We analyzed moose (Alces alces)–vehicle collisions (MVCs) in western Maine, USA, from 1992 to 2005 (n  =  8,156) using Geographic Information Systems to identify patterns of temporal and spatial distribution and develop predictive models based on road and landscape characteristics. We used chi-square and correlation analyses to assess temporal characteristics of MVCs, K-function and kernel analyses to identify spatial clusters of MVCs, and logistic regression to relate covariates for traffic, land-cover, land-form, and relative moose abundance to probability of MVC. We evaluated candidate models using Akaike's Information Criterion, area under the receiver operating characteristic curve (AUC), and the percentage of correctly classified observations. Most (81.6%) MVCs occurred from May to October, with peak monthly frequencies in June (18.6%). Moose–vehicle collisions were clustered spatially on roads at local (0–4 km) and regional scales (22–41 km and 45–54 km), but not at intermediate scales. Traffic-related covariates predicting MVCs included traffic volume and speed limit. For each additional 500 vehicles/day, odds of a location being an MVC increased by 57%. For each 8-km/hr increase in speed limit, odds of an MVC increased by 35%. Landscape composition covariates best predicted MVCs within a 2.5-km radius of the collision site. Mean percent cover within 2.5 km of MVCs was comprised of 36% more cutover forest, 10% more coniferous forest, 5% less deciduous–mixed forest, and 10% less nonwoody wetland than for random points. For every 5% increase in percent cutover and coniferous forest within 2.5 km of the road, predicted odds of MVC increased by 36% and 19%, respectively. Landscape configuration covariates best predicted MVCs within the 5.0-km radius. Moose–vehicle collisions were associated with areas of less interspersion of cover types; for each 5% increase in an index of interspersion–juxtaposition, predicted odds of MVC decreased by 11%. Our final model attained high predictive power (AUC  =  0.835) and validation accuracy (75.0%). The model also proved robust to physiographic variation, exhibiting high predictive power (AUC  =  0.828) and validation accuracy (68.8%). Managers seeking to prioritize resources for reducing MVCs or predicting future areas of high MVC probability should assess land-cover composition and configuration surrounding MVC hotspots at geographic extents out to 2.5–5 km and use this information to plan expensive roadside management practices such as fencing. The importance of traffic and landscape covariates in our modeling suggests that effective management to reduce MVCs will require a complex combination of driving speed reductions and modifications to forest management along roads.