Software defects can impact project cost and product delivery. Software Quality Assurance (SQA) activities (e.g., code review and software testing) have been exercised to ensure high quality of software systems, however, they can be challenging due to limited Quality Assurance (QA) resources. Defect models–a classification model that classifies whether software modules are defective– help developers identify the most risky modules to prioritise their limited QA resources. The interpretation of defect models also helps software managers understand defect characteristics to chart quality improvement plans. Unfortunately, a misleading interpretation of defect models may produce incorrect quality improvement plans, which ultimately misguide software development teams. This thesis hypothesises that: Experimental design impacts the interpretation of defect models. Empirical studies can provide a better understanding of such impact and guidelines to develop a more reliable interpretation of defect models. To validate the hypothesis, we formulate 3 research questions, i.e., (1) how do correlated metrics impact the interpretation of defect models?, (2) how can we automatically mitigate correlated metrics?, and (3) how can we identify the most important metrics from statistical and practical points of view? Through case studies of publicly-available open-source and industrial software systems, our results show that correlated metrics impact the interpretation of defect models and must be mitigated. Surprisingly, our results show that commonly-used automated feature selection techniques do not mitigate correlated metrics. On the other hand, our results show that our proposed feature selection technique, AutoSpearman, can automatically mitigate correlated metrics with little impact on model performance.