Dear Christopher,
> I have scans of two conditions in a set of subjects, but those
> subjects can be divided into two groups-- smokers and non-smokers. I
> was trying to phrase this for a more general case, i.e. looking to
> account for varinace due to gender or race or what have you. The
> question remains for me whether it is better to add one or two dummy
> variables for such.
>
> This question arose from an effort to look at my data in terms of a
> traditional two-way anova. That is, I have two groups, each with two
> conditions-- a split-plot design. while I can test for effects greater
> in one than the other (e.g. [-1 1 1 -1]), I am still struggling with
> the interpretation of such a plot, as it is not a a standard
> "interaction effect".
This contrast is a 'standard' interaction. It may be that you are used
to seeing this effect assessed with the F ratio. The T value assesses
the equivalent effect but retains sensitivity to the direction of the
interaction.
> I thought it would also be useful to look at images that approximate
> "main effect of condition" and "main effect of group". For the first, I
> used [-1 1 -1 1] and its inverse, looking for consistent increases or
> decreases across the groups. While SPM cannot model a group effect
> [using the fixed effects available in SPM96, at leaset], I have the
> idea that this contrast WILL acccount for the fact that there are two
> types of changes here-- is that overly optimistic?
In fact the group effect is modelled (by the subject-effects in the
confound partition of the design matrix). Your contrast for the main
effect of condition is eactly right. The contrast for the main effect
of group [1 1 -1 -1] although estimable requires a random-effects
analysis.
> For the second case, "effect of group", I had planned to compare all
> scans from one group with all of the other, in a
> single-subject/replications design. Given that each group contains
> repeated measures from the same subject, I had hoped to account for
> this by entering "condition" as a confound for each scan. an
> alternative, it seems, is to average the two scans from each subject
> and then compare these.
The latter is equivalent to a random-effects analysis and is the
appropriate way to proceed (i.e. reduce the data to one observation per
subject and simply compare the two groups.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|