I often find myself running regression analysis, using either a directly related behavioral index (such as task performance inside the scanner) or a related cognitive score (WM, etc) obtained outside the scanner. 

I have often noted that regression analyses give me clusters which appear to be inside white matter, or within or on the boarders of the ventricles. Now, in a standard GLM analysis I would suggest this may be motion artifact. I am not as sure it is in the case for the regressions. I am often regressing my behavioral measures on a first-level contrast (condition A - condition B), which means the detected regions show a relationship between the difference in these conditions and my behavioral variable. Furthermore, these activations are generally contrast specific (as in, I see it in A-B, but not C-D, or A-C, or whatever, and none of these conditions should evoke more motion than others). 

So positing a motion relationship would require a complex argument about the relationship between conditions, cognitive scores, and motion, which may be specific to certain kinds of trials. It is possible ('lower' cognitive functioning people move more, which may very well be true). But not a very satisfactory argument in terms of parsimony. 

Any other alternatives come to mind? Are regressions with behavioral scores more prone to false positive than standard GLM analyses? I imagine outliers is a greater issue, as I know from experience 1 or 2 outliers can have a dramatic effect on statistics in these analyses.So I am careful to exclude anyone who's behavior scores deviate to much from the norm. 

I am working on some regression data right now and have some nice results, and some white matter clusters which I find disturbing, and am interested in the opinions of others on the list as to how to deal with such an outcome. The white matter clusters really make me question the validity of my results. 

Since I am sure it will be asked, I am using p < 0.001 uncorrected voxel threshold, corrected (p < 0.05) with an extend threshold (48 voxels at 2x2x2 resampled, 4x4x4 original voxel size). I also attached an image of an example result, although note the extend threshold is not applied so some little 'specs' are still visible. 

Thanks for any advice or insight.