Rare Potential Poor Household Identification With a Focus Embedded Logistic Regression

Abstract

With the rapid development of poverty alleviation in China, multidimensional poverty identification has always been challenging. This paper adopted a focus embedded logistic regression (FeLR) to solve two types of difficulties–the rarity and hard-distinguishability, of the potential poor household (PPH) identification. The PPH identification was decomposed into two subproblems–the potential re-poverty household (PRPH) identification, and the potential unidentified poor household (PUPH) identification. The FeLR embedded a focal loss to deal with the hard-distinguishability, and adopted a weighting technique to address the rarity. The sample weight exponent was extended to negative values to overlook the hard negative samples. This setting significantly improved the recall of PPHs, compared with that using traditional logistic regression. A few indicators were critical to the incidence of PPH, especially the household income per capita, medical expenses for chronic diseases, and house structure. Local policy makers are suggested to pay more attention to the crucial indicators to against the poverty contrapuntally.

Keywords

• Focal loss
• logistic regression
• potential poor households
• rare events