I am probably repeating what has been said before, but in order to have a better idea about conjunctions, here are my thoughts. please correct them !!
The conjunction analysis assumes that the conditions are independent, which is not true for within-subject designs. In fact, one way anova can be used under the assumption that the conditions are independent. This assumption is true for between subject designs. For within-subject designs, one should be using a within-subject anova rather than a one way anova. In one-way anova, the contrasts such as 1 0 0 are not estimatable because of design matrix collinearity and betas are not unique. In pre-SPM5 versions, the error term was specified separately as effects of no-interest and it was possible to specify contrast such as 1 0 0 ones(1,n)/n, where n refers to number of subjects. The second alternative is to validate the assumptions of one-way anova. If each of the con_xxx images entered into an RFX model do not represent a task condition per se, but the difference between two conditions for a single subject, then the within subject variance is already accounted for. Hence one can use a one-way anova and run the contrasts.
It is still possibly a good idea to run the conjunctions outside the SPM results section using imcalc to circumvent the issues above. The assumption here, is that all the different SPM.mat models are equivalent with same degrees of freedom etc so that the minimum t-statistic can be obtained from thresholded maps.
Thanks
|