Dwass-Steel-Critchlow-Fligner pairwise comparisons

Discuss statistics related things
zafrin
Posts: 2
Joined: Fri Aug 23, 2019 8:02 am

Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by zafrin »

Good Morning everyone!

I was wondering if it is necessary to adjust alpha, after using Dwass-Steel-Critchlow-Fligner pairwise comparisons in the course of a Kruskal-Wallis-Test. I couldn't find any reference in literature if DSCF is correcting alpha simultaneously or if I should use Bonferroni or something like it instead to control the familywise error rate.

I'd appreciate any clue! Thank you!
User avatar
jonathon
Posts: 2613
Joined: Fri Jan 27, 2017 10:04 am

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by jonathon »

a quick test. we get a larger p-value when comparing only two levels, than when comparing three, so it looks like a correction is in play:
Tooth Growth.omv
(3.84 KiB) Downloaded 717 times
Screen Shot 2019-08-24 at 21.25.01.png
Screen Shot 2019-08-24 at 21.25.01.png (843.44 KiB) Viewed 15350 times
cheers

jonathon
coledavis
Posts: 43
Joined: Sun Jul 02, 2017 1:35 pm
Contact:

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by coledavis »

https://www.tandfonline.com/doi/abs/10. ... 9108830487

I believe that family-wise error rate is built in.
jpmaroco
Posts: 13
Joined: Fri Mar 24, 2017 4:24 pm

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by jpmaroco »

Dear all,
My 2c... I guess that either the Dunn's test or the Conover-Iman (similar to Dunn0s but using a t-distribution instead of the N(0,1) should be given (or having the possibility to select one like in the parametric ANOVA in jamovi).
I am using jamovi to teach 1st year students and actually have an example to support my suggestion:

[img]
image_2021-10-11_142456.png
image_2021-10-11_142456.png (20.13 KiB) Viewed 12204 times
[/img]

So p=0.028 for the omnibus test but none of the DSCF comparisons have p<0.05!...

Is it possible to add Dunn's test to the npanova?

Thanks!
User avatar
jonathon
Posts: 2613
Joined: Fri Jan 27, 2017 10:04 am

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by jonathon »

So p=0.028 for the omnibus test but none of the DSCF comparisons have p<0.05!...
i don't actually think that's a problem. it seems reasonable that you'd have enough data to establish that not all groups are 'equal', but not enough to establish which are not equal. it's possible to achieve this with a conventional ANOVA too.

if you've got the time to add these tests to the np anova, i'd be happy to merge them in.

https://dev.jamovi.org

cheers
jpmaroco
Posts: 13
Joined: Fri Mar 24, 2017 4:24 pm

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by jpmaroco »

Hi Jonathon,
Sure, sample size and power differences do explain the discrepancy (not easy to explain to students and even researchers in the social sciences).
...But, if the omnibus test reveals significant differences one can "afford" more liberal tests like the Dunn's or even LSD on ranks to detect these differences...

Ok, I will look into the dev.jamovi.org to see if I can incorporate Dunn's... there's a nice R package (FSA) that has a dunnTest function).
User avatar
jonathon
Posts: 2613
Joined: Fri Jan 27, 2017 10:04 am

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by jonathon »

cool cool.

let me know if you need any help with it. you'd be welcome to join our slack if you get stuck.

cheers
jpmaroco
Posts: 13
Joined: Fri Mar 24, 2017 4:24 pm

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by jpmaroco »

Cool Jonathon,

It was pretty easy (following the tutorial instructions):
(note: need to add the FSA library to the Imports: jmvcore (>= 0.8.5), R6, FSA in the DESCRIPTION file)
image_2021-10-12_100029.png
image_2021-10-12_100029.png (46.83 KiB) Viewed 12155 times


I do need help with two things:

1. How to incorporate the Dunn's test in the ANOVA > Non-parametric > One-Way ANOVA Kruskal-Wallis (I guess the base structure would be protected, ah?)
2. How to format the nice APA table results.
2.1. The problem is that the number of rows is k(k-1)/2 (where k is the number of groups in the group variable).
2.2. How to count the number of groups in the group variable?
2.3. How to pass that information to the row: attribute in .r.yaml file and to rowNo in the table function for .b.R file?

Thanks!
Last edited by jpmaroco on Thu Oct 14, 2021 10:19 am, edited 1 time in total.
User avatar
MAgojam
Posts: 421
Joined: Thu Jun 08, 2017 2:33 pm
Location: Parma (Italy)

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by MAgojam »

Hi @jpmaroco,
you could start by forking the moretests module, have a look at it (useful for answering your questions) and add the Dunn test.
https://github.com/jonathon-love/moretests

For Conover-Iman and Dunn test I suggest you take a look at Alexis Dinno's two packages in CRAN.
https://cran.r-project.org/web/packages ... index.html
https://cran.r-project.org/web/packages ... index.html

As you can see in this screenshot, I didn't inject any code into the moretests module, as suggested, but just copied the libraries into jamovi to use by Rj (some students needed them on the fly).
dunn_conover-iman_test.png
dunn_conover-iman_test.png (125.28 KiB) Viewed 12128 times
Cheers,
Maurizio
jpmaroco
Posts: 13
Joined: Fri Mar 24, 2017 4:24 pm

Re: Dwass-Steel-Critchlow-Fligner pairwise comparisons

Post by jpmaroco »

Thanks Maurizio,
Getting the Dunn module was pretty easy (still need to format the output table... when I get free time).
But, what I would really like is to have the Dunn (and Connover) included natively in the jamovi non-parametric one way ANOVA. @Jonathan said he could incorporate , if I programmed the module, but I am not clear if it will make it in a new release or it can be added by the user to this one... and how can I contribute? should I send the Dunn to the "moretests" module in github? How do I do that?
best regards from Lisbon!
Post Reply