Panel data: all dummies omitted because of collinearity

[loppayeah]

Hello all.

I'm working on a panel data to assess the determinants of debt/equity ratio in firms through the years. i'm analysing it with fixed effects technique. I did the LSDV model and the model doesn't present any issues, all the dummies i created (except very few ones) are specified.
however, when i try the Within group estimator via Linear regressions(FE,RA,BE) --> fixed effects, the output i have is unclear, as it says that all the dummy variables are omitted because of collinearity. as suggested on line, i dropped dummy number one right to avoid the multicollinearity issue which however comes up.
here is a screenshot.
http://minus.com/i/bcFjiTLzENgTT
does anyone know the reason? how can i resolve the issue?
thanks a lot

Leopold

[Maarten Buis]

The dummies are probably constant within grouping variable (A). Since the fixed effects estimator takes out all the variance at the group level, there is nothing left for those dummies to explain. If you only included those as controlls, then there is nothing you need to do: the output is exactly what you want. If you are substantively interested in these dummies then you have a problem.

[E. David Aja]

One way of thinking about fixed effects is that they're computed by measuring deviations from an entity-specific mean. If you try to include dummy variables for entities in a fixed effect estimation, you're attempting to estimate the difference between the identity of x_i in periods 1 through T. In other words, by selecting fixed effects, you've already "included" all the identity dummies, and so doing it again will result in the multicollinearity problem you're seeing.

I'm sure someone can give a much more technical explanation than that, but the takeaway is that if you're using fixed effects there's no need to include entity specific dummies.

[wanhaiyou]

Maarten Buis has provided an excellent explanation. In addition, the paper "Efficient Estimation of Time-Invariant and Rarely
Changing Variables in Finite Sample Panel Analyses with Unit Fixed Effects" written by Thomas Plu¨mper and Vera E. Troeger
provide another perspective on this issue.

[Kim Lam]

The dummies are probably constant within grouping variable (A). Since the fixed effects estimator takes out all the variance at the group level, there is nothing left for those dummies to explain. If you only included those as controlls, then there is nothing you need to do: the output is exactly what you want. If you are substantively interested in these dummies then you have a problem.

Hi,Maarten Buis.

I'm working on a project which faces collinearity problem. Unfortunately, my variables of interest is a set of dummies. May I know is there any solution to this tough problem?
Your suggestions would be much appreciated.

Best,
Kim

[Carlo Lazzaro]

Kim:
Others have already given full explanation of your problem.
Just two asides:
- in general terms, it is much more better to construct dummies via -fvvarlist- than by hand;
- are you sure that you can't go -xtreg, re-? Dis you perform a -hausman- specifcation test?

[Maarten Buis]

I'm working on a project which faces collinearity problem. Unfortunately, my variables of interest is a set of dummies.

What is the problem exactly? A set of indicator (dummy) variables in itself is not a problem. Are all indicator variables dropped, or just one, or a few but not all? What is the model you used: a linear regression, fixed effects panel data, ...?

[Kim Lam]

Dear Carlo Lazzaro , Maarten Buis,

Thank you so much for your reply and sorry for my simple description about my problem last time.

I have 18 years panel data(from 1995-2012).
Using the Lin-log regression, and based on the hausman test, I chose the fixed-effect model and introduce i.year

xtreg Y lnX1 lnX2 D1 D2 D3 D4 D5 i.year, fe vce(cluster id)

Dummies are:
D1=1 if year>=2006, otherwise 0
D2=1 if year>=2000, otherwise 0
D3=1 if year>=2003, otherwise 0
D4=D1*D2
D5=D1*D3

However, more than 1 year would be omitted because of collinearity (e.g:2009.year omitted because of collinearity except 1995). If I drooped D1, this problem does not exist. But D1 is one of the key variables, I had no idea to solve the collinearity problem after lots of trials. May I ask for your further suggestion? Thanks so much.

Best regards,
Kim

[Carlo Lazzaro]

Kim:
I fail to get why D1 (and so on) should live together with i.year in your model.

[Kim Lam]

Kim:
I fail to get why D1 (and so on) should live together with i.year in your model.

Dear Carlo:

I set D1(and so on) to study whether some events have influence on dependent variable, while using i.year to eliminate the yearly variation. I realized something's wrong when I used year dummies together with i.year. But how could I eliminate the rest yearly variation unrelated to dummy years?

There would be much appreciation for your help.

Best regards,
Kim

[Maarten Buis]

I realized something's wrong when I used year dummies together with i.year. But how could I eliminate the rest yearly variation unrelated to dummy years?

You cannot take out the year to year variation. In your design you assume that the yearly variation is due to the events you studied, so it is impossible to seperate the two. That is the weakness of your design, and there is little you can do about it. A partial solution would be to include year as a continuous variable to take out a trend, but that is not quite what you seem to be after.

Worry a bit about the weakness of your design, but don't worry too much: all designs have their weaknesses. Just think of your study as adding one piece of evidence. In time you and others will do other studies adding more evidence, and after a while someone will write a review article bringen all that evidence together. So each study individually does not have to solve all problems; the hope is that in time a sufficient body of evidence will develop to say something more conclusive.

[Kim Lam]

You cannot take out the year to year variation. In your design you assume that the yearly variation is due to the events you studied, so it is impossible to seperate the two. That is the weakness of your design, and there is little you can do about it. A partial solution would be to include year as a continuous variable to take out a trend, but that is not quite what you seem to be after.

Worry a bit about the weakness of your design, but don't worry too much: all designs have their weaknesses. Just think of your study as adding one piece of evidence. In time you and others will do other studies adding more evidence, and after a while someone will write a review article bringen all that evidence together. So each study individually does not have to solve all problems; the hope is that in time a sufficient body of evidence will develop to say something more conclusive.

Thank you so much for your help and encouragement!

[Friederike Osterndorff]

Hi,

I have a related question. I also use Panel data and 2 of my dummy variables are ommitted. This is the regression:

xtreg enrolment_rate_all csg_beneficiary age hhsize female mother_educ lhhincome Black traditional if age>=5 & age<=20, fe vce(cluster pid)

the hausman test "suggested" fixed effects. However, now the female and race(Black) variable are ommittted. Which makes sense, since they don't change over time, but I would still like to see if they affect the school enrolment rate.

Any tips will be highly appreciated.

Thanks

Friederike

[Carlo Lazzaro]

Friederike:
-hausman-outcome should not be considered as the Holy Bible.
If you're interested in time-varying predictors, you can go -re- and compare the results with the ones obtained with -fe-.
Just out of curiosity: did you run -hausman- before -cluster-ing your standard? Thanks.

[Friederike Osterndorff]

Hi Carlo,

thanks for the response. Yes I did run the hausman before cluster-ing, since the hausman did not allow the vce(cluster) option.