Controlling the familywise error rate in widefield optical neuroimaging of functional connectivity in mice

Statistical inference in functional neuroimaging is complicated by the multiple testing problem and spatial autocorrelation. Common methods in functional magnetic resonance imaging to control the familywise error rate (FWER) include random field theory (RFT) and permutation testing. The ability of these methods to control the FWER in optical neuroimaging has not been evaluated.

We attempt to control the FWER in optical intrinsic signal imaging resting-state functional connectivity using both RFT and permutation inference at a nominal value of 0.05. The FWER was derived using a mass empirical analysis of real data in which the null is known to be true.

Data from normal mice were repeatedly divided into two groups, and differences between functional connectivity maps were calculated with pixel-wise $$ -tests. As the null hypothesis was always true, all positives were false positives.

Gaussian RFT resulted in a higher than expected FWER with either cluster-based (0.15) or pixel-based (0.62) methods. $$ -distribution RFT could achieve FWERs of 0.05 (cluster-based or pixel-based). Permutation inference always controlled the FWER.

RFT can lead to highly inflated FWERs. Although $$ -distribution RFT can be accurate, it is sensitive to statistical assumptions. Permutation inference is robust to statistical errors and accurately controls the FWER.