In which cases should one prefer -xtreg, be- over -xtreg, re-

[Victoria Rogers]

After I ran a Hausman-test I got informed that I needed to use a fixed model instead of a random model. However, I want to test the effect of a time-invariant variable called Male.
I know that -xtreg, re- combines the effects of -xtreg, be- and -xtreg, fe- and -xtreg, fe- controls for all fixed effects.

-xtreg, re- doesn't control for all random effects if I'm correct

"xtreg, be answers the question about the effect of x when x changes between person. This can usefully be compared with the results of xtreg, fe, which answers the question about the effect of x when x changes within person. " Based on that I shouldn't use -xtreg, fe- and use -xtreg, be- instead because I want to know the effect of gender (variable Male)

However when I use -xtreg, be- instead of -xtreg, fe- I don't control for all fixed effects and therefore I would have an omitted variable bias.

Due to my above questions, I'm stuck after searching the Internet for 8.5 hours

[daniel klein]

Victoria,

I think we pretty much covered that question here. Essentially the suggestion was to use a hybrid model (or correlated random effects model, see Allision, 2009 for both) in which you can estimate both, withn-person and between-person effects. There is no model that can estimate between-person effects while fully control for (unobserved) heterogeneity.*

Nothing stops you from posting essentially the same question again, but potential helpers may add new ideas if they know about everything that has already been discussed. So you save time for yourself and for those willing to help, if you refer to things already discussed.

Best
Daniel


Allison, Paul, D. (2009). Fixed-Effects Regression Models. Thousand Oaks: Sage.


* Edit:

You may want to have a deeper look into instrument variable approaches, which is actually what you might had in mind, when suggesting xthtaylor. However, such models are, in my opinion, not very intuitive to understand. If I remember correctly, you are supposed to hand in a report, in which you are to explain the methods you used. This usually requires a good understanding of such methods on your part. Others might give much more advice on IV techniques (and the assumptions behind).

[Victoria Rogers]

The solution of that topic is:

Vitoria: If you want to allow all time-varying regressors to be correlated with the unobserved effects (the fixed-effects case), but are willing to assume that the time-varying regressor is uncorrelated with the effects you can do the following (assuming that the variable ID is your panel identifier):

global varying "alpha MRP SMB HML MOM"
global invariant "Male"

// generate time averages of all time-varying regressors
foreach variable of varlist $varying {
by ID: egen `variable'_mean = mean(`variable')
}

// standard fixed-effects estimation ($invariant will be dropped)
xtreg $varying $invariant, fe

// random-effects estimation augmented by time-averages of time-varying regressors
xtreg $varying $invariant *_mean, re

You will observe that the coefficients of the time-varying regressors will be exactly the same in both regressions. That means, even though you estimate the second specification with the random-effects estimator, you will obtain the fixed-effects estimates for them. In addition, you will get a random-effects estimate for the time-invariant regressor.

This is the essentially the Hausman-Taylor principle if all time-varying regressors are endogenous (with respect to the unobserved effects) and all time-invariant regressors are exogenous.

However, this leads to very bad outcomes/ insignificant results. Therefore, I want to consider other (basic) options as well. -xtreg, be- might already be the solution but I have to know the answers of my first post to be able to understand whether I should use it or not

[daniel klein]

However, this leads to very bad outcomes/ insignificant results.

No hard feelings, but if you are looking for a model that gives you the results you want to find, or in other words, a model that confirms what you already seem to know, this has pretty less in common with science, whatsoever. In that case answering methodological/statistical questions also seems like a waste of time, sine you seem to be willing to accept only those answers that give you results you are looking for, not answers with sound theoretical perspective. At this point I am out. Good luck anyway.

Best
Daniel

[Carlo Lazzaro]

Victoria:
as reported in http://www.stata.com/bookstore/micro...ata/index.html, page 260, between estimator (BE) is seldom used because OLS pooled estimator and -xtreg, re- are more efficient.
Besides, by definition, BE takes into account cross section variations only (and, in general, this is not the case for gender).
Again, after noticing that you have tried almost everything and are on a fast lane for a chair in panel data econometrics, I would think that the main reasons for seemingly "unsatisfying" results rest in your sample (as an Italian saying goes: "you cannot get blood from turnips") and in your boss's prior beliefs).


Kind regards,
Carlo

[Carlo Lazzaro]

Victoria:
as reported in http://www.stata.com/bookstore/micro...ata/index.html, page 260, between estimator (BE) is seldom used because OLS pooled estimator and -xtreg, re- are more efficient.
Besides, by definition, BE takes into account cross section variations only (and, in general, this is not the case for gender).
Again, after noticing that you have tried almost everything and are on a fast lane for a chair in panel data econometrics, I would think that the main reasons for seemingly "unsatisfactory" results rest in your sample (as an Italian saying goes: "you cannot get blood from turnips") and in your boss's prior beliefs.


