Abstract：Statistical downscaling is an effective way to downscale General Circulation Model (GCM) outputs to a finer temporal and spatial scale. Here, we proposed and trained a new station‐based non‐linear regression downscaling (SNRD) approach, and quantitatively demonstrated its better performance than linear regression model. In addition, we also compared the SNRD method with the Bias Correction Spatial Disaggregation (BCSD) method, which has been widely applied on downscaling precipitation. Results indicated that RMSEs and residuals of the BCSD method are 4.6–31 times as ones of SNRD in each region of China during validation period (2004–2015). Besides, with historical in‐situ observed monthly mean precipitation as the benchmark, SNRD downscaled model modified the over‐estimation (1.25–2 times as observed precipitation) flaw of BCSD model in rainy season (June–August) of China, which could provide assistance on the further researches on drought events in rainy season in future. Further, we also surprisingly found that the 2,500 m was the threshold height for statistical downscaling models. Both BCSD and SNRD downscaled precipitation displayed more steadily and accurately above 2,500 m, since there are less intense human activities in China. Thus, we recommend further studies should fully consider the height element in downscaling models to improve the accuracies.