A Comparison of Variable Selection Criteria for Multiple Linear Regression: A THIRD Simulation Study
Date of Award
3-1994
Document Type
Thesis
Degree Name
Master of Science
Department
Department of Operational Sciences
First Advisor
David R. Barr, PhD
Abstract
The goal of this thesis research is to introduce and study a modification of Millers Method which we call the Modified-Millers method MM method, study two step subset selection procedures, these of which apply Millers Method in the first step and employs another method Minimum MSE or Minimum Sp or Minimum Cp in the second step and these of which employ another method Minimum MSE or Minimum Sp or Minimum Cp in the first step and applies Millers Method in the second step. The results of all techniques will be compared including the results of the previous simulation studies done by Hanson in 1988 and by Woollard in 1993. In the researches mentioned above, Minimum MSE, Minimum Sp, Minimum Cp and Millers method were studied. Each of the techniques, studied in this thesis effort was applied on generated data with known multicollinearities, random predictors, variances, sample sizes and error terms. The performance of each method is evaluated by employing two performance measurements. The first one, denoted PM Performance Measurement, calculates the percentage of correct variables in a model. The second one, denoted FM viii Full Model calculates the ratio of perfect models, the set of which is Xl X2 X3 for this research, to the number of all models. The Response Surface Methodology technique was employed to set up a 26 full factorial experiment in order to study the effectiveness of each subset selection method with respect to the known data characteristics. Including the results of Hansens and Woollards study, eleven different subset selection techniques were compared. Considering the PM and FM measurements together, the CpMLR method which is a two stage subset selection method performs the best, and with a very small difference the SpMLR method which is another two stage subset selection procedure is close behind the CpMLR. The Modified-Millers Method, which is a one step subset selection method performs the third best.
AFIT Designator
AFIT-GOR-ENS-94M-08
DTIC Accession Number
ADA278678
Recommended Citation
Mutlu, Ertem, "A Comparison of Variable Selection Criteria for Multiple Linear Regression: A THIRD Simulation Study" (1994). Theses and Dissertations. 6768.
https://scholar.afit.edu/etd/6768
Comments
The author's Vita page is omitted.