Is it advisable to perform DK Standard Error Estimations for RE?

[Titiksha Das]

I have come across many papers which have performed DK-FE but none with DK-RE. If my Hausman Test statistic is proving RE the winner, then to prove robustness can I go for DK-RE instead of Robust RE?

[Titiksha Das]

I understood so from the Hoechle (2007) paper that its ok to use DK_RE but then why are there no papers with the same estimations?

[Carlo Lazzaro]

Titiksha:
take a look at the helpfile of the community-contributed module -xtscc- (rnethelp "http://fmwww.bc.edu/RePEc/bocode/x/xtscc.sthlp").

[Titiksha Das]

Titiksha:
take a look at the helpfile of the community-contributed module -xtscc- (rnethelp "http://fmwww.bc.edu/RePEc/bocode/x/xtscc.sthlp").

Thank you so much! I went through it which mentions that its technically possible but then I have not come across any journal papers or standard literature that has done the estimation of DK-RE. The most that papers have discussed are Robust RE to prove the robustness of RE but not DK-RE.

[Andrew Musau]

You seem to be conflating two issues: the choice of model (FE vs. RE) and the choice of standard errors (DK vs. cluster-robust). The choice between FE and RE is not arbitrary; it is determined by the sample data. Unless you have a randomized intervention, which is quite rare in the social sciences, you most likely cannot justify using RE. Therefore, with the exception of experimental studies, RE models are uncommon in the social sciences—not only in cases where DK standard errors are employed. On the other hand, the choice of standard errors depends on the nature and dimensions of your data. For an \(N \gg T\) panel, cluster-robust SEs address heteroskedasticity and arbitrary forms of serial correlation, while DK standard errors rely on large-\(T\) asymptotics.

[Titiksha Das]

You seem to be conflating two issues: the choice of model (FE vs. RE) and the choice of standard errors (DK vs. cluster-robust). The choice between FE and RE is not arbitrary; it is determined by the sample data. Unless you have a randomized intervention, which is quite rare in the social sciences, you most likely cannot justify using RE. Therefore, with the exception of experimental studies, RE models are uncommon in the social sciences—not only in cases where DK standard errors are employed. On the other hand, the choice of standard errors depends on the nature and dimensions of your data. For an \(N \gg T\) panel, cluster-robust SEs address heteroskedasticity and arbitrary forms of serial correlation, while DK standard errors rely on large-\(T\) asymptotics.

Thank you so much! My selection of RE model is based on the Hausman Test Statistic which tells me that RE is winning instead of FE. My sample panel data is for 33 years and 10 countries which makes it technically possible for operation of DK Standard errors

[Jeff Wooldridge]

It’s hard to justify using random effects in this situation. The Hausman test cannot tell you to use RE. The nulll hypothesis is that RE is consistent. Sometimes the null is not rejected when it should be. Aggregated data calls for FE. Moreover, with N only 10, we can’t trust N going to infinity asymptotics, which justifies RE. Plus, T is notably larger than N, and with serial correlation the usual Hausman test has an unknown limiting distribution. You should use fixed effects with DK standard errors. And T = 33 is maybe just enough to do it.

[Titiksha Das]

It’s hard to justify using random effects in this situation. The Hausman test cannot tell you to use RE. The nulll hypothesis is that RE is consistent. Sometimes the null is not rejected when it should be. Aggregated data calls for FE. Moreover, with N only 10, we can’t trust N going to infinity asymptotics, which justifies RE. Plus, T is notably larger than N, and with serial correlation the usual Hausman test has an unknown limiting distribution. You should use fixed effects with DK standard errors. And T = 33 is maybe just enough to do it.

Thank you so much for your input, Prof. Woolridge