The interannual variability (IAV) in soil respiration (R_soil) is crucial for learning the carbon budget in the global climate change, the factors and the processes controlling the soil respiration. Variability in climate factors may directly control changes in the R_soil, but also may indirectly control the R_soil by "function change" by changing the biological and ecological processes. We measured the soil respiration and climate factors (air temperature, soil temperature, soil water content and precipitation) on the winter wheat farmland in the Loess Plateau, China, for four growing seasons. The homogeneity-of-slope model (HOS model) was used to distinguish direct effect of IAV in soil respiration and the function change. The results indicated that the soil respiration had a very significant diurnal, seasonal and interannual variability. The diurnal variability was controlled by the soil temperature, reached the peak at 13:00-16:00. The stepwise regress analysis indicated that the seasonal variability in soil respiration was controlled by the soil temperature (T_soil, ranged from -3.8℃ to 30.7℃) and soil water content (SWC, ranged from 1.5% to 50.3%), both of them showed positive relationships with soil respiration that could be described by the function: R_soil =1.761+0.119T_soil - 1.30SWC,(R²=0.48,P < 0.0001). The precipitation had no significant relationship with R_soil. The mean annual R_soil ranged from 815.72 g C·m^-2·a^-1 to 980.12 g C m^-2 a^-1. The interannual variability of R_soil was controlled by the soil temperature. The results of HOS model analysis indicated that function change, interannual climatic variability, seasonal climatic variability and random error explained 10.6%、4.5%、58.4%、and 26.5%, respectively, of the observed variation in soil respiration. Function change account for more than twice the variance in soil respiration of interannual climatic variability, it suggests that the soil respiration in the future con not be predicted from the climatic variability alone, the function change must be considered into mechanistic models.