Non-stationary occurs due to non-isotropic (non-uniform) smoothness of VBM data. Because cluster size distribution varies depending on local smoothness some clusters tend to be large in smooth areas, while in rough regions, clusters tend to be small. This leads in particularly in VBM to invalid cluster size statistics and clusters sizes will be over- or underestimated.
Worsley et al. (1999) proposed to adjusts cluster sizes according to local roughness of images. This local roughness is provided in SPM as resel per voxel (RPV) image and is used to warp or flatten the image into an isotropic data space. The size of each cluster will be corrected according to the local smoothness values (using the RPV image) and the corrected sizes are indicated in the results table.
I have implemented the method of Satoru Hayasaka et al. (2004) and use some functions of his toolbox (thanks for the permission). I have integrated these functions in the VBM toolboxes. The non-stationary correction can be called via:
Toolbox|VBM5.1|Results with non-stationary correction.
In older versions I had integrated this functionality in the standard SPM results interface. However, the use of the same file names were causing several problems and now the correction is only available in VBM5.1 and VBM2. Furthermore, older versions have not used the function stat_threshold.m by Keith Worsley. This function accounts for the extra noise from estimating the smoothness locally, while spm_P.m assumes an almost perfect RPV image.
Additionally, you can also define the cluster size in terms of statistical significance. If you enter values < 1 as extent threshold these values will be interpreted as p-values of the cluster size statistic. You can either apply family-wise error (FWE) or uncorrected thresholds to define the cluster size. This correction is also implemented in the 'Threshold and transform spmT-maps' function.