How can we deal with atypical observations in SIR regression. This can be generalized to any predictive model.