Abstract. Software Quality Assurance (SQA) teams play a critical role in the software development process to ensure the absence of software defects. It is not feasible to perform exhaustive SQA tasks (i.e., software testing and code review) on a large software product given the limited SQA resources that are available. Thus, the prioritization of SQA efforts is an essential step in all SQA efforts. Defect prediction models are used to prioritize risky software modules and understand the impact of software metrics on the defect-proneness of software modules. The predictions and insights that are derived from defect prediction models can help software teams allocate their limited SQA resources to the modules that are most likely to be defective and avoid common past pitfalls that are associated with the defective modules of the past. However, the predictions and insights that are derived from defect prediction models may be inaccurate and unreliable if practitioners do not control for the impact of experimental components (e.g., datasets, metrics, and classifiers) on defect prediction models, which could lead to erroneous decision-making in practice. In this thesis, we investigate the impact of experimental components on the performance and interpretation of defect prediction models. More specifically, we investigate the impact of the three often overlooked experimental components (i.e., issue report mislabelling, parameter optimization of classification techniques, and model validation techniques) have on defect prediction models. Through case studies of systems that span both proprietary and open-source domains, we demonstrate that (1) issue report mislabelling does not impact the precision of defect prediction models, suggesting that researchers can rely on the predictions of defect prediction models that were trained using noisy defect datasets; (2) automated parameter optimization for classification techniques substantially improve the performance and stability of defect prediction models, as well as they change their interpretation, suggesting that researchers should no longer shy from applying parameter optimization to their models; and (3) the out-of-sample bootstrap validation technique produces a good balance between bias and variance of performance estimates, suggesting that the single holdout and cross-validation families that are commonly-used nowadays should be avoided.