Re: [Gretl-devel] difftest

Hi, following some off-list discussion, I took a closer look at the 'difftest' command. Here are some issues, I think some might be (small) bugs, others feature requests. Let's start with the easy stuff, feature requests :-) 1) Can we have a --quiet or --silent switch to access only the $test and $pvalue accessors in a script?

2) Can somebody later edit Wikipedia to tell the world that gretl has the Wilcoxon test, too?

3) Could the parametric t-test of differing means in (paired) samples be subsumed under difftest? (Currently there is no direct scripting way of performing it, I think.)

OK, now other issues: Consider first the corner case of all differences zero: <hansl> nulldata 10 series x = seq(1,10) series y = x difftest x y --signed-rank </hansl> which gives a somewhat "Missing values encountered" error.

Next, for a reduced sample n=5 (without ties this time) gretl reports "Sample too small for statistical significance". However, according to this page linked from Wikipedia http://vassarstats.net/textbook/ch12a.html for n=5 there is at least a 5% one-sided critical value (check the last table at the bottom). Furthermore, for small samples n=5..8, gretl appears to calculate the ingredients to what is called "original test" in the Wikipedia entry, namely "two sums of ranks of given sign", denoted by gretl with W+ and W-. However, the actual test statistic (according to https://en.wikipedia.org/wiki/Wilcoxon_signed-rank_test) is "the smaller of" those. But gretl never calculates that minimum-statistic explicitly, and never records it in $test. Here's an example: <hansl> nulldata 10 series x = seq(1,10) series y = -x smpl 1 8 difftest x y --signed-rank eval $test # gives NA wrongly eval $pvalue # understandable that this is NA </hansl>