LPG Subsidy Reform, Energy Compensation, and Social Risk in Bolivia: A Machine-Learning Agent-Based Microsimulation

This study evaluates alternative designs for reforming Bolivia's liquefied petroleum gas subsidy using a machine-learning agent-based microsimulation framework. The analysis harmonizes household survey data, expenditure data, demographic and health information, and monthly hydrocarbon production and commercialization series to simulate fiscal savings, poverty effects, energy substitution, administrative costs, and social risks under multiple reform scenarios. The model compares uncompensated subsidy removal, fixed energy transfers, full compensation for vulnerable LPG users, voucher-based compensation, maternal-child transfers, clean-energy transition kits, and hybrid policy packages. Machine-learning models are used to learn household vulnerability, fuel-use patterns, food insecurity risk, and behavioral propensities that feed into a monthly agent-based simulation. The results show that eliminating the subsidy without compensation generates the largest fiscal savings but increases poverty, extreme poverty, and pressure toward solid-fuel substitution. Full monetary compensation for Q1-Q2 LPG users substantially reduces social harm while preserving significant fiscal savings and dominates an equivalent voucher design under normal market conditions because of lower administrative friction. The most socially robust design combines targeted monetary energy compensation, maternal-child reinforcement, and clean-energy kits for households using solid fuels. The findings support a gradual replacement of the universal LPG subsidy with targeted, administratively lean, and behaviorally informed compensation mechanisms.