Mann-Whitney - Assumptions - SPSS

[@danushiwijekoon4883]

Thank you so much this is very helpful. I have a question..... What can we do when we come across a data set that does not meet the homogeneity of variance assumption' (is there another test we can run / how can we make sense of the data in this situation)

[@giorgiogiorgetti3939]

Ok, if i cannot use Mann-Whitney, how do i analyze not normally distributed data? THANKS IN ADVANCE

[@how2stats]

Use an independent samples t-test with bootstrapping as the estimation technique.

[@ducnguyen7700]

Can i transform the data before conducting Mann-whitney test so my they ta can meet this assumption?

[@simonchen8376]

​@@how2stats is it the correct link? kzbin.info/www/bejne/gmOaZpuvntaGjNU

[@ninjanj6148]

I think this can also help: "Bootstrapping in SPSS." how2stats, kzbin.info/www/bejne/b4fNq4OkpLp1gLM

[@ninjanj6148]

Thank you very much for your videos! As you suggested we can use independent sample t-test bootstrapping instead when the shapes of the distributions are not the same. Can this be also used for an ordinal data? I have a similar question about Kruskal-Wallis test. Is the assumption of the shapes of the distributions also considered for an ordinal data, so that we have to use median test instead?

[@doc2010]

Your videos are AWESOME! Thank you for doing this! You taught me statistics...

[@Elinsinba]

Thanks for your post! What does “similar nature” mean please?

[@catarinamanita5445]

Thank you so much for taking the time to do these videos! It really helps! all the best to you

[@xichinesetakeaways9834]

Thank you so much for your such clear instruction! Sequentially, could you please show us how would you report the MW test with this assumption and the effect size? Usually, on other resources, they just report r and p value without saying anything about the assumption and the effect size.

[@how2stats]

I have a "how2manuscripts" channel, so, I might report that sort of information there at some point.

[@xichinesetakeaways9834]

how2stats That would be great! Thank you so so much.

[@akanshaaggarwal6403]

Thank you for explaining it so well. Just a follow up question, which statistical test/ correction should be apply in this case (when the assumption of homogeneity of variance is not met?

[@wanderlustwernweh5313]

what did you do about it?

[@ChibuzorFOgamba]

Thank you so much for your video. Please what do you do the levenes's test looks good for most of the variables but the box plots and histograms are not looking the same when it comes to shape? Which do you go with? Also, in a data set with multiple dependent variables and independent variables, must every variable meet the homogeneity of variance criteria for the Mann Whitney test to be accurate? What happens if only one dependent variable or two have a significant levenes's test?

[@joshbridges4719]

Awesome video! Thanks for the description. It helps clear things up for me. Question though: I was under the impression the MW test can be run in a general sense without assuming equal variances. It's only when we want to use it to test a location shift that the distributions need to be the same. Is this accurate?

[@how2stats]

Technically, it's true, the assumption is homogeneity of distributional shapes for the MW test. However, testing the difference between distributional shapes is not straightforward. Consequently, people test homogeneity of variance, instead, because a difference in variances necessarily implies a difference in distributional shapes

[@leonzhao2108]

This is so helpful, could you maybe do a K-S test video as well at some point?

[@erfanfadaei]

Hi, great videos, thank you. My data appears to have completely different distributions - skewness of 2 groups are 0.935, and -1.386, while kurtosis is 0.004 and .389. I tested homogeneity of variance using method in your video and medians were not significant. Why might this be, and can I then confidently use the Mann-Whitney test? Thanks!

[@fadhilfp6]

I think you are talking about the shape of distribution ? Is not it?

[@nmmichalak]

You always have great videos, thank you! Question: I've seen the definition of M-W U described as the prob. a given value in 1 group exceeds a value in the other group. Why would this computation be affected by unequal variance? Also, if sample size is large enough, all tests of assumptions reject the null :)

[@how2stats]

Thanks! I can't say I've seen that definition before. I would say that the overarching null hypothesis of the MW is that two groups have equal mean ranks. Although the typically observed MW formula is not particularly intuitive, it is necessarily the case that there is a standard error of the difference between mean ranks. The standard error would be based on variability, which would be assumed to be equal across groups. I agree, with large N, all assumptions based on inferential statistics will almost always be statistically significant.

