AbstractWe assess monthly temperature and precipitation data derived by six statistically downscaled data sets for 14 general circulation models (GCMs) from the Climate Model Intercomparison Program Phase 5. We use a simple monthly water balance model to quantify and decompose uncertainties associated with the GCMs and statistical techniques in projections for four snow‐dominated regions in the western United States. The end‐of‐century projections from GCMs exhibit substantial variation in change over the regions (temperature range of 2.8–8.0 °C and precipitation range of −22–31%). The six downscaled data sets exhibit disparate high‐resolution representations of the magnitude and spatial patterns of future temperature (up to 2.2 °C) and precipitation (up to 30%) for a common GCM. Two data sets derived by the same downscaling method (Multivariate Adaptive Constructed Analogs) produce median losses of snow water equivalent over the Upper Colorado of 51% and 81%. The principal causes of the differences among the downscaled projections are related to the gridded observations used to bias correct the historical GCM output. Specifically, (1) whether a fixed atmospheric lapse rate (−6.5 °C/km) or a spatially and temporally varying lapse rate is used to extrapolate lower elevation observations to high‐elevations and (2) whether high‐elevation station data (e.g., SNOTEL) are included in the observations. The GCM projections are the largest source of uncertainty in the monthly water balance model simulations; however, the differences among seasonal projections produced by downscaled data sets in some regions highlight the need for careful evaluation of the statistically downscaled data in climate impact studies.