Resource Loss, Loneliness, and Depressive Symptoms Among Middle-Aged and Older Adults During COVID-19: A Moderated Mediation Analysis and LLM-Based Computational Replication

Public health emergencies impose lasting mental health burdens worldwide. Conservation of resources (COR) theory suggests resource loss increases depressive symptoms, yet how multiple resources relate to depressive symptoms remains unclear. Using 8485 respondents from the 2020 China Health and Retirement Longitudinal Study (CHARLS), we estimated a COR-guided moderated mediation model with robust OLS and bootstrap resampling. The 2018 data supplied baseline symptom information for agent initialization. Loneliness was examined as an indicator of relational deprivation and a statistical mediator. We also evaluated 300 DeepSeek-R1 virtual agents calibrated to the real data distribution to test whether large language model (LLM) agents could reproduce these associations. Higher work-time-change scores were associated with more depressive symptoms (B = 0.415, p < 0.001), whereas greater ease in making ends meet (B = −1.216, p < 0.001) was associated with fewer depressive symptoms. Loneliness was positively associated with depressive symptoms (B = 2.913, p < 0.001). Its interaction with intergenerational satisfaction was negative (B = −0.175, p = 0.015), and bootstrap intervals supported indirect associations that varied with relationship satisfaction. The agents reproduced several broad directions of association, but their score distribution and coefficient magnitudes differed from the survey data. These findings support a COR-guided statistical account of resource conditions during COVID-19 and identify the current limits of LLM-based computational replication.