I hope everyone is staying safe and sane out there. I'm running an analysis looking at brain structure differences associated with two separate environmental exposures in children (both dichotomous variables). I am controlling for sex and age, and I'm not interested in interaction effects. My go-to strategy does not match the documentation, but I'm not sure the documented solution using fsgd files will meet my needs. I was hoping I could check with the list about best practices. 

Because some participants have both exposures, my intuition was to use dummy coded variables so that the design matrix looks like this:

+1.00000 +3.11124 -0.48322 -0.51007 -0.53020
+1.00000 -0.20820 +0.51678 -0.51007 -0.53020
+1.00000 -2.06654 -0.48322 -0.51007 -0.53020
...

Where the columns are (mean, mean centered age, mean centered sex, mean centered exposure 1, mean centered exposure 2). A contrast for exposure 1 - exposure 2 would then look like this:

[0, 0, 0, 1, -1]

However, it's clear from the documentation that, given 3, 2 level factors, the recommended approach is to specify 8 classes (one for every combination of factors) and the covariate in an fsgd file. I can remove the interaction terms from the resulting matrix, but even so it looks very different than what I had above:

+1.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000 +3.11124
+0.00000 +1.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000
+1.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000 +0.00000 -2.06654

...

With a contrast with all classes with exposure 1 set to '1', and all classes with exposure 2 set to '-1' - with the order of the classes in the fsgd file it comes out as [-1 -1 1 1 1 1 -1 -1 0]. 

I'm not positive that this is the right way to model differences associated with each exposure given unequal #s of participants with neither/both/just one exposure, but I certainly don't have the expertise to be certain. Could someone give me some guidance on the best approach here? Thank you,

    -Matt

--