Re: [Gretl-devel] difftest
Am 29.11.2017 um 14:24 schrieb Allin Cottrell:
On Wed, 29 Nov 2017, Sven Schreiber wrote:
>>> 3) Could the parametric t-test of differing means in
(paired)
>>> samples be subsumed under difftest?
I'll just address this point for now. Yes, maybe it is worth
exposing
the (unpaired) two-means t-test that's in calculator.c as an option
under "difftest".
Argh, I hadn't read your answer carefully enough, concerning "paired"
versus "unpaired". Sorry!
Let me try to summarize gretl's status quo about testing
means/medians/central tendencies, maybe this list is useful for reference:
--------
1) Unpaired (AKA: unmatched, independent) samples:
a) 'meantest x y --unequal-vars' (parametric t-test)
b) 'difftest x y --rank-sum' (nonparametric Wilcoxon rank sum, AKA
Mann-Whitney)
c) mwu.gfn package (by Yi-Nung, also nonparametric Mann-Whitney)
2) Paired (matched, dependent) samples:
a) 'ols' of the manually created difference x-y on a constant
(parametric t-test)
b) 'difftest x y --sign' (nonparametric)
c) 'difftest x y --signed-rank' (nonparametric Wilcoxon signed rank)
3) One sample:
a) 'ols x const' (parametric t-test)
b) 'difftest x H0med --sign' with the manually created constant series
H0med (nonparametric)
c) 'difftest x H0med --signed-rank' with the manually created constant
series H0med (nonparametric Wilcoxon signed rank)
4) K>2 unmatched samples
a) 'anova', or 'ols ... --anova' (parametric F-test)
b) KruskalWallis.gfn package (by Jack, nonparametric)
5) K>2 matched samples
a) ?? : 'ols' with const as the artificial dependent variable and N-1
pairwise differences between the N variables as artificial regressors
(parametric F-test) ??
b) <Friedman nonparametric test not implemented>
--------
The cases 2a, 3b, 3c are suboptimal IMHO because they require some
manual intervention. The 'difftest' command might be extended to work
directly on a single series for cases 3b, 3c.
The extra package mwu.gfn is redundant in principle I guess, but the
devil is in the details.
The extra command 'meantest' in case 1a is the odd man out in some
sense. It could also be named 'difftest --normality' or something.
I think the nonparametric chapter of the user guide (ch. 37 currently)
should be extended to cover the existing tests. (This is not a demand,
just a reminder. Maybe something for the bug tracker.)
This feels like the time when I first used gretl for introductory
statistics teaching! Corrections very welcome.
thanks,
sven