On Mon, 1 Jun 2026, Artur T. wrote:

Sorry for capturing this thread for another thing:

1. It's not clear to me whether setting the lag to "-1" in the panel

case also yields the optimal lag selection according to Schwert

(1989). I am unsure as there is a separate "Panel data" section but

nothing is mentioned about this. Actually, the same question arises

for the "--test-down" option.

Hmm, no, the Schwert rule was not being applied in the panel case;

instead -1 was just being "translated" to 1. That's now fixed, but

with a caveat. The time-dimension of panel datasets is often quite

short, and if we apply Schwert's L_{12} rule unmodified it's quite

likely that the ratio of observations to lags will be too small to

compute the Im-Pesaram-Shin statistics. For now I've put in a

debatable hack: we compute the L_{12} lag order, but if the dataset is

a panel with T < 24 we actually take the lag as the minimum of the

L_{12} value and 4.

The --test-down option works fine with panel data. I've made some

improvements to the adf doc and to the working of the --verbose

option. Specifically, if you want to see exactly what's going on in

the panel case, you can give --verbose=max.