Estimated generalization error is the main index that indicates learning performance, but it is inadequate for further analysis. Bias-variance theory tries to overcome the limitation of analyzing learning performance, but the concept of bias-variance is still controversial when applied to the classification problem. In this paper, we propose a new alternative, simple and practical, analytical method called ?migration analysis? to analyze the learning results. We compare the properties of migration analysis to bias-variance framework, and use it to analyze two so-called ensemble learners: bagging and AdaBoost. The results not only explain these ensemble learners in different ways, but also shed light to the new promising learning algorithm.