The Typhoid Fever Surveillance in Africa Program: Geospatial Sampling Frames for Household-based Studies: Lessons Learned From a Multicountry Surveillance Network in Senegal, South Africa, and Sudan

Abstract

Background: Robust household sampling, commonly applied for population-based investigations, requires sampling frames or household lists to minimize selection bias. We have applied Google Earth Pro satellite imagery to constitute structure-based sampling frames at sites in Pikine, Senegal; Pietermaritzburg, South Africa; and Wad-Medani, Sudan. Here we present our experiences in using this approach and findings from assessing its applicability by determining positional accuracy.

Methods: Printouts of satellite imagery combined with Global Positioning System receivers were used to locate and to verify the locations of sample structures (simple random selection; weighted-stratified sampling). Positional accuracy was assessed by study site and administrative subareas by calculating normalized distances (meters) between coordinates taken from the sampling frame and on the ground using receivers. A higher accuracy in conjunction with smaller distances was assumed. Kruskal-Wallis and Dunn multiple pairwise comparisons were performed to evaluate positional accuracy by setting and by individual surveyor in Pietermaritzburg.

Results: The median normalized distances and interquartile ranges were 0.05 and 0.03-0.08 in Pikine, 0.09 and 0.05-0.19 in Pietermaritzburg, and 0.05 and 0.00-0.10 in Wad-Medani, respectively. Root mean square errors were 0.08 in Pikine, 0.42 in Pietermaritzburg, and 0.17 in Wad-Medani. Kruskal-Wallis and Dunn comparisons indicated significant differences by low- and high-density setting and interviewers who performed the presented approach with high accuracy compared to interviewers with poor accuracy.

Conclusions: The geospatial approach presented minimizes systematic errors and increases robustness and representativeness of a sample. However, the findings imply that this approach may not be applicable at all sites and settings; its success also depends on skills of surveyors working with aerial data. Methodological modifications are required, especially for resource-challenged sites that may be affected by constraints in data availability and area size.

Keywords: geospatial sampling frame; positional accuracy; satellite imagery; sub-Saharan Africa.

Figure 1.

Sampling frame of the study area and the administrative subunits (AdSubs) in Pikine, Senegal. Different colors depict the structures belonging to each AdSub. Illustration top left: enlarged illustration of enumerated structures for subunits 2, 3, and 6 (blue highlighted rectangle in main figure).

Figure 2.

Weighted-stratified random sampling of structures in Pikine, Senegal. ¹Selected structures (N₀) as per sample size calculation for the total survey area and each administrative subunit (AdSub) (flagged black) and replacement structures for the total survey area and each AdSub (flagged white). ²Selected structures (N₀) for the total survey area and each AdSub (flagged black). ³Replacement structures for the total survey area and each AdSub (flagged white). ⁴Identifiers (6–250, 6–304, 6–311) and the geographic coordinates (6–250: N14°44.702′/ W17°23.408′; 6–304: N14°44.632′/W17°23.284′; 6–311: N14°44.708′/W17°23.289′) obtained from Google Earth Pro.

Figure 3.

Normalized distances by administrative subunit (AdSub) in Pikine (Senegal), Pietermaritzburg (South Africa), and Wad-Medani (Sudan). Each individual box plot shows the range of normalized distances indicated as vertical line; bottom whisker (minimum normalized distance to first quartile; non-outlier), first quartile (25% of normalized distances/25th percentile), second quartile or median (50% of normalized distances/50th percentile), third quartile (75% of normalized distances/75th percentile), top whisker (third quartile to maximum normalized distance; non-outlier), and outliers plotted as circles. Senegal: The root mean square error (RMSE) of normalized distances by AdSub was 0.04, 0.06, 0.06, 0.06, 0.07, and 0.13 (ascending order). South Africa: The RMSE of normalized distances by AdSub was 0.32, 0.21, 0.40, 0.63, 0.29, 0.25, 0.32, 0.16, 0.32, 0.98, 0.56, 0.69, 0.28, 0.35, 0.38, 0.22, 0.10, 0.41, 0.35, 0.22, 0.54, and 0.31 (ascending order). Sudan: The RMSE of normalized distances by AdSub was 0.08, 0.11, 0.22, 0.10, 0.06, 0.43, 0.11, 0.06, 0.07, and 0.09 (ascending order).

Figure 4.

Normalized distances (meters) categorized into quintiles and graded accordingly by administrative subunit (AdSub) of each site. Each bar shows the frequency of normalized distances categorized into quintiles by AdSub and graded correspondingly as very good (lowest quintile), good, fair, poor, and very poor (highest quintile). Senegal: very good, 19.4%; good, 33.6%; fair, 36.8%; poor, 7.3%; and very poor, 2.9%. South Africa: very good, 20.4%; good, 14.6%; fair, 12.9%; poor, 22.2%; and very poor, 29.9%. Sudan: very good, 28.3%; good, 16.1%; fair, 27.7%; poor, 25.0%; and very poor, 2.9%.