SPSS - Kruskal Wallis H test with post-hoc
Vložit
- čas přidán 3. 04. 2017
- Instructional video showing how to perform a Kruskal-Wallis H test with SPSS, including a pairwise post-hoc test. Note that for the post-hoc you might get an error in SPSS, which can be resolved following the steps in the video • SPSS - Resolve pairwis... .
Companion website at PeterStatistics.com/CrashCourse
Donation welcome at Patreon: www.patreon.com/bePatron?u=19...
While other videos suggest to execute whitney mann u for every pair as a post hoc and since i am comparing 8 different groups with multiple factors, your solution just saved a few days of my life! thank you!
It was a great video. I had been struggling with this. Your video made it so simple! Thanks a lot.
You're welcome. Glad it helped.
Thank you! I got a headache from trying to figure out how to do it and you helped so much!!
No problem. Glad it could help.
hi mate, you've just helped me and my coursemates in our coursework. Legend!
Thanks, and glad it helped.On a side note, CZcams has cancelled my partner program since I don't have enough subscribers (currently around 754, should be 1000 😞), so it would really help if you subscribe (and all your coursemates) 😉
stikpet already done it!
This is what I am searching for .
Thank you ✨️
You’re welcome 😊 Glad it helped.
Thank you very much for this wonderful video!
Great video, this is very helplful! It helped me during my PhD journey! Many thanks!!
Glad it helped! You mentioned helped, so past sense, so assuming you got your PhD congratulations :-). I need to get myself one of those one day.
@@stikpet thank you - I went to your website it was also super helpful. Ah, no no. I am at the beginning of my PhD :-) if you dont get one I dont know who should :-)
Very helpful video.
thank for your clear explanation.
Thank you so much for the reply. Much Appreciated.
Thanks so much- very well explained!
thanks. Good luck with your research/analysis.
Great tutorial. Thanks!
thank you so much, it's very useful for our research
glad it helped. Happy holidays.
Thanks for this tutorial. It's really helpful
glad it helped.
Thank you so mch. Appreciate your reply.
Thank you very much!
Awesome! Thank you
You're welcome!
Thank you so much!!!
you're welcome, glad it helped.
easy to understand..... thanks
glad it helped.
Hi, thanks for your explanation. I do have one question, what happens when you run the Kruskal wallis and you get a significant difference, but then when you proceed to do the post hoc test in the model viewer screen you don't get any significant differences when comparing your p value to the adjusted sig?
How do you write that up in your report?
thanks!
Hi, this can indeed unfortunately happen. There is an overall effect of the nominal variable on the ordinal, but it cannot be pinpointed to a specific pair. So I would report it as something like that. So something like:
"A Kruskal-Wallis test showed that NOMINAL VARIABLE NAME had a significant effect on ORDINAL VARIABLE NAME, χ2(2, N = 54) = 21.33, p < .001, Ε2 = .40. Unfortunately a post-hoc test using Dunn's test with Bonferroni correction did not reveal any specific pair being significant different , p > .05."
If any pair is close you might want to add that pair mentioning it was close (also since the .05 threshold is somewhat arbitrary).
Also of interest perhaps: stats.stackexchange.com/questions/16665/how-can-i-get-a-significant-overall-anova-but-no-significant-pairwise-difference it's about a regular one-way ANOVA but applies similar to the Kruskal-Wallis.
Hope this helps.
Thank you very much!! I was starting to believe all my analysis was wrong because of this! I will consider your explanation!!
Can you explain the this question again, please?
l found that it did not significance after posthoc tests.
Beste Peter, dank voor uw video! Weet u toevallig of, en zo ja hoe ik een confidence interval kan berekenen om het mediane verschil heen? Alvast hartelijk dank!
Beste Puck. Ik ken helaas geen methode in SPSS om voor alle pairwise comparisons tegelijk een betrouwbaarheidsinterval te berekenen, het enige dat ik zou kunnen bedenken is om eerst met data - select cases ervoor te zorgen dat alleen twee groepen geselecteerd zijn, en dan via dezelfde route als in de video de optie 'Hodges-Lehmann estimate (2 samples)' aan te vinken (staat vrij onderaan in het scherm om 0:58). In de output dan dubbelklikken op de output en in de Model viewer onderaan bij 'view' de optie 'Confidence Interval Summary View' kiezen.
Can you kindly guide me on this. Is there a non parametric test equivalent to ANCOVA. My QOL scores are continous. And physical activity categories are three. Low, moderate and high. Thanks a lot. Your reply helped.
Unfortunately I'm not as familiar as I like to be with ANCOVA. You mention a continuous and ordinal variable, so perhaps a Spearman's rho might also be of interest.
