 |
 |
Yes, that's it exactly, Joe. So the only question that remains is why software doesn't support the Poline et al. or similar methods or if anyone knows of this software support, please enlighten me...thanks for that reference :-) Cheers, Darren ----- Original Message ----- From: "Joe Devlin" <[log in to unmask]> To: <[log in to unmask]> Sent: Saturday, May 03, 2003 6:12 PM Subject: Re: [FSL] Higher Level Analysis Thresholding Wisdom Needed > Hi Darren, > > > the z maps are an indication of whether or not the coefficient for a given > > functional model is greater than zero. > > Yes indeed -- z-maps show whether a linear contrast of parameter estimates > (ie the betas in the GLM) is greater than zero. If the contrast vector > contains just a single 1, then this test whethers the individual > coefficient is greater than zero. > > > Can the b values be converted to correlation coefficients or %variance > > accounted for (R^2)? > > I'm sure they can because I've seen an equation for converting t-values > into Pearson moment correlation (r-values) although I don't remember it > off-hand. Perhaps someone more statistically savvy could help out with > this? > > But because the two are equivalent, they provide exactly the same > information. In both cases, one is testing the fit between the model and > the observed data -- either as a t- or r-value. In fact, the early > Bandetinni et al (1993) paper proposed doing the analyses as a correlation > and some groups still report their results in that fashion. BTW, the > reason that Z-values are typically reported rather than t- or r-values is > that Z-values don't require degrees of freedom. A Z of 3.1 is always the > same whereas a t-value of 3.1 or r-value of 0.5 might be significant with > sufficiently high degrees of freedom. > > > That is, I wonder how to weigh up the spatial extent vs the > > degree of fit. Surely it is possible to have a very focal, isolated > > activation with a very high degree of fit. Maybe this is implicit in the z > > scores already, although a correlation coefficient might be more readily > > understood. > > This is an important issue. Cluster tests are based solely on spatial > extent and ignore the height of the voxels within the cluster. Voxel > stats (which is the z-value per voxel) are based solely on the degree of > fit at that voxel and ignore spatial neighborhoods. Jean-Bapiste Poline > and colleagues had a paper in NeuroIMage in 1997 (I think) where they > proposed a framework for combining these two types of information. I > thought it was pretty reasonable but I've never seen anyone use it and > none of the main software packages implement the idea. > > To illustrate the point you make, imaging that your doing an auditory > experiment and you expect activation -- among other places -- in the > medial geniculate nucleus of the thalamus. This is a small region where > you expect high activation (high z-values). The neighboring > thalamic nuclei, however, are not responsive to auditory stimuli so you > wouldn't expect much in the way of spatial extent. In this case, a > cluster statistic is almost certainly the wrong approach because you > simply won't find a big enough cluster given the volume of active tissue > -- even though there is good reason to believe the activation. This is > preceisely the type of problem Poline et al. tried to deal with. Note, > though, that reporting the correlation value here would be the same as > reporting the t-value (or Z-value). In both cases you're doing the right > thing in this particular case -- reporting a voxel-level statistic. > > Hope this was some help. > Cheers, > Joe >