Elsevier

NeuroImage

Volume 114, 1 July 2015, Pages 466-470
NeuroImage

Comments and Controversies
Towards a statistical test for functional connectivity dynamics

https://doi.org/10.1016/j.neuroimage.2015.03.047Get rights and content

Highlights

  • We consider window length choice when computing dynamic functional connectivity.

  • We provide statistical support for Leonardi and Van De Ville's 1/f rule of thumb.

  • We discuss limitations of Leonardi and Van De Ville's sinusoidal model.

  • We develop a test to identify non-stationary connectivity fluctuations.

Abstract

Sliding-window correlation is an emerging method for mapping time-resolved, resting-state functional connectivity. To avoid mapping spurious connectivity fluctuations (false positives), Leonardi and Van De Ville recently recommended choosing a window length exceeding the longest wavelength composing the BOLD signal, usually assumed to be ~ 100 s. Here, we provide further statistical support for this rule of thumb. However, we demonstrate that non-stationary fluctuations in functional connectivity can in theory be detected with much shorter window lengths (e.g. 40 s), while maintaining nominal control of false positives. We find that statistical power is near-maximal for window lengths chosen according to Leonardi and Van De Ville's rule of thumb. Furthermore, we lay some foundations for a parametric test to identify non-stationary fluctuations in functional connectivity, also noting limitations of the sinusoidal model upon which our work, and the work of Leonardi and Van De Ville, is based. Most notably, our analytical results pertain to covariances, as does our statistical test, whereas functional connectivity is more commonly measured using correlations.

Section snippets

Towards a statistical test

Leonardi and Van De Ville's 1/fmin recommendation corresponds to the smallest window length for which the sliding window covariance between two identical sinusoids is constant across all shifts in the window position; or in other words, the smallest window length for which there are absolutely no spurious fluctuations in covariance over time (see Fig. 1A). However, from a statistical viewpoint, some level of spurious fluctuations can be tolerated, particularly in the presence of system noise.

To

Limitations of the sinusoidal model

While Leonardi and Van De Ville's sinusoidal model is analytically tractable, it does not capture the BOLD signal's 1/f spectral distribution. The 1/f characteristic implies that covariance estimates are dominated by the lowest resolvable frequency, since it is this frequency that is of greatest amplitude. Dynamics arising from higher frequency BOLD fluctuations might therefore be overshadowed when using single-resolution approaches, or when modeling the BOLD signal as a sinusoid. In this

The need for generative null models

Regardless of the choice of window length, it is important to disambiguate fluctuations in connectivity dynamics of a neural origin from spurious dynamics arising from scanner drift, head movement (Van Dijk et al., 2012), variations in the respiratory volume/rate and cardiac rate (Chang et al., 2013) and non-stationarity in the fMRI data itself due to sleep, for example (Tagliazucchi and Laufs, 2014). Our test for stationary is indifferent to the origins of connectivity dynamics: The null

Conclusion

We concur with Leonardi and Van De Ville on the importance of recognizing and eliminating spurious connectivity dynamics due to inappropriate window lengths. From a statistical viewpoint, we suggest that their 1/fmin recommendation provides a good rule of thumb, but may be overly conservative in moderate SNR conditions. We contend statistical testing and appropriate surrogate data is crucial in this respect. We also contend that if dynamic fluctuations in connectivity are confirmed with

References (28)

Cited by (0)

View full text