On TalkStat I found this thread on nonparametric ANCOVA: www.talkstats.com/showthread.php/4502-Non-parametric-equivilent-to-ANCOVA they suggest to use a Quade test, which although not directly can be accomplished with SPSS according to the IBM answer given here: www-01.ibm.com/support/docview.wss?uid=swg21477497
If you like you could perhaps also use R. See this thread for a start: stats.stackexchange.com/questions/41270/nonparametric-equivalent-of-ancova-for-continuous-dependent-variables
Thank you!
You're welcome. Glad it helped.
excelent!! thank you very much
thanks, glad it helped.
Thanks for this explanation, If only I had found it earlier! I have a question still...I have been able to identify that a difference exists between three groups however I still don't know how to identify at what point does the difference exist. For example - in your example you have identified that a difference exists in general between three locations but I want to still know at what point (s) on the likert scale (Fully disagree, disagree, NA, agree, Fully agree) does the difference occur. Is it at the Fully disagree level or somewhere else? I may not be explaining myself very clearly! Sorrry
No problem. Glad it helped.
Not sure if I understand exactly what you would like to know, but to 'see' better how the groups differ a visualisation is often helpful. For multiple ordinal variables, I'd suggest to use stacked bar charts. I've a video on that here: czcams.com/video/djBtLpW0Bvk/video.html That should give a good view on what the differences might be. In general I suggest to first do a visualisation, that then gives an indication, then test if what you notice from the visualisation is also 'significant' by performing a test.
Hope this helps.
Hi, I seriously thank you for the video. It helps me a lot. :)
I just have one question, when I did run the test, I got significant p-value.
However, when I click pairwise comparison, none of the pair shows significant different. What should I do in this case?
Thank you in advance
Hi, in this case there seem to be some influence from the nominal variable on the ordinal overall, but differences between pairs are too small to specify any two particular groups. This could be because of the Bonferroni adjustment. Not much else I can think of that you could do. Hope it helps.
yeah, surely it helps. thank you so much. and thank you for your time too :)
Very helpful video!
I have a question. How do we present the results from the test with the results from post-hoc in APA? I can't find any reference to that...
APA actually doesn't say much about how to report results, only a little bit in general. With post-hoc tests I try to summarize it myself and would put the full output in an appendix. An example on how I'd do it can be found on my site at peterstatistics.com/CrashCourse/3-TwoVarUnpair/NomOrd/NomOrd3b.html. Hope it helps
Hello! How did you put an ordinal variable into test field? Because I cannot put them, only the scale ones.
I remember this issue with I think it was SPSS version 19. Not sure if they actually made a patch/fix for it, but the simple solution I used back then was to temporary adjust the measurement level of the variables to scale, then run the test, and adjust it back to scale after I was done.
Thank you for the helpful video... So the post hoc in your video is the Dunn's test right? I did Kruskal Wallis test and Mann-Whitney U test for multiple comparison, but many people said that Dunn's test is better than MWU and i just confused how to do Dunn's test in SPSS and this video made it clear...
Glad it helped. SPSS calls this the Dunn test with Bonferroni adjustment: www-01.ibm.com/support/docview.wss?uid=swg21479073.
MHU and KW just tell you if the items differ significantly from each other or not, they do not tell you which does. Hence the usefullness of the Post-Hoc test.
Thanks a lot for the video! Helps a lot! I just have one doubt. Is it correct to say the "Sig." value is according to Dunn's test and "Adj. Sig." is Dunn's test with bonferroni correction (or Dunn-Bonferroni approach)?
that is indeed also my understanding. Sig. = Dunn test, and then simply multiply this with the number of pairwise tests to get the adj. sig. which is the Bonferroni correction.
hey mate, great video! I've a question, is it possible to do a kruskal wallis with two groups? I want to know the interaction between groups and i cannot find anywhere.
The Kruskal-Wallis test is used with two or more groups, so yes. Although, when only having two groups it actually becomes a Mann-Whitney U test.
Hello! Great video. I have one question about reporting results. When I report it, do I report sig. or adj. sig.? For example, in your pairwise comparison Rotterdam-Haarlem, do I report p = 1 or p = .729? Thank you for your answer!
What I would do is to mention that an adjustment was used, and then still report as p = 1.000. For example: peterstatistics.com/CrashCourse/3-TwoVarUnpair/NomOrd/NomOrd3b.html this seems to be in line with for example www.bryanburnham.net/wp-content/uploads/2014/01/Reporting-Statistics-in-APA-Format.pdf
Hi,
Is the post-hoc test a series of Mann-Whitney U tests or is it Dunn's? If it is Dunn's, when writing this up in a report, would the test statistic be written as: z = ... , p = ... ? Also should this test be described as a multiple comparisons test rather than pairwise?
