Hi Giuseppe,
> If I understood it correctly, to run 1dSVAR semi-automatically, I could make a
> `custom' copy of the R script and substitute the actual answers for the
> corresponding readline() commands, right? I haven't thought about that, that
> seems pretty easy and it's a good occasion to get some familiarity with the code.
readline() seems to be useful when you read a line in an interactive mode. I was thinking about the following. Go through all the commands for one subjects (without no graphical stuff such as plotting nor model checking), and save all the command lines into a file called myCmds.R. Then make as many copies as the number of subjects, and replace those input/output files for each specific subjects. Then on the terminal (outside of R) run
R CMD BATCH myCmds.R myDiary &
> As for the numerical solution issue when loops are present, is this a problem
> even with the direct approach, in the sense that even though the program
> doesn't crash, its results may not valid?
If the numerical solution converges when loops are present, you're fine. The problem is when the numerical solution fails to converge. And this issue is not just about SVAR, but also occurring in SEM.
> if you have a set of ROIs whose network structure you're interested in
> exploring (and, of course, you don't actually know a priori who's connected
> to whom), would you think it's a legitimate strategy to examine first the
> simple partial correlation matrix (using 1ddot, as suggested by Bob in a
> previous e-mail on the list) and then make a connection matrix for 1dSVAR
> where the edges to estimate (NA values) are only chosen among those that
> had a partial correlation coefficient significantly different from zero?
This is tough. Such a correlation calculation seems to ignore any effect from the past history of each ROI and other regions.
Gang