公积金提振家庭消费的微观机制与异质性研究——基于可解释机器学习与因果森林算法

A Study on the Micro-Level Mechanisms and Heterogeneity of How Housing Provident Funds Boost Household Consumption—Based on Explainable Machine Learning and Causal Forest Algorithms

摘要: 本文基于CFPS 2022微观家庭数据，构建“基准预测–机制解析–因果效应估计”的递进建模框架，系统分析公积金潜在流动性与家庭消费之间的关联机制，并在双重机器学习识别假设下估计其平均处理效应。研究首先运用随机森林模型评估公积金在消费预测中的相对重要性，继而通过SHAP可解释机器学习揭示缴存强度与消费的非线性关系及收入交互效应，最后采用双重机器学习框架下的因果森林方法，在控制收入、人口特征与家庭禀赋等混杂因素后，测算公积金对消费的平均处理效应与个体异质性。研究发现：第一，公积金变量对家庭消费具有较强解释力，但明显低于家庭纯收入，表明公积金主要是依托收入渠道发挥作用的流动性补充手段。第二，SHAP分析显示公积金缴存强度与消费存在显著非线性关系，月缴存额在1000~2000元区间对消费的边际拉动提升较快，超过3000元后边际效应逐步放缓。第三，因果森林估计表明，在控制可观测混杂因素且识别假设成立的条件下，模型估计显示公积金月缴存额增加与家庭消费之间存在显著正向关联，对应平均处理效应约为13.84元。年化消费响应系数约为1.15。第四，公积金消费效应存在显著异质性，中高收入、城镇户口和年龄较高家庭的消费响应更强。第五，公积金对文教娱乐等发展型消费的拉动强于食品等生存型消费，有助于推动消费结构升级。

Abstract: Based on micro-household data from CFPS 2022, this paper constructs a progressive modeling framework of “benchmark forecasting-mechanism analysis-causal effect estimation” to systematically analyze the mechanism linking the potential liquidity of housing provident funds to household consumption, and estimates its average treatment effect under the dual machine learning identification hypothesis. The study first employs a random forest model to assess the relative importance of the housing provident fund in consumption forecasting. It then uses SHAP-based explainable machine learning to reveal the nonlinear relationship between contribution intensity and consumption, as well as the interaction effects with income. Finally, using the Causal Forest method within a dual machine learning framework, the study estimates the average treatment effect of the housing provident fund on consumption and individual heterogeneity after controlling for confounding factors such as income, demographic characteristics, and household endowments. The study found the following: First, the housing provident fund variable has strong explanatory power for household consumption, but it is significantly lower than that of household net income, indicating that the housing provident fund primarily functions as a liquidity supplement through the income channel. Second, SHAP analysis shows a significant nonlinear relationship between the intensity of housing provident fund contributions and consumption; the marginal increase in consumption rises rapidly when monthly contributions fall within the 1000~2000 yuan range, but the marginal effect gradually slows after exceeding 3000 yuan. Third, causal forest estimation indicates that, under conditions where observable confounding factors are controlled and identification assumptions hold, the model estimates a significant positive association between an increase in monthly housing provident fund contributions and household consumption, with a corresponding average treatment effect of approximately 13.84 yuan. The annualized consumption response coefficient is approximately 1.15. Fourth, the consumption effect of the housing provident fund exhibits significant heterogeneity, with stronger consumption responses observed among middle- and high-income households, those with urban household registration, and older households. Fifth, the housing provident fund exerts a stronger stimulatory effect on development-oriented consumption—such as culture, education, and entertainment—than on subsistence-oriented consumption—such as food—thereby helping to promote an upgrade in the consumption structure.