Counterfactual crop-yield attribution to seasonal forecast advisories via doubly robust estimation: an artificial intelligence evaluation of adoption-selection bias in extension targeting

Abstract

Seasonal forecast advisories are promoted as a climate service that raises crop yields, yet extension agencies typically target farmers who already differ from non-adopters in ways that also affect yield. the present analysis formalises the resulting adoption-selection bias and evaluates a doubly robust estimator for counterfactual yield attribution under a declared data-generating process. The design is a replication-based evaluation: 600 data sets of N = 500 were drawn from a process with intercept 12.0, exposure 0.42, group -3.1, severity 1.35 and error SD 4.2, and the ordinary least squares estimator was characterised across replications. The estimated bias was -0.04629 for the intercept and 0.00086 for the exposure coefficient; the root mean squared errors were 0.95661 and 0.01664 respectively, and empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. The results indicate that under correct specification the estimator recovers the exposure effect with negligible bias and conservative coverage, but that the intercept is the fragile quantity when group composition is unbalanced. Doubly robust attribution is defensible for advisory evaluation only when the selection model is reported alongside the outcome model, and it derives the conditions under which extension targeting can be interpreted as a counterfactual yield effect rather than a description of adopters.