A Study of Multivariate Statistical Analysis Techniques for Computer Performance Evaluation.
Abstract
Computer performance analysts have relied too much on the use of linear regression analysis because of a lack of other data analysis techniques. Now, however, there are a number of easily available multivariate statistical techniques that need to be tested for applicability to Computer Performance Evaluation (CPE). This thesis presents some examples of how these techniques might be applied in this area, as well as some conclusions on their usefulness and applicability. Ridge Regression appeared to be useful when the independent variables were multicollinear and an explanatory model was desired. Automatic Interaction Detection was helpful in detecting the structure of data, as was Cluster Analysis. Canonical Correlation Analysis was capable of reducing the dimensionality of data while relating two sets of variables, but was considered only somewhat applicable to CPE. Factor Analysis is similar in that it can also reduce the dimensionality of data by combining variables into common factors. Both techniques work best when the variables are multicollinear. Factor analysis was also judged to be of limited applicability to CPE. Discriminant Analysis seemed to generally be non-applicable to CPE because it was designed to work with nominally scaled data and CPE data tends to be intervally scaled. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1982
- Accession Number
- ADA124899
Entities
People
- Gregory Magavero
Organizations
- Air Force Institute of Technology