On Fri, 28 Nov 2008 05:02:28 +0000, Amit Etkin <[log in to unmask]>
wrote:
>Hi all,
>
>I've recently run a 2x2x2 ANOVA, where I find very significant results through
>SPM, but have very confusing results when extracting values for significant
>voxels. The ANOVA (see attached design matrix) has 2 within subject
factors,
>and 2 groups of subjects (N=17 and 16). I'm trying to look at the effect of
>group and specified the following contrast vector:
>
>ones(1,17)/17 -ones(1,16)/16 1 -1 0 0 0 0 .5 .5 -.5 -.5 0 0 0 0 .5 .5 -.5 -.5
>
>Doing so gives me a pattern of results that are echoed in effects of group
>when I look at ANOVAs across only one within subject factor (ie 1/2 of the
>current ANOVA), as well as when I look at t-tests between groups across
one
>level of a single factor....so that all makes sense. The strange part is that if
I
>extract signal from a peak voxel or cluster in the group2>group1 contrast,
>using the eigenvariate (VOI) function, I get data that makes a lot of sense
(4
>values per subject, significant ANOVA results when put into SPSS, etc).
When
>I extract, however, a highly significant peak voxel or cluster from the
>group1>group2 contrast, in the same way as I did the reverse, the data no
>longer make any sense....there's no longer any significance at the t-test or
>ANOVA level when put into SPSS.
>
>It seems I must be doing something wrong, but not sure what it is. Any and
all
>thoughts welcome!!
Are you sure that it's not an effect of thresholding? Maybe it's nearly
significant but not quite so when you put it into SPSS.
Also, there are minor differences between SPSS and SPM. E.g. SPSS will
probably do multiple comparison corrections for the number of tests in an
ANOVA; SPM won't. (SPM will do corrections for the number of voxels, but
not number of contrasts.)
Also, SPM (and the way it's typically used) uses a so-called pooled variance
model in ANOVAs with both between- and within-subject factors. Usually not
the way SPSS etc work in the default setup, as far as I know. (Not quite sure
that's at issue with factors with only two levels and effect of subject
included.)
These issues might make a difference, particularly in the context of
significance, where there's the additional impact of thresholding.
>
>thanks,
>Amit
>
|