Dear Christine,
> At equal CORRECTED thresholds (P=0.05), there were (more and
> especially) larger clusters with spm96 (identical smoothing!).
> altogether, activation appeared to be more robust with spm96.
>
> and taking into account that the spatial preprocessing and thus
> smoothness of the data was really identical, it seems to me, that the
> smaller clustersize can only be due to a more stringent correction
> algorithm in SPM99. However, Matthew pointed out in his tutorial that
> excess false positives in spm96 especially occur at low degrees of
> freedom (<40). In my special case this was 56.6 (spm96) and 59.4
> (spm99) respectively, and this should not be to small?
Your analysis is absolutely right. The only difference between SPM96
and SPM99 (in terms of estimtation and inference) is that the spatial
smoothness estimator has been upgraded. This upgrade accommodates
non-stationary spatial correlations among the errors when correcting
for the search volume using tests based on peak height (but not on
those based on spatial extent, at this stage). This advance (due to
Keith Worsley) allows one to be more confident about corrected
inferences in situations where the smoothness may change from region to
region (e.g. VBM). Practically speaking the only thing that will
change is the corrected P value and the effective FWHM of the component
fields (found in the footnotes of the results tables). From your point
of view you will find that (i) this FWHM is smaller or (ii) the search
volume is bigger.
> Shall I conclude from this that spm96 is not more sensitive to real
> activation (even if it looks reasonable) but more susceptible to false
> positive activation and should thus not be used any longer? Is there a
> general recommendation what to do if I find significant activation at a
> corrected level with spm96 (which seems reasonable), which does not
> appear in spm99 with equal significance?
Having dealt with this question internally with every new version of
SPM our recommendation is simple. Analyze your data with, and only
with, the latest version. The latest version is [hopefully] the most
robust. Differences among versions can be ascribed to differences in
robustness to violations of the assumptions on which SPM is
predicated. These are:
For SPM96
1) Multivariate Gaussian distribution of the error terms.
2) Differentiable spatial autocorrelation function
3) Data conform to a reasonable lattice representation of an
underlying field (i.e. voxels are small relative to smoothness)
4) Stationary autorcorrelations among errors
For SPM99 the last assumption can be dropped. I suspect the difference
in your case is related to this. I should note that SPM99 and SPM96
have been compared and, with no violotations of the above, give the
same results.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|