Mission Impossible? Exploring the Promise of Multiple Imputation for Predicting Missing GPS-Based Land Area Measures in Household Surveys

Research has provided robust evidence for the use of GPS as the new, scalable gold-standard in land area measurement in household surveys. Nonetheless, facing budget constraints, survey agencies often measure with GPS only plots within a given radius of dwelling locations. It is, subsequently, common for significant shares of plots not to be measured, and research has highlighted the selection biases resulting from using incomplete data. This study relies on nationally-representative, multi-topic household survey data from Malawi and Ethiopia with near-negligible missingness in GPS-based plot areas to validate the accuracy of a Multiple Imputation (MI) model for predicting missing GPS-based plot areas in household surveys. The analysis randomly creates missingness among plots beyond two operationally-relevant distance measures from the dwelling locations, conducts MI for each artificially-created dataset, and compares the distributions of the imputed plot-level outcomes, namely area and agricultural productivity, with the distributions of their true, observed counterparts. MI procedure results in imputed yields that are statistically undistinguishable from the true distributions with up to 82% and 56% missingness, respectively for Malawi and Ethiopia, for plots located more than 1 kilometer away from dwellings. The study highlights the promise of using MI for reliably predicting missing GPS-based plot areas. Acknowledgement : The authors thank Tomoki Fujii and Alberto Zezza, Heather Moylan for their comments on the earlier versions of this paper.