gretl 2026d, Windows 11 Pro 64-bit.
The User's Guide, section 22.5 ("Panel data specifics", p. 218), says two things about panel data:
--cluster for two-way clustering;$unit and $time.In 2026d only the exact form --cluster=$unit,$time clusters two ways. If either name in the pair is a series (firm,year, $unit,year, firm,$time), the standard errors are those of one-way clustering on the second variable. ols on panel data does the same. Nothing warns the user:
The attached cluster-named-series.inp uses grunfeld, where firm equals $unit and year equals 1934 + $time in every row:
open grunfeld
panel invest const value kstock --pooled --cluster=$unit,$time
panel invest const value kstock --pooled --cluster=firm,year
panel invest const value kstock --pooled --cluster=year
Standard errors of const, value and kstock:
--cluster=$unit,$time 19.71668032 0.01639515 0.07954319 (two-way)
--cluster=firm,year 10.27288943 0.00790949 0.03867233
--cluster=$unit,year 10.27288943 0.00790949 0.03867233
--cluster=firm,$time 10.27288943 0.00790949 0.03867233
ols --cluster=firm,year 10.27288943 0.00790949 0.03867233
--cluster=year 10.27288943 0.00790949 0.03867233 (one-way by year)
Expected: all the two-way forms give the first line. Three one-way fits, V(firm) + V(year) − V(firm × year), each with gretl's own factor, give it to every digit (19.71668032, 0.01639515, 0.07954319). So only the parsing of the named pair seems to be at fault. On cross-sectional data a pair is refused, as the guide says.
Thanks for the report. We're working on it.
This problem is now fixed in git.