Automatic selection of indicators in a fully saturated regression (Q626211)

The problem of mean and variance estimation from an i.i.d. sample contaminated by outliers is discussed. The authors consider a linear regression model in which the presence of outliers is modeled by impulse dummy variables (indicators) added for some subset of the sample. Only significant indicators are to be included into the model. The resulting model is obtained as a combination of models derived by different subsets of the sample. Asymptotics of the estimate for the mean and variance are investigated. Results of Monte Carlo simulations are presented.