Thanks.
It should be the Dunn's test. IBM claims it is at: www.ibm.com/support/pages/can-spss-perform-dunns-nonparametric-comparison-post-hoc-testing-after-kruskal-wallis-test .
The difference is nicely mentioned "The ranks of the data on which the tests are based change if they are reranked in a
pairwise fashion. Dunn’s (1964) insight was to retain the rank sums from the omnibus test and to approximate a z-test statistic to the exact rank-sum statistic. Dunn’s test is the appropriate procedure following a Kruskal-Wallis test" (Dino, 2015, p. 293, journals.sagepub.com/doi/pdf/10.1177/1536867X1501500117)
There is no official rule on how something must be reported (up to the journal I suppose) but it seems that z and p-value are indeed the norm: www.researchgate.net/post/How_do_I_report_a_Kruskal_Willis_test_with_Dunns_multiple_comparison_post_hoc_Is_it_in_the_same_way_ANOVA_where_we_use_Fa_b_n_and_the_p_value
It is a paired-multiple-comparison, since we are pairing two categories each time, and doing this multiple times.
Hope this helps.
Can the Bonferroni test be used after the Karuskal-Wallis test? Does the Bonferroni test require that the data depend on a normal distribution or not?
The Bonferroni correction is done with the post-hoc test, not the Kruskal-Wallis. The post-hoc test shown is what IBM calls a Dunn test (www.ibm.com/support/pages/can-spss-perform-dunns-nonparametric-comparison-post-hoc-testing-after-kruskal-wallis-test). It is a good question to ask about the requirements to use this post-hoc and unfortunately I couldn't find much information on it. On reddit I did find someone mentioning it is no problem: www.reddit.com/r/AskStatistics/comments/niyw64/help_with_kruskal_wallis_test_and_posthoc_dunns/
Thanks. it was useful for me. i passed my exam.
great, glad to be of help and congratulations.
Hurray for online SPSS teachers like you :D
Hi, I`m sorry, I`m completly new in this world, and I can't understand what kind of post hoc test did you do?
Actually, I`m try understand because I have to do this Kruskal Wallis H test for my thesis and I don´t know what´s "the statistical origin" or the test that justifice the pairwise comparison, Something like Mann-Whitney, actually I don´t know.... Please help!!!!
SPSS uses a so-called Dunn's test (or Dunn-Bonferroni) test for the post hoc analyses. You might also want to have a look here on my site: peterstatistics.com/CrashCourse/3-TwoVarUnpair/NomOrd/NomOrd3b.html
Hope this helps.
Hi, thanks for your explanation. The post hoc test is Bonferroni or Dunn test?
Glad you enjoyed the video.
It is a Dunn-Bonferroni method. IBM (who own SPSS) mention this here: www-01.ibm.com/support/docview.wss?uid=swg21479073
On a side note, CZcams has cancelled my partner program since I don't have enough subscribers, so it would really help if you subscribe (and all your friends) 😉
Hello,
first of all: great video. exactly what I looked for. ;-b
but I've got a question: is it somehow possible to change the post hoc method to sidak?
Glad you liked the video. Unfortunately to my knowledge it is not possible to change to any other post-hoc method (except for the one-way ANOVA) in SPSS. The Sidak adjustment can actually easily be calculated manually 1-(1-unadjusted p-value)^k , where k is the number of comparisons. Since the unadjusted p-values are given in the SPSS output it is not too complex to determine the Sidak adjusted versions.
Thank you! :)
Hi, I couldn't open the model visor so I can't make the Bonferroni test. Do you know if there is another way to do theBonferroni test in SPSS after making the K-W test? I mean, there is another option to access to the test?
Odd that you cannot open the model viewer. You could try to click on 'Edit' in the menu-bar then 'Options', then select the 'Output' tab, and at ' Output Display' select 'Pivot tables and charts'. Then perform the analysis again, but now the results should show up as pivot tables, and not in the model viewer anymore.
thank you
glad it helped.
Thank you so much for this. Can you help me with the name of that post hoc test?
A long time ago someone asked the same question. Had to dig up my response from back then: "Indeed it is annoying when they don't name the test. However after some digging according to SPSS themselves, they use a Dunn test with Bonferroni adjustment (www-01.ibm.com/support/docview.wss?uid=swg21479073)."
Thank you. But I wonder, what is the name of the "pairwise comparison" that is default in SPSS as a post-hoc test? Is it Mann-Whitney or Dunn's test? Or something else? I have not found this information anywhere and I think it is strange that SPSS does not describe what test they use.
