A comparison of alternatives to regression analysis as model building techniques to develop predictive equations for software metrics

Abstract

The almost exclusive use of regression analysis to derive predictive equations for software development metrics found in papers published before 1990 has recently been complemented by increasing numbers of studies using non-traditional methods, such as neural networks, fuzzy logic models, case-based reasoning systems, rule-based systems, and regression trees. There has also been an increasing level of sophistication in the regression-based techniques used, including robust regression methods, factor analysis, resampling methods, and more effective and efficient validation procedures. This paper examines the implications of using these alternative methods and provides some recommendations as to when they may be appropriate. A comparison between standard linear regression, robust regression, and the alternative techniques is also made in terms of their modelling capabilities with specific reference to software metrics.