The RATS Software Forum

Dear all,
I would like to estimate a Panel VAR on my dataset (20 countries and 42 observations for each country, the dataset is perfectly balanced). I have three variables (X1, X2 and X3) and two lags in the VAR. I included also as deterministic variables:
- a country-specific linear trend (variable ‘trend’ in the code)
- a country fixed effect (variable ‘dummies’)
- a time fixed effect (variable ‘time’)
I have two problems/questions :
- When I display the result (see the PVAR_results file), one can see that RATS does not estimate the dummy and the trend of the last individual (i.e. the individual 20). Is it normal or did I write something wrong in my codes?
- The list of deterministic variables is (in order): time trend dummies. When I modify the order, the results change also (and in some cases the changes are significant). Do I have to respect any order when I specify the deterministic variables?
Thanks,
Best,
Romain

Your dummy collection is by construction collinear in two directions. The sum of the time dummies is equal to the sum of the individual dummies. And 1 x time1 + 2 x time 2 + ...+ T x time T adds up to the sum of the individual trends across time. When RATS runs across collinearity like this, it zeros out the regressor at which it detects collinearity, so if you choose a different order, it's a different dummy that triggers that.

Thanks Tom for bringing this to my attention.
Do you have any idea to fix that problem (because I really need to have country fixed effects, country-specific linear trends and time fixed effects) ? Do you think that taking a log-trend or a quadractic trend can help to resolve the problem.
Thanks in advance.
Romain

RomainR wrote:Thanks Tom for bringing this to my attention.
Do you have any idea to fix that problem (because I really need to have country fixed effects, country-specific linear trends and time fixed effects) ? Do you think that taking a log-trend or a quadractic trend can help to resolve the problem.
Thanks in advance.
Romain

You're getting those. The lag coefficients should be the same regardless of the ordering. However, because of the collinearity in the dummies, you can't estimate the separate effects. A quadratic trend or log trend wouldn't help.

Thanks Tom for your answer. I appreciated your help.
Romain