Counterfactual-Yield Inference for Computer Agriculture Seed-Variety Recommendations Under Unobserved Farmer Adoption: An Artificial Intelligence Instrumental-Variable Bias Bound

Seed-variety recommendation systems in computer agriculture must estimate counterfactual yield under farmer adoption that is observed only selectively, so that the recommended variety is confounded with the management, soil and climatic conditions under which farmers choose it. The bound is derived from a linear potential-outcomes model with an additive adoption equation, and it is expressed as a function of the first-stage strength, the exclusion restriction and the residual variance of the yield equation. A replication-based evaluation of the ['Intercept', 'exposure', 'group', 'severity'] estimator was conducted: 600 data sets of N = 500 were drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2), and the ordinary least squares estimator was characterised under the stated process. Across the 600 replications the estimated bias for the intercept was 0.05399 and for exposure -0.00128; the root mean squared errors were 0.88516 and 0.01625 respectively; and the empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. The bound implies that a recommendation rule which ignores adoption selection can reverse the sign of the yield contrast when the first-stage F statistic falls below a threshold that the paper derives. The contribution is a computable diagnostic that a seed-recommendation system can run before deploying a variety ranking, together with a statement of the conditions under which the ranking is not identified from the available data.