Exploring the complementarity between traditional econometric methods and machine learning – an application to adoption and disadoption of conservation practices

This study explores the potential complementarity between traditional econometric methods and machine learning in analysing the adoption and disadoption of a key conservation practice in agriculture, cover crops. While the adoption of conservation practices has been widely examined, the literature on their disadoption is limited. Using a unique longitudinal panel survey of Iowa farmers, we compare logistic regression models with a Random Forest algorithm to examine factors driving conservation adoption and disadoption. Our findings show that while traditional logistic regression models offer interpretability grounded in economic theory, Random Forest provides superior predictive power and reveals complex, non-linear relationships among key factors such as past adoption behaviour, cost-share participation, and farmer perceptions. SHAP (SHapley Additive exPlanations) analysis identifies adoption scale, past adoption behaviour, and environmental factors as primary drivers of disadoption. Farmers with larger previous cover crop acreage and consistent adoption history are significantly less likely to disadopt, while the long-term impact of cost-share programmes on continued use appears limited. By combining machine learning’s predictive power with the interpretability of traditional econometrics, the study provides a deeper understanding of the drivers behind conservation decisions, which are crucial for informing policy design that promotes more sustainable adoption of conservation practices.