Software defect prediction performance varies over a large range. Menzies suggested there is a ceiling effect of 80% Recall [8]. Most of the data sets used are highly imbalanced. This paper asks, what is the empirical effect of using different datasets with varying levels of imbalance on predictive performance? We use data synthesised by a previous meta-analysis of 600 fault prediction models and their results. Four model evaluation measures (the Mathews Correlation Coeficient (MCC), F-Measure, Precision and Re- call ) are compared to the corresponding data imbalance ratio. When the data are imbalanced, the predictive performance of software defect prediction studies is low. As the data become more balanced, the predictive performance of prediction models increases, from an average MCC of 0.15, until the minority class makes up 20% of the instances in the dataset, where the MCC reaches an average value of about 0.34. As the proportion of the minority class increases above 20%, the predictive performance does not significantly increase. Using datasets with more than 20% of the instances being defective has not had a significant impact on the predictive performance when using MCC. We conclude that comparing the results of defect prediction studies should take into account the imbalance of the data.