[Gretl-users] Re: Question on "meantest" command

Am 11.12.2024 um 07:57 schrieb Artur T.: Good point, thanks for reminding us. I've done some research in my email archives and for the record I will reproduce some points here that are useful to know in this context, I believe: - The Mann-Whitney test is the same thing as the Wilcoxon rank-sum test. Unfortunately, it is a bit misleading that in mwu's help text it only says "Mann-Whitney U test is similar to non-parametric Wilcoxon rank sum test". (That's because there the parametric t-test is also mentioned, so it's not wrong, but doesn't mention the equivalence of Mann-Whitney and rank-sum.) - Maybe not universally known: gretl does the Wilcoxon rank-sum test on x and y via the command "difftest x y --rank-sum", but also quite accessible in the menu via Tools/Non-parametric Tests/ Difference Test (and then there's the option button Wilcoxon rank sum test). - What's the remaining purpose of the mwu.gfn package then? It assumes a different data layout. Instead of having the two variables (or "samples") in the two variables x and y, it takes the data to be in one single series, plus a binary indicator series d which says whether an observation belongs to the first sample or the second one. Below I'm attaching a hansl script with a working demo and comparison. - Maybe we could have such a dummy-based interface as an option in the standard Difference test dialog window as well? The extension would be to accept a binary dummy as Variable 2, and if that's the case, then split Variable 1 in two subsamples and run the test on that, like mwu.gfn does. Perhaps an additional tick box would be needed to actually convey the meaning of the dummy, don't know. - (Technical note: the dummy trick shown below works nicely at the hansl level with the --rank-sum option, but fails due to missing values with the --sign and the --signed-rank options. Some more data handling would be needed there, but that's perhaps slightly off-topic here.) cheers sven ------ <hansl> # Alternative function mimicking functionality of the mwu package function void ranksumtest_withdum(series x, series d "indicator dummy")     errorif(!isdummy(d), "need dummy")     series group_d1 = d ? x : NA     series group_d0 = !d ? x : NA     difftest group_d1 group_d0 --rank-sum end function open denmark # artificial grouping for illustration series dummy = randgen(b, 0.3, 1)    # Bernoulli ranksumtest_withdum(LRM, dummy) ## compare with mwu include mwu.gfn mwu(LRM,dummy) </hansl>