Negative binomial regression has been proposed as an approach to predicting fault-prone software modules. However, little work has been reported to study the strength, weakness, and applicability of this method. In this paper, we present a deep study to investigate the effectiveness of using negative binomial regression to predict fault-prone software modules under two different conditions, self-assessment and forward assessment. The performance of negative binomial regression model is also compared with another popular fault prediction model—binary logistic regression method. The study is performed on six versions of an open-source objected-oriented project, Apache Ant. The study shows (1) the performance of forward assessment is better than or at least as same as the performance of self-assessment; (2) in predicting fault-prone modules, negative binomial regression model could not outperform binary logistic regression model; and (3) negative binomial regression is effective in predicting multiple errors in one module.