I have a set of fMRI data from around N=30 participants. Each participant has 3 runs of around 150 3-sec TRs per run. Each trial lasts around 3 TRs and there are 3 trial types. In the 2 most important (most frequent) trial types participants are performing a task in which they are imagining giving money to between 10 and 25 known individuals. I have questionnaire data which gives a quantitative measure of how much the subject likes/dislikes each of these individuals, and how much the individuals like/dislike each other (i.e. this is matrix data). Now I would like to use this questionnaire data to create 2 or 3 (3 max) regressors for BOLD signal change. 

There are many different ways I can define these 2-3 variables, as I will compute them from between 3 and 8 sub-variables. One caveat is that they must be binary (I cannot assume the variables from which they are composed sum linearly). To help me decide how to define my regressors, I have been looking at the sparsity of the composite variable matrices for each participant and across the group, and how sparsity changes as I alter the variable definitions. For example, one variable definition has sparsity (density) [sparsity=N nonzero elements/N elements] between 0 and .14 (mean .02). I guess this is too sparse. Another definition has sparsity from .34 to .59 (mean .43), which I guess is too dense. 

So my question is: 

- is there a rule of thumb for deciding how sparse an fMRI regressor should be? 

- if not, I guess some model fitting step is needed. So, any advice on this would be welcome. 

Thanks