An underlying assumption of the above parametric approach is that the process is a = Gaussian field, i.e., its statistical characteristics, including its roughness parameter (or its reciprocal, the smoothness ) ), are the same at each point in the image. The FWHM of the process should be constant in all directions and across all voxels in the image. While these assumptions are reasonable for functional imaging data, they are likely to be violated for structural imaging data. Binary structure masks, for example, are constant across large regions, and even after smoothing the signal changes more rapidly at the edges of structures . The distribution of cluster sizes that occur by accident is therefore considerably skewed towards larger cluster sizes in smooth regions of the image, resulting in more false positives (and false negatives in rough regions) than predicted by formulae for stationary fields. To address this, , suggested a statistical flattening approach in which the data are warped into a new space, which may have higher dimension than the data, so that in the new space the smoothness of the normalized residuals of the statistical model is stationary. The " -value for cluster sizes above a threshold can then be applied using size measurements in the new space, or by estimating the effective resolution of the field directly from the normalized residuals .