Evaluating the accuracy and feasibility of gridded population sampling to overcome bias due to missing populations in household surveys

Abstract

In low- and middle-income countries (LMICs), household survey data are a main source of information for planning, evaluation, and decision-making. LMIC survey methods have not changed substantially in forty years, while population trends and technologies have. This mismatch is likely excluding an increasing number of vulnerable and mobile populations and leading to erosion in LMIC household survey data accuracy. When populations are under-represented in surveys, they are less likely to benefit from funds allocated by governments and international agencies to promote health and social equity. This thesis evaluates the accuracy of using gridded population sample frames as an alternative to census sample frames, and the feasibility of implementing gridded population surveys in a real-world setting.
To compare accuracy of census and gridded population sample frames, a realistic “true” population was simulated in Khomas, Namibia for 2016. The simulated population was derived from Namibia’s 2011 census and 2013 Demographic and Health Survey comprising several household outcome indicators and latitude-longitude coordinates digitised from actual building locations in satellite imagery. Sixteen versions of realistic outdated-inaccurate censuses were simulated by (a) removing households in buildings not present in satellite imagery in 2011, 2006, and 2001, and (b) randomly removing a percentage of rural, urban slum, and urban non-slum households based on a literature review of LMIC census undercounts. For each realistic census, a gridded population sample frame was modelled using WorldPop methodologies. Accuracy was then assessed by drawing repeated samples from the simulated censuses and gridded population datasets, and comparing them to the “true” population. To evaluate feasibility, a gridded population survey was conducted in Kathmandu, Nepal and focus group discussions and observation were made with survey implementers.
In the accuracy assessment, I found that the main source of error in gridded population datasets was not outdated, inaccurate census inputs, but rather, the WorldPop practice of using average population density from aggregated census units to predict population density in finer grid cells. Despite inaccuracies, I found that WorldPop methods disaggregated population in or around settlements, and that the distribution of population estimates in grid cells within urban areas were +/-20% the “true” population distribution. Gridded population sample frames derived from outdated or inaccurate census data produced more accurate survey estimates than the original outdated, inaccurate census EAs frames. In the Kathmandu feasibility assessment, gridded population sampling proved to be cost-effective and feasible, and use of an area-microcensus design identified more vulnerable and mobile households than a comparable two-stage sample.

More information

Identifiers

ORCID for Dana Thomson: orcid.org/0000-0002-9507-9123

ORCID for Nikolaos Tzavidis: orcid.org/0000-0002-8413-8095

Catalogue record

Export record