Indeed it is annoying when they don't name the test. However after some digging according to SPSS themselves, they use a Dunn test with Bonferroni adjustment (www-01.ibm.com/support/docview.wss?uid=swg21479073).
@@stikpet Thank you for a prompt and detailed reply! Much appreciated! :)
Helo. It was a great video! This is Dunn Bonferroni test right? How i make nominal (supercript) with this test? Thank you
Glad it helped. SPSS calls this the Dunn test with Bonferroni adjustment: www-01.ibm.com/support/docview.wss?uid=swg21479073. Not sure what you want to make nominal, the data should already be nominal.
Any idea on what to do if pairwise consumptions doesn't appear on the post hoc analysis drop down?
you're not the first with this question, so I'll copy my answer I posted a while ago:
Two reasons I can think of.
1. your grouping variable only has two categories. In this case there is no point for a post-hoc comparison, since there are only two groups to compare. So the Kruskal-Wallis test result already indicates if there is or isn't a difference between those two groups.
2. If the Kruskal-Wallis test is not significant (i.e. the significance is above 0.05) the pairwise comparisons will not show, since none of them will be significant.
Hope any of the two options above resolves your issue. Let me know otherwise.
Hallo! Duidelijke video! In mijn dataset heb zelf een 20tal variabelen (Likert van 1-5) die samen een schaal vormen. Ik heb de variabelen gesommeerd om een schaal te vormen en het meetniveau is dus van ordinaal in scale veranderd. Kan ik, als de afhankelijke var op scale niveau is, de Kruskal wallis uitvoeren ? Alvast hartelijk dank voor uw reactie!
Als je een scale en een nominale variabele hebt, zou je kunnen kijken of een one-way ANOVA (enkelvoudige variantie analyse) geschikt is. Die heeft vaak dan de voorkeur. Mocht je echter niet voldoen aan de voorwaarden voor een one-way ANOVA, dan wordt vaak juist weer de Kruskal-Wallis toets aangeraden. Het mag dus wel, maar wellicht kan het dus beter met een one-way ANOVA.
De variabele is niet normaal verdeeld, tevens werk ik vrij weinig cases, dus zal ik voor de Kruskal gaan. Hartelijk dank voor uw reactie!
Thank you for your video. I have a doubt. The table has 2 "p": "sig" and "adj.sig." What's the difference between them? Which one should I choose? Thanks again,
Usually the adj. sig. is the one to go for. It is adjusted to control for all the different tests. With each pairwise test you run a risk of 5% to make an incorrect decision, and with multiple tests the chances of at least one of them making the wrong decision increases very fast. So various people have come up with various methods to control for this (often called control for family wise error). I explain this a bit also on my site at peterstatistics.com/CrashCourse/3-TwoVarUnpair/NomOrd/NomOrd3b.html
Hope this clarifies things.
Hey, thanks for your help with this video
I just had one problem if i click on view: pairwise comparisons i get pairwise comarisons view encounterd some problem(s) and will not be show.
Is there a way that i can still see them?
At the end of the video I mention another video (also in the description) in case you get an error for the pairwise comparisons view not showing: czcams.com/video/vbYuV5TlMNU/video.html. Should probably mention that earlier in the video , but better late then never. Hopefully that will resolve also the problem for you.
Ok, thanks!
How can I change the decimal places in the model viewer? I need values i.e. 1.3E-7, not just 0.000
It is indeed annoying that the results in the model viewer cannot be selected or adjusted in lay-out. However I just checked for at least SPSS 23 and SPSS 25 (but perhaps also other versions) you can click on 'Edit' in the menu-bar then 'Options', then select the 'Output' tab, and at ' Output Display' select 'Pivot tables and charts'. Then perform the analysis again, but now the results should show up as pivot tables, and not in the model viewer anymore. You can then double click on the table, and then again double click on the sig. value to see the scientific notation.
Hope this helps.
I have a question. I am comparing the difference in Quality of life scores based on physical activity categories. If I have a high number on the mean rank in the post hoc test, I have a high mean rank on the high category physical activity . can I interpret participants in high category of Physical activity have higher Quality of life scores.
Since you are thinking of using a Kruskal-Wallis test I assume your quality of life score is ordinal, and since you mention in the interpretation 'high category of physical activity' it might be that also physical activity is ordinal. In those cases Goodman-Kruskal Gamma might be of interest. On my site you find some more info on this: peterstatistics.com/CrashCourse/3-TwoVarUnpair/OrdOrd/OrdOrd3.html
If you only have two categories with the physical activity, then indeed I'd go with the Kruskal Wallis test (which actually is a Mann-Whitney U test if you only have two categories). If you coded 1 = low quality of life and 5 = high quality of life, then indeed a higher mean rank for one group would also suggest a higher quality of life for that group. Note that this flips if you had coded 1 = high quality of life to 5 = low quality of life. Then a higher mean rank would indicate a lower quality of life.
