Variable selection in multivariate methods using global score estimation (Q2430229)

The authors consider methods of variables subset selection for such multivariate methods (MMs) as principal components (PC), and factor or correspondence analysis. For these techniques no external response variable is present whose prediction could be used as a criterion of the quality of a subset. So the authors propose to use the global scores of MMs as such criterion. E.g., for PC they take the global scores derived from the full set of variables and try to predict these scores by subsets. The best subset is obtained by stepwise procedures (backward, forward, etc.). The proposed technique is illustrated by a case study of Japanese environmental management survey data.