• California;
  • climate change;
  • PRISM;
  • species distribution modeling


Empirical species distribution models are widely used to predict the effects of climate change on biodiversity distribution but rely on multiple assumptions about the certainty of the locality and climate data. Here, we assess the effect of historical climate data variability when forecasting geographic responses of California mammals to 20th century climate change. We first used two methods to derive gridded climate surfaces from weather station data (ANUSPLIN and PRISM) representing two sampling eras: historic (1900–1940) and current (1980–2005). We then used the two sources of climate data in conjunction with a maximum entropy algorithm (MAXENT) to predict both the historic and current distributions of all major mammal species vouchered historically in California. Results indicate that levels of disagreement between the two climate datasets are considerably greater in the historical era than in the current era. For the bioclimatic variables used in modeling historical mammal distributions, precipitation variables were less concordant than temperature variables. These discrepancies are reflected in the low agreement between historic mammal range predictions and further propagated when the historic models are projected to present day. Nonetheless, some common patterns exist across mammal species and climate estimates. Range stability is the most common prediction between the two eras, followed by expansion and contraction. Jepson ecoregions with relatively high levels of range stability include parts of the Great Central Valley and Sierra Nevada, while other parts of the Central Valley, the Sonoran desert, and Central- and Southwestern California yield predictions of range shifts. Historical species distribution modeling can greatly inform studies attempting to describe how species will continue to move geographically in response to future changes in climate. We suggest that alternative estimates of historical climate and their uncertainties are ultimately required in order to provide a quantitative measure of the confidence in predicted changes in distribution.