[Jeroen vanDam]

Thank you for your quick reply and helping me out once again Carlo! I think I know everything to continue with my analysis!

[Carlo Lazzaro]

Jeroen:
you are welcome.
All the best for your research project.
Dag!

[Jeroen vanDam]

Dear all, Dear Carlo,

I continued analysing my data and interpreted the results. I presented one overall model, next to that I thought I could deal with the time-invariant variables (and therefore omitted by STATA) by creating two seperate models that focused only on one of the two values of the dummies (so a model with if dummy==1 and one with if dummy==0). However, the N of both models becomes quite low (N=16 and N=23), thereby reducing the statistical power and reliability of my results. I'm doing research on the influence of TMT and supervisory board nationality diversity on cross-border M&A activity, thereby controlling for the corporate governance structure of the firm by using a dummy variable (1 for firms with a one-tier structure and 0 for firms with a two-tier structure). Another (control) variable that is omitted by STATA is the industry a particular firm is operating in, this variable doesn't change as well during the research period. According to my supervisor there are some options to include those time-invariant variables referring to this website: http://stats.stackexchange.com/questions/90754/how-to-keep-time-invariant-variables-in-a-fixed-effects-modelhttp://stats.stackexchange.com/questions/90754/how-to-keep-time-invariant-variables-in-a-fixed-effects-model. However, it remains unclear how I can precisely deal with those omitted variables, is the Hausman-Taylor Estimator a good option in this case? Or are there particular options that justify my choice of making two seperate models next to the overall model? I checked a lot of other forums on this website but it still remains quite vague what to do in this case. I hope someone can help me out with this!

Kind regards,

Jeroen van Dam

[Carlo Lazzaro]

Jeroen:
I don't think that running two diffferent models will fix your problem: both kind of firms will keep their time-invariant predictors as they are (regardless their dummy values, their industry, for instance, is not expected to change across years).
Besides, you have focused on -fe- specification only, so far: why did you rule out the -re- option, if you're interested in estimating teh coefficient of time-invariant predictors, too?

[Jeroen vanDam]

Carlo:

I ran two different models to make a comparision between the one- and two-tier structure since in this case it's my main interest. I ran the Hausman test for multiple hypothesis in my study. For this particular analysis it was 'allowed' to use the random effects model. However, this was not the case for all the hypothesis/relationships, for consistency I thought it was a good idea to stick to to the fixed effects option for all hypothesis. Now I realize that this might be wrong. One option is to deal switch to the -re- option. Question considering this: heteroskedasticity and autocorrelation is still present. Am I right that in this case the right command is , re vce(cluster Owncode ) 'Owncode' refers to the code I gave each individual firm in my sample. The second question that arrises is how to deal with the remaining hypothesis where it is not allowed to use random effects? In this specific hypothesis it is less important that both time invariant variables are included, however it would still be good to have them. Next to that, is this allowed using two different way of analyzing panel data in your research? I hope you can help me out!

Kind regards,

Jeroen van Dam

[Carlo Lazzaro]

Jeroen:
- I would go -re-; I do not follow you about te remaining hypotheses for which -re- would seem unappropriate;
- you code for clustered standard errors is correct;
- you may want to present both -fe- and -re- specifications and comment on the difference; however, since you have time-invariant predictors I would stick with -re- model only,

[Jeroen vanDam]

Carlo:

First of all, thank you with the comments considering the -re- option. Let me eleborate on the hypothesis issue:
I have 3 hypothesis: H1 & H2 have the same dependent variable. However, H3 has another dependent variable. The other variables remain quite the same. Therefore I ran two Hausmann tests (since the analysis is different from ). For H1 and H2 the -re- option is justified. This is not the case for H3, the Hausman test suggest using the fixed effects model. Due to this I'm a bit confused. Is it justified to use the -re- option in at H3? Or is it better to stick to the fixed effects model? Thank you in advance!

Kind regards,

Jeroen

[Carlo Lazzaro]

Jeroen:
thanks for making things clearer.
I see three options:
- running -re- for dataset H1 and H2;
- going -fe- for H3;
- presenting both -fe- and -re- for H3 if you're interested in estimating coefficients of time-invariant predictors notwithstanding -hausman- verdict.

[Jeroen vanDam]

Dear Carlo,

Thank you for your quick reply once again. I had one remaing question, do you perhaps know if it is suitable to use the Hausman-Taylor estimator in case of the fixed effects model? This model can control for time-invariant variables as I understood (basically a adjusted random-effects model) I read some literature about this, but it looks quite complicated to use though. I agree with the options you suggest, thank you!

[Carlo Lazzaro]

Jeroen.
as far as I know, the -xthtaylor- regression model is appropriate for -re- model in which the absence of correlation between individual effects and the vector of predictors is violated.

[Alex Karl]

Hi respected members,

I would like to have your suggestion about the use of Stata command for fixed effect GLS or "FEGLS" estimation. What should this look like? Many thanks in advance

[Kalle Bernhardson]

Carlo: Do you know where I can find the source for the message "large N, small T -> xtreg vce (robust) and small N, large T -> xtscc"?
I have 49 time periods and 59 panels. So i think its equal which one I take.
But I need a source for my working paper. I already looked at xtreg pdf.stata.
Thank you

[Mariam Malik]

Hello Carlo! Could you please tell me that which command i will use in order to remove heteroscedasticity and serial correlation in panel data. Can i use Newey-West standard error? In my study N=5 T=20

[Thiago Vizine]

Dear Carlo, I'm facing the same problem as my coleagues, and would like to ask for your help. I also have a dataser with large N and small T. For what I could understand, the comand

xtreg, fe vce (robust)

should solve heteroskedasticity and/or autocorrelation problems, right?. Can I use vce (robust) for random effects either, like

xtreg, re vce (robust)?

One more question, for what I read, first difference also solve heteroskedasticity and/or autocorrelation problems. I my test, first difference option presented better numbers for Standard Erros (small S.E) than the comand

xtreg, fe vce (robust)

Should I use first diference, or stay with the fe option?

Kind Regards

[Carlo Lazzaro]

Thiago:
welcome to this forum.
1) Correct: -xtreg, fe vce (robust)- and -xtreg, fe vce(cluster clusterid)- both take heteroskedasticity and/or autocorrelation into account;
2) Correct: -xtreg, re vce (robust)- and -xtreg, re vce(cluster clusterid)- both take heteroskedasticity and/or autocorrelation into account;
3) I would stay with -xtreg,fe- (with cluster-robust standard error) option.