Sampling bias in baseline endemicity classification surveys for lymphatic filariasis in sub-Saharan Africa

Abstract

Mapping surveys were conducted prior to the scale up of national programs to eliminate lymphatic filariasis in order to define populations-at-risk. Usually, relatively few locations were surveyed for this process and some sites were purposively sampled from areas with high LF transmission risk. This survey design potentially led to sampling bias in mapping data that, when used in secondary analyses, may lack representativeness. We present the first quantitative assessment of sampling bias in LF mapping surveys across multiple countries. First, a list of 17 socio-ecological covariates commonly used in secondary analyses was compiled. Then, simulations were performed to determine covariate distributions characteristic of a simulated set of randomly selected communities. These were compared to distributions from observed mapping survey locations to test for evidence of systematic sampling bias. Finally, using country mapping datasets, the probability a location was sampled was modeled as a function of socio-ecological covariates using boosted regression trees. We found statistically significant covariate pattern differences potentially indicative of sampling bias in Benin, Cote d’Ivoire, Ghana, and Togo. Non-random differences in sampling along these commonly used covariates suggest these locations may not be representative of inhabited areas in these countries and use of these data in secondary analyses should explore the potential impact of bias.