Kruskal-Wallis - SPSS (part 2)

I perform and interpret the output for a Kruskall-Wallis in SPSS, including the estimation of effect size and post-hoc testing (i.e., multiple comparisons). I also demonstrate how to test the assumption of homogeneity of variance based on a non-parametric version of Levene's test. It is a misconception that the Kruskall-Wallis test and the Mann-Whitney U test do not assume homogeneity of variance. They do not, however, assume normally distributed data.

The reference for post-hoc testing without using a Bonferroni correction in the case of analysing 4 or less means is:

Keselman, Games, & Rogan (1979).Protecting the overall rate of Type I errors for pairwise comparisons with an omnibus test statistic. Psychological Bulletin, 86(4), 884-888.

The reference for calculating eta squared in the Kruskall-Wallis case is:

Green, S. B. & Salkind, N. J. (2005). Using SPSS for Windows and Macintosh: Analyzing and understanding data (fourth edition). New Jersey:Pearson

[NhungLe-qmyr]

I really LOVE your videos! They are so helpful! If I am allowed to make a suggestion, I would that it would be great if the links to videos that you mentioned in this video were also included below this video. That way, the audience doesn't have to go through your page to search for them. Plus, the ones they found might not be the right ones. But I still appreciate these lectures a lot!!!

[taylorccoburn]

You have saved my master's thesis! :D Thanks so much for your awesome, peppy, eays-to-understand, helpful videos!

[purityrima]

Thank you very much. In deed you are God sent! Your videos are always very handy! The best I have come across. My homogeneity of variance still turned out to be significant here..precisely p=0.000. However, thank you for the additional insights via the suggestions to same questions below. God Bless you more!

[juanestebandiaz]

It is a nice video; however, there is a mistake that might confuse some people. Before performing the non-parametric levine's test, you should compute mean ranked scores not means from the raw data when you agreggate. In your video you mention that, but you don't do it. You use drinks and not Rdrinks during the aggregation, that is wrong and should be corrected.

[NaijaGirl]

Thank you soo much, you have no idea how helpful this was. God bless you and your family!!

[howstats]

Yes, I've got a video on Pearson chi-square. It's called 2 x 2 Contingency Table Analysis.

[Antgn]

Hi ! Thank you for uploading all this material, it has been a valuable tool for me. I believe the (notlegacybox) allows you to do pairwise Kruskall-Wallis tests in multiple samples in a faster way than selecting for the two groups to be compared each time (as you just did here). Please let me know if you find that to be the case. Thanks again for the great job!

[awfominaya]

This is an excellent video. Thank you for uploading it. I am actually analyzing 4 variables and only one of them violated the nonparametric homogeneity of variance assumption. My question is, should I correct for multiple comparisons when reviewing the significance values for the homogeneity test. If so, my single variable is no longer violating the assumption.

[hayhacker]

my sample is non-normally distributed, so i will use Kruskal-wallis test, but my result of test the Equality of variances is 0.02 which is below p-value 0.05 then reject the null hypothesis, so i cant use Wallis test now !, what test should i do then ? im working on my Master thesis. thanks

[emilymac]

I am incredibly grateful for your videos! Thank you for helping me with my Master's project. :)

[tropicaltina]

What should you do if the non-parametric version of Levene's test (the ANOVA you do here) ends up being significant and fails???  What's the alternative then?

[ellemason]

First, thank you so much for this SPSS guide. I know that this is old, but you clicked the raw variable in calculating the rank mean for groups. You clicked the "raw" drinks variable instead of the "rank" of drinks per group. Was it intentional by design?

[howstats]

You might be stuck in this case. If your sample sizes are equal, perhaps you can rely upon Howell (2007) who wrote that ANOVA is robust to violations of homogeneity of variance, so long as the ratio of the largest to smallest variance is not greater than 4. This is a stretch though, I think. Perhaps you can categorise your variables so that you can do a Pearson chi-square contingency table analysis. You'll lose power, though. There's always a price to pay.

[leonachun]

Hi, thank you for uploading these videos! they have been extremely helpful. Keep up the good work!

[nikhil]

Thanks a ton for great video. oneway_abs found to be significant less than 0.05, now what to do? how to proceed? can i still do kruskal wallis test?

[luftballon]

When you do the non-parametric Levene's Test, I thought you found the mean of the RANKS, rather than the mean of the raw data? At least that's what you did in the Levene's Test Part 3 video, right?

[raulspike]

Be careful, he says calculate rank means but then in the aggregate data window he chooses not ranked data by mistake.

[dylanparker]

is our null hypothesis for the ANOVA that we have homogeneity of variance in the original (un-ranked) data or that we have homogeneity of variance in the ranked data please?

since we're using ranked data, it feels as though it should be the latter (but i suspect it's actually the former)?

[janssensbarbara]

Is it possible you didn't create the mean rank but just the continuous mean by group? Is the way you test this assumption still valid in this case?

[jamesmaye]

What do i do if i need to do a two-way within subjects ANOVA but the data in not normally distributed???