Bayesian Analysis of Bug-Fixing Time using Report Data

Abstract

Background: Bug-fixing is the crux of software maintenance. It entails tending to heaps of bug reports using limited resources. Using historical data, we can ask questions that contribute to better-informed allocation heuristics. The caveat here is that often there is not enough data to provide a sound response. This issue is especially prominent for young projects. Also, answers may vary from project to project. Consequently, it is impossible to generalize results without assuming a notion of relatedness between projects. Aims: Evaluate the independent impact of three report features in the bug-fixing time (BFT), generalizing results from many projects: bug priority, code-churn size in bug fixing commits, and existence of links to other reports (e.g., depends on or blocks other bug reports). Method: We analyze 55 projects from the Apache ecosystem using Bayesian statistics. Similar to standard random effects methodology, we assume each project’s average BFT is a dispersed version of a global average BFT that we want to assess. We split the data based on feature values/range (e.g., with or without links). For each split, we compute a posterior distribution over its respective global BFT. Finally, we compare the posteriors to establish the feature’s effect on the BFT. We run independent analyses for each feature. Results: Our results show that the existence of links and higher code-churn values lead to BFTs that are at least twice as long. On the other hand, considering three levels of priority (low, medium, and high), we observe no difference in the BFT. Conclusion: To the best of our knowledge, this is the first study using hierarchical Bayes to extrapolate results from multiple projects and assess the global effect of different attributes on the BFT. We use this methodology to gain insight on how links, priority, and code-churn size impact the BFT. On top of that, our posteriors can be used as a prior to analyze novel projects, potentially young and scarce on data. We also believe our methodology can be reused for other generalization studies in empirical software engineering.