Variable definition and transformation.

Understanding how local fiscal spending shapes rural income is central to China’s rural revitalisation strategy. Using panel data from 17 prefecture-level cities in Henan Province for 2010–2023, this study investigates the non-linear and heterogeneous effects of fiscal expenditure on farmers’ per-capita income. A hybrid econometric–machine learning framework is developed, combining city–year fixed effects with a residual XGBoost learner and SHapley Additive exPlanations (SHAP) to capture both the linear baseline relationships and complex non-linear interactions among fiscal items while retaining interpretability. Model evaluation based on repeated nested cross-validation and 500 permutation tests demonstrates high predictive reliability (out-of-sample R2 = 0.924, p = 0.002). SHAP-based analysis reveals that healthcare and education spending are the dominant determinants of rural income, jointly accounting for over 40% of model influence. Partial-dependence plots uncover clear threshold effects: education and health expenditures exhibit inverted-U shapes with turning points at approximately ¥1,800 and ¥1,050 per rural resident (2015 prices), respectively. Infrastructure investment shows consistently positive but diminishing returns, while social-security transfers produce concave yet non-negative effects. Heterogeneity analysis further indicates that low-capacity cities derive greater benefits from technology and transport spending, whereas high-capacity cities gain more from health and urbanisation budgets. Robustness tests across alternative learners (Random Forest, LightGBM) and variable definitions confirm the stability of these non-linear thresholds. The results highlight the importance of optimising fiscal composition rather than merely increasing total spending, suggesting that city-specific expenditure ceilings—particularly for education and health—could raise rural incomes by 2–3% while enhancing fiscal efficiency.

厘清地方财政支出如何影响农村居民收入，是中国乡村振兴战略的核心议题。本研究采用河南省17个地级市2010—2023年的面板数据，考察财政支出对农民人均收入的非线性与异质性影响。本研究构建了计量-机器学习混合分析框架，将城市-年份固定效应与残差XGBoost学习器、SHapley可加解释（SHAP）相结合，既能够捕捉财政支出项目间的线性基准关系与复杂非线性交互作用，同时保留模型的可解释性。基于重复嵌套交叉验证与500次置换检验的模型评估结果显示，该模型具备优异的预测可靠性（样本外R²=0.924，p=0.002）。基于SHAP的分析表明，医疗卫生与教育支出是影响农村居民收入的核心驱动因素，二者合计占据模型解释力的40%以上。偏依赖图揭示了显著的门槛效应：教育与医疗卫生支出均呈现倒U型关系，其拐点分别约为每农村居民1800元与1050元（2015年不变价）。基础设施投资始终呈现正向影响，但边际收益递减，而社会保障转移支付则呈现凹性且非负的影响效果。异质性分析进一步表明，财政实力较弱的地级市可从科技与交通支出中获取更多收益，而财政实力较强的地级市则更多受益于医疗卫生与城镇化相关预算。通过更换学习器（随机森林、LightGBM）与变量定义方式开展的稳健性检验，证实了上述非线性门槛效应的稳定性。本研究结果凸显了优化财政支出结构而非单纯扩大总支出规模的重要性，表明针对各地级市制定差异化的支出上限（尤其是教育与医疗卫生领域），可在提升财政效率的同时，推动农村居民收入增长2%至3%。