7.5 Regression as a Metric Validation Tool

7.5 Regression as a Metric Validation Tool

Regression analysis will be one of our primary tools in the validation of new software metrics. One of the primary criteria for the use of a new metric in our metric suite will be that it identifies a unique source of variation in one or more criterion measures. If, for example, our criterion measure is software faults, then a new metric under consideration will contribute significantly to the regression ANOVA. We will notice this improvement in one of two ways: (1) the adjusted R-square statistic will increase, and (2) the regression F statistic will increase.

Sometimes, the effect of a new metric will be very difficult to isolate in a regression model with raw metrics because of the problems with multicollinearity. We noted that the use of the orthogonal domain metrics in the metric validation process would eliminate this problem. In this case we would first perform the PCA on the raw metrics, including the new metric that we wish to validate. The resulting factor scores (domain metrics) would then be used as the independent variables in a regression analysis with the fault measure as a dependent variable.

Whether our regression analysis is performed on the raw metrics or the domain metrics, we will not consider using any metric in our working metric set that does not enhance our understanding of the criterion variable(s). In the context of regression analysis, we will see this enhancement by our ability to explain ever-increasing amounts of the variation in the dependent variable(s).



Software Engineering Measurement
Software Engineering Measurement
ISBN: 0849315034
EAN: 2147483647
Year: 2003
Pages: 139

flylib.com © 2008-2017.
If you may any questions please contact us: flylib@qtcs.net