The potentially confounding effect of class size on the ability of object-oriented metrics to predict change-proneness: A meta-analysis

Recent research shows that class size has a strong confounding effect on the ability of object-oriented (OO) metrics to predict change-proneness and hence suggests that it should be considered as a confounding variable. Otherwise, misleading analysis results would be obtained. However, this conclusion is drawn from only one software system and it is not clear whether it can be generalized to other systems. To attack this problem, based on 102 systems, this paper employs statistical meta-analysis techniques to examine the potentially confounding effect of class size on the associations between 55 OO metrics and change-proneness. For each metric, we first compute its degrees of association with change-proneness under controlling/not controlling for class size on individual systems. Then, we employ random-effect models to compute their average degrees of associations under these two cases over all systems. Finally, we apply statistical methods to test whether class size has a confounding effect. Our experimental results indicate that the confounding effect of class size in general exists and hence confirm that we should consider it as a confounding variable when validating OO metrics on change-proneness. ©, 2015, Science Press. All right reserved.