Causal forest estimation of irrigation-adoption effects on maize yield under climate stress: an artificial intelligence counterfactual evaluation of smallholder resilience

Smallholder maize systems in the West African savannah face rising heat exposure, yet the yield return to irrigation adoption remains poorly identified because adoption is endogenous to unobserved farm management and because treatment effects are likely heterogeneous across the covariate space. The reanalysis records a mean temperature of 26.414 C, a standard deviation of 0.943 C and a warming trend of 0.1553 C per decade, which establishes the heat-load context within which irrigation adoption is evaluated. The estimator is a generalised random forest that recursively partitions the covariate space by maximising heterogeneity in the local treatment effect, with honesty and subsampling used to obtain asymptotic normality and valid confidence intervals. The paper proves pointwise consistency and asymptotic normality of the local average treatment effect under unconfoundedness, overlap and a Lipschitz response surface, and derives a calibration diagnostic for the forest's out-of-bag residuals. The contribution is a discipline-specific identification strategy and a formal guarantee for counterfactual yield evaluation, not a report of field estimates. The framework permits smallholder advisories to target irrigation investment where the predicted treatment effect is largest, and it clarifies the conditions under which such targeting improves resilience rather than merely redistributing risk.