Abstract
Background: Like any research discipline, software engineering research must be of a certain quality to be valuable. High quality research in software engineering ensures that knowledge is accumulated and helpful advice is given to the industry. One way of assessing research quality is to conduct systematic reviews of the published research literature.
Objective: The purpose of this work was to assess the quality of published experiments in software engineering with respect to the validity of inference and the quality of reporting. More specifically, the aim was to investigate the level of statistical power, the analysis of effect size, the handling of selection bias in quasi-experiments, and the completeness and consistency of the reporting of information regarding subjects, experimental settings, design, analysis, and validity. Furthermore, the work aimed at providing suggestions for improvements, using the potential deficiencies detected as a basis. Method: The quality was assessed by conducting a systematic review of the 113 experiments published in nine major software engineering journals and three conference proceedings in the decade 1993-2002.
Results: The review revealed that software engineering experiments were generally designed with unacceptably low power and that inadequate attention was paid to issues of statistical power. Effect sizes were sparsely reported and not interpreted with respect to their practical importance for the particular context. There seemed to be little awareness of the importance of controlling for selection bias in quasi-experiments. Moreover, the review revealed a need for more complete and standardized reporting of information, which is crucial for understanding software engineering experiments and judging their results.
Implications: The consequence of low power is that the actual effects of software engineering technologies will not be detected to an acceptable extent. The lack of reporting of effect sizes and the improper interpretation of effect sizes result in ignorance of the practical importance, and thereby the relevance to industry, of experimental results. The lack of control for selection bias in quasi-experiments may make these experiments less credible than randomized experiments. This is an unsatisfactory situation, because quasi-experiments serve an important role in investigating cause-effect relationships in software engineering, for example, in industrial settings. Finally, the incomplete and unstandardized reporting makes it difficult for the reader to understand an experiment and judge its results.
Conclusions: Insufficient quality was revealed in the reviewed experiments. This has implications for inferences drawn from the experiments and might in turn lead to the accumulation of erroneous information and the offering of misleading advice to the industry. Ways to improve this situation are suggested.