The decline of woodland caribou (Rangifer tarandus caribou) has been attributed to anthropogenic landscape disturbances, but critical distance thresholds and time lags between disturbance and extirpation are unknown. Using a database of caribou presence and extirpation for northern Ontario, Canada, geo-coded to 10 × 10-km cells, we constructed logistic regression models to predict caribou extirpation based on distance to the nearest of each of 9 disturbance types: forest cutovers, fires, roads, utility corridors, mines, pits and quarries, lakes, trails, and rail lines. We used Akaike's Information Criterion to select parsimonious models and Receiver-Operating Characteristic curves to derive optimal thresholds. To deal with the effects of spatial autocorrelation on estimates of model significance, we used subsampling and restricted randomizations. Forest cutovers were the best predictor of caribou occupancy, with a tolerance threshold of 13 km to nearest cutover and a time lag of 2 decades between disturbance by cutting and caribou extirpation. Management of woodland caribou should incorporate buffers around habitat and requires long-term monitoring of range occupancy.