Identifying persons with the same stance in topic documents that contain competing viewpoints can help readers construct the background of a topic and facilitate topic reading. In this paper, we propose an unsupervised method for identifying topic persons with the same stance. Specifically, we employ a model-based Expectation-Maximization (EM) method to cluster individuals into positively correlated groups. In addition, we utilize an off-topic block elimination technique and a weighted correlation coefficient to remove off-topic text blocks and alleviate the text sparseness problem. We also present an effective initialization algorithm that generates appropriate EM initializations. Our experiment results demonstrate that the proposed method clusters topic persons with the same stance correctly and outperforms many well-known clustering methods. Moreover, the initialization algorithm yields accurate and stable stance identification results.