Income, psychological security, and subjective well-being in urban China: a machine learning analysis with SHAP...

Subjective well-being has become a core indicator for measuring social progress and policy effectiveness. However, the "Easterlin Paradox" remains prevalent, and this paradox refers to the disconnect between economic growth and improvements in well-being. While income, as a traditional economic indicator, exhibits the characteristic of diminishing marginal effects, psychological security and its mechanism of influence on subjective well-being have not been fully explored-here, psychological security is a multidimensional psychological construct. Existing studies mostly adopt linear models, which struggle to capture the nonlinear relationships and interactions among variables, and lack in-depth interpretation of the underlying mechanisms that influence subjective well-being.

Based on a nationwide survey of 1,369 urban residents in China, this study employed machine learning (ML) methods (including 11 algorithms such as LightGBM, XGBoost, and Random Forest) to predict subjective well-being, combined with SHapley additive exPlanations (SHAP) for interpretability analysis. Data preprocessing included missing value imputation, standardization, and SMOTE-Tomek sampling to address class imbalance. Model parameters were optimized through tenfold cross-validation and grid search, with evaluation metrics including AUC, accuracy, recall, F1-store.

The LightGBM model performed best (AUC = 0.854), significantly outperforming traditional linear models. SHAP analysis revealed that general Psychological Security was the core factor influencing subjective well-being, ranking first in importance across all subjective well-being level groups. The impact of income on subjective well-being exhibited a "threshold effect," with middle-income levels having the strongest promotive effect, and psychological security significantly enhanced the marginal effect of income. The formation mechanisms of different subjective well-being levels varied: the high subjective well-being group followed a "psychological resource dominance-structural resource synergy" mechanism; the medium group relied on a "psychological resource compensation-structural resource balance" mechanism; and the low subjective well-being group was trapped in a negative cycle of "dual deficiency in psychological and structural resources."

Psychological security is the most stable and core factor influencing subjective well-being, and the synergistic effect between income and psychological security is significant. Machine learning models outperform traditional methods in capturing complex nonlinear relationships, and SHAP analysis provides an interpretable quantitative basis for understanding the formation mechanisms of subjective well-being. The findings offer targeted recommendations for optimizing well-being enhancement policies, emphasizing the need for differentiated intervention strategies based on the resource-matching characteristics of different subjective well-being groups.