That all sounds correct. The number of steps of IRF's doesn't have to match with the lags at all, but I probably wouldn't put much stock in them beyond about five simply because you have only 20 time series points, not because of the number of lags of instruments.levierfr wrote:Tom,
No worries, we are all in the same boat here.
So let me get this straight: I obtain the first difference of the endogenous variables to get rid of the individual effects and essentially to use all the past information of Y(i,t) as instruments, which I do in the ablags step (restricting the max. no, of lags). Then I run a regression using these instrumental variable, then later use the GMM to estimate these regressions in a panel setting (hence the lwindow=panel). By then grouping these six linreg lines of code, I essentially replicate the VAR in a panel setting, and by running the SUR on it I get the covariance/correlation matrix of residuals. From here can I directly generate the appropriate IRFs and FEVDs using the IMPULSE and ERRORS commands? The steps should equal 5 (i.e. the max. lag) or 10 (given in textbooks).
BTW, because you have 20 data points, and thus not a really small T, you can also try estimating the equations by standard fixed effects (on the undifferenced data)---the dynamic panel bias diminishes with T. That can be done by PREG one equation at a time, or by SYSTEM with a set of individual dummies as DETERMINISTICS. Those estimates would be (slightly) biased, but may have lower overall MSE than the GMM approach. A comparison of the results might be interesting.