An Approach to Estimating the Number of Defects from lines of Code, with Censored Poisson and Binomial Models

Software defect prediction plays a vital role in ensuring software quality and optimizing resource allocation in development projects. However, the defect density, measured as defects per thousand lines of code (KLOC), is often not directly observable. Instead, software development teams typically record the presence of defects as a Boolean value, indicating whether a module has one or more reported defects. This paper proposes a novel method to estimate the number of defects in a software project using probability models and censoring techniques, based on the available lines of code (LOC) and defect presence data. We investigated Poisson and Binomial models, implemented with different tools such as JAGS and Metropolis, and obtained consistent results across the models. The experimental results show the proposed approach can provide valuable insights for project managers, enabling them to allocate resources effectively and prioritize quality assurance activities, ultimately leading to improved software reliability and customer satisfaction.