Drylands comprise 40% of Earth's land mass and are critical to food security, carbon sequestration, and threatened and endangered wildlife. Exotic weed invasions, overgrazing, energy extraction, and other factors have degraded many drylands, and this has placed an increased emphasis on dryland restoration. The increased restoration focus has generated a wealth of experience, innovations and empirical data, yet the goal of restoring diverse, native, dryland plant assemblages composed of grasses, forbs, and shrubs has generally proven beyond reach. Of particular concern are shrubs, which often fail to establish or establish at trivially low densities. We used data from two Great Plains, USA coal mines to explore factors regulating shrub establishment. Our predictor data related to weather and restoration (e.g., seed rates, rock cover) variables, and our response data described shrub abundances on fields of the mines. We found that seeded non-shrubs, especially grasses, formed an important competitive barrier to shrub establishment: With every one standard deviation increase in non-shrub seed rate, the probability shrubs were present decreased ~0.1 and shrub cover decreased ~35%. Since new fields were seeded almost every year for > 20 years, the data also provided a unique opportunity to explore effects of stochastic drivers (i.e., precipitation, year effects). With every one standard deviation increase in precipitation the first growing season following seeding, the probability shrubs were present decreased ~0.07 and shrub cover decreased ~47%. High precipitation appeared to harm shrubs by increasing grass growth/competition. Also, weak evidence suggested shrub establishment was better in rockier fields where grass abundance/competition was lower. Multiple lines of evidence suggest reducing grass seed rates below levels typically used in Great Plains restoration would benefit shrubs without substantially impacting grass stand development over the long term. We used Bayesian statistics to estimate effects of seed rates and other restoration predictors probabilistically to allow knowledge of the predictors' effects to be refined through time in an adaptive management framework. We believe this framework could improve restoration planning in a variety of systems where restoration outcomes remain highly uncertain and ongoing restoration efforts are continually providing new data of value for reducing the uncertainty.