Hi! I have a cointegrated relationship of top income shares and bank debt/GDP in Sweden for 1903-2012. The graphs move jointly, and seems to have a long run positive relationship, as can be seen in attached pic. However, the sign of the coefficient of the theorized explanatory variable is negative (see 2nd pic), indicating a long run negative relationship. How can this be?
Can you check how your software handles normalization. For example, E-Views comes up with a coint. vector of negative sign when actual sign is positive. Therefore, by multiplying with -1, you will get the correct sign. I believe this is the issue here.
Thanks for clarification! Since I get the same result with both Eviews, Stata, and Gretl, this seems to be the case. But in the list of cointegrating equations, it seems as if it normalized on one (but with a parenthesis with -1). I attach a pic from Eviews.
Yes, your coefficient is positive. It is normalization issue so where u see 1.000 should be -1.000 and your estimated coefficient should 2. Something and so on. Problem solved
Great! Thanks a lot! I've been struggling with this for several days! :)
No problem. That's what friends are for. You can now go outside for fresh air!!
I didn't go out for fresh air. Instead, I just encountered a new problem, hehe. Would you mind looking in to it? I have a cointegrating relation between the top decile income share of the population, and the private debt-to-GDP ratio (not the same variable as the one I used before, now mortgage institutes are included). I use two structural breaks as exogenous variables (as proposed by Johansen(2000), and put them inside the matrix. Now, the coefficient on the non-normalized variable (top income share) is not significant, but one of the breaks is. How could this be interpreted? Should I normalize on top income instead of debt/GDP - which goes against my theory? Or is it ok that the coefficient of the top income share is insignificant as long as some other variables (the structural break) are?
It is a good result because you have allowed the dummy to be treated outside the cointegration space. See, Doornik, Hendry and Nielsen (1998) Journal of Economic Surveys "concerning the treatment of impulse indicator variables, we strongly recommend that these be entered unrestrictedly if they are used >>>>>We certainly advise against their restriction to the cointegration space " p562. It will be nice if you can associate the dummy with any relevant historical evidence.
Concerning your results, I like the statistical significance of the error correction term which signals cointegration.
What remains for the soup to be cooked -- I wonder what will happen if you lowered your lag order to 2, 4 or 5. Please pay attention to your residual correlation.
Thanks for your input! Yes, with 2 lags the coefficient on the non-normalized variable (top income share) is significant. I base my lag selection on AIC (suggests 6 lags) and BIC (suggests 2 lags), and present both models in my paper. I check for autocorrelation in residuals of the underlying VAR (with both 2 and 6 lags), and come out "clean". But didn't perform any post-estimation tests of the residuals on the VECM. Should I do that also?
And by the way, I do have relevant historical evidence justifying the break dates: They correspond to the start and the beginning of credit regulations in Sweden. There are very persuasive accounts of why this period should be set to 1952-1986. :)
- Badges
- Science topic
- Similar topics
- Geoscience
- Asia
- India