Dear John,
Thank you for your quick reply and useful suggestions.
Please let me ask you several questions regarding part
of your comments.
> 1) Also use a symmetric weighting image for the spatial normalisation.
> This can be specified by modifying the spm_defaults.m file at
> around line 191 so that it uses a weighting image that has also
> been left-right flipped and averaged.
I understand this can be done by setting sptl_MskObj = 1
in the spm_defaults.m file. But I am not sure which image
should be chosen as "object masking image". An average
image between L-R flipped and unflipped images of the
object? Should this image scaled so that the contained
values will be between 0 and 1?
> 2) Remember that rotations, zooms and translations are also included
> in the deformation fields - not just shape. In order to discount
> the effects of subject positioning in the scanner, you will need
> to somehow normalise the deformation fields for subject position.
> Also think about whether you want to normalise out head size from
> the deformations.
These factors can be discounted by applying two-step
normalization, namely, performing affine transformation
to all images first, and then performing non-linear
normalization on these affine-transformed images.
Correct?
> 4) There is a hidden option within SPM99, that although not yet fully
> tested, is nevertheless there in anticipation of doing multi-variate
> stats. It computes a voxelwise Wilk's Lambda, which is transformed
> to a Chi^2 statistic. For the random field theory, the transformed
> statistic field is assumed to be a Chi^2 field (I think), with the
> corrected p values based on these assumptions. The multivariate
> version is accessed by typing the following into Matlab (after
> invoking SPM):
> spm_spm_ui
Great. Yet, as far as I can see, the basic stat menu has
only 1-way option for MANOVA. Are there any tricks to
implement 2-way MANOVA with this routine?
> Note that a Wilk's Lambda field transformed to a Chi^2 field is
> not exactly the same as a pure Chi^2 field. The corrections for
> numbers of resels is therefore not exact, but I'm sure it's still
> very close.
Just a casual idea. In special cases of 1-way MANOVA,
a Wilk's Lambda field can be exactly transformed to a
F field (Johnson & Wichern, Applied multivariate
statistical analysis, 4th ed., 1998, p323). I guess
Worsley's theory can be exactly applied in those cases.
(I am not a statistician. Please someone correct me if
wrong.)
Many thanks, in advance.
Kota KATANODA
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|