Functional neuroimaging is an important tool in psychiatry. While collecting a sufficient number of subjects is a challenge for any medical research domain, researchers in psychiatry face additional challenges; they must find patients who will volunteer to be injected with radioactivity (PET) or be contained in a coffin-like space (fMRI). The result is that analyses are often performed with low degrees of freedom at each voxel.
The standard approach to the multiple comparisons problem is to use the theory of random t fields to find a threshold which controls the familywise error (FWE) over the whole brain. I present evidence that such an approach is quite conservative for low degrees of freedom, even when the assumptions appear to be satisfied. An alternative, nonparametric approach is to permute the data to build empirical null distributions of interest. While parametric tests are usually more powerful than corresponding nonparametric tests, I find that the permutation test consistently outperforms the random t field results for analyses with low degrees of freedom.
|