This thesis empirically compares statistical prediction models using fault count data and fault binary data. The types of statistical models that are studied in detail are Logistic Regression for binary data and Negative Binomial Regression for the count data. Different model building approaches are also evaluated: manual variable selection, stepwise variable selection, and hybrid selection (classification and regression trees combined with stepwise selection). The data set comes from a commercial Java application development project. In this project special attention was paid to data collection to ensure data accuracy. The comparison criteria we used were a consistency coefficient and the estimated cost savings from using the prediction model. The results indicate that while different model building approaches result in different object-oriented metrics being selected, there is no marked difference in the quality of the models that are produced. These results suggest that there is no compelling reason to collect highly accurate fault count data when building object-oriented models, and that fault binary data (which are much easier to collect) will do just as well. (Abstract shortened by UMI.)
展开▼
