#345 Rank-sum reports n1 = n2 = full sample size and z = nan

[Artur Tarassow]

Hi,

here some report on the difftest --rank-sum test:

Version / OS: gretl 2026e-git (gretlcli), ubuntu

Description:
difftest <series> --split-by=<dummy> --rank-sum does not split the sample into the two groups. Both n1 and n2 are reported as the size of the whole current sample. The variance of the rank sum is then wrong, and on keane.gdt the z statistic comes out as nan. The parametric counterpart meantest ... --split-by splits correctly.</dummy></series>

Minimal script:

open keane.gdt -q
smpl status==3 && year==87 --restrict --permanent
summary wage --by=manuf --simple             # n = 948 (manuf=0), 290 (manuf=1)
meantest wage --split-by=manuf               # OK, meaningful t-test result
meantest wage --split-by=manuf --unequal-vars   # OK
difftest wage --split-by=manuf --rank-sum    # WRONG

open mrw.gdt -q
summary gdpgrow --by=nonoil --simple
difftest gdpgrow --split-by=nonoil --rank-sum   # WRONG

Actual output:

keane.gdt: n1 = 1269, n2 = 1285, z = (1.20379e+06 - 1.62115e+06) / -nan = -nan

mrw.gdt: n1 = 121, n2 = 121,   z = (10807.5 - 14701.5) / 544.5 = -7.15152

Expected output: n1 and n2 equal the group sizes (948 and 290 for keane.gdt; 19 and 98 for mrw.gdt), and z is finite. For mrw.gdt, z should be 1.3635 with a two-sided p of 0.1727.

Note: meantest wage --split-by=manuf works. It gives t(1236)-type output with a valid p-value (Welch: t(467) = -1.43, p = 0.153). The problem seems limited to the --rank-sum branch of difftest with --split-by.

Best
Artur

[Allin Cottrell]

Thanks, Artur. The back-end for difftest was seriously messed up. That's now fixed in git.