Compare 2 groups, omitted because of fixed effects

[Fabian Neeskens]

Hello Statalist,

I'm comparing a list of ''Best Companies for Working Mothers'' with a peer group that never have been in that list (with the best suitable match, constructed by hand)
Want to compare the financial performance of the two groups ( ROA is the financial performance measure)
Y variable= ROA, controls are= FirmSize logemployees Risk ResearchandDevelopmentExpense, variables of interest are dummies= HighCompensation SupportforEducation
Listed YN=dummy variable, is a company family friendly=1, 0 otherwise.
Therefore a company has always a 1 when listed, and 0 when never be listed (no variety, so omitted)
I would like to compare the ROA, and see if listed has a positive/negative effect. However, it is omitted since a company will always receive the same value (1 or 0).
How can I solve it that I can use Listed yes/no to see the effect on ROA, with fixed effects? Or is there another way?

Hereby the stata codes:

Stata 13: I have 11 years of panel data, 1 observation per year, 1219 oberservations total.

destring HighCompensation SupportforEducation
xtset GlobalCompanyKey DataYearFiscal
xtreg ROA ListedYN FirmSize logemployees Risk ResearchandDevelopmentExpense HighCompensation SupportforEducation, fe cluster (CompanyName)


Thank you very much for helping!

Fabian

[William Lisowski]

Welcome to Statalist.

I'm sorry to advise you that any variable that does not change with time (which apparently is the case for your ListYN) cannot be included it in a fixed effects model. If you allow a fixed effect for a firm, and if ListYN has the same value for every observation of that firm, there is no information to estimate how much of the firm's effect is due to ListYN and how much is due to the fixed effect.

[Carlo Lazzaro]

Fabian:
welcome to the list.
As an aside to William's hepful insight, you may want to test via -hausman. which is the preferable specification for your dataset, keeping in mind that, as reported by any decent econometrics textbook, both -fe- and -re- have their methodological weaknesses (as Wiliam pointed out for fixed effect; for -re- is the often untenable assumption of correlation=0 between the vector of predictors and the individual effects).

[Fabian Neeskens]

Dear William and Carlo,

Thanks a lot for such a fast reply on my question!

I would like to explain my problem again then, because my initial explanation is probably wrong.

I have panel data for 100 firms who are or were in a list called 'Best Companies for Working Mothers'. For all of these companies I have found a peer based on industry, market cap, etc. Now I would like to test whether the companies who are in the list outperform their peers. I want to control for size and d/e ratio and some social dummy variables. What regression(s) would you advise me instead?

I already did the hausman test as you advised. The result of the test for me was to use a model with fixed effects.

I have got dummies which indicate whether a company is a 'Working Mother Company' or a peer, and whether the company or the peer is indeed in the list in that particular year.

Thanks in advance!

Kind regards,
Fabian

[Carlo Lazzaro]

Fabian:
what does the F-test at the foot of -xtreg, fe- outcomem table tell you?

[Fabian Neeskens]

Dear Carlo,

Hereby:

F(11,81) = 4.79
corr(u_i, Xb) = -0.5321 Prob > F = 0.0000


sigma_u | .07459418
sigma_e | .08120718
rho | .45763123 (fraction of variance due to u_i)

I hope this is what you are looking for

[Carlo Lazzaro]

Fabian:
the result of the F-test tells you that -xtreg, fe- is "better" than pooled OLS.
However, the issue concerning the time-invariant effects you're interested in still remains; they are cancelled out via -fe- specification.
I would try -re-, despite -hausman- test result.
Eventually, I do hope that http://journals.cambridge.org/abstra...49847014000077 can offer you some guidance.

[Fabian Neeskens]

Dear Carlo,

Someone came with the suggestion to regress with only xtreg (so without ,fe or ,re at the and) but with industry and year fixed effects.
What do you think about that suggestion?

[Rich Goldstein]

if you don't put an re or fe or anything else, you will get an re as that is the default - see the help file

[Fabian Neeskens]

Dear Rich,

Thanks for your answer.
What is better, random effects or a normal regression with the industry and year fixed effects? Since the hausman test gave the result that I have to use fixed effects, but to see the difference between to groups I simply can't use fixed effects.

Thanks!

[Carlo Lazzaro]

Fabian:
I would go -xtreg, re- with vce(robust) option,

If [you] suspect that there is heteroskedasticity or within-panel serial correlation.

(quoted from Stata 13.1 .pdf manual, -xtreg- entry, page 372).

[William Lisowski]

The Stata FAQ on Fixed-, between-, and random-effects and xtreg has nicely worked examples that seem to bear on this. In particular, there are three examples for the fixed effects models, which produce identical results, with the third example being the use of the regress command (rather than xtreg) with dummy variables for the individual units (firms, in Fabian's case). The difference between xtreg, fe and regress with dummy variables seems to come down to whether or not you want to see the actual unit-specific effects (and I suspect to efficiency in the underlying calculations; with under 200 firms that shouldn't be an issue for Fabian's data).

With that said, I am not an expert on fixed- an random-effects models and may well have overlooked something. I defer to those, like Carlo, with more experience.

[Carlo Lazzaro]

I second William's skillful remark and referring to the Stata thread that he mentioned, probably Fabian's query rests on the following sentence

People generally want to use the random-effects model because they wish to estimate the variables that are constant within unit

.
That's why, despite -husman. results, I would go -xtreg, re-.
However, as it is always the case, the last call about what specification fits data "better" than the alternatives is Fabian's most crucial task, also in light of what is disseminated throughout the literature in his research field.

[Fabian Neeskens]

Dear William and Carlo,

Your explanations are clear, thank you very much. I would like to use cluster on a company level:
xtreg ROA everinlist FirmSize logemployees Risk rndlog HighCompensation SupportforEducation WorkforceReductions EmployeeRelations WorkLifeBenefits EmployeeInvolvement $ind $yd if peermirroringFFlist ==1, cluster ( CompanyName)

It works properly, but however when I want to do regress the model with only the control variables (HighCompensation till EmployeeInvolvement are dummies):
xtreg TobinsQ everinlist FirmSize logemployees Risk rndlog $ind $yd if peermirroringFFlist ==1, cluster ( CompanyName)

I receive the message: panels are not nested within clusters.

The strange thing is that it works with the regression with all the variables included, and doesn't with only the control variables (when there is just one variable of dummies included it still works).

Do you know what is wrong?

Thanks for thinking along!

[Fabian Neeskens]

To add something to my last post, is the regression still okay when adding nonest?
xtreg TobinsQ everinlist FirmSize logemployees Risk rndlog $ind $yd if peermirroringFFlist ==1, cluster ( CompanyName) nonest

It gives the same result, but it defenitely changes something.. Since you gave me the advice of vce (robust) I'm wondering what to do..