> > You make the point that
> > smoothness estimation does not matter since
> you're
> > using using bootstrap methods to determine
> > significance, but the ad-hoc method will raise
> > concerns that the results will never hold up to
> > replication.
>
> Are you concerned about replicability of which
> nodes are connected to each other across
> hemispheres?
No, I am just concerned that reviewers would suspect you built this entire procedure to get this result to be significant. You will need to justify why it is OK to stitch the hemispheres at all. There is an argument to be made that you're stitching nodes that would be sampled by one voxel only, for example, as if you are using a voxelized gray matter ribbon. But I don't see a clean way to deal with this given that your analysis is actually on the surface.

>
> - the method is deterministic so in principle
> replicable if used on the same data.

I was thinking of new data replication

>
> > Would other correction methods,
> > such as FDR
>
> I have never understood why people use FDR for
> whole-brain correction - in my opinion the method
> is flawed.
> But that's another topic.

I don't understand why that is that objectionable, assuming you have p values worth something.

>
> > and FWE fail also?
>
> Do you mean Bonferroni correction? That is very
> conservative, no?

No I meant Family Wise Error correction, using a clustering threshold. I see you are using the threshold-free approach, but I don't know what to tell you there because I don't know how well it works under different situations, particularly since you don't have the data.

>
> Currently we use Threshold-Free Cluster
> Enhancement (custom implementation on surfaces),
> which is also quite conservative but has the
> advantage, compared to traditional cluster-based
> correction, that there is no 'free' parameter for
> an uncorrected (node-wise) threshold.
>
> > If your maps are
> > quite clean and the blobs fall quite nicely in
> > regions supported by prior independent
> findings,
> > and you can't get more subjects, you can show
> the
> > maps without multiple comparison corrections
> and
> > throw yourself at the mercy of the reviewers.
>
> Of course - but this sounds like double dipping to
> me: first see where the blobs are, then find
> lterature that mention those blobs, and then use
> that to justify these 'a priori' ROIs with no need
> for multiple comparison correction.

I have to differ with you on that. You always have a model for what you're looking for somewhere in your mind. You cannot call anytime you are citing converging evidence double dipping. The sequence of actions you mentioned is not what I was recommending. If your blobs fall where your hypothesis (generated from prior findings) suggests they would, then that is reason to think you have something real at hand.
Note that I did not suggest the finding would be statistically significant, if that were the case, but only that one be straight forward about it with the reviewers.

> In my opinion
> that's even worse than a somewhat related trick,
> 'small volume correction'.

I agree with you that 'small volume correction' should not be used. However I really differ with you that what I suggested is worse. I proposed you show the entire dataset at the uncorrected threshold and if 'activation' regions fall where they make sense, and not elsewhere, then you can suggest that you may be on the right track but under powered or using an inappropriate multiple correction method.
This is not the same as picking what parts of the data you want to show.

> Sure, you could argue that the same applies to
> this ad-hoc approach of connecting hemispheres,
> but arguably this is a lot less arbitrary and
> possibly even justifiable.

So what I proposed was actually the least ad-hoc. I just proposed you show the data as is, as a last resort, but it is not a way to establish significance. For what you propose to fly you have to motivate it by identifying why it is that merging the two hemispheres at the medial wall makes sense.

cheers,
Ziad