Hi
On 13 Feb 2008, at 22:19, Joseph T. Devlin wrote:
> Hi Yvonne,
>
> Since no one more statistically savvy than me has answered as yet,
> I'll poke my head out and say "no" -- FSL doesn't offer an
> implementation of the "minimum statistic" implementation of
> conjunction analyses. To be honest, though, there are real
> problems with interpreting those statistic values, as pointed out
> by Tom Nichols and colleagues. In many cases it is sufficient to
> use inclusive masking. If the contrasts are truly orthogonal, then
> you can reasonably multiply the p-values together to get a
> "conjunction" p-value in a given voxel. For instance, if you were
> interested in areas in Group 1 where A>B and masked it with Group
> 2's A>B (at p<0.001 uncorrected to pick a random threshold...),
> then any areas at p<0.001 in the G1 A>B contrast that are also
> present in G2 A>B are unlikely to be there by chance given that the
> corresponding p-value would be p<10^-6.
>
Just to note that even this is a bit misleading, because p<10^-6
refers to the null hypothesis that NEITHER of the activations were
present - you certainly can't claim a conjunction with this p-value.
I agree with Joe that just showing the overlap is fine.
T
> Personally, I wouldn't feel the need to see a computed p-value --
> it would be sufficient to know that the "conjunction" contrasts
> were truly orthogonal and know the level that each one was
>
|