Assessing Agricultural Support Policy Impact through Traditional versus Machine-Learning Techniques

This paper analyses the impact of agricultural subsidies using traditional quasi-experimental research design that combines genetic matching procedure with regression analysis and causal forests, an adaptation of the random forest algorithm of Breiman (2001) for treatment effect estimation. The study is based on a structured orchard farm survey conducted in Albania, an EU candidate country. By employing both traditional and machine learning methods, the comparative methodological approach represents a notable contribution by enhancing the robustness of the findings, while highlighting the advantages of the random forest algorithm. The research results indicate that policy support significantly increased on-farm investments by approximately 4.7 million ALL (representing a 39% increase relative to the sample average investment for the analysed period), and direct apple revenues by about 2.48 million ALL (a 29% increase relative to the sample average revenue). Moreover, the policy had a substantial impact on altering the variety cultivation structure - beneficiary farmers replaced lower-quality apple varieties with higher-quality, market-demanded varieties, while non-beneficiaries showed no significant changes in their variety structure. As a result, the policy support enabled beneficiary farmers to better align their production structure with market demand, potentially boosting their future competitiveness.