[Richard Williams]

Here is a very old post by Bill Gould on such issues:

http://www.stata.com/statalist/archi.../msg00193.html

Without seeing actual output (and maybe not even then) I suspect nobody is going to be able to help you further. But I've been wrong before.

[Mattia Calisi]

Sorry for my bad English, i am an italian student and I have the same problem:
I would like to use the regression with fixed model but 9 variable are omitted because of collinearity.
8 variables are dummy and one is a number (age of firms).
My model is RETURN ON EQUITY = AGE+ SIZE + 8 DUMMY that describe the 9 sector of activity in which firms operate (Eletric,manufacturing,health care ecc).

SO how I can include the 9 variable that are omissed from fixed model ?

[Clyde Schechter]

The answer probably is that you can't. If each firm remains in the same sector at all times in your data, then the sector variables will necessarily be collinear with the fixed effect, and they cannot be estimated in a fixed effects model. Look into -xtreg, be- to see if that will give you what you are looking for.

[Mattia Calisi]

thank you very much! sorry Professor Schechter but I have more question:
1)The variable AGE is omissed because of collineary but it's a numeric variable. Every firm has got a number that it's his age
2)How you suggest to me? To regress my panel data using only random effect?

[Clyde Schechter]

1. This is surprising. Are you sure age is calculated correctly? Do you have multiple observations for each firm--and the age changes from one observation to the next? If age is constant within firm, then it, of course, will be dropped just like other time-invariant effects when you use a fixed-effects model. But it is unusual in a panel data set (which I assume you have, though you don't actually say that) for age not to vary. So I wonder if you either have just one observation per firm, or have non-varying (and hence, probably incorrect) values for age. If this suggestion doesn't resolve your question, in your next post please show us, in a code block (see FAQ for how to do that if you don't know), the exact command you gave and the exact output you got from Stata. That is best done by copying and pasting from the Results window into the code block: do not retype it yourself.

2. I am loathe to advise you on this. The econometrics and finance communities have strong opinions about when it is permissible to use random effects models, typically being when a Hausman test says it's OK. While I don't disagree with the reasoning behind that, suffice it to say that this is something that is fairly routinely disregarded in the health-related fields, where random effects models are pretty routinely used without verifying consistency (perhaps because we typically have a multi-level model and the use of multi-level fixed effects is challenging at best). So, lacking much experience with that aspect of things, I will abstain from making a recommendation of that nature here. You might want to consult a colleague in your discipline about this.

That said, I still recommend you look at -xtreg, be-, as it may give you the parameter estimates you are looking for, though they will not be adjusted for other time-invariant firm attributes.

[Mattia Calisi]

Thanks you Professor Schechter ! Your post is very useful, I will look at xtreg, be and consult with some my collegues

[Mattia Calisi]

Test of Hausman say that it's better using random effect, there area other test that i can use? To better understand if my model is a good model?