Subset selection using the total least squares approach in collinearity problems with errors in the variables (Q1822478)

The linear model \(Ax=b\) describes the relationship between a vector b of one response variable and an \(m\times n\) matrix A, containing n predictor variables. In the situation in which there is a nearly exact linear relationship among the predictions in A the ordinary least squares tend to be inflated. Ridge regression, shrinking factors and reducing the rank of the matrix are techniques for handling the collinearity problem. In many applications it is desirable to reduce the dimensionality of the space and also to reduce the number of variables. How to pick these variables is the problem of subset selection. A new subset selection algorithm based on total least squares is presented and compared with the algorithm of \textit{G. H. Golub}, \textit{V. Klema} and \textit{G. W. Stewart} [Rank degeneracy and least squares problems. Techn. Rep. TR 456, Dept. Comput. Sci., Univ. Maryland, College Park (1976)] with respect to its stability in subset choice and its accuracy in parameter estimation and prediction, when errors occur in all variables.