Advances and Applications in Statistics
Volume 4, Issue 3, Pages 265 - 286
(December 2004)
|
|
ON SUBSET SELECTION AND BEYOND
Erhard Reschenhofer (Austria)
|
Abstract: This paper
considers methods of variable selection that
are based on common statistics like the
adjusted R-squared statistic, the t-statistic,
or the F-statistic and proposes various
modifications for the case of non-nested
models. The resulting model selection criteria
are somehow related to the risk inflation
criteria proposed by Foster and George [Ann.
Statist. 22 (1994), 1947-1975] and George and
Foster [Biometrika 87 (2000), 731-747]. Next,
the final prediction error criterion is
modified in a similar way so that it can also
be used for subset selection. Finally, a
universal modeling procedure is discussed that
can be used for the simultaneous selection of
the model class, the criterion for variable
selection, and the method for the estimation
of the model parameters. |
Keywords and phrases: AIC, model selection criteria, risk inflation criteria. |
|
Number of Downloads: 359 | Number of Views: 1098 |
|