Sara:
I think your confusion is coming from 1) t-testing the wrong output from 3dDeconvolve and 2) misunderstanding the t-test program.
Based on your explanation of your experiment, it appears you have some contrasts that you have coded for in 3dDeconvolve, probably with the -glt option (?) or you may be looking at the ouputs calculated for your regressors (F-value, coefficient, etc, depending on what you asked for).
Regardless, of your 3dDeconvolve run ( I will get to that later), you are probably interested in using 3dttest to test the brain response to condition 1 versus condition 2 (if there are more than 2 conditions you should be using 3danova). To do this you want compare condition 1 in subjects 1 and 2 to condition 2 in subjects 1 and 2. You are correct in saying that 3dttest calculates the mean in that it will test the differences between the mean of condition 1 and 2 given the amount of variability present within the subjects.
In this case you would use a command line like:
3dttest -set 1 subject1_condition1+tlrc subject 2_condition1+tlrc -set 2 subject 1_condition2+tlrc subject2_condition2+tlrc -paired -prefix Condition1_versus_Condition2
The paired option was used because it sounded like you had a within subjects design where conditions 1 and 2 were presented to each subject. Think of set1 being condition 1 and set2 being condition 2.
So now to the issue of 3dDeconvolve. You probably want to do the ttest on a subrick that directly reflects only the activity in the brain in response to each condition, not necessarily the response versus the amount of noise (such as would be the case if you t-tested t-values). I would recommend using the Coefficient for the regressor that coded for condition 1 and the Coefficient for the regressor that coded for regressor 2. If maxlag was set to >0, however, you will have multiple coefficients. In that case I recommend 1) either picking the lag where you suspect greatest activation or 2) creating a glt to calculate area under the IRF for each regressor. Take these values to talairach space and blur them before entering them into the ttest.
Finally, another approach that would answer your question about what regions people have in common for condition 1 and what regions they have in common for condition 2, is to create a conjunction map. To do this,
you want to extract "active" regions for condition 1 and condition 2 for each subject. First, threshold the data the same way you did before using 3dmerge.
Next, create a mask of active voxels for each condition for each subject using 3dcalc.
3dcalc -expr 'step(a)' -a subject1_condition1_thresholded_data -prefix subject1_condition1_thresholded_data_MASK
Repeat this for each condition and subject.
Next, use 3dmerge to count the number of 'hits' at each voxel.
3dmerge -gcount -prefix condition1_gcount
subject1_condition1_thresholded_data_MASK+tlrc
subject2_condition1_thresholded_data_MASK+tlrc
You can then see voxels/clusters that are active in both subjects.
If you are interested in knowing which voxels belong to subject 1, which to subject 2, and which overlap, you need to use 3dcalc.
3dcalc -expr '(a+(b*2))' -a subject1_condition1_thresholded_data_MASK+tlrc -b subject2_condition1_thresholded_data_MASK+tlrc -prefix condition1_subject_overlap
By multiplying the second subject by 2 you can interpret the resultant map in terms of what the voxels add up to.
1 = subject 1
2 = subject 2
3 = both subjects.
You can also see that this conjunction map approach can also be used if you used a glt in 3ddeconvolve that contrated conditions 1 and 2 within subject. You could threshold that result, mask it, and find the overlap between subjects that way too. You may want to blur your masked datasets before doing the conjunctio step.
Hope that helps,
Christine Smith