Am 06.03.2023 um 19:21 schrieb Cottrell, Allin:
> On Fri, Mar 3, 2023 at 10:52 AM Sven Schreiber > <sven.schreiber@fu-berlin.de> wrote: >> >> I suspect that the lrvar function (long-run variance estimation) is >> slightly wrong in the panel case. Consider the following example >> [...] > > Well, it's more than slightly wrong: the current function is only > applicable to straight time-series data and should not be allowed > for panel data. > > As a stop-gap measure I've put in a block against using it on panel > data, with a specific error message. (In git, will shortly be in > snapshots.) But what, if anything, _should_ we return for panel > data? An average (weighted or unweighted) of the per-unit long-run > variance values (as per your script), or a vector of per-unit values, > or ...?

Thanks, Allin. Yes, with panels it gets more complicated, as always. The two possibilities that you mention are the most natural ones, I guess. (To be clear, my script didn't consider different per-unit true values, but used the average of the per-unit statistics to estimate the homogeneous value.)  In principle one could also impose homogeneous means instead of doing a fixed-effects type demeaning.

cheers

sven