The Millennium Development Goals (MDGs) have prompted an expansion in approaches to deriving health metrics to measure progress toward their achievement. Accurate measurements should take into account the high degrees of spatial heterogeneity in health risks across countries, and this has prompted the development of sophisticated cartographic techniques for mapping and modeling risks. Conversion of these risks to relevant population-based metrics requires equally detailed information on the spatial distribution and attributes of the denominator populations. However, spatial information on age and sex composition over large areas is lacking, prompting many influential studies that have rigorously accounted for health risk heterogeneities to overlook the substantial demographic variations that exist subnationally and merely apply national-level adjustments.Here we outline the development of high resolution age- and sex-structured spatial population datasets for Africa in 2000-2015 built from over a million measurements from more than 20,000 subnational units, increasing input data detail from previous studies by over 400-fold. We analyze the large spatial variations seen within countries and across the continent for key MDG indicator groups, focusing on children under 5 and women of childbearing age, and find that substantial differences in health and development indicators can result through using only national level statistics, compared to accounting for subnational variation.Progress toward meeting the MDGs will be measured through national-level indicators that mask substantial inequalities and heterogeneities across nations. Cartographic approaches are providing opportunities for quantitative assessments of these inequalities and the targeting of interventions, but demographic spatial datasets to support such efforts remain reliant on coarse and outdated input data for accurately locating risk groups. We have shown here that sufficient data exist to map the distribution of key vulnerable groups, and that doing so has substantial impacts on derived metrics through accounting for spatial demographic heterogeneities that exist within nations across Africa. © 2013 Tatem et al.; licensee BioMed Central Ltd.
The AfriPop project (http://www.afripop.org) has recently completed construction of 2010 and 2015 estimates of population distribution for continental Africa and Madagascar at approximately 100 m spatial resolution. Full details are provided in Linard et al.[31] and on the project website (http://www.afripop.org). Briefly, a GIS-linked database of census and official population estimate data was constructed, targeting the most recent and spatially detailed datasets available, given their importance in producing accurate mapping [31-33]. Detailed maps of settlement extents were derived from Landsat satellite imagery through either semi-automated classification approaches [33,34] or expert opinion-based analyses [31]. These settlement maps were then used to refine land cover data, while local census data mapped at fine resolution enumeration area level from sample countries across the continent were exploited to identify typical regional per-land cover class population densities, which were then applied to redistribute census counts to map human population distributions at 100 m spatial resolution continent-wide [31,33,35]. Where available, additional country-specific datasets providing valuable data on population distributions not captured by censuses, such as internally displaced people or detailed national surveys, were incorporated into the mapping process [36]. In order to examine the effects of utilizing subnational data on age and sex structures of populations, 2010 national-level data were first obtained to provide a baseline for comparison. These were obtained from the United Nations Population Division’s World Population Prospects 2010 publication [27] and are derived from national-level demographic models built upon census data. These national-level proportions were then used to adjust the gridded population dataset described above to produce separate five-year age group gridded datasets by sex, following approaches used in many previous studies that assume demographic homogeneity across countries (e.g. [6,9,10,28,30,37]). Data on subnational population compositions from the last 20 years were obtained from a variety of sources for all mainland African countries, plus Madagascar (Additional file 1: Protocol S1). Contemporary census-based counts broken down at a fine resolution administrative unit level generally provide the most reliable source for population composition mapping, due to the large sample sizes providing reliable information summarized for small areas. Where age and sex data reported at subnational levels were available for censuses undertaken within the last two decades, these were obtained for this study (Additional file 1: Protocol S1). An addition to the aggregated full census data are large samples of household-level records derived from censuses (census microdata) that provide age and sex structure, reported generally by administrative level 1 (e.g., province) or 2 (e.g., district). Census microdata on subnational age and sex proportions by subnational regions for African