Do I need to cluster the standard errors?

[Melina Selia]

Dear all,
I have a panel data set of a study where 69 individuals were observed over 8 points in time with regards to several questions (unbalanced panel with an average of 7.1 observations per individual, number of obs = 492). According to the Hausman test, I should use the random effects model over the fixed effects model. What I was wondering now is whether I should use -vce(cluster participant)- in order to relax the independence assumption i.e. that observations of the same individual may be correlated over time.

Can somebody help me with this issue?

Thank you very much in advance!

Melina

[Dario Maimone Ansaldo Patti]

Hi,

If you estimate a panel using robust standard errors, actually Stata clusters them by id to account for observations nested within groups.

[Melina Selia]

Hi Dario,
Thanks for your quick reply. I am aware that with -xtreg- vce(cluster, clustervar) and vce(robust) do the same thing. But do I understand your answer correctly, that you recommend to use robust/clustered standard errors for my analysis?

Basically I am deciding between the two:

xtreg self_eff neg_ext neg_int, re vce(cluster, participant)
xtref self_eff neg_ext neg_int, re

whereas self_eff is a measure for self-efficacy, and neg_ext resp. neg_int are measures for negative context influencing self-efficacy.

Thanks!

[Dario Maimone Ansaldo Patti]

Yes. In my opinion you should do it to overcome the possible sources of heteroskedasticity and within serial correlation you may have. Using re and fe do not change the need of having robust standard errors (or clusteerd ones in this case) if this is your concern.

[Melina Selia]

Ok. thank you very much!
With regards to heteroskedasticity I have another question. Namely, with -xtreg, re- it is not possible to use xttest3 to check for heteroskedasticity and I could not find another heteroskedasticity-test for a random effects model.
But when I run -xttest3- on my data after using a fixed effects model, I get a clear sign of heteroskedasticity. Is it possible to infer from this, that there is also heteroskedasticity in the random effects model?
I am pretty sure there is but I need some kind of proof for my thesis...

thanks in advance for all the help, much appreciated!!

[Andrew Musau]

If you cluster your standard errors, you do not need to test for heteroscedasticity. The reason is that cluster-robust standard errors are valid both in the presence and in the absence of heteroscedasticity.

[Eric de Souza]

Just seen this thread. If I understand well, you first test for FE versus RE before clustering using the Hausman test,, and then you ask whether you should cluster.
But if you needed to cluster the Hausman test is invalid.
You should first cluster and then use the "robust" Hausman test which is not the same as the -hausman- test.
The robust Hausman test is based on the "correlated random effects model" or the Mundlak model.

[Melina Selia]

Andrew, thank you for clarifying. In that case I will cluster my errors anyways.

Eric, thanks also for your message. I indeed used the robust test -xtoverid-, sorry for not being specific enough there.

[Melina Selia]

Hi all,
I have posted another question in the forum but have not received any answer yet. As your advice was very helpful for this question and the questions are somewhat related I wanted to ask you whether you could take a look at that other question? Would be so much appreciated!!

Please find it with this link: https://www.statalist.org/forums/for...ar-and-xtlogit

thank you very much and have a good day everyone!