The DHS Program

The DHS Program Modeled Map Surfaces: Understanding the Utility of Spatial Interpolation for Generating Indicators at Subnational Administrative Levels (English) 

Document Type

Spatial Analysis Reports

Publication Topic(s)

Geographic Information

Language

English

Recommended Citation

Gething, Peter W. and Clara R. Burgert-Brucker. 2017. The DHS Program Modeled Map Surfaces: Understanding the Utility of Spatial Interpolation for Generating Indicators at Subnational Administrative Levels. DHS Spatial Analysis Reports No. 15. Rockville, Maryland, USA: ICF.

Download Citation

RIS format / Text format / Endnote format

Publication Date

August 2017

Publication ID

SAR15

Abstract:

As interest in describing subnational variation in demographic and health indicators grows, The DHS Program has commenced a plan of work to enable the routine creation and dissemination of spatially modeled indicator surfaces for a set of key indicators. This will be done based on model-based geostatistical (MBG) techniques. To further understand the performance of these models and the factors influencing that performance, this study has investigated (1) what common factors drive MBG modeled surface accuracy and (2) how accurately MBG models can predict aggregated estimates at a finer spatial scale than the first subnational administrative unit level (SNU1). To explore factors influencing model performance, 15 DHS indicators were modeled across 16 countries. Certain indicators performed consistently well across all countries, and others consistently less well. The amount of cluster-level variation in each indicator and the extent to which that variation was spatially autocorrelated were the two most important factors in determining the accuracy of modeled surfaces. It was possible to predict how accurately a given indicator could be mapped based on simple attributes of the raw input data. To explore the ability of MBG models to provide accurate indicator estimates below the SNU1 level, experiments were conducted using the 2014 Kenya DHS, which is unique in having been sampled with approximately four times greater density of clusters than a standard DHS. Analyses based on artificially thinned versions of this data demonstrated that MBG surfaces and aggregated estimates performed progressively better as survey sample size was increased. We found that the use of a geostatistical model to estimate aggregated indicator estimates tends to yield more precise estimates than the default approach of directly calculating weighted means of survey data. On average across the indicators and performance metrics, the use of a geostatistical model based on a standard DHS survey to estimate indicators at “standard” SNU2 level (i.e., at a level of geographical aggregation equivalent to SNU2 in most countries) yields results of accuracy equivalent to a survey three times larger in the absence of a geostatistical model. The integration of geospatial methods in the survey design and subsequent data analysis stages should be considered for future DHS surveys, especially if a desired outcome is to provide precise estimates below the SNU1 level. This integration will provide more precise estimates below the SNU1 level and do so with fewer requirements for large sample sizes.