Multicollinearity and fixed effects model

[Joshua Miller]

Hi guys,

I am running a balanced panel data analysis of 62 countries over 11 years with 682 observations. To establish the appropriate model I have run the White test for heteroscedasticity to see whether OLS model is appropriate. The results indicated that there is heteroscedasticity and therefore my understanding is the OLS model is inappropriate.

From there I ran the Hausman test for fixed or random effects and established that fixed effects is appropriate.

Now do I/can I run a test for multicollinearity? As I can see from my correlations matrix I have some relatively high values (0.7, 0.6). I know the VIF test establishes multicollinearity however this cannot be run for xtreg?

If this is the case can I run an OLS regression using country and time fixed effects and use standard robust errors for heteroscedasticity and then run VIF?

If so how would I go about generating the dummy variables for country/time effects in STATA,i'm not sure if this is right but would this be the code:

reg y x1....xn icountry iyear, robust

[Carlo Lazzaro]

Joshua:
-you do not report how you -xtset- your data before -xtreg-;
- I fail to get why you are seemingly moving from OLS to panel data analysis and back again to OLS. Why don't simply stay -xtreg,fe vce(cluster panelid)?
- event though what is high or low may well depend on your research field, your correlation values do not seem that high (please, see -help estat vif- and related paragraphs in Stata 13.1 .pdf manual, -regress postestimation- entry.

[Joshua Miller]

Hi Carlo,

Thank you for your response.

I apologise if I am not very clear I am very new to econometrics!

I used xtset country year

By using the xtreg, fe command I cannot check for multicollinearity is this correct? Is this something I can ignore with a fixed effects model?

What is the difference between xtreg, fe vce (cluster panelid) and xtreg, fe, robust as I have run both versions and they produced identical results.

Kind Regards,

Joshua Miller

[Carlo Lazzaro]

Joshua:
thanks for providing more details.
As you already included year in -xtset-, it may happen that Stata will drop some year dummy due to collinearity.
You are right: you can't directly check for multicollinearity after -xtreg-, but take a look at two before-forum Stata threads on this topic:
http://www.stata.com/statalist/archi.../msg00018.html
http://www.stata.com/statalist/archi.../msg00482.html.
No difference indeed between the two -vce()- options (but this doesn't hold for -regress-, for instance). That's why I prefer vce(cluster panelid) vs -vce(robust)-..

[Joshua Miller]

Hi Carlo,

Thank you for your hasty response and for clearing up the differences in vce(cluster) and robust.

From looking at the two threads posted I would need to create my own dummy variables which can be done via the tabulate command; I would need to create 62 country specific dummy variables and 11 time specific dummy variables?

I think I am mis-understanding how this process is done butin Layman's terms am I supposed to have 62 country and 11 time variations which I can compute in to STATA i.e. a dummy variable based on country region and so on?

I had assumed (and hoped) STATA can calculate the dummy variables as it does in the fixed effects model but I do not understand how!

Kind Regards,

Joshua Miller

[Carlo Lazzaro]

Joshua:
the best way to create dummies is relying on -fvvarlist-.
As an aside: is it worthy to bother yourself that much with multicollinearity (http://www.stata.com/statalist/archi...sg00018.html)?

[Joshua Miller]

Hi Carlo,

is this the same method as I had used above?

reg y x1....xn i.country i.year

Unfortunately that link is not working for me.

[Carlo Lazzaro]

Joshua:
if you paste the links in your browser, you'll get the threads.
Otherwise, search on the web with the string: -multicollinearity and pane-l and they will show up along with tons of entries on this topic.

[Joshua Miller]

Great, thanks for your help Carlo.

[Sebastian Kripfganz]

As a further comment: The traditional Hausman test is not robust to heteroscedasticity or serial correlation in the idiosyncratic disturbances. A robust regression-based variable addition test can be obtained by adding time averages of your variables to the model and then by testing for joint significance of all those time averages. See the following presentation slides by Jeff Wooldridge for details:
Correlated Random Effects Panel Data Models

The following thread might provide useful information as well:
http://www.statalist.org/forums/foru...riant-variable