Hausman Test / FE or RE with invariant variables

[Theresa Miller]

Hi everyone,
my name is Theresa Miller (I have submitted a request to change this...).

I am an absolute stata and econometrics newbie, and even though I know this topic has been discussed before, I do not understand the explanations that are present in this forum - therefore sorry that I am asking again.
Maybe somebody can explain for a complete beginner/idiot like me.

I have panel data (60 firms 10 years) to test whether the accounting figure ctl and total assets tas influences the log annualized stock price volatility lvol. On top of that I have several time-invariant variables like sector sec, and country cty, and one called trs.

After setting the data to panel data (xtset) I have ran different regressions (pooled ols, FE and RE):
reg lvol ctl tas sec cty trs
xtreg lvol ctl tas sec cty trs, fe
estimates store fe
xtreg lvol ctl tas sec cty trs, re
estimates store re
>> Now I ran a hausman test to determine whether I have to use FE or RE model<<
hausman fe re

--> the result was p value = 0.0000, i.e. the Hausman test is significant and that means I have to use FE model.
BUT the problem is: in FE model, my timeinvariant variables are always dropped, and therefore I don't get an explanation whether sec, cty and trs have an effect on my depvar lvol .

Now my question is: what model (some hybrid thing?) to I need to use in this case if I need the timeinvariant variables to be taken into account?

Agreed this question has been asked before, but I really know very little econometrics (this here above is my success of two days of watching tutorials...), so maybe somebody can explain it very simple? I don't need to become a pro, I just need to know which model is the correct one to use...

THANKS SO MUCH IN ADVANCE!!!

[Marcos Almeida]

Hello Theresa,

Welcome to the Stata Forum.

As you correctly remarked, there are several threads on this subject. But I suggest you start by taking a look at this one:http://www.statalist.org/forums/foru...-effects-model.

I gather that Hausman test "suggests" you to avoid a random-effects model, but it is not always an absolute obligation.

That said, it seems you should take a look at the "be" option of - xtreg -, which deals with between-effects model. I guess it is what you wish for your data.

By the way, you may find the interpretation in the Stata Manual:http://www.stata.com/manuals13/xtxtreg.pdf.

Oftentimes, even if we "just need to know which model is the correct one", we cannot help but delve into the realms of the matter so as to clearly understand it. Reading the Manual on Panel Data will surely pay back any effort we take.

Finally, I kindly ask you to provide commands under CODE delimiters, as recommended in the FAQ. Moreover, displaying the output is a great step to entice helpful comments.

Best,

Marcos

[Carlo Lazzaro]

Theresa.
as Marcos pointed out, theoretical ground more than -hausman- test result should drive the choice between -fe- vs -re- specification.
This remark is well explained in Hsiao's http://www.cambridge.org/us/academic...ta-3rd-edition, pages 48-50.
Coming to your second question (and echoing in part Marcos' helpful reply), there's no such thing as a correct model (please see, if interested, https://en.wikipedia.org/wiki/All_models_are_wrong): you should weight -fe- specification (which should be chosen according to -hausman- test result) against the expected payoff from investigating time-invariant variables (that requires for -re- specification).
As a closing-out aside, as per -hausman- test machinery choosing between -fe- and -re- specification boils down to a matter of correlation.

[Theresa Miller]

Hi Marcos and Carlo,
Thanks a lot for your help. I think I have understood now (even though I have not read the links yet so just very basic understanding).
In my paper, I will now run both tests and explain that it would be preferable to use RE due to the time-invariant variables, then I will run the Hausman test and conclude that it suggests using FE. Then I will explain that there is no one right model and interpret both results.

@Carlo I am not entirely sure what you mean with your last sentence: Do you mean that the Hausman test only looks at the correlation of the results or did I get that wrong?

Thanks again to both of you, I will now read the links

[Carlo Lazzaro]

Theresa:
my mistake, I was too brief.
I meant the correlation between individual effects and the vector of regressors:
-re- assumes no correlation (often an untenable assumption), whereas -fe- allows any correlation,
Both -xtreg- entry in Stata .pdf manual and http://www.stata.com/bookstore/fixed...ssion-models/v are worth reading on the topic(s) you're interested in.

[Sebastian Kripfganz]

I suggest that you have a closer look at the so-called "hybrid" or correlated random-effects model. It is not as complicated as it might look, so do not be afraid of having a look at it. The link given by Marcos to an earlier Statalist post is an excellent starting point.

[Marcos Almeida]

Hi Theresa,

Panel data analysis is a tricky matter. Please don't harsh too much into conclusions before getting to grips - at least to an introductory level - with the main assumptions.

I believe what Carlo and I was saying was basically related to the caveats of the Hausman test.

I kindly suggest you start by reading the manual (I provided in #2).

Also, as Sebastien reinforced, an hybrid model is something to think about.

Additionally, I pointed out to a betweeen-effects model. There is an example in the manual. Actually, the very first example. Maybe it is not frequently used, but I feel it could give some information lacking in the FE model of yours.

Finally, there are many details/suggestion/solutions that Forum members cannot spot unless you present the results of your "tentative" model.

Best,

Marcos

[Theresa Miller]

Hi,
thanks again for all the advice

I have never studied statistics or econometrics beyond knowing what a standard deviation is, therefore even the manuals are very difficult / impossible to understand.
I will try to use the BE or correlated RE Model if I manage to understand the basics, but I won't dig any further. If I realize I have some time left after that I can work on understanding that stuff

Thanks everyone!


PS: In case any of you every decide to publish examples with example stata files... make sure its not huge. I can barely open any file because I have only the student version... and why would an example have to be so huge anyway

[Carlo Lazzaro]

Theresa:

If I realize I have some time left after that I can work on understanding that stuff ...

I presume you're under time constraints, but usually the reverse works better: trying to grab the core theoretical matter via examples and then get ready for your analysis.

[Theresa Miller]

I agree, but as you say correctly my time is verrrryyyy limited. The perks of working fulltime and studying at the same time
Thanks again.

[Carlo Lazzaro]

Theresa:
kudos on you for your double effort:
i do hope that you can grab the core stuff of the matter you're interested in anyway.
As an aside, I would prefer -re- vs .-be- specification.