Dear Paul,
> I have a few questions concerning conjunction analyses. I have read the
> mail archives and actually am now more confused. I would really
> appreciate any input that you may have.
>
> I have a study with 9 subjects. I would like to do a conjunction
> analysis but have several questions.
>
> 1. Should corrected or uncorreted p values be used for the analysis.
> The mailbase continually refers to uncorrected p values.
In SPM99 a conjunction SPM comprises the minimium t values of the
component SPMs. These minimum t values have their own distributional
approximation which allows one to compute both corrected and
uncorrected p values, just like ordinary SPMs. The criteria for using
corrected or uncorrected inference is exactly the same as for any other
SPM.
> 2. In relation to question 1, should a height threshold be used. I have
> worked with SPM and realize that extent thresholds cannot be used with
> conjunction analyses.
The distributional approximations for the spatial extent of a
conjunction SPM are not known (at present) and therefor inference based
on spatial extent is precluded. Consequently height is currently used
to specify thresholding. This can be corrected or uncorrected and both
pertin to the final significance of the conjunction SPM (Pconj) (not the
components - in SPM99b the uncorrected height threshold refered to the
components).
>
> 3. What p value is most appropriate? A recent mail from Karl stated
> that between .5 and .05 is most approptiate. At what level (i.e. simple
> threshold or height threshold) is this entered? Also, could you explain
> why such a p value is appropriate.
Thresholds are entered in the results section after specifying which
contrasts are to enter into the conjunction (by holding down the
control key). The recommendations above probably refered to uncorrected
p values for the component SPMs (Pcomp) (in SPM99b). In SPM99 a
corrected p value of 0.05 or an uncorrected p value of 0.001 would be
sensible. These might correspond to 0.5 or even more from the
perspective of the component SPMs. Note that for uncorrected p values
Pconj = Pcomp^n, where n = number of [orthogonalized] contrasts.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|