The relationship between VC and the number of branch test cases is weak; however, the combination of validations that we used, taken together, provide compelling evidence for this approach.
One reason for the weak relationship may be due to the lack of sensitivity of VC. A large number of the methods that were examined had ratios of 1. We partially addressed the lack of sensitivity issue by multiplying the ratio by the number of unique objects associated with a method. We will further discuss the sensitivity issue in the future work section of the paper. The lack of sensitivity is not surprising given the stage in the development process at which we wish to apply the metric. Early models, of necessity, contain less exact information than later models, such as code; however, estimates of project size and effort must be made as early as possible with the best information available.
A post hoc analysis revealed an interesting relationship. The agreement between measures for the base class and the internal visibility is very high (for one set of classes the correlation was 1.0). This agreement declines as classes lower in the hierarchy are added to the computation.