September 21, 2009

Well, it had to be said.

Just one sentence from the 2003 paper in Human Brain Mapping on "Model Assessment and Model Building for fMRI" by M. Razavi, T.J. Grabowski, W.P. Vispoel, P. Monahan, S. Mehta, B. Eaton, & L. Bolinger:
Thus, while in the current fMRI time series analyses [i.e., what pretty much everybody else is doing, pretty much all of the time] the model that leads to the highest number of activated voxels is selected, proper modeling necessitates selecting [instead] the model with the highest quality, i.e., a model that is valid and has the ... highest goodness of fit, even if the number of activated voxels is lower; and discard the model that is not valid or has a statistically insignificant or lower goodness of fit, even if the number of activated voxels is higher.
It is remarkable that this had to be said at all. Even more remarkable is that so much of what is done in fMRI still happens without consideration of this issue.

No comments:

Post a Comment

Post a Comment

ShareThis

Blog Archive