Validation of Software Testing Experiments: A Meta-Analysis of ICST 2013

Researchers in software testing are often faced with the following problem of empirical validation: does a new testing technique actually help analysts find more faults than some baseline method? Researchers evaluate their contribution using statistics to refute the null hypothesis that their technique is no better at finding faults than the state of the art. The decision as to which statistical methods are appropriate is best left to an expert statistician, but the reality is that software testing researchers often don´t have this luxury. We developed an algorithm, Means Test, to help automate some aspects of statistical analysis. We implemented Means Test in the statistical software environment R, encouraging reuse and decreasing the need to write and test statistical analysis code. Our experiment showed that Means Test has significantly higher F-measures than several other common hypothesis tests. We applied Means Test to systematically validate the work presented at the 2013 IEEE Sixth International Conference on Software Testing, Verification, and Validation (ICST´13). We found six papers that potentially misstated the significance of their results. Means Test provides a free and easy-to-use possibility for researchers to check whether their chosen statistical methods and the results obtained are plausible. It is available for download at coest.org.