EViews.com

Hi,

I realized that eViews 11 comes with switching VAR option in the VAR estimation, something very helpful for me. I want to run a model with panel data (yearly data of 40 firms over 35 years) with 6 endogenous variables and a bunch of exogenous ones. I specified the markov switching type and ticked the lagged endogenous box.

However, running the model results in NAs for the standard errors and significance. The Warning tells me that "WARNING: Singular covariance - coefficients are not unique
Failure to improve objective (non-zero gradients) after 0 iterations".

A regular VAR is estimated without any issues. So my questions are:
- Is it possible to estimate a switching VAR for panel data (the help file suggests otherwise, but the item is not greyed out)
- Are there limitations on the number of endogenous or exogenous variables
- Interestingly, I get some outputs for less endogenous variables for a specific seed value. The same set of variables produces NAs for another seed value. Would you have any insights why this happens?

Thanks for any help. Best.

1. I'll have to look at the greying out issue, but Markov switching VARs are problematic for panel data specifications. I'm suspect that it should be erroring for that data configuration, but perhaps I am not remembering what we are doing. I'll have to take a look.
2. No in principle, but yes in practice as you'll get convergence and singularity issues.
3. These are highly nonlinear models that are notoriously difficult to estimate, hence the inclusion of randomized starting values and other tools. Even in non panel singularities and sensitivity to starting values are quite common.

The algorithm for estimating the switching model goes through after correctly accounting for the panel lags in the VAR, but I think that the handling of the state initializations isn't ideal and probably misleading. You could make an argument that what the algorithm is producing is valid, but I think that I wouldn't, and I'll go in an make it so that this is non-estimable. Sorry for the confusion.

Thank you very much for the explanations. Now, I can be sure that it is not possible to estimate a VAR switching model with panel data.

May I ask another question that directly connects here? An alternative would be to run simple switching regressions for each variable separately. But it seems that I cannot account for the panel structure in eViews either. Is it correct?

Thanks and best.

Correct on the panel front. The issue is that the probabilities need to be handled on an individual basis, and this is quite tricky.