Confusion w/ panel data, fixed effects, dummies, clustered standard errors etc

[Tobias Maier]

Hey STATA community,

I currently do a project and now I am stuck in the data analysis. Therefore, I need your help AND sorry for opening a new thread with these basic questions. I just want to understand to STATA command once that I can apply it in the future without major questions..

I have firm-level data over multiple years. The literature I am following uses sector and year dummies, pseudo fixed effects and clustered standard errors on industry-level. Now I wonder, how I can fit all that into one model.

I set my panel using:
xtset companynum DataYearFiscal, yearly
Whereas companynum is the company identifier and DataYearFiscal is the year.

So far, I managed to build the following model:
xtreg DepVar IndVar1 IndVar2 ... IndVarn i.DataYearFiscal i.companynum, vce (cluster Industry)
I have some questions to better understand the command.
Does is fit the above-mentioned requirements?
i.DataYearFiscal and i.companynum are dummies, aren't they? Are they used to cover time-variant effects and firm-variant effects respectively?
vce (cluster Industry) says clustered standard errors on industry-level to account for differences between industries (e.g. manufacturing vs retail). What happens if I exclude it?
If I include fe, i.companynum is ommited because of collinearity.

The above-mentioned model, however, has an R-Square between = 1.0000. That seems odd to me. Why is this so? Where did I take the wrong exit?

Thank you very much!
Tobi

[Carlo Lazzaro]

Tobias:
welcome to this forum.
Some comments about your query:
- your -xtset-looks OK;
- -i.companynum- is redundant, as you have already included this fixed effect in -xtset-. As expected, it is wiped out by -fe- machinery, being a time-invariant predictor;;
- your take about -i.DataYearFiscal and- i.companynum- is right. However, I would recommend you to take a look at any decent textbook on panel data econometrics;
- cluster/robust option accounts for heteroskedasticity and/or autocorrelation;
- your R_sq looks actually strange. What is the outcome of -xttest0- after -xtreg,re-?

[Tobias Maier]

Hi,

Thank you for your quick response!

I do not need to include either i.companynum or fe because I use a panel regression!? But what does "pseudo fixed effects" refer to? I am a bit confused by the "pseudo".

I will consult a text book to check how I apply them correctly.

Ok. I do understand the concept of heteroskedasticity. However, does is make sense the do that on industry-level if the panel is on firm-level?

Regarding the R-square issue, my outputs are:

R-sq:
within = 0.0764
between = 1.0000
overall = 0.1672

Wald chi2(14) = .
Prob > chi2 = .

xttest0

Breusch and Pagan Lagrangian multiplier test for random effects

DepVar[companynum,t] = Xb + u[companynum] + e[companynum,t]

Estimated results:
| Var sd = sqrt(Var)
----------+--------------------------------------
DepVar | 1773.943 42.11821
e | 1739.027 41.70165
u | 0 0

Test: Var(u) = 0
chibar2(01) = 0.00
Prob > chibar2 = 1.0000

I have honestly no clue what this tells me now. *facepalm*

[Carlo Lazzaro]

Tobias:
as suspected from the sky-rocketing between R-sq, the outcome of -xttest0- tells you that your dataset does not show evidence of panel-wise effect; hence, you should switch to pooled OLS.
On pseudo-panels (ie, cohorts of individuals instead of individuals followed-up over time), you may want to take a look at: https://www.insee.fr/en/statistiques...uillerm_EN.pdf.

[Tobias Maier]

Hi Carlo,

I read through the document. However, I can't get my head around the idea, why I have a pseudo-panel rather than a normal one. My data is on individual firm level and I do follow individuals for the whole time period. Is it because my panel is unbalanced?

Do unbalanced panel require different methods/commands or controls? Or do I have to drop the firm-years with missing variables? I thought, STATA ignores them anyway...

Thank you very much!
Tobi

[Carlo Lazzaro]

Tobias:
if you have a sample composed of the same units (firm, in your case) that you follow-up over years, you have a panel dataset.
The fact that your panel is unbalanced has nothing to do with pseudo-panel definition.
Eventually, as you correctly surmise, Stata can handle both balanced and unbalaced panel datasets with no problem (and no difference in the -xt- suite commands to be used for your regression).
That said, as per -xtttest0- outcome, your panel dataset does not show evidence of panel-wise fixed effect; hence, it should be estimated vi pooled OLS.

[Tobias Maier]

Hi Carlo,

Sorry for not getting back to you! Thank you very much for your explanation. It helps tremendously!

All the best