The estimation of SVAR

[res84529]

Hi there!

I found that sometimes different coefficient estimates are obtained when using cvmodel instruction to estimate the non-recursive SVAR. The situation is most troublesome particularly when I want to obtain the standard errors for those coefficient estimates which were obtained early by doing more piters in the preliminary estimation.

For example, in the output file attached, I estimate the model first by "cvmodel(parmset=kimroubi,pmethod=genetic,piters=1000,iters=2000,factor=afactor) vmat ab" and obtain the coefficient estimates which are closest to reasonableness. But when I reduce the number of piters to 100 in order to obtain the standard errors for those coefficent estimates, I found it's very hard to obtain the same coefficient estimates.

Could you please help me to solve this problem?

Code, data and output files are attached

Best Regards

[TomDoan]

Your model isn't identified. How is it that you plan to distinguish the first two equations in the SVAR?

[res84529]

Hi Tom!

Yes, I know the model hasn't been properly identified. I'm still in the process of experimenting some variation in the non-recursive identification restrictions to see if ms shock, md shock, and e/r shock can be identified. At this stage, I need to obtain correct coefficient estimates (with expected sign and reasonable magnitudes) and their standard errors for each trial.

Yet I found it sometimes is impossible to obtain the standard errors for those coefficient estimates which were found most reasonable by doing more piters in the earlier estimation. When I switch to use the instruction "cvmodel(parmset=kimroubi,pmethod=genetic,piters=100,iters=2000,factor=afactor) vmat ab", RATs gave different coefficient estimates (although the signs remain unchanged, the magnitudes are different) from that using cvmodel(parmset=kimroubi,pmethod=genetic,piters=1000,iters=2000,factor=afactor) vmat ab". But I still want RATs to deliver the same coefficient estimates which I obtained with the latter instruction together with the associated standard errors. This is the problem I met. Could you please give some advice to solve this problem?

Best Regards

[TomDoan]

You do realize that not identified means that there are at least two modes with identical log likelihoods, don't you? It's quite possible that one of the two is on a very small "hill" and the other is on a bigger one, so almost any initial conditions will end up at the latter rather than the former.