Recently, we proposed a two-step adaptive strategy for the statistical analysis of brain connectivity that is based on a first screening at the subnetwork level and a filtering at the connection/node level. The method was shown to guarantee strong control of type-I error through rigourous statistical proofs. In addition, the gain of power obtained by this method is considerable especially with an appropriate decomposition of the global network. Here, we discuss the extension of the two-step methods to multivariate statistics and we compare its performance against both standard methods and univariate two-step methods. We present as well a practical example of detecting topological nodal differences between functional connectivity matrices of resting state and movie-watching, respectively.