Practical assessment of the models for identification of defect-prone classes in object-oriented commercial systems using design metrics

The goal of this paper is to investigate and assess the ability of explanatory models based on design metrics to describe and predict defect counts in an object-oriented software system. Specifically, we empirically evaluate the influence of design decisions to defect behavior of the classes in two products from the commercial software domain. Information provided by these models can help in resource allocation and serve as a base for assessment and future improvements. We use innovative statistical methods to deal with the peculiarities of the software engineering data, such as non-normally distributed count data. To deal with overdispersed data and excess of zeroes in the dependent variable, we use negative binomial (NB) and zero-inflated NB regression in addition to Poisson regression. Furthermore, we form a framework for comparison of models' descriptive and predictive ability. Predictive capability of the models to identify most critical classes in the system early in the software development process can help in allocation of resources and foster software quality improvement. In addition to the correlation coefficients, we use additional statistics to assess a models' ability to explain high variability in the data and Pareto analysis to assess a models' ability to identify the most critical classes in the system. Results indicate that design aspects related to communication between classes and inheritance can be used as indicators of the most defect-prone classes, which require the majority of resources in development and testing phases. The zero-inflated negative binomial regression model, designed to explicitly model the occurrence of zero counts in the dataset, provides the best results for this purpose. (C) 2002 Elsevier Science Inc. All rights reserved.