Hello John, Darren,
I think your remark about "people who know a bit more" scared pretty
much everyone off ;) so I start this with the disclaimer that I voice an
opinion despite, not because of, this remark.
I would agree, though, that using only one of the groups is a problem in
general. I think everyone should be treated as equally as possible
before entering an analysis, to the point that "equally unfair" may be
preferrable. I see Darren's point that you want to see differences "with
regard to only that group", but I would still be worried that you end up
comparing apples and oranges (i.e., data processed this and that way,
and that resulting differences arise from the differences in
processing). I would thus be hesitant to do this.
The effect of customized priors are difficult to predict a priori,
although I must admit that I do not know many exhaustive simulations on
this. Simon Keller concluded several years ago (PMID 15528086) that
custom templates do not have huge effect on the resulting SPM's, but one
of my contributions to the paper was to mention that this may not be the
case for unusual populations, such as children (my case) or elderly
subjects (your case). Also, the newer segmentation approaches have been
substantially improved since then, so it is hard to use that data as a
guidance today. Along those lines: one of the issues you naturally are
concerned about in the setting of an unusual population is the effect of
your priors on tissue segmentation. An elegant way out of this may be to
use the vbm8 approach that will not use priors for segmentation at all,
but will only use them to achieve spatial normalization. This way, you
can at least minimize the impact of your priors on the processing,
whichever way you decide to go.
Finally, having written the paper (PMID 18424084), I naturally think it
is worth a try to feed all the data into the TOM toolbox and then
extract those features that best statistically model age and gender
across your whole population. Having "children" in the title was
probably the worst decision when writing the manuscript, as the approach
is applicable in all cases when you have a population that deviates from
the "standard population" used to construct your prior. With the numbers
you mentioned, the result should be pretty stable.
So much for my contribution to the confusion :)
Cheers,
Marko
John Ashburner wrote:
> Hi Darren,
> I could be wrong, but isn't it generally better to treat all scans
> equally (ie not use any of the information about the columns of the
> design matrix in which you're interested) if the intention is to make
> statistical inferences about it? If a template is generated from some
> subjects (selected based on a categorical variable of interest), but not
> others, this would imply that the data are not all treated the same.
>
> I would like to hear what do others who know a bit more about stats
> think about this, particularly in regards to approaches such as lesion
> masking for spatial normalisation.
>
> All the best,
> -John
>
>
>
>
> On 17 October 2013 04:59, Darren G <[log in to unmask]
> <mailto:[log in to unmask]>> wrote:
>
> Hi
>
> I have some questions about creating a custom template for a VBM
> study. Reading through the list, I see that the typical advice is to
> use all the subjects for creating the template, as opposed to just
> using the controls, for example. I think this makes sense if
> subjects are largely from similar groups with equal numbers of
> subjects represented, but I am wondering what to do if I have about
> 90 healthy controls and about 150 patients. Since the patients do
> have some atrophy and both groups span a broad age range I think a
> custom template would help the analysis.
>
> However, I fear that using all the subjects will bias the template
> towards the patients. I could take a subset of the patients, but
> because of the broad age range the choice of which patients to pick
> is even more difficult.
>
> Why not use just the controls to create the template? I understand
> this will bias the template towards the controls, but since I am
> trying to detect differences between patients and controls this
> seems appropriate.
>
> What are the arguments for not just using the controls and is there
> any literature on this that has studied different ways of making a
> customized template, i.e., controls only vs. controls + patients?
>
> Any other suggestions for creating a template with this imbalanced
> group?
>
> Thanks
> Darren
>
>
--
____________________________________________________
PD Dr. med. Marko Wilke
Facharzt für Kinder- und Jugendmedizin
Leiter, Experimentelle Pädiatrische Neurobildgebung
Universitäts-Kinderklinik
Abt. III (Neuropädiatrie)
Marko Wilke, MD, PhD
Pediatrician
Head, Experimental Pediatric Neuroimaging
University Children's Hospital
Dept. III (Pediatric Neurology)
Hoppe-Seyler-Str. 1
D - 72076 Tübingen, Germany
Tel. +49 7071 29-83416
Fax +49 7071 29-5473
[log in to unmask]
http://www.medizin.uni-tuebingen.de/kinder/epn/
____________________________________________________
|