Legitimate Techniques for Improving the R-Square and Related Statistics of a Multiple Regression Model
Abstract
Cost and analysts and DOD contractors frequently use regression analysis to develop cost estimating relationships, production relationships, and various forecasting equations. Invariably, those regression equations are presented in the text of the final report along with the statistical properties -- i.e. the R-Square, the Standard Error of the Estimate, the Durbin-Watson Statistic, etc. These statistics are often presented as evidence of the validity and accuracy of the resulting equation. The higher the R-square the bolder the print and the more prominently displayed. Unfortunately, high R-square's, favorable Durbin-Watson statistics, etc., can be artificially or inadvertently inflated to appear more favorable. In reality, the equation with good statistical properties may not reflect a valid causal relationship to explain variations in the dependent variable. In many cases the regression equations prove to be of little value in forecasting or explaining the relationships with new data. This paper discusses techniques for artificially raising the R-square and related statistical properties of regression equations. These techniques are presented for the benefit of analysts who are trying to improve the statistical properties of their equations and for the benefit of managers who must approve payment for such analysis. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1981
- Accession Number
- ADA109370
Entities
People
- Edwin J. Curle