Hello Everybody,
Working on our MBA thesis regarding, the determinants of carbon emissions and have created panel data consisting
of 14 EU countries for a period of 20 years over several variables. For some of the variables we got missing data but
not more 3-4% out of 280 observations and only for some variables.
The steps we have followed so far are as follows:
1. Test for stationarity using Levin-Lin-Chiu test for the variables that contain balanced data. For the ones that
are unbalanced we used Fisher test. For the variables that had unit root we did the first differences transformation and
that transformed data to stationary.
2. We tried panel regression considering random effects then ran xttest0 (LM Test Breush Pagan) which failed to reject
the null thus the random effects seem not appropriate.
3. We tried panel regression considering fixed effects then observed the F-statistic which indicates that fixed effects
seem not appropriate as well. At this stage we also tried the fixed effects model with dummies for the individual countries
and the F statistic was as expected the same with the model without the dummies.
4. When we run the Hausman test we noticed that the order "hausman fixed random" returns Prob > chi2 = 0.0643 which is
more than 0.005 and fails to indicate for fixed effects. We also run "hausman fixed random, sigmamore" that returned
chi2(10) = 17.38 and Prob > chi2 = 0.0664 thus failing to reject the null hypothesis "Difference in coefficients not systematic".
5. When we run the Hausman test we noticed that the order "hausman random fixed" returns negative chi2 and a warning
of failure to meet the asymptotic assumption of the Hausman test.
6. We also checked for time fixed effects and we got F( 18, 159) = 0.77 | Prob > F = 0.7291 which is higher than
0.05 thus reject the null hypothesis of having time fixed effects. (maybe this was not necessary)
7. As this was not conclusive for random effects we considered it as an Indication to go with Pooled OLS knowing that
this is the less preferable method.
8. After that we run a simple vif test which looks good but the hettest is not good.
After this detailed description of the steps followed so far i would need a hand with some queries we are
struggling around with my thesis partner:
- Do you think that running a Pooled OLS with the robust option would eliminate any additional test for heteroskedasticity?
- Are there any other tests that could be done after the Pooled OLS regression to argue about a solid estimation model?
- After running the initial model we noticed that some introduced control variables are affecting severely the significance
of other variables thus removed from the model. Since we were following a process of elimination we removed the ones
with the higher P value in each step. Is this a rational approach or are there any suggestions?
Of course any comments are more than welcome! I read a great saying in the forum that "we are all beginners" but
noticed that there are some very experienced ones in here!
Grateful in advance!
Working on our MBA thesis regarding, the determinants of carbon emissions and have created panel data consisting
of 14 EU countries for a period of 20 years over several variables. For some of the variables we got missing data but
not more 3-4% out of 280 observations and only for some variables.
The steps we have followed so far are as follows:
1. Test for stationarity using Levin-Lin-Chiu test for the variables that contain balanced data. For the ones that
are unbalanced we used Fisher test. For the variables that had unit root we did the first differences transformation and
that transformed data to stationary.
2. We tried panel regression considering random effects then ran xttest0 (LM Test Breush Pagan) which failed to reject
the null thus the random effects seem not appropriate.
3. We tried panel regression considering fixed effects then observed the F-statistic which indicates that fixed effects
seem not appropriate as well. At this stage we also tried the fixed effects model with dummies for the individual countries
and the F statistic was as expected the same with the model without the dummies.
4. When we run the Hausman test we noticed that the order "hausman fixed random" returns Prob > chi2 = 0.0643 which is
more than 0.005 and fails to indicate for fixed effects. We also run "hausman fixed random, sigmamore" that returned
chi2(10) = 17.38 and Prob > chi2 = 0.0664 thus failing to reject the null hypothesis "Difference in coefficients not systematic".
5. When we run the Hausman test we noticed that the order "hausman random fixed" returns negative chi2 and a warning
of failure to meet the asymptotic assumption of the Hausman test.
6. We also checked for time fixed effects and we got F( 18, 159) = 0.77 | Prob > F = 0.7291 which is higher than
0.05 thus reject the null hypothesis of having time fixed effects. (maybe this was not necessary)
7. As this was not conclusive for random effects we considered it as an Indication to go with Pooled OLS knowing that
this is the less preferable method.
8. After that we run a simple vif test which looks good but the hettest is not good.
After this detailed description of the steps followed so far i would need a hand with some queries we are
struggling around with my thesis partner:
- Do you think that running a Pooled OLS with the robust option would eliminate any additional test for heteroskedasticity?
- Are there any other tests that could be done after the Pooled OLS regression to argue about a solid estimation model?
- After running the initial model we noticed that some introduced control variables are affecting severely the significance
of other variables thus removed from the model. Since we were following a process of elimination we removed the ones
with the higher P value in each step. Is this a rational approach or are there any suggestions?
Of course any comments are more than welcome! I read a great saying in the forum that "we are all beginners" but
noticed that there are some very experienced ones in here!
Grateful in advance!
Comment