural underutilisation of RCHD in SSA [87]. Although embedding standard data entry procedures and automated quality assessment tools within electronic HMIS may alleviate some quality issues, continued efforts to strengthen the manual, paper-based data capture processes and tools used by health workers should remain a priority [21, 24]. Few studies captured both public and private HFs, which has rarely been possible in SSA owing to suboptimal reporting by the private sector [88, 89]. There is a need for additional policy measures targeted to eliminating this gap so that the entire HF network can be factored into routine monitoring and decision-making processes.

Imprecision was evident in translating catchment areas to population denominators. Most publications followed one of two broad approaches (Fig. 3). The first, applied where catchments were aligned with established administrative units, estimated denominators by aggregating nationally- or locally-produced population counts. These counts were typically projected using objective growth rates, and, in one case, were ‘downweighted’ in line with distance decay [72]. Recent advances in the production of spatially disaggregated demographic data have enabled an alternate, GIS-based approach intersecting catchment boundaries with fine spatial-scale gridded population surfaces. Though this may improve the precision of denominators associated with non-standard administrative/spatial units, such as catchments [13], the gains may be attenuated if small-area population demographics are unknown, necessitating subgroup approximation as a proportion of the total cell count [47]. Similar advances in temporal granularity are also needed; reliance on temporally coarse data, such as the decennial census, has meant that catchment denominators are effectively treated as static counts despite fluctuating in response to short-term population movement [90], individual travel behaviours [91], seasonal conditions [43, 78] and disease epidemiology [57]. Aggregated mobile phone call records have shown promise for tracking spatio-temporal population dynamics [90,91,92] and could contextualise longitudinal service utilisation patterns discerned from RCHD, speaking to the potential of hybrid methods drawing upon multiple data sources. Further methodological enhancement would be needed, however, to address the selection biases associated with HF utilisation [87] and mobile phone ownership [93].

Fig. 3

Outline of the inputs and methods for denominator estimation identified by the review

Having taken HF catchment areas as the spatial unit of interest, this review has a distinct focus to much of the research literature on small-area and subnational population estimation and contributes to the fields of public health and spatial demography. The breadth of the review was a strength, having employed a search strategy bridging geospatial, epidemiological and demographic methods for the estimation of HF catchment populations, and gathering peer-reviewed and grey literature from several sources. It is acknowledged, however, that relevant publications utilising such methods may have been omitted if substantive methodological content was absent from titles or abstracts. Moreover, by limiting the review to the health sector methods unique to education, or other public services [20], may have been excluded.

Conclusion

This review found that most studies implemented estimation methods using data from a single or subset of HFs only. Such methods are unlikely to be generalisable if benefitting from well-developed and robust data infrastructure unrepresentative of the wider health system, underlining the need for investment in methods that can be scaled to national-level HF networks. Whilst considerable methodological variation was observed, standardised and scalable methods could be achieved by leveraging data sources that are readily available at national scale, such as RCHD, nationally representative household surveys and spatially disaggregated demographic data. Many publications focussed on indicators related to malaria, but RCHD could also help to fulfil the need for population denominators in respect of other heath conditions. Although quality concerns have historically resulted in underutilisation of RCHD in SSA, emphasising their value for catchment population estimation could accelerate quality improvement initiatives and efforts to improve private sector reporting rates. Future methodological development should move away from using accessibility measures in isolation towards geostatistical approaches uniting spatial characteristics of health service supply with the broader range of supply-side, individual-level and environmental factors that may exert an influence on health-seekers’ choice behaviour. In particular, explicitly accounting for inter-facility competition in catchment estimation could help to overcome the commonplace, but likely invalid, assumption of attendance to the nearest facility. Future research should also consider the potential of adapting innovative approaches utilised in other sectors, disciplines or high-income countries for HF catchment population estimation in SSA.