I am creating the deconvolution script and matrix files for an experiment and was hoping to receive some clarification as to how matrix files work, as I have heard two different accounts. In this experiment there were three different conditions (stimulus types), with each broken down based on correctness of the behavioral response. I would like to look at correct trials across conditions (Stimtype A correct – Stimtype B correct) and also run the same contrast independent of response (Stimtype A all – Stimtype B all). Looking at the correct and incorrect responses separately is a straightforward procedure, but looking at the all conditions has led to some confusion. There have been three options recommended.
1). Use a separate deconvolution script with new .1D files that specifically reflect the “all” conditions (e.g., Stimtype A all .1D, Stimtype B all .1D, & Stimtype C all .1D). In the matrix file, a Stimtype A all – Stimtype B all contrast would like like: 00111100 ... 00-1-1-1-100.
* a possible downside is having fewer conditions in our deconvolution (i.e., not taking into account correct versus incorrect, so not accounting for systematic variance based on response accuracy)
2. Within the original deconvolution script create matrix files giving both the correct and incorrect trials of Stimtype A “1”s and the correct and incorrect trials of Stimtype B “-1”s.
* there is some concern about whether correct + incorrect conditions with an unequal number of trials will be weighted appropriately in the “all” contrast
3. Within the original deconvolution script (6 conditions) create matrix files that “weight” the correct and incorrect responses based on behavioral responses. In the example of a Stimtype A all – Stimtype B all contrast, if a participant had 75% accuracy for Stimtype A trials, in the matrix files one denotes Stimtype A corrects with “.75s” and Stimtype A incorrects with “.25s”
* there is some concern about whether this weighting procedure (instead of 1 and –1 in the matrix file) will have the appropriate effect on analysis, taking into account possible differences between correct and incorrect trials
Can you please address which of this approaches will yield the best results and why?