The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent.

Background: Household surveys are the main source of demographic, health and socioeconomic data in low-and middle-income countries (LMICs). To conduct such a survey, census population information mapped into enumeration areas (EAs) typically serves a sampling frame from which to generate a random sample. However, the use of census information to generate this sample frame can be problematic as in many LMIC contexts, such data are often outdated or incomplete, potentially introducing coverage issues into the sample frame. Increasingly, where census data are outdated or unavailable, modelled population datasets in the gridded form are being used to create household survey sampling frames. Methods: Previously this process was done by either sampling from a set of the uniform grid cells (UGC) which are then manually subdivided to achieve the desired population size, or by sampling very small grid cells then aggregating cells into larger units to achieve a minimum population per survey cluster. The former approach is time and resource-intensive as well as results in substantial heterogeneity in the output sampling units, while the latter can complicate the calculation of unbiased sampling weights. Using the context of Somalia, which has not had a full census since 1987, we implemented a quadtree algorithm for the first time to create a population sampling frame. The approach uses gridded population estimates and it is based on the idea of a quadtree decomposition in which an area successively subdivided into four equal size quadrants, until the content of each quadrant is homogenous. Results: The quadtree approach used here produced much more homogeneous sampling units than the UGC (1 × 1 km and 3 × 3 km) approach. At the national and prewar regional scale, the standard deviation and coefficient of variation, as indications of homogeneity, were calculated for the output sampling units using quadtree and UGC 1 × 1 km and 3 × 3 km approaches to create the sampling frame and the results showed outstanding performance for quadtree approach. Conclusion: Our approach reduces the manual burden of manually subdividing UGC into highly populated areas, while allowing for correct calculation of sampling weights. The algorithm produces a relatively homogenous population counts within the sampling units, reducing the variation in the weights and improving the precision of the resulting estimates. Furthermore, a protocol of creating approximately equal-sized blocks and using tablets for randomized

Enumeration Areas (EAs) are the operational geographic units for the collection and dissemination of census data and are often used as a national sampling frame for various types of surveys. In many poor or conflict-affected countries, EA demarcations are incomplete, outdated, or missing. Even for countries that are stable and prosperous, creating and updating EAs is one of the most challenging yet essential tasks in the preparation for a national census. Commonly, EAs are created by manually digitising small geographic units on high-resolution satellite imagery or physically walking the boundaries of units, both of which are highly time, cost, and labour intensive. In addition, creating EAs requires considering population and area size within each unit. This is an optimisation problem that can best be solved by a computer. Here, for the first time, we produce a semi-automatic mapping of pre-defined census EAs based on high-resolution gridded population and settlement datasets and using publicly available natural and administrative boundaries. We demonstrate the approach in generating rural EAs for Somalia where such mapping is not existent. In addition, we compare our automated approach against manually digitised EAs created in urban areas of Mogadishu and Hargeysa. Our semi-automatically generated EAs are consistent with standard EAs, including having identifiable boundaries for field teams to follow on the ground, and appropriate sizing and population for coverage by an enumerator. Furthermore, our semi-automated urban EAs have no gaps, in contrast, to manually drawn urban EAs. Our work shows the time, labour and cost-saving value of automated EA delineation and points to the potential for broadly available tools suitable for low-income and data-poor settings but applicable to potentially wider contexts.