Hope this helps.
Help me please!, trying to write the results, i don't know what test is the post hoc of kruskal wallis, is mann whitney? or what? Thank you !!
On my website I also show how results can be reported. The full result could look like peterstatistics.com/CrashCourse/3-TwoVarUnpair/NomOrd/NomOrd4.html The post-hoc test is according to SPSS a Dunn test.
Sir, help me, please. I would like to do post hoc with non-parametric test as i do with anova. How can i do?
Not sure what you need. The non-parametric alternative for a one-way anova is the Kruskal-Wallis H test, and the corresponding post-hoc is shown in this video.
Sir, may I know that is the value of the test statistic equal to the mean difference? As you can see Rotterdam (minus) Haarlem is (-1.747) but the table in 3.00s shown 1.747. Sorry, I am confused about that. Sincerely appreciate it if you could give me some feedback on this matter.
indeed the test statistic is the mean difference in ranks. It took me a while to figure out if it is doing Category 1 - Category 2 or Category 2 - Category 1. It appears that it does this based on the original coding of the variables. In my example I had as values for the location variable 1 = Diemen, 2 = Haarlem and 3 = Rotterdam. So even though it shows Rotterdam-Haarlem, it will actually do then Haarlem minus Rotterdam, simply because Haarlem is coded as 2 and Rotterdam as 3. I don't know this for sure, but checked with another variable with more categories and that seems to be it.
If you are interested in all calculations, the video on how to do this with Excel shows more or less the step by step calculations (czcams.com/video/ca64XSXF1CU/video.html).
Hope this helps.
@@stikpet Sir, thank you so much for the time and help! Much appreciated it =)
Hello, Please help!
For post-hoc tests....Under view I dont have the option to select Pairwise Comparisons. I only have "Categorical Field Information" and "Continuous Field Information"
Thanks in advance! And great video!
Dear Erica,
Two reasons I can think of.
1. your grouping variable only has two categories. In this case there is no point for a post-hoc comparison, since there are only two groups to compare. So the Kruskal-Wallis test result already indicates if there is or isn't a difference between those two groups.
2. If the Kruskal-Wallis test is not significant (i.e. the significance is above 0.05) the pairwise comparisons will not show, since none of them will be significant.
Hope any of the two options above resolves your issue. Let me know otherwise.
Thank you so much!!
Thank u
You'r welcome. Glad the video helped.
I have a significance of p = 2.0 but still, the summary says "Reject the null hypothesis", can anyone tell why is that so? When I open the summary, there is a significance between 2 of 3 groups..
A significance of 2.0 should not be possible. A significance is a probability, so should always be between 0 and 1. Most likely if you double click and look in the test result section, you should see the correct p-value. I'm guessing it will be 0.002 or something.
Hi, great video.
In the Node view I have no labels, just the sample average rank so I'm unable to distinguish which samples differ significantly with ease.
Any suggestions?
Glad the video was helpful. Not sure what might cause the labels to disappear, but one possible solution might be to use:
Analyze - Nonparametric Tests - Legacy Dialogs - K independent samples
Move your ordinal variable to the 'Test variable list' and the nominal variable (the one that determines the groups) to the 'Grouping variable', then click on 'Define Range...'
Enter the value you assigned the first group at 'Minimum' and the last one at 'Maximum'. Click 'Continue',
At the 'Test Type' section Make sure the 'Kruskal-Wallis H' option is checked and then click on 'OK'.
In the output you should now see the a table called 'Ranks' (as the first table) which shows the categories and in the last column their 'Mean Rank'. These should match the ones shown in your diagram.
Hope this helps.
@@stikpet Excellent thanks, It appears the variable that was meant to be ordinal was set to nominal.
Thanks for getting back, and happy it worked. Actually that is strange, since if you try to move a nominal variable to the 'test field' section (in the version used in the video) it should even give an error that 'nominal fields cannot be placed in the list'.
Hi,
The option model viewer isn't availble anymore in the newest version of spss (spss 27). Do you know how I can get the pairwise comparisons table even when the p-value isn't significant?
