A technique to develop and validate simulation models

Abstract

The development and validation of simulation models requires access to extensive and detailed datasets on the subject to be modelled. The availability of such data is often limited. Consequently, this paper explores the use of resampling techniques in the estimation of parameters for and the assessment of accuracy of simulation models. A description of jackknife and cross-validation techniques is presented, as well as an application of these techniques to the fitting of a crop simulation model to a limited dataset. The example concerns the application of a simulation model for the growth of maize (Zea mays) in northern Australia. Jackknife techniques were applied to the estimation of the potential kernel number and the potential kernel growth rate of a specified maize hybrid, and the prediction accuracy of the estimation of grain yield was assessed by cross-validation. The jackknife estimates were found to differ from the estimates obtained from a single fit either to all the data or to subsets of half the data sampled from the datasets. In addition, a more reliable estimate of the prediction variance was found from the cross-validation step compared to that found in the more traditional validation method of using one half of a dataset independent of the estimation half. The use of the jackknife and cross-validation techniques permitted a limited dataset to be used in both the parameter estimation and validation processes of model development.