[@nmmichalak]

Up front, I'm no expert, but I've been particularly obsessed with this test lately; everything I'm saying is based on informal online reading. That said, I'm curious about different ways to calculate standard error for this test. For example, the welch test uses a different formula for standard error and degrees of freedom in order to adjust for equal variance violations. I'm aware that there are "Mann-Whitney versions" of this, but I've had difficulty reading the papers on them. I'm curious if you've come across anything like a welch M-W U test that's as effective (i.e. controls false positive rates) as the welch test?

[@how2stats]

I can't say I'm familiar with an adjusted Mann-Whitney test. However, it is a little known fact that the independent samples t-test on ranked data and the Mann-Whitney U test yield *very* similar results. In fact, the independent samples t-test on ranks can protect the Type I error rate better than the Mann-Whitney in some cases (and give better power, too; see Conover & Iman, 1976; Zimmerman & Zumbo, 1993). So, if one were worried about non-normality, heterogeneity of variances (and even unequal sample sizes), one could conduct Welch's t-test via bootstrapping. It's about as full proof as you can get and keep power respectably high. In my opinion, there's no need to conduct the Mann-Whitney U test, anymore. Zimmerman, D. W., & Zumbo, B. D. (1993). Rank transformations and the power of the Student t test and Welch t-test for non-normal populations with unequal variances. Canadian Journal of Experimental Psychology/Revue Canadienne de Psychologie Expérimentale, 47(3), 523-539.

[@mayaskjevling888]

Hi! Could I ask which sources you used when making this tutorial?

[@imenzribi956]

Hi first of all thank you for the video it really helps me .Please I have one question I did all the steps that was shown in the video but when I run SPSS , the result don't show the test of homogenity of variance (the table) . I don't know what's the problem ?

[@charlesbourgoigne2130]

Is an independent samples t-test with bootstrapping better than Mann-Whitney-U-test? Can I use bootstrapping all the time instead of looking whether assumptions are made or not?

[@how2stats]

That's what I argue in my (free) textbook (Chapter 16): www.how2statsbook.com

[@SkidMcmarxx]

Does that mean that the results from the previous video don't really tell us anything useful?

[@estherchan3062]

I have a very small sample and it does not meet the assumptions for both t-test and Mann-Whitney, and it is too small to be bootstrapped. Is there anything else I can do?

[@how2stats]

First, simulation research suggests that the t-test (including the rank transform t-tests, which is very similar to the Mann-Whitney U) can be expected to yield valid resutls with sample sizes as small as N = 2 in each group (De Winter, 2013). So, sample size is not an issue. Secondly, what assumptions are violated? Mann-Whitney doesn't assume any level of normality. De Winter, J. C. (2013). Using the Student's t-test with extremely small sample sizes. Practical Assessment, Research, and Evaluation, 18(1), 10.

[@drmohgladiator]

Great video .. Then what to do when the assumption of unequal distributions is violated ?

[@how2stats]

I'd probably use bootstrapping.

[@alfredwong3790]

Do you mean using bootstrapping with the T test? Would this be the preferred method for comparing groups than the MW test?

[@1006caraqueno]

So if your continuous variables are non parametric BUT also violate the required assumptions of the Mann-Whitney U how are you supposed to compare the two?

[@how2stats]

Independent sample-test with bootstrapping.

[@liquidpaper1720]

What a great video! I have a question, though. Would it be any better if we tested homogeneity of variance with the non-parametric Levene test (the one that uses the ranks)?

[@how2stats]

If you're testing the assumption associated with Mann-Whitney, the non-parametric version of the Levene test seems to make more sense.

[@syuhaidahalimi3532]

Hi, in my decriptives, why there is no interquartile range value?

[@how2stats]

Do you have a standard deviation?