Statistical downscaling is one of the tools used most frequently to estimate the future climate change. Many studies have investigated which methods and potential predictors, if applied to real data or GCM simulations of present climate, fit the observations best. In applications to GCM simulations of future climates, the methods and predictors performing best in present climate conditions have usually been directly applied, without considering their capacity to capture the changes due to increased greenhouse gas concentrations and the uncertainty induced by the selection of a single method and predictor.

In the contribution, we examine daily mean temperature values in winter at a network of stations in central and western Europe, downscaled from the 2xCO2 run of the CCCM2 GCM. Warming rates at individual sites are intercompared between different downscaling methods (regression of gridded predictor values, regression of principal components of predictors, canonical correlation analysis), different numbers of principal components and canonical pairs, and different sets of predictors (500 and 1000 hPa heights, 850 hPa temperature, 1000/500 hPa thickness, and their combinations). Predictors are defined on a grid covering Euroatlantic midlatitudes.

We show that different choices of method and predictors may lead to drastically different warming rates. We argue that for a proper application of the downscaling to construction of climate change scenarios, the predictors must be selected that will be affected in their magnitude by the climate changes to come. For this reason, the use of 1000 hPa heights and sea level pressure as the only predictors in downscaling is not appropriate. The differences among methods are also discussed, and doubts are cast on the use of the regression of predictor's principal components because the climate change estimate appears to be highly sensitive to the number of components involved in the regression.