Hi Michael,
One subtlety that may help inform your thinking is that how you model
gender will have no impact on the differential contrast between groups.
So, if that is the only contrast (cope) that you are interested in, then
it doesn't matter if you model gender as (1, -1) or (1, 0), or either of
those with the mean subtracted. How you model gender only affects the
interpretation of the "mean" signal associated with each individual
group (i.e., EV1 or EV2 by themselves in your example, but not EV1-EV2)
Eugene's final post to the initial thread that you referenced explains
this nicely.
cheers,
-MH
On Thu, 2011-06-16 at 07:19 +0100, Stephen Smith wrote:
> Hi
>
> On 15 Jun 2011, at 17:38, Michael B wrote:
>
> > Hi experts,
> >
> > I'm setting up a design matrix, whose aim is to model a group
> > comparison, while controlling for a single covariate. I'm making a
> > single design matrix for each of several covariates. Some of the
> > covariates are continuous (age, disease duration, education, IQ)
> > while others are categorical [gender (male, female), disease sub-
> > type (A, B, C), medication (yes, no), depression (yes, no)]. I'm
> > confident with my matrices with the continuous variables (e.g., I
> > demeaned each subject's age with the mean age of all subjects and
> > included these demeaned ages in a single EV). However, I'm not sure
> > about my matrices with the categorical variables.
> >
> > Presently, my matrices with categorical variables look like this:
> >
> > GENDER: male = 1, female = -1.
> >
> > EV1 (group 1) EV2 (group 2) gender
> > 1 0 1
> > 1 0 -1
> > 1 0 -1
> > 1 0 -1
> > 0 1 -1
> > 0 1 1
> > 0 1 -1
> > 0 1 -1
> >
> > However, I recently read the following on the forum:
> >
> > "The non-zero means of the sex and hand evs will affect the
> > interpretation of the contrasts for A and B (because some of the
> > mean is accounted for by these extra evs). Use -1s instead of 0s
> > for the sex ev entries so that it's mean only reflects the imbalance
> > between genders. " (Demeaning covariates in an unpaired two group
> > comparison; Fri, 18 Feb 2011 17:40:24)
> >
> > The subjects that constitute each group are patients of the same
> > disease. Importantly, this disease is about 2 times as prevalent
> > among women than it is among men. In my sample, I have 29 males and
> > 55 females. Does having two times as many female subjects as male
> > subjects influence the way in which I should code for gender? That
> > is, is my goal to make the mean of the gender covariate equal to 0?
> > For example, should male = 2 and female = -1?
> >
>
>
> Right - but you need to make it *exactly* zero mean - just subtract
> the mean value from all the original values.
>
>
> > DISEASE SUB-TYPE: A = 1, B = 0, C = -1
> >
> > EV1 (group 1) EV2 (group 2) disease sub-type
> > 1 0 1
> > 1 0 1
> > 1 0 1
> > 1 0 0
> > 0 1 1
> > 0 1 0
> > 0 1 1
> > 0 1 -1
> >
> > I read the following in the same message:
> > "For handedness, if you only have one or two lefties you might want
> > to keep that one as is (so it just models away the effect of left
> > handedness). This won't have an effect on the A-B contrast. "
> > (Demeaning covariates in an unpaired two group comparison; Fri, 18
> > Feb 2011 17:40:24) - In the user's design matrix, handedness was
> > coded as follows: right = 1 and left = 0. The user was advised to
> > maintain this coding, instead of changing left to -1, if there were
> > few left-handed subjects.
> >
> > There is a very uneven distribution of disease sub-types in my
> > sample, which is a relatively accurate approximation of the
> > distribution in the patient population. In my sample, I have 60
> > subjects with A, 18 subjects with B, and 6 subjects with C. Is the C
> > group so small that I should "just model away the effect" of this
> > sub-type? Again, is my goal to make the mean of the disease sub-type
> > covariate equal to 0? For example, should A = 1, B = -3, C = 0?
> >
>
>
> The most general model is to have one EV for each sub-type - so one EV
> for every combination of group and disease sub-type.
>
>
> If you think that 6 in group C is too low, maybe either exclude those,
> or merge them with a different sub-type.
>
>
> Cheers
>
>
>
> >
> > I would really appreciate some guidance!
> >
> > Regards,
> > Michael
> >
> >
>
>
> ---------------------------------------------------------------------------
> 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
> ---------------------------------------------------------------------------
>
>
>
>
>
>
>
|