Introduction

The 2030 round of censuses is likely to be one of the most challenging with which to provide timely, robust and comparable statistics. Census response rates continue to decline across many countries and regions resulting in significant undercounts. Combined with the legacy of the COVID-19 pandemic which saw many countries postpone their national census, have issues with recruiting field staff, and international funding for population data collection and analysis disappearing, these factors have combined to create a complex and challenging environment within which to design, implement and disseminate a census. According to the UN Economic Commission for Africa (UNECA), undercounts and census delays due to COVID-19, conflict or financial limitations have resulted in an estimated one in three Africans not being counted during the 2020 round of censuses.

Despite this challenging environment, the availability of new data sources and technologies offer National Statistics Offices (NSOs) alternative solutions to the traditional census approach. Digital data collection for the census through tablets, smartphones and other online resources is already making census processes cheaper, more efficient and improve census quality. Recent advances in satellite imagery and Artificial Intelligence (AI) technology have further enhanced capacities in the use of alternative methodologies to complement census data collection efforts. For example, buildings mapped from satellite imagery using AI, together with counts of people obtained at small area unit level, can help to create detailed population estimates that can either support more effective and efficient census planning and implementation, or provide alternative population data in hard to count areas where a census cannot be fully enumerated.

These statistical modelling approaches have been developed to estimate populations based on the correlation between population density and geospatial covariate data. In countries where a census is not possible, these methods can be implemented at low cost (although they require a minimum level of technical capacity) and may provide alternative ways of deriving more recent population estimates to inform census planning and implementation.

Combining geospatial and statistical data will allow NSOs to produce very-high resolution gridded population data (typically 100m or 1km grid cells) with full national coverage. A significant advantage of this approach over more traditional census dissemination is that is that it provides the flexibility to aggregate grid cells to estimate the total population within any defined boundary, but with no additional disclosure risk. There are two broad methodologies that can be applied to produce high-resolution gridded population estimates:

Top-down Method

The top-down method disaggregates known population totals for each coarser administrative unit (e.g. states or local government areas) into finer 100 m gridded population estimates. Such reliable population totals may be obtained from national population and housing census results; thus, the top-down population estimation approach is handy where recent census totals are available. Within this method, gridded population estimates are produced using Machine Learning algorithms (e.g., random forest models) to disaggregate population totals based on relationships between population density and a set of geospatial covariates such as building density, distance to city centre, or nighttime lights intensity. The disaggregation can be applied across the entire country (unconstrained) or limited only to areas where settlements have been mapped (constrained).

The top-down population disaggregation approach allows for both classical and Bayesian statistical inference approaches, thereby, providing users with a means of quantifying uncertainties in parameter estimates. 

Bottom-up Method

The bottom-up method takes geolocated household survey data from a sample of locations and fits it to a statistical model to estimate population sizes for unsampled areas based on relationships with geospatial covariates. This approach applies customised statistical models to make the best use of available model input survey data, and the use of Bayesian statistical inference provides probabilistic estimates of uncertainty in the parameter estimates. This suitable approach requires the availability of a recent geolocated nationally representative household survey dataset or partial census data and can be used to support census preparations and implementation. It is a useful tool for filling census data gaps in countries where recent census data have not been undertaken, or where the existence of hard-to-reach areas meant that some places could not be visited during a recent census thereby leading to incomplete census data. Note that while this approach provides an opportunity to quantify uncertainty with ease, it also requires more detailed and reliable input of population data and requires more time for model development, testing and implementation.

The methodology adopted by an NSO will be determined by the availability of the input data to the model. Where a country has complete census coverage that it wants to try and disaggregate down to a high-resolution (100m or 1km) grid, the top-down methodology will be applied. The top-down approach has been applied in Mali and Burkina Faso previously. Where a country has incomplete or no census data available, the bottom-up method will be applied. This bottom-up approach has been applied previously in multiple countries, including the Demographic Republic of the Congo, Zambia and Nigeria. It should be noted that as the different methods require different inputs, the gridded population estimates that they produce will have different characteristics.

These approaches to estimating population also align with the methods used by the UN Population Division in the production of their World Population Prospects report.  These methods are therefore mature and trusted enough to provide consistent and comparable denominator data where a census cannot be fully enumerated.

Comparison on use of alternative approaches in the non-digital versus digital census era

In the era before widespread digital technology, alternatives were constrained by the physical limitations of paper, logistics, and manual data processing. The goal was primarily to reduce the immense cost and time of a full enumeration. Digital technology, data science, and increased connectivity have unlocked powerful new alternatives. The focus has shifted from simple cost-saving to creating a dynamic, integrated data ecosystem.