Understanding Household Welfare in Bhutan: A Regression and Classification Approach to Consumption Expenditure and Poverty

Abstract

Poverty analysis in Bhutan has remained largely descriptive, with limited use of predictive modelling. This study applies a combined regression and classification framework to the Household Consumption and Expenditure Survey 2025 Public Use File, covering 5,883 households across all 20 dzongkhags. Four survey-weighted regression models predict log per-capita food expenditure, interpreted through SHapley Additive exPlanations; weighted XGBoost performs best (cross-validated R ² = 0.705), with total expenditure, household size and a low-expenditure indicator the dominant determinants, consistent with Engel’s law. For poverty classification against the national poverty line (Nu. 6,204 per capita per month), we train five models, including a survey-weighted logistic baseline, on low-cost structural characteristics only, deliberately excluding expenditure-derived features to prevent target leakage. The linear support vector machine and the logistic baseline achieve receiver operating characteristic–area under the curve (ROC-AUC) = 0.960 under repeated stratified cross-validation, and a tuned random forest attains 0.855 recall for the poor class on held-out data. A single expenditure-aggregate feature inflates ROC-AUC to 0.993, illustrating the risk of leakage. Poor households are larger and concentrated in rural and eastern dzongkhags. The results support a proxy means test instrument for social-protection targeting under Bhutan’s 13th Five-Year Plan. JEL Codes C45, C53, D12, I32, I38, O12