Causal Machine Learning Reveals Divergent Effectiveness of Governance Modes for Sustainable Agricultural Transformation

Agrifood structural transformation underpins progress toward Sustainable Development Goals, yet whether the state should withdraw, deregulate, or inject this transition remains contested. We evaluate three governance modes across over 2,700 Chinese counties and two decades, ap- plying Causal Forest with double/debiased machine learning to three policy reforms—the 2006 agricultural tax abolition (withdraw), the 2016 supply-side reform (deregulate), and the 2014 targeted poverty alleviation campaign (inject). Governance design, not fiscal magnitude, deter- mines effectiveness, and each mode’s impact is conditional on local institutional context: with- drawal accelerates the shift out of agriculture where pre-reform distortions bind most tightly; deregulation diversifies cropping structures; yet injection produces no significant average ef- fect, masking offsetting heterogeneity along fiscal-capacity lines. Data-driven targeting robustly outperforms administrative allocation in out-of-sample validation while disproportionately re- taining the poorest counties. Sustainable agricultural transitions thus depend on the political economy of policy execution, not on spending magnitude.