The RATS Software Forum

Hi,

I'm trying to estimate a panel structural var based on Pedroni, Peter (2013) "Structural Panel VARs", Econometrics, 1 (2),180-206. I got the code from Prof. Pedroni, and transformed my dataset (unbalanced panel) into the format the code required (balanced panel with NAs). However, when I try to estimate the model, I always get the following error.

processing panel member 1
chosenlag is 1
## MAT14. Non-invertible Matrix. Using Generalized Inverse for SYMMETRIC.
The Error Occurred At Location 968, Line 52 of loop/block

I don't know what went wrong. Could you please have a look at the data and code attached and help me if possible?

Individual 2 has only 10 usable data points. That's probably not enough to do the analysis that's required on each individual. (It estimates a separate 1 lag VAR on each individual.)

TomDoan wrote:Individual 2 has only 10 usable data points. That's probably not enough to do the analysis that's required on each individual. (It estimates a separate 1 lag VAR on each individual.)

Hi Tom,
Thanks for your reply. This makes sense. I actually already removed some short series before trying the svar. Do you have any recommendation about the minimum time periods of each individual for a panel svar? I have 4 endogenous variables, which means 5 parameters per equation, and 20 in total for the system. What is the safe minimum number of observations for running a panel var in my case?

TomDoan wrote:Individual 2 has only 10 usable data points. That's probably not enough to do the analysis that's required on each individual. (It estimates a separate 1 lag VAR on each individual.)

Hi Tom,
On a second thought, the short time series shouldn’t be the major reason for the error. VARs are estimated equation by equation, which means I never estimate more than 5 parameters in a single regression. So it should be feasible given the length of the shortest countries in my panel.

Except that it runs the VAR on differences, so that loses one data point. Plus the one lag loses another, so you're down to 8 data points. Then it is freely estimating a 4 x 4 covariance matrix, which, in effect, means three extra free coefficients (for the contemporaneous cross effects). All told, that means with 10 data points and 4 variables, you're getting a perfect fit. With 11 data points, you wouldn't have a perfect fit, but the results would be quite suspect due to overfitting. I probably would want closer to 20 data points per individual to be safe.