Measuring and predicting software productivity: A systematic map and review

0950-5849/$ see front matter 2010 Elsevier B.V. A doi:10.1016/j.infsof.2010.12.001 ⇑ Address: School of Computing, Blekinge Institute 372 25, Sweden. E-mail addresses: [email protected], kai.petersen URLs: http://www.bth.se/besq, http://www.ericsso Context: Software productivity measurement is essential in order to control and improve the performance of software development. For example, by identifying role models (e.g. projects, individuals, tasks) when comparing productivity data. The prediction is of relevance to determine whether corrective actions are needed, and to discover which alternative improvement action would yield the best results. Objective: In this study we identify studies for software productivity prediction and measurement. Based on the identified studies we first create a classification scheme and map the studies into the scheme (systematic map). Thereafter, a detailed analysis and synthesis of the studies is conducted. Method: As a research method for systematically identifying and aggregating the evidence of productivity measurement and prediction approaches systematic mapping and systematic review have been used. Results: In total 38 studies have been identified, resulting in a classification scheme for empirical research on software productivity. The mapping allowed to identify the rigor of the evidence with respect to the different productivity approaches. In the detailed analysis the results were tabulated and synthesized to provide recommendations to practitioners. Conclusion: Risks with simple ratio-based measurement approaches were shown. In response to the problems data envelopment analysis seems to be a strong approach to capture multivariate productivity measures, and allows to identify reference projects to which inefficient projects should be compared. Regarding simulation no general prediction model can be identified. Simulation and statistical process control are promising methods for software productivity prediction. Overall, further evidence is needed to make stronger claims and recommendations. In particular, the discussion of validity threats should become standard, and models need to be compared with each other. 2010 Elsevier B.V. All rights reserved.