Accurate predictions of ecological restoration outcomes are needed across the increasingly large landscapes requiring treatment following disturbances. However, observational studies often fail to account for nonrandom treatment application, which can result in invalid inference. Examining a spatiotemporally extensive management treatment-- post-fire seeding of declining sagebrush shrubs across the semiarid U.S. over two decades -- we quantify drivers and consequences of selection biases in restoration, using remotely sensed data. Treatments were disproportionately applied in more stressful, degraded ecological conditions. Failure to incorporate nonrandom treatment allocation led to the conclusion that costly, widespread seedings were unsuccessful; however, after considering biases, restoration positively affected sagebrush recovery. Treatment effect sizes varied with climate, indicating possible prioritization criteria for interventions. Our findings revise the perspective that widespread post-fire sagebrush seedings have been broadly “unsuccessful” and demonstrate how selection biases can pose substantive inferential hazards in observational studies of restoration efficacy and development of restoration theory.