Hi,
On 29 Jun 2007, at 00:32, Anna Engels wrote:
> Hello,
> I have several questions about setting up a HLA with covariates (of
> interest). I have a single group, with scores from 3 different
> questionnaires (A, B, & C). I would like to know which brain areas
> are
> associated with responses on each questionnaire (for a given lower-
> level
> contrast). Questions below:
>
> 1) Am I correct that each questionnaire score needs to be demeaned
> (separately for each questionnaire)? If I am not interested in the
> group
> mean (EV1), is it necessary to demean the scores for each
> questionnaire,
> or can I input the raw questionnaire scores?
Indeed - if you don't orthogonalise EV2 wrt EV1 then it is only the
fitting of EV1 that is affected. Hence you don't need to demean them
- though we generally recommend doing that kind of thing anyway as
it's often useful to be able to interpret the group mean EV fitting
as a sanity check or as a pre-mask on the other EVs to increase
sensitivity. It makes no difference to the EV2 significance testing
whether you orthogonalise (in this case demean) it or not.
> Assuming I would need to demean the scores, this is how I would set
> up the
> EVs and contrasts (using a small number of subjects as an example).
>
> EVs:
> Group EV1(group mean) EV2(Ques A) EV3(Ques B) EV4(Ques C)
> 1 1 1.4 3.2 3.2
> 1 1 5.4 .2 1.2
> 1 1 2.4 -1.8 -3.8
> 1 1 -4.6 3.2 -3.8
> 1 1 -4.6 -4.8 3.2
>
>
> Contrasts:
> Contrast EV1 EV2 EV3 EV4
> Questionnaire A 0 1 0 0
> Questionnaire B 0 0 1 0
> Questionnaire C 0 0 0 1
Looks good.
> 2) If in addition I am interested in looking at the interaction
> effects, I
> think I would add 4 more EVs to this model (consisting of the
> following
> multiplied scores) set up the orthogonalizations as specified
> below, and
> add 4 corresponding contrasts. Does this look like the correct way
> to set
> up the orthogonalizations (and interactions)?
>
> Add 4 more EVs:
> EV5 = QuestionnaireA * QuestionnaireB
> EV6 = QuestionnaireA * QuestionnaireC
> EV7 = QuestionnaireB * QuestionnaireC
> EV8 = QuestionnaireA* QuestionnaireB * QuestionnaireC
>
> Set up orthogonalizations:
> Orthogonalize EV5 wrt EV2, EV3, & EV4
> Orthogonalize EV6 wrt EV2, EV3, & EV4
> Orthogonalize EV7 wrt EV2, EV3, & EV4
> Orthogonalize EV8 wrt EV2, EV3, EV4, EV5, EV6, & EV7
I think I would simply demean them before multiplying them together,
and then not worrying about any further orhogonalisation.
> Add 4 more contrasts:
> Contrast EV5 EV6 EV7 EV8
> QuestionnaireA*B 1 0 0 0
> QuestionnaireA*C 0 1 0 0
> QuesionnaireB*C 0 0 1 0
> QuestionnaireA*B*C 0 0 0 1
>
>
> 3) If the questionnaires are not on the same scale, is it necessary to
> transform the scores to z-scores? If we do not z-score the
> questionnaire
> scores, are the interactions interpretable?
It is not necessary for them to be on the same scale UNLESS you
intend to carry out contrasts BETWEEN them. Everything you've said
above is unaffected by arbitrary scaling of any of the EVs.
Cheers.
>
>
> Thanks for your help
>
> ~Anna
------------------------------------------------------------------------
---
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
------------------------------------------------------------------------
---
|