Comparison of machine learning predictions of subjective poverty in rural China

Purpose - Despite rising incomes and reduction of extreme poverty, the feeling of being poor remains widespread. Support programs can improve well-being, but they first require identifying who are the households that judge their income is insufficient to meet their basic needs, and what factors are associated with subjective poverty. Design/methodology/approach - Households report the income level they judge is sufficient to make ends meet. Then, they are classified as being subjectively poor if their own monetary income is inferior to the level they indicated. Second, the study compares the performance of three machine learning algorithms, the random forest, support vector machines and least absolute shrinkage and selection operator (LASSO) regression, applied to a set of socioeconomic variables to predict subjective poverty status. Findings - The random forest generates 85.29% of correct predictions using a range of income and non-income predictors, closely followed by the other two techniques. For the middle-income group, the LASSO regression outperforms random forest. Subjective poverty is mostly associated with monetary income for low-income households. However, a combination of low income, low endowment (land, consumption assets) and unusual large expenditure (medical, gifts) constitutes the key predictors of feeling poor for the middle-income households. Practical implications - To reduce the feeling of poverty, policy intervention should continue to focus on increasing incomes. However, improvements in nonincome domains such as health expenditure, education and family demographics can also relieve the feeling of income inadequacy. Methodologically, better performance of either algorithm depends on the data at hand. Originality/value - For the first time, the authors show that prediction techniques are reliable to identify subjective poverty prevalence, with example from rural China. The analysis offers specific attention to the modest-income households, who may feel poor but not be identified as such by objective poverty lines, and is relevant when policy-makers seek to address the “next step” after ending extreme poverty. Prediction performance and mechanisms for three machine learning algorithms are compared.