Causal Effects of Climate-Smart Agriculture Adoption on Smallholder Welfare and Food Security in Zambia: Evidence from Endogenous Switching Regression, Double Machine Learning, and Bayesian Estimation

Abstract Climate-smart agriculture (CSA) is promoted across sub-Saharan Africa to raise smallholder productivity, resilience, and welfare, with food security the ultimate objective, yet rigorous causal evidence for Zambia remains limited. Using 2,116 smallholder households from the Water and Soil Accelerator (WASA) survey across four provinces, this study estimates the causal effects of CSA adoption on agricultural income and crop yield, treating both as proximate determinants of household food access and availability. Three complementary estimators are applied. Multinomial Endogenous Switching Regression (MESR) models selection into four ordered adoption regimes (T0 to T3) and yields stratum-specific effects on the treated (ATTs). Double Machine Learning (DML) estimates population-average effects through Neyman-orthogonal, cross-fitted regression with LASSO and Random Forest learners. A Bayesian DML extension corroborates these, with the posterior probability of a positive effect at 1.000 for all outcomes and posterior and frequentist means differing by at most 0.041. A Stable Unit Treatment Value Assumption (SUTVA) test confirms spillover below the 10 percent materiality threshold. DML places the income premium at approximately 33 percent and the yield premium at approximately 47 percent for the smallholder population, both significant at the 1 percent level. MESR traces a consistent positive gradient, proportionally larger for subsistence than for commercial farmers and falling at the highest adoption intensity; after a full two-stage bootstrap, only the subsistence yield gains at dual- and multi-practice intensity reach individual significance, so identification rests on DML and BDML. The results support targeted CSA promotion for subsistence households and caution against universal multi-practice mandates.