Kind regards,
Carlo

[Victoria Rogers]

No hard feelings, but if you are looking for a model that gives you the results you want to find, or in other words, a model that confirms what you already seem to know, this has pretty less in common with science, whatsoever. In that case answering methodological/statistical questions also seems like a waste of time, sine you seem to be willing to accept only those answers that give you results you are looking for, not answers with sound theoretical perspective. At this point I am out. Good luck anyway.

Best
Daniel

I just want to compare the results of using different legitimate models. If I would only care about the results, then it's possible to change observations but I'm never going to do that. Besides that, I want to learn as much as possible about the possible options of researching the effects of gender. There's hardly any information about -xtreg, be- (between effects) if you compare it with the information about fixed and random effects.

So, my question remains. Why would anyone use -xtreg, be- instead of -xtreg, re- considering that -xtreg, re- is more efficient. and controls for more bias.

[Victoria Rogers]

Victoria:
Besides, by definition, BE takes into account cross section variations only (and, in general, this is not the case for gender).

Why isn't that the case for gender in general? Based on the following quote I found online and based on the fact that researched persons hardly ever undergo a sex change operation, I don't understand why anyone would need to use -xtreg, re- instead of -xtreg, be- because the difference between -xtreg, re- and -xtreg, be- is -xtreg, fe- which effects don't seem the matter based on the quote below. (it remains strange that -xtreg, re- includes -xtreg, fe- but -xtreg, re- doesn't control for all fixed effects like -xtreg, fe- does.)

"xtreg, be answers the question about the effect of x when x changes between person. This can usefully be compared with the results of xtreg, fe, which answers the question about the effect of x when x changes within person."

[Carlo Lazzaro]

Victoria wrote:

There's hardly any information about -xtreg, be- (between effects) if you compare it with the information about fixed and random effects.

as reported in http://www.stata.com/bookstore/micro...ata/index.html, pages 260 BE is covered for completness of information although it is seldom used because OLS pooled estimator and -random effect estimator are more efficient.

Kind regards,
Carlo

[Victoria Rogers]

Victoria wrote:


as reported in http://www.stata.com/bookstore/micro...ata/index.html, pages 260 BE is covered for completness of information although it is seldom used because OLS pooled estimator and -random effect estimator are more efficient.

Kind regards,
Carlo

I was just reading on your website but didn't find the related information on page 260, therefore I'm now reading the section about Between estimator on page 254 (page 294 of pdf version). The version I have showed your quote on 254.
So according to you -xtreg, be- is never used without trying -xtreg, re- or -xtreg, fe-? Then could you please explain why is -xtreg, re- more efficient than -xtreg, be- in case of an important time-invariant variable?

with BE, it's not possible to use time dummies. So, RE tends to be better. However, I still need to control for all fixed effects to make the omitted variable as small as possible. Does anyone know a method which is different than the method Sebastian used (Sebastian's method is posted by myself at 04:22 Today in this topic.


I would like to use a model with controls for all fixed effects but still states the coefficient of my time-invariant variable called Male. Sebastian's method seems to be the best option so far, but are there any other methods which could lead to better/significant results?

[Carlo Lazzaro]

Victoria:
the link I quoted actually refers to A. Colin Cameron and Pravin K. Trivedi. Microeconometrics Using Stata, Revised Edition. College Station, TX: Stata Press, 2010.
Authors reported (page 260) that -xtreg, be - is seldom used, because:
- individual invariant regressors cannot be estimated, because BE estimator focuses on cross section variations and perform a regression on group means. (As a partial correction to my previous reply, -xtreg, be- can estimate a coefficient for gender, but may report a higher standard error than -xtreg, re-);
- pooled OLS estimator and random effect estimator are more efficient.
Besides (pages 260-261):
- if -xtreg, fe- is the right specification, -xtreg, be- estimates are inconsistent;
- if xtreg, re- is the righ specification, -xtreg, be- is less efficient than -xtreg,re-, as its overall_R2 is usually worse;
- xtreg, be- does not allow an explicit option for heteroskedasticity-robust standard errors (SE). You can only replace default SE with bootstrap SE, via -vce(bootstrap).

Kind regards,
Carlo

[Nick Cox]

People see their local time on posts. 4:22 is right for you only if you are in the same time zone as Victoria. She is referring to post #3 in this thread.

[Carlo Lazzaro]

Victoria:
good if -xtreg, be- works for your resarch purposes (but, especially if your results shall be disseminated outside the company, consider Daniel's remark).
However, I would investigate if any relevant change in terms of significance of coefficients occurs if default SE are replaced with bootstrap SE.
Kind regards,
Carlo

[Nick Cox]

Victoria: You are referring to a pirated copy of copyrighted works. Doing that publicly is poor practice. Doing that privately is poor practice, but we can't see you do it.

[Victoria Rogers]

@Nick, I didn't know it was a pirated copy. I just searched for the title of the book on Google and within 10 seconds (literally) I found that website with the pdf of the book. I only referred so that other people could use the free version of 2009. I assumed that 2009 became free due to the release of 2010. Now that I think more about it, the versions were similar so that would not make sense.