Mapping populations at risk: improving spatial demographic data for infectious disease modeling and metric derivation

The health divide between urban and rural populations has been well documented [54], as have the increasing levels of urbanization around the world [55]. While much has been done in the past 10 years to refine population data that delineates urban areas [19, 56, 57], much less is known about populations within cities even within relatively coarse divisions (city center, suburban, peri-urban) or by slum dwellers and others (much also remains to be learned about population distribution within rural areas). Africa, in particular, will undergo rapid urbanization in the coming decades [55], yet the data record to understand the demographic variation and health conditions in these cities, let alone changes in disease transmission that may result from urban change, is largely absent. If it is important to understand what is happening within urban areas, even the currently available cluster-level data of the DHS program (Figure 2(c)) is inadequate. While the DHS program and some other surveys have focused on collecting larger urban samples, there remains a need for large samples of urban populations to permit city-specific analyses. Additionally, the definition of urban is not standard across the DHS countries. There is reason to believe that even greater heterogeneity of health and socioeconomic characteristics exist within urban areas.

As with the demographic datasets discussed already, there are a variety of disparate spatial datasets on urban populations that could be brought together to get a better perspective. City population counts for cities with populations of 100,000 and above are produced by the UN Population Division World Urbanization Prospects [55]. Alternative sources of city and settlement population sizes include the City Population website (http://​www.​citypopulation.​de), while different projects are focused on mapping city extents, which these counts could be matched to (e.g., [19, 58], http://​www.​afripop.​org). There still exist significant gaps, however, such as time-series of urban spatial extents, which would facilitate the development of ways to forecast changes in urban extents. Also, information on properly defined neighborhoods within cities is important, such as within-household and within-neighborhood population density, but so are other contexts (e.g., schools).

Most low-income countries do not produce population projections, or forecasts, at a subnational level. Even the United Nations Population Division’s urban population projections [55] do not produce city-level population projections. Yet the demographic inputs for generating subnational estimates and projections are increasingly becoming available. Subnational projections are now being undertaken at least for very large countries (for example, India and China) and for small and large cities in the developing world [59]. For the latter, the forecasting method departs from the traditional cohort-component method and instead uses longitudinal data on cities and subnational estimates of demographic rates (urban fertility, mortality, and migration) derived from survey and census microdata in an econometric model of city growth [60]. These new approaches depend on harnessing old data in a spatial framework. The spatial framework allows disparate units to be linked in new ways, yielding new estimates and projections. Because these methods are largely probabilistic and derived from modeling exercises, the uncertainty associated with these estimates should also be characterized.

The variety of ages, spatial resolutions, and sample sizes of input demographic data translates to great variations in accuracies and uncertainties of any output gridded demographic data products, and this is rarely acknowledged [1]. The most basic level of quantification and communication of this uncertainty to users involves the provision of information on input datasets and methods used in construction, such as is undertaken for GPW, GRUMP [19], and AfriPop [21] (http://​www.​afripop.​org). Ideally, a more rigorous quantification of the uncertainty inherent in output gridded demographic datasets should be undertaken. The rigorous handling and propagation of uncertainty through a mapping process is now regularly undertaken in disease risk mapping within a Bayesian framework (e.g., [2, 16, 17]), resulting in full posterior prediction distributions for each grid cell, providing flexibility in the derivation of differing uncertainty metrics and enabling the production of accompanying uncertainty maps. Undertaking an equivalent approach for deriving accompanying uncertainty maps for demographic datasets would require consideration of the input datasets and output requirements. For instance, this could take the form of estimating the uncertainty in gridded population distribution mapping from census data summarized by administrative unit. Here, two of the major sources of uncertainty are the age of the census data in relation to the output prediction year and the size of the administrative units relative to the population sizes within them and output grid cell size. While uncertainty in temporal projection of census data is relatively well studied, the spatial aspect remains unexplored. For example, gridding a 50,000 km² administrative unit containing 1 million children under age 5 to 30 arc second resolution results in greater uncertainty about the population size and composition residing in each grid square than does the same resolution gridding of a 1,000 km² administrative unit containing 10,000 children under age 5. Based on simulating all possible permutations of grid square composition, bound by the limits imposed by the original administrative unit size and vulnerable population totals, per-grid square measures of spatial uncertainty in composition that relate directly to the gridding methodology could be derived. Secondly, the availability of the DHS cluster locations opens up the possibility of estimating surfaces of variables with associated uncertainty, and this is discussed below.

The availability of the GPS coordinates of DHS clusters (Figure 2(c)) has prompted several studies to utilize geostatistical approaches to derive continuous estimated surfaces of variables of interest. However, survey data are generally collected to be nationally representative and, as such, their sampling frames may not lend themselves to finely resolved geographic grids. The DHS program has been a leader in collecting and providing geocoded information of the survey clusters in addition to their standard data files in which data can be tabulated by first-order subnational regions as well as urban/rural classification. Early examples have demonstrated the value of such approaches for deriving continuous maps of variables of interest from geolocated DHS cluster data. For example, Gemperli et al. [61] investigated spatial patterns of malaria endemicity as well as socio-economic risk factors on infant mortality in Mali using a Bayesian hierarchical geostatistical model. Meanwhile, Soares and Clements [16] used a similar approach for anemia mapping. However, these approaches did not take account of the DHS sampling design or the random spatial displacement that cluster data undergo, and overcoming these issues should be a priority for future applications [62]. Apart from utilizing the cluster location, the subnational regions supply information that can be used with more finely resolved grids. One approach that uses spatial coverage of census aggregates combined with the attribute breadth in survey data is that of the Poverty Mapping efforts [63]. But this is not necessarily the only approach to consider.

Very little is known about migration and mobility within countries, which may occur seasonally and periodically as well as permanently, except through case studies and qualitative place-specific analyses. Disease modeling and health metric derivation, as well as demographic analyses, increasingly require information on migration and mobility [64‐66]. These data are the weak link of the demographic record – even the stock estimates of subnational migration have been largely ignored. Disease modelers often want to know about daily movements rather than decadal ones, but the decadal moves may be important to evaluate for changes to place-specific vulnerability of residents. Decadal moves should be examined more closely with existing survey and census microdata; characterizing more frequent moves will require data collection methods that depart from the standard demographic tool kit. Use of new data, such as spatial locations derived from GPS tracking devices [67] and cell phone usage [68] may show promise. However, to be useful, methods for using these data in combination with more standard demographic data will be necessary.