Regression-based multisource conditional domain adaptation for policy outcome prediction

Many practical problems in the social sciences require the prediction of policy outcomes to support decision-making. However, existing research still lacks suitable methods for forecasting unknown policy outcomes, and theoretical guarantees of the prediction error bounds after policy implementation are lacking. Moreover, prior studies have primarily emphasized classification tasks, leaving regression problems underexplored. To bridge these gaps, we propose a conditional unsupervised domain adaptation (CUDA) framework. Within an agnostic PAC learning setting, our analysis allows us to derive generalization error bounds for the target model, thereby establishing prediction error guarantees for regression tasks in the postpolicy period and ensuring both the reliability and stability of predictions. Specifically, given pretreatment outcomes in the target domain, our framework jointly minimizes our empirical risk function on the basis of pretreatment outcomes and our empirical error function of domain adaptation in the posttreatment period, producing estimators with strong generalizability. Our theoretical analysis further reveals two key challenges: (i) the distribution shift between the source and target domains after policy implementation and (ii) the unlimited number of regression labels due to the continuous label space, which complicates unsupervised domain adaptation for regression tasks. Motivated by these insights, we develop an adversarial learning algorithm titled the conditional unsupervised multisource domain adversarial network (CUMDAN) to address unsupervised domain adaptation in regression for policy outcome prediction. Finally, we conduct experiments on four real-world datasets concerning judicial reforms in China to predict their effects on firms and capital market development. Comparative analyses confirm the effectiveness of the proposed method.