Indeed, all now simply shows in the output. Still no post hoc will be shown if the overall test is not significant. Not sure if it is possible to change this :-(
@@stikpet Do you now if is correct to use a mann withney u test instead? I just want to see if there is a significant difference between my 3 groups. Since one group is very small, i use non parametric tests.
instead of the Kruskal-Wallis or instead of the post-hoc test done by SPSS (Dunn test)? If you only have two groups, the Kruskal-Wallis test gives the same result as a Mann-Whitney U test, and there won't be a need for a post-hoc test. If it is for a post-hoc test, according to www.real-statistics.com/one-way-analysis-of-variance-anova/kruskal-wallis-test/pairwise-mann-whitney-tests/ this can be done, but it will give some different results. There is also some more technical discussion on stackexchage: stats.stackexchange.com/questions/461045/can-i-use-the-mann-whitney-u-test-sequentially-pairwise-when-i-have-three-grou . Hope this helps.
Hello! Do you have any idea why when I use the "legacy dialog" K-W test that I get a different result?
That is new to me. Will have to look into the SPSS algorithms to see if they use a different formula. I'm on holiday now so will take some time. Can you perhaps indicate what is different? (the sig. the test statistic, or something else)
Can't seem to find anything on Legacy dialogs in the SPSS algorithms. Without some more info, it's going to be like searching for a needle in a haystack. Could you explain what exactly is different?
Hi! I have a question: I am running into the problem of not being able to see the data in my pairwise comparison. There are so many tests, I cant see the numbers. I've tried coping it to the original SPSS output, and transferring to Word but neither of these work. Suggestions?
Hopefully I understand the situation. The Model Viewer is a bit user-unfriendly since you cannot click and adjust parts. One solution is to let the results not show in the model viewer, but in the regular output. This can be done by the following: click on 'Edit' in the menu-bar then 'Options', then select the 'Output' tab, and at ' Output Display' select 'Pivot tables and charts'. Then perform the analysis again, but now the results should show up as pivot tables, and not in the model viewer anymore. Hope this helped.
@@stikpet Thank you so much for that.
Goedemiddag, heel erg bedankt voor de duidelijke video! Ik had na deze video nog 2 vragen:Ik heb twee variabelen (1 categorisch, bestaande uit 4 groepen, de ander continue). Mijn samples sizes van de groepen verschillen nogal (34, 30, 6 en 14). Mijn data is daarnaast niet normaal verdeeld en er is ook geen homogeneity of variance. Ik heb hierbij de volgende vragen: 1. mag ik wel een Kruskal Wallis test doen? Indien het antwoord nee is, welke test dan wel? En 2: is het goed om daarna met de Bonferroni Dunn test te corrigeren voor "multipele comparisons" omdat ik de verschillende groepen wil vergelijken?? Alvast bedankt voor de moeite!
Beste M. de Boer,
Het voordeel van een niet-parametrische toets (zoals de Kruskal-Wallis), is dat er wat minder vereisten zijn. Zo hoeft er niet gekeken te worden naar de homogeneity of variances. Wel is het zo dat als de verdelingen onderling dezelfde vorm hebben de K-W toets in principe de mediaan van elke groep vergelijkt, maar anders nog steeds naar de gemiddelde ranking kijkt. Wat meer info over de aannames is te vinden op: statistics.laerd.com/spss-tutorials/kruskal-wallis-h-test-using-spss-statistics.php
De K-W houdt net als een gewone one-way ANOVA rekening met de verschillen in aantallen tussen de groepen, en er is dus ook geen vereiste dat deze allemaal hetzelfde zouden moeten zijn.
SPSS geeft reeds de aangepaste significanties weer onder adj. sig. in de post-hoc analyse, dus hoef je niet zelf nog een keer te doen.
Wellicht ook interessant om even te bekijken: peterstatistics.com/CrashCourse/3-TwoVarUnpair/NomOrd/NomOrd0.html
Hopelijk zo genoeg input.
Dank voor het uitgebreide antwoord en de link! Het blijkt dat mijn data niet onderling dezelfde vorm hebben, is er nog een andere test die ik dan kan doen of moet ik die aanname van de K-W gewoon negeren?
Beste M. de Boer,
Wellicht had je het over het hoofd gezien, maar op de site van Laerd, staat "However, if your distributions have a different shape, you can only use the Kruskal-Wallis H test to compare mean ranks". Je kunt dus nog steeds de K-W toets gebruiken, alleen zal deze niet meer toetsen of de mediaan van elke groep hetzelfde zal zijn, maar of de gemiddelde ranking gelijk zullen zijn.
Succes.
Great video, but a question. I thought KW H test was only used to compare 2 or more categorical independent to a continous non parametric data set?
Thanks. The KW is often seen as the non-paremetric version of the one-way ANOVA, and used if the the normality assumption is violated. However, it is also often mentioned to be suitable for ordinal data. See for example www.researchgate.net/post/For_what_kind_of_data_is_the_Kruskal-Wallis_test_used or statistics.laerd.com/spss-tutorials/kruskal-wallis-h-test-using-spss-statistics.php#:~:text=The%20Kruskal%2DWallis%20H%20test,continuous%20or%20ordinal%20dependent%20variable.