countries within the last twenty years were obtained where available (Additional file 1: Protocol S1). While census data are often readily available for high-income countries, for African countries census data with subnational reporting of age and sex structure can often be either unavailable or substantially more than a decade old. Alternative national household survey data sources were therefore exploited to provide the most contemporary and spatially detailed estimates as possible of age and sex proportions, given the constraints of their sampling frameworks. Here, national household survey data on population age and sex compositions were obtained from the most recent Demographic and Health Survey (DHS), Malaria Indicator Survey (MIS), or AIDS Indicator Survey (AIS) [38], or from Multiple Indicator Cluster Surveys (MICS) [39], for all countries where such surveys have been undertaken. Summaries of subnational population structure by sex and five-year age groupings from either full national census summaries, census microdata, or household surveys were obtained for 47 of the 50 countries in mainland Africa, plus Madagascar. Where multiple datasets from similar time periods were available, the census or census microdata were given priority for use, due to the larger sample sizes. For four countries (Libya, Eritrea, Western Sahara, and Equatorial Guinea), no subnational estimates of age and sex structures were found, and for these countries the UN national estimates and projections for the 2000-2015 period [27] were obtained and used in the mapping. The relatively small sample sizes for household survey data and census microdata compared to those from full census data mean that age and sex proportions derived from them are more uncertain. To ensure that age proportions derived from these datasets were representative of those derived from census data, instances where (i) national household surveys were undertaken in the same year or within one year of a national census and (ii) census microdata samples and the full census that each was derived from were collated and statistical comparisons undertaken, which showed consistent and strong correlations (Additional file 1: Protocol S1). Once datasets on numbers and proportions of individuals by age and sex had been collated for as many subnational units as available within the last two decades, using sample weights where applicable to household surveys, these were matched to corresponding GIS datasets showing the boundaries of each unit. Africa-wide GIS-linked data on proportions of individuals by age and sex and by administrative unit were created for as close to 2010 as was available (Figure 1, further datasets are provided in Additional file 1: Protocol S1). Spatial demographic datasets for mainland Africa and Madagascar. (a) The estimated proportion of children under 5 years old subnationally; (b) the estimated proportion of women of childbearing age subnationally; (c) the Africa-wide 1km spatial resolution gridded dataset of numbers of children under 5 years old in 2010, with close-ups showing 100m spatial resolution detail for southern Ghana and Luanda, Angola. The production of spatial population datasets for Africa has previously relied on simple interpolation between census-derived timepoints where available or, more commonly, the application of UN Population Division national-level growth rate estimates [27]. For 45 of the 50 countries in mainland Africa plus Madagascar, subnational growth rates derived from either censuses or official national estimates were obtained (see Additional file 1: Protocol S1 for details). Additionally, separate growth rates for urban and rural areas nationally were obtained for those countries and time periods for which subnational growth rate data were not available [40]. Finally, estimated population sizes for named African cities [40], and the urban extents dataset used in the construction of the Global Rural Urban Mapping Project (GRUMP) [26] were obtained. The urban extents matching those African cities for which individual population totals are estimated in the UN World Urbanization Prospects [40] were identified, and the totals for 2000-2015 matched up. The GIS unit-linked age and sex subnational proportions dataset described above was used to adjust the existing AfriPop 2010 spatial population datasets [31], to produce estimates of the distributions of populations by sex and five-year age group across Africa in 2010. The datasets were then adjusted to ensure that national population totals by age group, specific city totals and urban/rural totals matched those reported by the UN [27,40]. For the analyses outlined in the remainder of this paper, the summation of the datasets representing males and females in the 0-5 year age group was undertaken to produce a 2010 distribution dataset of children under 5 years old, and the summation of datasets representing females in the 15-49 year age groups was undertaken to produce a 2010 dataset of women of childbearing age. The application of subnational growth rates to produce 2000, 2005, and 2015 datasets is described in Additional file 1: Protocol S1. To examine the effects on health and development indicators through use of the new subnational characterizations of