Spatially disaggregated population estimates in the absence of national population and housing census data

Population data for a sample of areas across the area or country of interest are needed as a primary input to bottom-up population estimation. These data may come from a partial census, census-like population survey (i.e., where a survey is designed to provide population counts), or a specifically designed microcensus survey. The design of the microcensus (or other) survey used to provide training data is a vital step to ensure accurate and unbiased population predictions. The survey design should capture as best as possible the range of densities, demographics, and environments that exist across the area of interest, providing a representative sample. Consideration (for example using stratified sampling) should be given to geographical, socioeconomic, or environmental factors that may influence population densities within the specific context in which the work is being conducted. For example, urban areas are more densely populated than rural; average household sizes may vary according to ethnic, religious, or cultural groups; and the ruggedness of terrain may also influence settlement patterns and population density. The importance of these and other factors will vary by setting, so survey design should be conducted carefully, building on a detailed understanding of the specific context. Another key requirement for these data are robust georeferencing of the geographical areas where population data has been acquired.

Population enumeration can be conducted within administrative units (e.g., census tracts) or within other arbitrarily designed polygons, as long as the population data are explicitly linked to the correct geographical area. Printed maps can be used as guides to identify the correct geographical areas, based on physical landscape characteristics (e.g., roads, mountains, rivers, buildings). However, technological advances now offer a range of more sophisticated methods for the accurate geographical positioning of enumeration activities. GPS can be used to ensure enumeration is occurring in the correct location, and the inclusion of GPS technology in smart phones and tablet computers can integrate navigation and the recording of geographical coordinates into enumeration activities, minimizing locational error and human effort.

The covariates used for bottom-up population estimation should be (i) strongly correlated to population density and (ii) available consistently across all areas where the population estimation is required. Population estimation approaches have previously been categorized as utilizing the relationships between population and covariate data representing the following: (i) built-up areas, (ii) areas of specific land use types, (iii) counts of dwelling units, (iv) satellite-derived measures such as spectral radiance, or (v) socioeconomic or physical characteristics (34, 35, 40). In practice, integration of multiple of these elements is likely to result in the best predictive performance, as they each capture different facets relating to how population numbers and densities vary spatially. Although access to high-quality, spatially comprehensive datasets representing some of these characteristics has traditionally been difficult in resource-poor settings, advances in image-processing techniques, computational power, and the increasing availability of very high-resolution satellite imagery (of the order of <10-m spatial resolution) means that the production of high-quality covariates for many settings is increasingly feasible (41, 42). The mapping of human settlements and even individual buildings from a new generation of satellite imagery and aerial photography is providing detailed geospatial data on human settlement patterns, a key input (as settlement areas or dwelling counts within specified areas) for bottom-up population estimation (12, 43). Furthermore, semantic detail can also be used, such as the density of dwellings or inhabitants (35, 37), types of buildings (e.g., residential/nonresidential) (44), or types of settlement patterns (e.g., informal/formal), which can be distinguished using computer-vision and machine-learning approaches (43).

Other ancillary spatial datasets capturing features related to how humans are distributed on the landscape are already widely used to improve the accuracy of dasymetric mapping. A comparison of top-down population density models in 32 low- and middle-income countries found that covariates related to defining settlements, climate, topography, and ecology consistently explained the most variation in population density (45). Datasets representing factors, such as distance to roads, elevation, slope, and night-time lights (4) can also be integrated for the improvement of predictive accuracy for bottom-up population mapping (35). There has been substantial growth in the availability of volunteered geographic information (VGI), such as through OpenStreetMap, which can help address these issues by providing data such as settlement extents, building footprints, or the locations of roads or facilities (46, 47), and VGI has been utilized in dasymetric mapping previously (11, 48, 49). However, VGI data can be prone to spatial bias in completeness, with a tendency toward better data availability in urban areas and wealthier countries so should be used with care (50). Finally, novel data sources, such as those derived from mobile communications and social media, show potential for not only providing additional spatial covariates to improve the accuracy of both top-down and bottom-up modeling approaches (51), but also as a way to capture the daily, seasonal, and annual dynamics of populations (52–54).

