Potential for energy poverty reduction by error decomposition with machine learning

Abstract

Energy poverty remains a pressing global challenge, with approximately 685 million people lacking electricity access and 2.1 billion without clean cooking fuels. Traditional metrics often fail to capture the multidimensional nature of energy deprivation, prompting the adoption of frameworks like the World Bank’s Multi-Tier Framework (MTF). However, existing approaches overlook the concept of potential to identify where and how energy poverty reduction efforts can be most effective. This study bridges this gap with an error decomposition framework that analyzes whether households-level or regional-level interventions should be prioritized. Using machine learning’s (ML) XGBoost, we develop a predictive model of multidimensional energy poverty for Nepal, Myanmar, and Cambodia that helps avoid misspecification problems and outperforms traditional econometric methods, achieving test accuracies of up to 0.78 when incorporating spatial fixed effects. The error decomposition reveals systematic underperformance in certain regions and demographic groups, highlighting latent opportunities for policy intervention. Key findings indicate that energy poverty is shaped by both household-level characteristics and systemic regional factors, with urban-rural and ethnic disparities playing significant roles. In Nepal, marginalized ethnic groups exhibit persistent energy deprivation despite high socioeconomic status, while Myanmar’s urban areas suffer from unreliable supply despite high connection rates. Cambodia’s rural households remain underserved, emphasizing the need for decentralized energy solutions. By distinguishing between reducible and irreducible error components, our framework provides actionable insights for targeted policy interventions, advancing progress toward Sustainable Development Goal 7 (SDG 7).