Hi Sanjana,
Sure, modeling that in just the last 4 runs makes sense. I had initially gathered that all runs were in a single model.
Regarding running an initial regression before running a secondary regression on the residuals, there are 2 distinct issues with it.
1. Assuming the betas in the first regression are not important (such as for motion in this case), a secondary regression is okay *if* the regressors in the secondary step are first orthogonalized with respect to the initial regressors. So in this case, the S1,S2,S1xS2 regressors should have motion, etc. projected out as well (as for the input). If this is not done, any motion components in the regressors are orthogonal to the input, and so the fit to regressors of interest will be distorted as the regression tries to minimize the least squares of the residuals (which will subsequently contain that motion). My guess as to the impact of this is that it would drive those betas closer to zero (as a collection) as the amount of motion in the regressors increases.
But in any case, this is not appropriate. So rather than pondering the effect, it would be better to avoid the problem. Either use a complete regression step or project motion (anything else? censoring?) from those regressors. Having -polort in both means you do not need to project out the polort from the regressors.
2. A different problem arises if you actually care about the betas from the first regression (such as if you were to first model S1 and S2, and then model S1xS2 in the residuals, which I understand you are not doing). In this case, the S1 and S2 betas would likely be unfairly larger. The amount of S1xS2 in them will inflate the betas by the amount of S1xS2 in the original input. It would be identical to projecting S1xS2 out of S1 and S2 (which some groups do by habit).
Anyway, the most accurate way to go is to simply put this all in a single model.
- rick