hi Gang

thanks so much for your advice. I just want to clear up a couple of points:

Quote

1) To avoid any double-dipping accusation, it would be better to just run voxel-wise analysis on the whole brain instead of focusing on individual ROIs. With a one-way repeated-measures (or within-subject) ANOVA with three levels, you can simply run three separate paired t-tests, which can also be obtained through -adiff or -acontr in 3dANOVA2.

OK, so from an implementation standpoint we run the 3dANOVA2 command like so:

3dANOVA2 -type 3 -alevels 3 -blevels 21 \
-dset (bunch of dsets here) \
-adiff 1 2 NCvsLC \
-adiff 2 3 LCvsHC \
-adiff 1 3 NCvsHC \
-fa Context \
-bucket ANOVA_results

So now we look at the main effect of Context and we see a cluster in the angular gyrus. The question remains, what is causing the significant F. All three conditions could differ from one another or it could be one condition driving the effect. I would (really!) like to provide a valid answer to that question, but I still don't see how without doing some averaging over the voxels in the cluster identified by the omnibus F. My plan was to compare the differences between conditions in the cluster via averaging but I am getting the impression that I can't do this? Specifically, the hypothesis we actually had was that HC > LC, and the comparison of LC vs NC is really meant as a type of sanity check. I get that defining a cluster by HC > LC at the whole-brain level biases the LC vs NC test but I feel as though the cluster from the ANOVA is not inherently biased in the same way since really any difference between means could have caused it.

Quote

2) The phrase "significantly different" in your description should be interpreted in the statistical, not practical, context. What this means is that "HC and LC are significantly different from 0" is meaningful because there is a significance level (like a p-value) associated with it. However, "the differences between LC and NC are not significantly different from 0" is *not* meaningful even though everybody seems to know what you mean! What I'm trying to say is that, the practical difference LC and NC could be big (e.g., 0.65% signal change) but that difference might have failed to reach a desired significance level due to power issue or some other reasons. So it would be difficult to statistically show the equality part in the relationship "HC > LC = NC".

Yes, this is what I had in mind. I am not trying to make a strong assertion about the null. It's just that it seems to me there are two interesting possibilities (among other ones) and I would like to tell them apart. One is that a region responds to both types of context in a linear fashion (and since HC as more context than LC you get an increase such that NC < LC < HC), and the other is that it responds specifically to the information in the HC condition.

Quote

it seems that you want the interaction effect from a 2x2 within-subject ANOVA with two factors, condition and region. Then you can claim the contrast between the two conditions at one region is greater than the other region at a designated significance level. Again be careful about statements like "LC vs NC nonsignificant in one cluster".

This is where I get confused. To run this ANOVA I would have to average within the cluster. So why can I not average within the cluster to compare conditions after the whole-brain ANOVA above but can average to run this ANOVA? I feel as though I am missing something.

thanks again

James