Hello panel devotees, I believe I'm seeing a bug in cumulated sample restrictions in the panel context. Before I can set up a minimal example, let me describe hopefully in enough detail what I'm doing, what's failing, which workaround exists, and so on. I can't supply the dataset (yet) because it's big. The full data is structured and recognized by gretl as follows (I believe the translated German terms are self-explanatory in this context): Voller Datenbereich: 1:001 - 5860:900 (n = 5274000) This is monthly, in general heavily unbalanced, the index variables stemming from $unit and $obsdate are "stationindex" and "iso" (for ISO date format). For certain data cleaning reasons I'm first applying the following restriction: (1) smpl STATIONS_ID==291 --restrict --replace which gives me: Voller Datensatz: 5274000 Beobachtungen Aktuelle Stichprobe: 1288 Beobachtungen So far, so good. Then I want to keep only those obs where a certain variable "temp" is non-missing, so I do: (2) smpl ok(temp) --restrict which then yields: Voller Datensatz: 5274000 Beobachtungen Aktuelle Stichprobe: 1:001 - 2:156 (n = 312) Notice that gretl by itself figured out that this subsample has a valid and feasible panel structure, I didn't tell it that. Does this side effect matter? I don't know. At this point let me demonstrate that there are no missings in this subsample; in the console, asking for "=sum(missing(dataset))" yields 0. Finally I want to retain only the most recent time period, and since "iso" is my time variable, I try to apply this restriction: (3) smpl iso==max(iso) --restrict In theory there should be at least one observation where any given variable takes on its own max (and in fact in this case there are two such obs) -- however, now I'm getting an error message: "No obs would be left!" (retranslated) This can't be true, can it? So I tried the following workaround before getting to step (3), i.e. starting from scratch and applying restrictions (1) and (2), then I create a helper series: series checkiso = iso==max(iso) And indeed, "=sum(checkiso)" gives me the correct answer 2. Now I try the equivalent workaround restriction: (3b) smpl checkiso --restrict ... and it works nicely, as expected, leaving only two obs in the active sample! So why does step (3) not work? I believe it's a bug, but maybe I'm missing something. Ah, one more remark: For other values in the step-(1) restriction I'm seeing that there won't be a recognized panel dataset after step (2), and apparently in those case there is no problem! That's why I mentioned it above, maybe it's relevant. Sorry for a long message, I hope I was clear enough. thanks sven