类规模对面向对象度量易变性预测能力的潜在混和效应:一个元分析

The Potentially Confounding Effect of Class Size on the Ability of Object-Oriented Metrics to Predict Change-Proneness:A Meta-Analysis

最近的研究表明,类的规模对面向对象(OO)度量的易变性预测能力存在很强的混和效应,因此需要将其作为一个混和变量来考虑,否则有可能会得到误导性的结果.然而,先前的研究仅仅分析了一个软件系统,因此不清楚这个结论是否可以推广到其他系统上.为解决此问题,文中在102个Java软件系统的基础上利用元分析技术检查类的规模对55个OO度量和易变性之间关联关系的潜在混和效应.对每一个OO度量,首先在单个系统上分别计算在控制规模和不控制规模的两种情况下它与易变性的关联强度.然后,利用随机效应元分析模型计算在所有系统上且分别在这两种情况下它与易变性的平均关联强度.最后,在此基础上利用统计方法检测类规模的潜在混和效应.实验结果表明类规模的混和效应是广泛存在的,因此在验证OO度量的易变性预测能力时确实需要将其作为一个混和变量来考虑.

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 recta-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 meth- ods 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.