children under 5 and women of childbearing age compared to undertaking national-level age adjustments using the UN data [27], two sets of illustrative analyses were undertaken. Firstly, Africa-wide estimates of the number of children under 5 years old residing in different Plasmodium falciparum malaria prevalence classes were calculated, and secondly, estimates of the number of women of childbearing age residing at different travel times from the nearest major settlement (population >50,000) across Africa and nearest health facility for countries with open access geolocated datasets of facilities were estimated. In each case the focus was on the size of the change in output metrics through accounting for demographic spatial heterogeneity, rather than the estimates produced and their fidelity. One component of MDG 6 is an aim to halt and begin to reverse the incidence of malaria [1], with targets focused on those under 5 years of age, upon whom the greatest burden from the disease falls. To assess achievement of these targets, and the derivation of malaria metrics in general, maps of malaria prevalence are increasingly being used in combination with spatial population datasets to estimate numbers at risk and burdens (e.g. [5,6,41,42]). The Malaria Atlas Project (http://www.map.ox.ac.uk) has recently published a mapped distribution of the intensity of P. falciparum transmission in 2010 based upon infection prevalence among children aged 2 to 10 years (PfPR2-10) [5]. Here, the estimated distribution of prevalence by classes that have been proposed in the selection of suites of interventions at scale to reach control targets at different time periods [43,44] (Figure 2a) was used to extract estimated numbers of children under 5 years old per country residing in these different prevalence classes from (i) the AfriPop 2010 population dataset [31] adjusted to represent children under 5 using UN national proportion estimates [27] as described above, and (ii) the dataset of the 2010 population under 5 constructed from subnational data described above. For both datasets, national population totals were adjusted to match UN reported numbers [27] to ensure that any differences seen in numbers at risk were due solely to the addition of subnational information on under-5 proportions. Further details are provided in Additional file 2: Protocol S2. P. falciparum malaria prevalence in Africa and the effects on metrics of accounting for subnational age structure. (a) Predicted prevalence classes for P. falciparum malaria in Africa [5]. (b) The absolute percentage changes in estimated numbers of children under 5 years old residing under the three prevalence classes through changing from using UN national proportions [27] to produce per grid cell estimates of numbers under 5 years to using the subnational proportion data assembled here (Additional file 1: Protocol S1). (c) The changes in estimated numbers of children under 5 years old residing under the three prevalence classes through changing from using UN national proportions [27] to produce per grid cell estimates of numbers under 5 years to using the subnational proportion data assembled here (Additional file 1: Protocol S1). In (b) and (c), data values are only plotted when a transmission class encompasses >10% of the population of a country. Improving access to and for remote populations is an important priority for many of the MDG targets, such as those focused on eradicating extreme poverty, achieving universal primary education, and developing a global partnership for development [1]. Moreover, each health-related goal is dependent upon accessing populations to provide interventions, principally delivered through health facilities, and the difficulty in traveling to these facilities has been consistently highlighted as a barrier to treatment in rural populations, particularly in maternal health [45,46]. The measurement of accessibility or “remoteness” of populations is therefore of importance in measuring progress toward achieving these goals, and increasingly, approaches based on GIS-derived travel times are being applied [45-51]. A recently developed continent-wide travel time dataset [52,53] was used here to map those regions estimated to be greater than five hours from the nearest settlement of population size greater than 50,000. This dataset was used as an illustrative proxy for health system access, since reliable continent-wide datasets on health facility locations do not currently exist. To demonstrate the size of the variations achieved when using actual health facility data, for eight countries with open-access datasets of health facility locations (Additional file 2: Protocol S2), maps representing estimated travel times to the nearest facilities were constructed following previous approaches [50-55] (Additional file 2: Protocol S2). The accessibility datasets were used to extract estimated numbers of women of childbearing age per country residing in different travel time classes from (i) the AfriPop 2010 population dataset [31] adjusted to represent women of childbearing age using UN national proportion estimates [27] as described above and (ii) the 2010 distribution dataset of women of childbearing age constructed from subnational data described above.