Abstract: Functional magnetic resonance imaging (fMRI) allows for the indirect measurement of whole brain neuronal activity using local blood oxygenation level. Functional connectivity, i.e., the correlation between the temporal activity of remote regions, may be used to track brain reorganization while, for example, a subject learns a new skill. However, testing the significance of changes in functional connectivity is challenging for individual data, because fMRI time series exhibit dependencies in both space and time that may not be properly captured by classical parametric models. To address this issue, we propose a new statistical procedure in a bootstrap hypothesis testing framework after various strategies were implemented to take temporal dependencies into account. These alternatives were evaluated on Gaussian and non-Gaussian Monte-Carlo simulations of space-time processes, as well as on a longitudinal study of motor skill learning. The results demonstrated that neglecting the temporal dependencies or modeling them as an autoregressive process of order 1 may lead to poor control of the false positive rate, i.e. to liberal tests. The version of the procedure based on a circular block bootstrap achieved robust, satisfactory performances in all settings.
Key words and phrases: Block bootstrap, correlation, data-driven block length selection, double bootstrap, fMRI, functional connectivity, hypothesis testing.