Accounting for spatial scale is essential for understanding habitat selection, but few studies have used spatial statistics to reveal the characteristic scale at which organisms respond to their environment. We studied habitat selection by GPS-tracked red deer ( Cervus elaphus L., 1758) in the Pyrenees Mountains, France, by applying a geostatistical model that compares autocorrelation of a resource between used and available sites to uncover the scale at which animals assess habitat. Using an artificial landscape, we demonstrated that the model can handle discrete habitat classes. Based on conventional hierarchical analysis, deer selected for open habitat, especially meadow, and avoided coniferous forest, more strongly at the coarse level of the home range than GPS locations. Home ranges exhibited generally lower autocorrelation in elevation and meadow habitat than random locations within the population range, indicative of preference for high habitat heterogeneity. Mean maximum discrepancy in autocorrelation, which was more pronounced at the level of the home range than GPS locations, occurred at 830 m for meadow habitat and at 1511 m for elevation, suggesting that red deer responded to their environment at this scale. Our study demonstrates how spatial statistics can serve as an instructive complement to conventional approaches to habitat selection.