Outlier Robust Prediction

Outlier robust prediction describes methods for estimating the population total when the sample contains representative outliers, i.e. true values that are extremely unlikely under the working model, but which cannot be considered to be unique, in the sense that non-sampled population units with similar extreme values quite likely exist. Outlier robust methods of parameter estimation are described, as is an outlier-robust bias correction for the contribution of the non-sampled outliers to the population total. The extension to outlier robust non-parametric prediction is described, and empirical evidence presented for the gains that can be achieved when these outlier robust methods are applied in a realistic farm survey scenario. The chapter concludes with a discussion of how sample design can help outlier robustness.