@@stikpet I see. That makes sense. Thank you for your advice. If it ids please possible are you able to contact me on h1995bs@gmail.com? It's regarding this data set I have been given but having issues with it on SPSS. Will be very happy to donate to a patreon account if you have one or advertise your CZcams channel on my social media pages. The videos are very useful!
@@lucyjones3734 email send :-) And just in case anyone else reads this. I do have a patron account at www.patreon.com/bePatron?u=19398076 . Not required though, but hey always welcome a small (or big) support :-)
Dear Sir, I really appreciate your efforts. I wonder if we can use this post-hoc method for two ordinal variables? I'm a bit confused..For instance, I performed Kruskal Wallis to determine whether there are significant differences in the rankings of the likert statements (ordinal) among the 3 groups of firms with different size (ordinal), and I found statistically significant differences in a few statements. Is it appropriate to perform Dunn's test?
Thank you in advance..
If you have two ordinal variables, you might look into a peterstatistics.com/CrashCourse/4-TwoVarPair/OrdOrd/OrdOrdPair3a.html
Or if you have more than two: peterstatistics.com/CrashCourse/5-ThreeVarPair/ordinal/MultipleOrdinal3a.html
Videos on how to do each of those tests are also available on the same site.
Hope this is what you were looking for.
@@stikpet Thank you so much for your reply. I watched the videos which expanded my knowledge, i greatly appreciate. However my variables are unpaired. I need to investigate if the level of agreement (regarding the likert type statements) differed between different group of firms. These firms were classified under small, medium and large size companies.
So the firm size are the groups, which are ordinal but with only three seen as the nominal variable for the Kruskal-Wallis test, with the Likert items as the ordinal variable. Don't see a reason why you wouldn't then be allowed to do a Dunn's test (but I also don't know everything :-) ).
You mention you are interested in the level of agreement. There are actually measures of agreement in statistics. I'm not really familiar with them but a quick scan of this article www.ncbi.nlm.nih.gov/pmc/articles/PMC5654219/ suggests that for your situation a Fleiss' kappa might be of interest. It might be something to look into.
Hope this helps
@@stikpet I reviewed the article and the methods and i learn so much thanks to you. My data does not meet the assumptions of the Fleiss' kappa. Actually I’ve reviewed so many articles so far, especially those in journals indexed in SCI. Kruskal-Wallis has been widely performed in similar studies. After all of your comments and suggestions, now i think Kruskal-Wallis and Dunn’s test are the most appropriate :)
Thank you so much four your interest and time.
Hi can I have some instructions or guidance from you in this test? I could not insert the variables (either nominal or ordinal) into the test field.
A while ago someone had the same issue. I'll plagiarise my response there: I remember this issue with I think it was SPSS version 19. Not sure if they actually made a patch/fix for it, but the simple solution I used back then was to temporarily adjust the measurement level of the variables to scale, then run the test, and adjust it back to scale after I was done.
Hope this solves it for you as well.
@@stikpet if I were going to adjust the measurement level to scale, all my values remains?
@@sandragoh7292 you should be able to adjust it, perform the test and without any problem change it back again. Just to be safe make a copy of your datafile (.sav).
Why pairwise conparison option does not appear on that button?
in the newer versions there is no 'model viewer' anymore and all output is simply shown in the output window. Pairwise comparisons will most likely not show if the overall test is not significant. If you are using version 22 or 23 and the button is not there, and the overall test (Kruskal Wallis) is significant, you might need to follow the steps from this video: czcams.com/video/vbYuV5TlMNU/video.html
Hope this helps.
HI WHAT VERSION OF THE SORTWARE ARE YOU USING?
No need for the all caps :-) I guess I made the video using SPSS 22 or 23. It should work I think with any version from at least SPSS 19. The only version I have never fully used is the subscription version, that will look quite different. If you are having trouble, which version do you use?
What is the name of post hoc test, which is
used here
SPSS uses a Dunn test with Bonferroni adjustment (www-01.ibm.com/support/docview.wss?uid=swg21479073).
On a side note, CZcams has cancelled my partner program since I don't have enough subscribers, so it would really help if you subscribe 😉
Thank you so much for the clear explanation! Just had a question. My data is composed of 5 species of butterflies found in one lawn. I've labelled my butterflies species from 1 to 5. In the pairwise comparisons, I see that one species has the highest rank. But I'm not on an ordinal scale as you are. Can I still accept that the butterfly with the highest rank as the most common on the lawn? Many thanks.
