Dear research community,
After running several tests (including F-Test, Breusch and Pagan’s (1980) Lagrange Multiplier (LM) Test and Hausman (1978) Test) I came to the conclusion that a fixed effects model is the most appropriate one for my data.
To ensure that the estimates are efficient I run a couple of diagnostic tests.
Following a modified Wald statistic the idiosyncratic errors seem to be heteroskedastic. However there is no evidence of serial correlation following the test proposed by Wooldridge (2002).--> here the sub-question if it is correct to run the command "xtserial" after:
egen countrynum = group(Country)
xtset countrynum
xtset countrynum Year, yearly
xtreg DV IVs, fe
The main questions is whether I can make use of robust (sandwich) estimators to correct for heteroskedasticity even though there seems to be no autocorrelation problems?
Thanks for your help in advance.
Best
Frederic
Dear Frederic,
Yes you can use the robust (sandwich) estimators to correct for heteroskedasticity error. For more details please read page 38 of below attachment. https://www.princeton.edu/~otorres/Panel101.pdf
Thank you for your reply Dilesha,
actually I came to the conclusion that the non-clustered robust estimator is known to be invalid for panel data. I have used the clustered standard error as it is, itself, also robust to heteroskedasticity.
Apparently, if you, incorrectly, give Stata the command -xtreg DV Ivs, fe vce(robust)-, Stata will actually disobey this command and substitute vce(cluster country_num) without even adverting to it. As such, results are identical for both commands.
Yes, you can use it. Furthermore, if you know there is no serial correlation, you can cluster the standard errots at a lower level (unit-time), increasing your degrees of freedom.
I am running a Fixed effect model and detected heteroscedasticity. Now after clustering the standard errors, my all results became insignificant...what should I do now?