This dataset contains two predictive lek (breeding site) persistence raster layers covering the U.S. greater sage-grouse distribution. In the United States, locations where males display and breed with females (i.e., leks) are often monitored annually by state wildlife agencies, providing valuable information on the persistence of birds in the surrounding areas. A U.S. range-wide lek database was recently compiled for greater sage-Grouse (O’Donnell et al. 2021), providing a standardized source of information to build statistical models to evaluate environmental characteristics associated with lek persistence. The compiled lek database classified a subset of leks as being either active (leks currently used for breeding activities) or inactive (leks no longer used for breeding activities) based on count data collected over a 20-year monitoring period. We fit the outcome of a lek being active or inactive as a function of environmental predictors characterizing surrounding conditions in a logistic regression model. Covariates included sagebrush cover, pinyon-juniper cover, topography, precipitation, point and line disturbance densities, and landscape configuration metrics. We included the Bureau of Land Management habitat assessment areas (termed mid-scales) as regional random effects in the form of random intercepts and random slopes (for a subset of covariates). The final model included 13 covariates. We predicted conditional probabilities of lek persistence across the U.S. occupied range using the covariate layers and regional mid-scales, which we make available here as a 30-meter resolution continuous raster dataset. The predictions were conditional because they were specific to each mid-scale factor level (i.e., pixel predictions were influenced by the regional mid-scale polygon they fell within via the associated mid-scale intercept and random slope deviations). We applied sensitivity thresholds (capturing percentage of leks correctly classified as active) to the continuous probability layer to bin persistence probabilities into high, medium, low, and marginal areas of persistence, which we make available here as a 30-m categorical raster dataset.