Not sure if I fully understand your data structure. If you 'simply' measured how many butterflies there were of each species, then you could use a Pearson chi-square test of goodness-of-fit. It can indicate if one or more species occur more often than others. If so you can then follow up with a post-hoc test to figure out which one. On my site you can find more on this at peterstatistics.com/CrashCourse/2-SingleVar/Nominal/Nominal-2a-Test.html it also has links to my CZcams videos on that test, and for the post-hoc just click on the left menu on, well, 'post hoc'. Hope this helps.
@@stikpet thanks for your detailed reply. Actually I have 20 species with uneven counts for each. So for example I have 5 data entries for butterfly A, each entry reporting the number of butterfly A recorded (from 2 to 50), then 9 data entries for butterfly B and so on .. I want to know which species are the most significant in terms of their quantities found on the lawn. Therefore I conducted a KW test with pairwise comparison. I obtained a p-value
@@guillaumebillard1943 But if it is only the counts then in my opinion I don't think the KW would make sense. The KW is used with an ordinal scale, and I don't see any ordinal scale in your data, only counts. As long as you have sufficient data points in each category, it shouldn't be a problem to use the Pearson Chi-square, otherwise an exact multinomial test might also be possible. In the end it is of course your decision, you know your data better than I do :-).
Good luck with the analysis.
@@stikpet Thank you! Just one thing: KW can also be used for continuous data as well i.e. quantitative. Here I have quantities of butterflies by species (n species=20; 2800 butterflies in total). When I look at your video of chi-square test goodness of fit, the data you use if for example martial status (married, divorced etc). I can't really breakdown my data like that. And sometimes I have not many data points, example of butterfly species "E" has 3 data points, with a total of 8 butterflies :) Sorry if I'm not being clear :)
Indeed the KW can also be used with continuous data, it is sometimes referred to as the non-paremetric alternative of a one-way ANOVA, which is usually preferred with continuous data.
If I understand correct. You have 'measured' butterflies. So actually you have only one variable: 'species'. Then 2800 butterflies, so 2800 rows in SPSS. Now you want to know if species X occurs significantly more or less often than lets say species Y. If so then the Pearson chi-square test of goodness-of-fit is I think the one to use. If some of the categories have extreme low counts, you could perhaps exclude them and report so that you did, or look into the exact multinomial test. Not sure if this can be done with SPSS, with Python I got it to work.
I am having difficulty on what letter/symbol I should use in my APA Format report. Should I use H or F? How do I do it?
Actually APA does not specify in detail for each test how to report it, but only gives some generic guidelines and a few examples. The Kruskal Wallis value is using a chi-square distribution, so that would be I guess the way to go. I give an example on my website at: peterstatistics.com/CrashCourse/3-TwoVarUnpair/NomOrd/NomOrd3a.html
Hope this helps.
Thank you! Means a lot!
please help me sir
I want to use Kruskal-Wallis H test but the variable that represents the group is multi-choice (I've unloaded in spss program with yes or no code ) and now I do not know how to perform Kruskal-Wallis H test
I am very frustrated
Hard to tell but if I'd had to guess you asked something like 'which group(s) do you belong to?' A, B, C and/or D, and then another question like a Likert scale. If so you can check each group one by one. Are people from group A giving different answers than the other groups? People from group B, etc.
Otherwise I'd need a bit more info to understand what it is you want to establish.
Hi sir
I am sorry my English languge is not good
My problem is that I have a set of items designed on the Likert scale and distributed to engineers in a different job position (designers, project managers, site engineers, consultants).Because one person may be have two positions or more , I have designed this question as a multiple choise. They can choose more than one (designer and consultant at the same time, for example). Then I unloaded answers in SPSS as a multiple-choice question (coded by yes and no) to calculate the ratios for each group. But when I wanted to apply Kruskal-Wallis H test t I did not know how
Your English is better than my Arab(?) :-) . Splitting the multiple answers up into separate yes/no variables is good, so you did not do anything wrong there. I'd suggest you simply run the K-W test four times (one for each job position). You don't need a post-hoc test then anymore since there are only two groups each time (those who have that position and those who don't). You'll actually then be doing a Mann-Whitney U test each time.
Hope this is clear and solves your problem.
Thank you very much sir
You are man!
Yes, I'm a man :-) Guess you meant 'you are the man', for which thanks 😀
Thank you very much!!!
You're welcome! Glad it helped and good luck with your study/research.
Thank you!
You're welcome! Glad it helped.
thank you
You're welcome
Thank you!
glad it helped.
On a side note, CZcams has cancelled my partner program since I don't have enough subscribers (currently around 637, should be 1000 😞), so it would really help if you subscribe (and all your friends) 😉