Hi Jeremy,
I assume that you did a contrast of Task 2 > Task 1 where Task 2 has longer mean RT and found voxels that showed greater activation.

In general, there are two basic approaches.  You can add an impulse regressor to your design matrix that has its height modulated by the demeaned RT.  Alternatively, you can incorporate RT directly into the Task regressors by making each trial a boxcar that has a duration equal to the RT for that trial.  Assuming all the standard assumptions of linearity, both methods should account for most RT-related variance in your data and allow a contrast of Task 2 > Task 1 without time-on-task being a confound.

WRT performing a contrast of RT regressors... in the first method, you only have one RT modulated regressor so the contrast is against baseline, and in the second method, RT is already incorporated into the Task regressors. 

WRT making an exclusive mask... you can do that for the first method, but that will be overly conservative because it would exclude voxels that have both a significant difference between tasks and are modulated by RT.  There is a high probability that voxels significant for Task 2 > Task 1 are a subset of voxels that are modulated by RT.

As for the temporal offset hypothesis, I guess the reviewer is suggesting that the magnitude of the response for both tasks is identical but Task 1 is offset later in time relative to Task 2 and the contrast is giving you a false positive because of the model's misfit for Task 1.  Since you incorporated the temporal derivative in your design matrix, this is highly unlikely.  The purpose of the temporal derivative is to account for small shifts in timing.  However, maybe the time shift is so large that the temporal derivative is incapable of accounting for all the variance.  To demonstrate that this is not the case, you should plot the peri-stimulus plots for each condition and show that any offset that might be present is small, and thus, can be accounted for by the temporal derivative.
cheers,

jack

On Thu, 26 Jan 2012 22:14:57 +0000, Jeremy Elman <[log in to unmask]> wrote:

>Hello all,
>
>I have a question about comparing two tasks with significantly different mean reaction times (1946ms vs. 2341ms). A reviewer brought up the point that it is unclear whether differences found using our current analysis (using fixed duration and impulse convolved with the double-gamma HRF and temporal derivative) are due to magnitude differences or simply caused by temporal offsets/differences in task duration. While we have reason to believe our region of interest is not sensitive to variability in RT, I agree that this needs to be addressed and want to make sure I go about this in a satisfactory way.
>
>I've looked through previous posts and it seems one approach is to include an RT regressor (with either variable duration or impulse). If this modulated regressor is included, would I be safe in interpreting significant regions resulting from contrasts of the original (unmodulated) regressors as likely reflecting differences of response magnitude between the tasks? Further, would it make sense to run the same contrasts on modulated regressors in order to obtain an RT-sensitive map, which could then be used as an exclusive mask for the unmodulated contrasts?
>
>It seems that the above may be useful to do regardless, I might not be able to make specific conclusions about whether my current results are due to magnitude differences or temporal offset. In order to do this, would it make sense to use FIR basis functions and examine how the tasks differ within each time bin? Or alternatively, generating average peri-stimulus plots for each condition and examining those for differences?.
>
>I understand that there is probably no "correct" answer, but I would greatly appreciate any input as to what kind of interpretations can be made from the varous approaches. 
>
>Thank you for your help,
>Jeremy