Hi,
On 10 Feb 2009, at 10:50, Colm Connolly wrote:
> Hi everybody,
>
> I'm about to perform several large VBM randomise (from fsl-4.1.2)
> runs and would like to clarify a few questions before I dive
> headlong in do this.
>
> I've got m+n subjects: m controls and n drug users. I'd like to look
> at what effect measures such as duration of drug use and education
> may have on GM volume. The drug users can be carved into two
> subgroups: short and long term abstinent users.
>
> I had intended to examine separately for each group the correlation
> of the variables of interest and GM volume.
>
> I've set up my regressors as follows
> Short Long Ctrl Age Short Age Long Age Ctrl DoU Short DoU Long Edu
> Short Edu Long Edu Ctrl Abst Short Abst Long
> 1 0 0 40 0 0 12 0 1 0 0 12 0
> 0 1 0 0 43 0 0 0.29 0 2 0 0 48
> 0 1 0 0 32 0 0 4 0 1 0 0 56
> 1 0 0 43 0 0 17 0 3 0 0 16 0
> 0 1 0 0 44 0 0 2 0 3 0 0 44
> 0 0 1 0 0 24 0 0 0 0 2 0 0
> 0 0 1 0 0 29.9 0 0 0 0 3 0 0
> 0 0 1 0 0 44.4 0 0 0 0 2 0 0
> 0 0 1 0 0 37 0 0 0 0 4 0 0
>
> The problem I have is that some of the variables of interest make no
> sense for the control users (they are effectively N/As), namely
> duration of use and length of abstinence. Hence there are two
> columns missing above, 'DoU ctrl' and 'Abst ctrl'. Should I include
> these columns in the matrix and if so what what values should I put
> in them? Should the values I put in these N/A columns for the
> control users be the mean value of all the users since the effect of
> a covariate is dictated by the it's distance from the mean?
You don't need EVs for covariates that don't make sense - just make
sure that the EVs for the other group are demeaned before being padded
with zeros for the control subjects.
>
>
> If I understand this correctly, the first 3 columns, encoding group
> membership model the mean. Can I therefore avoid demeaning the other
> covariates and not put the -D option on the randomise command line?
That's right.
>
>
> Also, I have education encoded as an ordinal, namely
> Did not finish high school 1
> High school grad 2
> Some college 3
> College grad 4
> Graduate school 5
>
> will the constitute a problem, especially if I have to demean the
> variables?
The problem is whether this works well in linear regression with these
values - presumably the correlation will be much more complex than
just linear, but to a first approximation it may be informative. The
alternative (which is probably overkill given the likely underlying
true correlation) is to have one EV for each of these values.
Cheers.
>
>
> Thanks for your time,
>
> Bye,
> --
> Dr Colm G. Connolly
> School of Psychology and Institute of Neuroscience
> The Lloyd Building
> University of Dublin
> Trinity College, Dublin 2, Éire
> Fax: +353-1-671-3183
>
> Please note that electronic mail to, from or within the Trinity
> College Dublin, may be the
> subject of a request under the Freedom of Information Act.
>
---------------------------------------------------------------------------
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
---------------------------------------------------------------------------
|