Hi Gang,
Thanks for the suggestion. It made me realize that there are a number of scenarios that would "count" as a linear increase across three conditions, that would be missed by doing a conjunction of t-tests. Part of the problem is that with a conjunction analysis, the two components that are combined are thresholded at "nearly" significant (e.g., p = 0.10 for a resultant p value of 0.01 [0.10 X 0.10 = 0.01]). In this case, it forces the two t-tests to be nearly significant and even this requirement is a much stronger test than the test for linear increases as carried out with 3dRegAna.
What is needed is a 3dMEMA version of 3dRegAna. :)
In any case, I think I am then going to return to the idea of coding a single vector with condition number (i.e., 1, 2, or 3), convolving it with a HRF, and then using 3dDeconvolve to get a beta and t-score for this vector. Then I would put those betas and ts into a one-sample, t-test using 3dMEMA.
What do you think of this approach?
Is it OK to convolve a vector with "real" numbers? In my case, the values 1, 2, and 3 reflect the level of confidence a participant has in his/her response. Or is there a better way to code these differences that can get around the problem of -1, 0, and 1 effect coding you mentioned earlier?
Christine
