Granular computing is an effective method to deal with imprecise, fuzzy and incomplete information. Commonly, it consists of three popular models: fuzzy sets, rough sets and quotient space. The main interest of the first two methods is to deal with the problem with uncertainty information and that of the latter is to implement the multi-granularity computing. In particular, a quotient space which has a hierarchical structure will be divided into different granules by equivalence relations. In this paper, such hierarchical quotient space is applied to propose a new feature selection method. Specifically, the feature subset is selected by calculating the dependency in the position region of such hierarchical quotient space. The experimental results demonstrate that the performance of the proposed approach outperforms those attainable by typical feature selection methods, in terms of both the size of reduction and classification accuracy.