Dear Lena,
I don't like the current approach. As recently argued https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;22c0ac2e.1507 it should be okay to take some summary estimate from significant clusters of e.g. the Flexible factorial and to conduct post-hoc tests, trying to explain the interaction.
However, if you have the feeling it's necessary to control for other variables (age, gender, ...) you should do so right from the beginning. Right now you run an ANOVA with SPM, detect some clusters & extract some average estimate, forward this into another ANOVA, this time with covariates, and only then the post-hoc tests, which seems to be quite exploratory / circular in part. And you might miss clusters in the whole-brain analysis associated with an interaction or main effects as you don't control for these other variables.
As long as you account for multiple testing properly (in this case the initial voxel threshold would have to be adjusted to take into account the number of models, e.g. instead of .001 you would go with .001/3 for the three "interaction models", see below) you could go with a series of separate models:
- For main effect group: One-way ANOVA based on the average contrast estimates = (2s + 5s + 8s)/3, the F contrast would be
[1 -1 0;
0 1 -1]
- To mimick main effect condition you could set up three One-way ANOVAs (factor group) based on the contrast estimates (2s - 5s), (2s - 8s), (5s - 8s), the F contrast would be
[1/3 1/3 1/3]
- Within these models you could also test for "interactions", e.g. in the One-way ANOVA based on contrast estimates (2s - 5s) the F contrast
[1 -1 0;
0 1 -1]
would be about group (A, PC, HC) x condition (2s, 5s) and so on.
In these One-way ANOVAs you could also add covariates to control for e.g. age, gender, ...
> I couldn’t include IQ as a covariate in the SPM or SPSS analysis
Why?
> I showed that the findings remained significant
As you analyze only a subset of subjects it doesn't mean this would hold if you looked at the whole group of subjects.
Best
Helmut
|