Abstract
Abstract
Climate-smart agriculture (CSA) is an increasingly important climate action in building resilience to extreme weather shocks and enhancing sustainable development goals, such as food security and poverty reduction. Thus, increasing CSA adoption is a top international development priority. However, adoption remains low in developing countries due to resource constraints such as land, labor, and capital. Can machine learning and econometric techniques combine to explain CSA adoption? We respond to this question by applying supervised machine learning techniques (including Decision Tree, Probit, and Random Forest) along with a control function econometric approach to predict CSA adoption in southern Malawi using agroforestry fertilizer trees (Faidherbia albida) as a proxy for CSA. We utilize primary survey data from 808 households in a CSA-related intervention area across southern Malawi, where a bundle of CSA practices was promoted from 2009 to 2014 by Catholic Relief Services in Malawi. Our approach accounts for the endogeneity of CSA program participation and selection bias associated with CSA adoption. We found that machine learning models accurately predicted agroforestry adoption by 64%, 62%, and 61%, respectively, while the econometric model predicted adoption by 33%. Our results suggest that machine learning models can provide superior predictability of CSA adoption compared to standard econometric approaches. The findings provide an important first step in comparing the predictive power of artificial intelligence to enhance CSA adoption and thereby contribute to environmental development outcomes in rural areas of Malawi and similar contexts elsewhere