Statistical shape models are commonly used in various applications of computer vision. Nevertheless, these models are not well adapted to hierarchical structures. This paper proposes a solution to this problem by presenting a general framework to build multilevel statistical shape models. Based on multilevel component analysis, the idea is to decompose the data into a within-individual and a between-individual component. As a result, several sub-models are deduced and can be treated separately, each level characterizing one sub-model. In this paper, we present a multilevel model of the human spine. The results show that such a modelization offers more flexibility and allows deformations that classical statistical models can simply not generate.