The goal of bottom-up models is to predict populations across large areas where data exist for only a small subset of the area. Less emphasis is placed on explaining the processes producing population distributions, as might be done in models projecting fertility, mortality, and migration rates. It is important to distinguish this goal because it has implications for understanding the methodological approach. First the associations with the covariate datasets are not interpreted causally, which might lead to simplistic environmental determinism arguments. The initial covariate selection, as discussed above, does identify covariates that are expected to be correlated with population density. While some ancillary data could be associated with population density or growth rates, the relationships are also seen as indirect markers of the variation in population, which reflects how human settlements modify and are constrained by the environment. Second, the statistical methods used in bottom-up models must be capable of drawing together multiple sources of data to build the model. These multiple data sources provide information not only from the data space of relationships between population and covariates, but also from their positions in space. Specifically, observations taken closer together in space tend to be more similar than those further apart. This feature, known as spatial autocorrelation, is common to our first-hand experiences of the world—populations cluster together in similar groups—but it violates assumptions of independence that underlie classic statistical methods. Explicit consideration of spatial structure, on the other hand, can substantially improve predictive outputs (55, 56) because spatially structured variation between microcensus areas that is not fully explained by the covariates is still a source of useful information for predictions. Future research in this area should aim to explicitly address spatial autocorrelation, which should provide improved predictive accuracy. Several options are available to ensure that residual spatial autocorrelation is appropriately dealt with. For example, spatially autoregressive models may be used for the predictive modeling of areal data, with a spatial adjacency matrix being used to ensure that the probability of values estimated is conditional on the level of values in neighboring areal units (57). Alternatively, where population data can be represented as spatial points (e.g., using the centroid of small areal units with linked population data), geostatistical methods that model residual autocorrelation as a function of distance between points may be applied (58).

There are few examples in the literature of areal-based spatially explicit models where prediction into unsurveyed locations is the main aim, but none, as far as we are aware, that have applied these methods to population estimation. Nevertheless, currently available software does include this functionality (59), and where applied, the inclusion of spatial dependency has resulted in improved predictive performance (60). Predicted populations from these types of areal-based methods would be explicitly associated with areal units, and the subsequent spatial disaggregation of predicted population counts (along with associated confidence intervals or SEs) can then be used to provide high spatial-resolution population estimates and an indication of predictive uncertainty. While there are no specific examples in the literature of these methods being applied for population estimation, the extrapolation of household survey data by these means has been used to provide bottom-up spatial predictions of population age structures (61). These geostatistical methods have utilized data from specific field surveys and household survey data (e.g., Demographic and Health Survey data), where spatial coordinates refer to the location where the survey was conducted (e.g., a point within a village, or a health center location), or the (normally spatially displaced) survey cluster centroid for national household survey data. Application for the prediction of population size or density would require careful consideration of how population data are represented most accurately as points: explicit geographically linked enumeration (i.e., using GPS supported enumeration hardware to geolocate all households enumerated) provides the ideal basis for this, enabling enumeration within small, uniformly sized, and well-defined areas. Predictive outputs from geostatistical methods could be provided on a grid at the same spatial resolution for which covariate datasets are available (or at a coarser spatial resolution if required), thus removing the need for subsequent spatial disaggregation.

The previous examples of bottom-up population estimation discussed have focused on the application of linear regression models of raw population counts, population densities, or the natural log of one of these values. For the modeling of population counts, the application of Poisson, negative binomial, or quasi-Poisson regression may be more appropriate, given that population counts are inherently positive integers. Note also that the spatial modeling approaches we discuss here are more commonly discussed in ecological studies to predict the abundance of plant and animal species (62) than in the demography or population studies literature.

Because the primary objective of bottom-up estimation is prediction in nonsampled areas, validation of the spatially disaggregated population estimates is rare. Where results have been reported previously they indicate a good correlation between predicted and observed total population or population density values, particularly when considering larger administrative units. For example, an R² value of 0.72 (squared correlation coefficient for observed vs. predicted counts, using independent testing data) was obtained by Harvey et al. (34), highlighting good predictive performance. Similarly, population predictions from a linear regression model developed using data from 10% of available census units in the Netherlands, along with building floor space or volume data, produced predictive errors (calculated as median absolute percentage error, based on the remaining 90% of available census units) of 18.3% at the smallest administrative level, 9.3% at the largest administrative level, and 0.5% at the national level (44). Several previous studies have identified persistent overestimation of population density in rural areas or areas where buildings are largely nonresidential (e.g., industrial sites) and underestimation of population density in high-density urban settings, particularly where multistory buildings are common (34, 35, 44). This suggests that contextual information, such as residential vs. nonresidential buildings or building height information, may improve predictions. The utility of such additional information has been highlighted in the Netherlands, where building footprint areas, building floor space areas (which incorporates multiple floors per building), and building volume (which incorporates building heights) were used as covariates in a predictive linear regression approach, with building floor space found to produce the most accurate predictions (44). However, not all studies assessed predictive performance against an independent testing dataset, and in the majority of examples, values were not back-transformed before accuracy assessment (i.e., validation was based on population density or the log of population density values rather than population counts) (27, 35). Both of these scenarios result in an overestimation of predictive accuracy and differences in accuracy assessment protocols make meaningful comparison of different modeling approaches impossible.

Future research should incorporate the comparison and validation of different methodologies to provide a clear understanding of predictive accuracy. While limited applications to date have indicated good predictive performance, robust testing and validation should be conducted in a range of settings. In particular, validation studies performed in countries with comprehensive and reliable census data coverage should be a priority. Population estimates can also be compared with other administrative data, such as vaccination records (61) or local population projections. This type of testing and validation is required to strengthen the evidence that bottom-up approaches can be considered as an appropriate means of generating robust population estimates in areas where complete census coverage is not possible.