AbstractSpatial organization plays prominent roles in disease transmission, genetics, and demography of wildlife populations and is therefore an important consideration not only for wildlife management, but also for inference about populations and processes. We used hierarchical agglomerative clustering of a spatial graph network to partition Wind Cave National Park (WICA) into five regions used by 163 female elk (Cervus elaphus) marked with global positioning system collars during 2005–08 and 2011–13. We grouped elk based on differential use of the five regions, developed a priori models for inter‐group variation in the occurrence of chronic wasting disease (CWD), and used Akaike's information criterion to compare models and stratify regions. Previous descriptions of elk population structure, which have been based on social contact or overlap of individual ranges, have distinguished spatially disjunct population subsets. Constructing hierarchical partitions of the landscape enabled us to also discern and describe overlapping and nested subsets. During 2016–18, apparent park‐wide prevalence of CWD was 0.18 (90% CI = [0.146, 0.182]); however, prevalence within three spatial strata used primarily by different elk ranged from 0.03 ([0.008, 0.074]) to 0.29 ([0.211, 0.375]). In context with published estimates of recruitment, predation, and anthropogenic mortality, such differences in prevalence equate to increasing local abundance of elk in southwestern WICA, stable to declining abundance in the west/northwest, and rapidly declining abundance in the east. Despite the modest size of WICA (11,357 ha), park‐wide averages conflate effects of elk distribution and disease, obscuring spatial patterns with profound implications for study and management of elk and CWD. Graph networks have been used widely in ecology to describe such phenomena as social relationships, connectivity of habitat patches, animal movements, and the spread of disease. Extension to partitioning of geographic range is straightforward but entails different considerations. We discuss allocation of sampling effort, construction of an initial partition, specification of a model for graph cohesion, selection of a clustering algorithm, and identification of useful partitions.