[Gretl-users] Re: panel regression HAC

Am 06.09.2023 um 11:05 schrieb Alison Loddick: Hi Alison, thanks for fighting on behalf of gretl! As mentioned in section 22.4 of the user guide, the Arellano cov estimator for panel data can be seen as HAC (at least for small T), and it says that an alternative denomination would be “clustered (over entities)”. Actually, I'm looking at the Eviews program (13) and their help files to check out the situation there right now. There's no explicit mention of HAC in the panel options, but you have "period cluster", "cross-section cluster", and "two-way cluster" (all combined with the "White" name). So it looks like gretl's "Arellano" option is conceptually the same as Eviews's "White period (cross-section cluster)". What do your colleagues mean exactly when they say they use HAC for panel regression? It's not obvious. A related way of looking at this issue is that gretl's native panel tools do not handle cross-sectional dependence (CSD, contemporaneous error correlation). From your description it sounds as if the time-dimension (within) correlation is all you want to take into account or allow, and then you're good to go with core gretl. If, however, you also worry (or want to teach) CSD, then I would point you to the "CSDpanel" contributed function package written by Jörg Breitung and myself, currently at version 0.2 from three years ago. Among some other tools, we offer the two-way clustered standard errors suggested by Driscoll and Kraay (1998). (I guess this should correspond to Eviews's two-way cluster option, but I don't know why they don't mention the names of Driscoll/Kraay or whether there are any other subtleties.) So it seems that the only option we don't have available in the gretl ecosystem is "White cross-section (period cluster)". Not sure how relevant that is. Everybody is welcome to correct me if I'm misinterpreting something in this area. cheers sven