This paper uses a conceptual rainfallrunoff model, the Xinanjiang model, to evaluate regionalization approaches using 210 catchments in southeast Australia as examples. Three regionalization approaches are compared: spatial proximity, physical similarity and regression. The results show that: spatial proximity provides the best regionalization solution; the physical similarity approach is intermediary; the regression approach is the least satisfactory. A new method, integrated similarity, is proposed by combining the spatial proximity and physical similarity. It turns out the integrated similarity further improves runoff prediction in free of observational data catchments. It is likely that taking into account the parameter uncertainty will improve the performance of predictions in free of observational data catchments. 