Predicting Crop Yield via Partial Linear Model with Bootstrap

We construct partial linear models to predict Minnesota corn and soybean yields by county. Climate variables, such as monthly precipitation and temperature measures, are included as covariates. However, fitting a standard linear regression is inadequate, and hence, an arbitrary nonparametric function over time is included for superior prediction performance. In a novel approach, the nonparametric component is approximated using an increasing sequence of orthonormal basis functions of the appropriate function space. We use different bootstrap schemes to produce prediction bounds and error estimates for the model, since the noise terms appear to be heteroscedastic and non-normal in the data. Results are presented and caveats and extensions to the model are also discussed.