Phenological observations are a valuable source of information for investigating the relationship between climate variation and plant development. Potential climate change in the future will shift the occurrence of phenological phases. Information about future climate conditions is needed in order to estimate this shift. General Circulation Models (GCMs) provide the best information about the future climate change. They are able to simulate reliably the most important mean climate features in large-scale, but they fail in regional scale because of their low spatial resolution (about 2.5°). Researchers dealing with impact studies need, as an input, estimation of potential climate change on a regional or even a local level. A common approach to bridge the scale gap between the large and regional, or local-scale is statistical downscaling. The basic idea behind statistical downscaling is to use the observed relationships between the large-scale climate parameter (predictor) and local-scale climate or climate dependent parameter (predictand), for the projection of GCM results on a regional or local-scale. In our case, beginning of flowering (BF) of dandelion (Taraxacum officinale) on 21 locations in Slovenia was related to average January (TJ), February (TF) and March air temperature (TM) on the 1000-mb pressure level across Central Europe. Statistical models were developed and tested with data for the time period 1958-1999. Before developing statistical models, Empirical Orthogonal Functions (EOF) Analysis was performed on climatological and phenological data. The EOF Analysis was used for outlier detection in the data, and for the reduction of input (predictor) data in the developed statistical models. Multiple linear regression (MLR) models were used to relate the BF with expansion coefficients of first three EOFs for TJ, TF and TM. Criteria for testing the quality of statistical models were variability of predictand data explained by the models, residual analysis and cross-validation. The models explain on average 68% of predictand variability. The correlation coefficient between observations and cross-validation estimations is on average 0.71. Absolute values of studentized residuals are with few exceptions lower than the critical value (2.75 for alpha = 0.01) and have no typical patterns. None of the unusual residuals represent an influential point - Cook's distance does not exceed the value of 0.43. The results show a strong relationship between predictand BF and predictors TJ, TF and TM. Developed models can be used for the estimation of missing BF values in historical records or to downscale the GCM results in regional climate change studies.