Dear Victoria,
the contrast vectors are correct, except the third one for combined runs: Mathematically it's [(A + B)/2 - C]/2 and results in a vector [1/4 1/4 -1/2 ... 1/4 1/4 -1/2 ... ].
You don't have to rerun models based on single runs only. For those subjects with two good runs set up a single model with two sessions. If you want to look at the beta estimates for the runs seperately just define some more contrasts in the model with two runs like [1 ... 0 ...] for first run and [0 ... 1 ...] for the second run. Averaging these two beta estimates outside of SPM should lead to equal results (plus minus some rounding errors) as those for the contrast vector collapsing across conditions like [1/2 ... 1/2 ...].
By the way, when looking at the statistics on whole brain, differences can be expected. The contrast [1 ... 0 ...] in the combined model will probably lead to larger activations than the contrast [1 ...] in a model for the first run only, as the combined model has more DOFs (due to the two runs). The beta estimates are (should be ;) the same though.
If you go with the contrast vectors for combined runs for those with two runs and the contrast vectors for single runs for the others, as mentioned before, then the data should be scaled correctly and you can perform some analyses. However, how would you treat these subjects with a single good run, leaving aside it's fMRI? If you have to throw away half of the data in a behavioral study due to some technical reasons then I would exclude them as well, simply because half of the data can't be analysed. I would set the threshold to something like at least 67 or 75 %. But this is up to you of course.
Best,
Helmut
|