Spatially restricting analyses

Although in a perfect world all interesting results would reach a whole-brain corrected level of significance, in practice, this is often not the case. One productive way of dealing with this issue is to incorporate prior knowledge into our analyses in the form of spatial restriction. Performing fewer statistical tests in this restricted space increases sensitivity for detecting effects, because fewer comparisons need to be corrected for. Although this approach is necessarily limiting, the fact that it requires a more specific anatomical hypothesis can also be a helpful encouragement in developing anatomically specific theories related to our experiments.

It’s a very good idea to spend some time thinking about spatial hypotheses, and how to formalize them, before conducting analyses. This helps guard against the bias that is necessarily introduced by having a peek at whole-brain results.

Using a an explicit mask to constrain voxelwise analysis

In a voxelwise analysis, reducing the number of voxels reduces the number of comparisons that need to be controlled for, and can thus increase your sensitivity. You can do this by specifying an explicit mask during model estimation, results reporting, or as a small-volume correction (SVC). Although conceptually similar, these methods differ somewhat in practice (see below).

Explicit masks for voxelwise analyses can be generated in several ways:

  1. A gray matter mask to restrict analyses to gray matter voxels (see Creating an explicit mask).
  2. A region (or combination of multiple regions) from a standard-space atlas.
  3. A thresholded statistic map that is independent of the current analysis.


The different methods of explicitly restricting an analysis all limit the voxels examined; however, they also affect the smoothness estimation that SPM does, which will affect the cluster-level corrections. If an explicit mask is specified during model estimation, the voxels outside of the mask are never considered, and so smoothness is just estimated based on the in-mask voxels. If the mask is specified during results reporting or as an SVC, the smoothness is based on the entire image.

Conducting a region of interest (ROI) analysis

Averaging over voxels in a region is another way of approaching spatial restriction. One common application is when the region in question is one that is functionally defined (e.g., the “fusiform face area”), and in which an independent “functional localizer” session can be run to identify this region in each participant. However, ROIs can also be defined using anatomical atlases or previous results. MarsBaR ( is a very handy toolbox for conducting ROI analyses in SPM.

Supplementing restricted analyses with whole-brain analyses

The advantage of conducting analyses that don’t consider the whole brain is that these analyses are potentially more sensitive due to the decreased number of tests that need to be corrected for. The disadvantage is that because they are spatially restricted, they can miss significant results. This might happen for a number of reasons, including incorrect assumptions that went into the way to restrict an analysis. Thus, it may be beneficial to conduct an additional whole-brain exploratory analysis to supplement restricted analyses. By using a more lenient threshold and/or reporting effect sizes, this can help identify other possible areas for future inquiry and add some context within which to interpret restricted results.

Avoiding nonindependence and circularity

Whenever you restrict your analysis, it is essential that you do not do it in a way that might bias your results. That is, the results you use to restrict an analysis must be independent of that analysis. Using anatomical regions based on an atlas, or results based on an independent set of data, are both straightforward ways to do this. See [Kriegeskorte2009] for more.