Fixed-Effects model: cluster, absorb or ....

[Daller Rowney]

hello

I have a panel data, where I set panelvariable

-xtset ID Year-

-xtreg X Y, fe robust -

gives me the same results as

-xtreg X Y, fe i(ID)-

and

 -areg X Y , absorb (ID).

Thats ok =)


So my fixed effects model is controlling for ID and TIME. Because the IDs (a company) seldom change industry, the firm fixed effects controls also for industry effects.

Now, when I want control for industry anyways, I try:

xtreg X Y, fe i(INDUSTRY)

warning: existing panel variable is not INDUSTRY

My problem is now, thru excluding the variable (which is even not a panel variable set with -xtset- ID Year), I get strong significant coefficients t >2.7 and a big difference, although the industry effects shuld be covered through -fe robust- in the first equation. Shuld I try to mix ID with INDUSTRY, and then set

-xtset ID*INDUSTRY Year-

Can this work??


Uff, thank you

[Jorge Eduardo Perez Perez]

So my fixed effects model is controlling for ID and TIME.

No, your fixed effects model is controlling for id level heterogeneity, and the identifying variation comes from time variation in X and Y within firm.

Because the IDs (a company) seldom change industry, the firm fixed effects controls also for industry effects.

Right, but this means that you will not be able to control for industry, because it will be collinear with the firm fixed effects.

I want control for industry anyways,

Then try

xtreg X Y i.INDUSTRY, fe

But if industry does not change within firm across years, then the industry dummies will not be estimated because they will be collinear with the firm fixed effects.

[Daller Rowney]

No, your fixed effects model is controlling for id level heterogeneity, and the identifying variation comes from time variation in X and Y within firm.

Right, but this means that you will not be able to control for industry, because it will be collinear with the firm fixed effects.

But if industry does not change within firm across years, then the industry dummies will not be estimated because they will be collinear with the firm fixed effects.

Thank you for the comment! I checked the panel and no firm (t=15y) is changing EVER its industry. But of course, this is also depended on the classifier and the level of classification (DJ,ICB..)

When i try i.Industry, Stata doesnt generate Dummys in the Data Browser, but the coefficient is different.

But what if i think more of a cluster in the Industry , i may try

xtreg X Y, fe cluster(INDUSTRY)

its gives again better results (even in line of the literature!) than i.industry without the error code and it feels now industry controlled...My group number drops now from ID (150) to a nice industry clustered group of 47.

Do I control now the wrong relationship now with the fixed effects model? test-stat(xtoverid, hausmann) leads me to FE.

treg X Y, fe i(INDUSTRY)
warning: existing panel variable is not INDUSTRY

[Jorge Eduardo Perez Perez]

Clustering by industry means that you are assuming that errors may be correlated between different firms in the same industry, and that they are independent across industries, net of the firm fixed effects. Think if this assumption makes sense for your model. There should be no difference between the coefficients between the clustered and not clustered versions, just in the standard errors.

47 may be a small number of clusters to use a cluster robust variance: this is only consistent with a large number of clusters. Consider a bootstrapping scheme for robustness.

[Carolin Kl]

Hello everybody,

I have a similar "problem". I also have panel data from 2008-2010 and I set

xtset ID Year

for company and time

In my model I want to control for industry and year fixed effects.

I run a hausman test, which says I should use the fixed effects model, but there, some variables were omitted and the results a much worse (e.g. not significant), while running a random-effect-regression my results are significant/better (mostly).

 xtreg var1 var2 var3, re i(SIC)

 xtreg var1 var2 var3, fe i(SIC)

I also included i(SIC) after re and fe, my results are significant then. SIC means SIC-Code for the Industry (4 digit), but then I also get: warning: existing panel variable is not SIC
Is this a problem or can I ignore that message in here? This would be my "favorite" model, because of the pretty good results. Or have I include i.Year in here?
I do not know if it is neccessary to include i.Year if I also set xtset ID Year?

Or is it possible to

xtreg var1 var2 var3 i.Year i.SIC, re

oder

xtreg var1 var2 var3 i.Year i.SIC, fe

I do not know what I should do now, I would be very happy if someone could give me an advise here, because this is my firts own empirical study and STATA is also quite new for me.

THANKS A LOT!