Benchmarking Evaluation of Causal-Transportability Estimators for Computer Agriculture Yield-Gap Prediction Under Unmeasured Confounding: An Artificial Intelligence Bias-Bound Audit for Fertilizer-Subsidy Allocation

Causal-transportability estimators are increasingly proposed for yield-gap prediction in computer agriculture, yet their finite-sample behaviour under unmeasured confounding remains poorly characterised for fertilizer-subsidy allocation. Six hundred data sets of N = 500 were drawn with seed 916459064, and estimator bias, root-mean-square error and nominal 95% confidence-interval coverage were computed for each coefficient. The intercept exhibited estimated bias of -0.03321 with RMSE 0.85764, while the exposure coefficient exhibited estimated bias of 0.00054 with RMSE 0.01595; empirical coverage of the nominal 95% intervals was 1.0 for both the intercept and the exposure coefficient. The results indicate that under the stated process the estimator is approximately unbiased for the exposure coefficient and that nominal coverage is achieved, but the intercept bias and the wide RMSE for the intercept reveal that transportability of the baseline yield level is the binding constraint for subsidy targeting. The paper derives a bias bound for the exposure coefficient under unmeasured confounding and shows that the bound is not identified from the observed data alone, which implies that benchmarking audits of causal-transportability estimators must report coverage alongside bias to avoid overstating transportability.