This study estimates the site-specific crop yield response function using varying coefficient models. It is widely recognized that the parameters of yield response function vary dramatically across space and over time. Previous studies usually capture this variability of response by using locational and time dummy variables. While that approach reveals the existence of the response variability, the exact pattern of the variability is unknown, and the capacity of ex ante prediction of such models are limited. This study takes a step forward to explicitly explain how the response varies with the actual site characteristic variables, such as soil, water, topography, weather, and other factors that are commonly available to producers. By using the varying coefficient model, the parameters of the response function are specified to change continuously with those site variables. Based on a simulation data set, the varying coefficient model is demonstrate to outperform the site-dummy model by creating better variable rate application (VRA) fertilizer prescriptions. We further propose to apply the model to large sample of high resolution production data, and create ex ante spatially explicit optimal VRA fertilizer recommendations. The ultimate goal is to develop a precision decision system which can statistically turn the soil testing and weather forecasting information into input application prescriptions for producers.