Computations and analysis in robust regression model (Q1979096): Difference between revisions

An essential task in many regression problems is to screen a large number of potential explanatory variables to select one or a few subsets of them which fit best the information contained in the response variable. Another important task is to see how the above screening process is affected by outliers in the data. Namely, the set of selected models should be robust in the sense that they are indifferent to radical change of a small portion of the data or a small change in all of the data. The principle of minimum description length and the associated stochastic complexity methods, newly developed from the probabilistic theory of information and coding, provide a promising approach to model selection in robust regression. This paper gives an exposition to computations and some application issues of the stochastic complexity criterion. Specifically, the methods of computing the robust parameter estimates, the weight functions and the criterion function that are involved in the model selection procedure are addressed. A package of \(S\)-language programs written for the computations is also introduced. Two complicated but important cases in applications of model selection are investigated: namely, that when no clear-cut best model can be selected and when there are too many candidate models so that the exhaustive selection is not computationally feasible. Finally, some examples and a simulation study to illustrate and assess the proposed methods are presented.