A new approach is presented, in which rough set theory is applied to select attributes as nodes of a decision tree. Initially, dataset is partitioned into subsets based on different condition attributes, then an attribute is chosen as a node for branching when the size of its corresponding implicit region is smaller than that of all other attributes. This approach is compared to the entropy-based method on some datasets from the UCI Machine Learning Database Repository, which instantiates the performance of the rough set approach. Statistical experiments showed that the proposed approach is feasible for decision-tree construction.