Conclusion

The proposed testability metric estimates the likelihood that an object will reveal its faults. The metric is sensitive to changes in the number of data attributes and the depth of the inheritance. By making the metric dependent on the information content of the class, as represented by the signatures of the methods, rather than the structure of the class, the metric may be used in the early phases of the development cycle.

Although the sensitivity of the metric is less than it would be if applied later in development when the code is available, it is our opinion that the metric provides a sufficient basis for estimating the resources and effort necessary for adequate testing of the code.

There are several possible extensions of this work. These include:

• Sensitivity studies - The metric as currently defined is not very sensitive. A large number of methods cluster around metric values of 0, 1 and infinity. Possible redefinitions could include weights for the individual attributes of a class, such as its methods and attributes to reflect their participation in the state of the object.
• External visibility - The external visibility is analogous to the internal visibility for a method/class. The external visibility does not include the implicit attributes of the class. The relationship between this measure and a specification-based testing technique is being investigated.
• Variability of the measure - The difference between accessor and modifier methods was obvious and treating them differently was easily justified. Other factors that affect the testability may be worthy of study. For example, in our data analysis, we noted informally that the visibility measure decreased as the depth of the class in the inheritance hierarchy increased. These observations may lead to better understanding of other less obvious relationships.

Although this metric has not been tested on a full scale software project, our preliminary results from applying the metric on a wide range of C++ classes have been successful both theoretically and experimentally. It has been defined in such a way that it will fit into our framework for iterative metric management[17]. In this framework an early measure, such as VC, is paired with measures that can only be applied later in the life cycle, such as Voas's information loss metric[24] or the simpler McCabe's Cyclomatic Complexity[14] metric. With this framework and automated metric collection tools, a manager can estimate early and measure exactly later in the life cycle, to provide a continuous picture of the project.