Kruskal Wallis stata vs wizard

[Tristan Lane]

Hi Evan and everyone.

Same data.
Different outcomes for wizard (significant) and stata (non significant).
22 per group.
Prism gives same as wizard.

Allowing for data being right is the reason exact vs approximate values?

[Evan Miller]

Hi Tristan,

Wizard (and I guess Prism) are using the chi-squared approximation for Kruskal-Wallis with 3+ groups. For 2 groups (Mann-Whitney) Wizard will compute the exact version when N<40 per group.

Exact vs approximate has more to do with computational complexity than anything else. Wizard's implementation follows this paper for the Mann-Whitney U:

https://www.jstor.org/stable/2347656?seq=1

I can look into providing an exact version for 3+ groups - which I guess is what Stata is doing.

How many groups are in your test?

Evan

--
You received this message because you are subscribed to the Google Groups "Wizard User Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to wizard-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/wizard-users/34da1daf-7a19-4670-96ef-3d1d78f75d37%40googlegroups.com.

[Tristan Lane]

Ah cool.

Will read.

Also annoyingly it’s because of a weird sorting in the data...

Weirdness over, sorry!