R - Two-Way Between Subjects MANOVA

[@piyushashah1]

Thanks, if I use this method in a paper, what are the ideal citations?

[@StatisticsofDOOM]

For MANOVA? Probably Tabachnick and Fidell 2012 book.

[@oneadd5012]

Thank you for sharing! I want to know how to calculate confidence interval after manova. Could you show me the steps or upload a new video to instruct us? Thanks!

[@StatisticsofDOOM]

Check out the package "emmeans" which will do marginal means for you.

[@oneadd5012]

@@StatisticsofDOOM Thank you very much! I will give it a try.

[@oneadd5012]

Hi, I used emmeans to calculate confidence intervals, but it is more difficult to me. I tried effectivesize. I did it in two different ways: one done after following procedures in this video, the partial eta sqaure is different from the result in the video by using 1-lambda; another done based on directly using manova to the data without doing lm, but the partial eta square is also different from the value of (1-lambda). But the two results are similar in partial eta squared. Again I use SPSS to check the manova results, finding this result inculding partial eta square is similar to new methods used. Above all, I applied two methods to the overall results of manova rather than the post hoc individual anova test. So could you elaborate how to get partial eta squre and confidence interval for your "Manova" model? Thanks!

[@StatisticsofDOOM]

@@oneadd5012 oh, you are doing CI for effect size not the overall means. You can try MBESS or MOTE for partial eta squared noncentral confidence intervals.

[@oneadd5012]

Thanks! But why there's a difference between "partial eta square" values by using different tools or measures? Could the value by formula (1-lambda) be used always? Or I report the value by other measures (e.g. SPSS). The two might lead contrary interpretation of results due to their big discrepancy. Thank you! I'll try MBESS and MOTE.

[@sannebast]

I have a question about this video: in another video you made named 'R - Data Screening 4 Assumptions' you said to use Levene's test for ANOVA and Box M test for MANOVA. Why did you still choose to do Levene's here? And can you always opt for that as well with MANOVA, or only sometimes? Thank you for all your videos. You really helped me out big time :) and you have a very nice way of lecturing!

[@StatisticsofDOOM]

I think it's useful to know which ANOVA might be the issue!

[@_Anonymous_9]

Great video. Can we still use MANOVA if we have a continous IV also? For example, if you measured age and included it as a covariate in the above analysis? Edit: I know technically that is a MANCOVA, but can we use the manova() command to do it?

[@StatisticsofDOOM]

I am unsure actually - it's been a while since I used this package. I would just try it and see what happens!

[@_Anonymous_9]

@@StatisticsofDOOM Hey thanks for your reply 😊. Well it works, but I was thinking of if there are any extra steps one should take. For example, typically ANOVA is robust when using categorical variables and looking at interactions. I couldn't find any information on including a continous variable (e.g. age) and looking at interaction effects. Textbooks say dichotomise, but every professor I know (and I think in many journal outlets) frown upon it for discarding data. Have you ever done an ANCOVA and had your categorical IVs interact with a continous IV? (rather than simply server as a control variable). Is that "allowed"? . Thanks once more 🙏

[@StatisticsofDOOM]

@@_Anonymous_9 Yes, I have several videos on ANCOVA actually. I would recommend just doing it traditional "regression" style. You'd have to do each DV separately, though.

[@Nibiil152]

I'm also doing a MANCOVA with age (and also sex) as covariates. I've been following the steps of this video and of the video on ANCOVA and I have doubts regarding the "contrasts=list" you wrote inside the lm function here (while not in the ANCOVA - why not there, actually?). When I try to do this for the IVs and for the CVs, it works for the IVs but for the CVs it only works for sex and not for age, and I get an error telling me that contrasts apply only to factors. So, shall I apply this only to the IVs, as even sex wouldn't strictly be a factor, if I'm not really interested in its effect?

[@soyoungwoo8831]

Thank you for the great video! Can I use this method with 3 IVs? I have a 2x2x3 design.

[@StatisticsofDOOM]

In theory yes! In practice, MANOVA may not be a great choice when ANOVAs could serve just as well.

[@soyoungwoo8831]

@@StatisticsofDOOM Oh so running ANOVAs for each DV? My DVs are somewhat related so I chose MANOVA but not sure if this is the right test. Thank you for your help :)

[@StatisticsofDOOM]

@@soyoungwoo8831 Sure! That would allow you to test each separately. Or if you wanted to control for that correlation, you could use a multilevel model to combine them together (which would be better than MANOVA).

[@soyoungwoo8831]

@@StatisticsofDOOM Thank you for your advice! I will try the tests :)

[@eurosonly]

You are conducting a study to examine the impact of both diet and exercise on weight loss. Participants are randomly assigned to be on a strict raw diet or no diet. In addition, participants are randomly assigned to complete 30 mins of cardio daily or to not workout at all. After 6 month, participants record the number of pounds they have lost.

[@StatisticsofDOOM]

Ok, that sounds like a repeated measures ANOVA, not a MANOVA though.

[@vrlwz666]

If i'm only interested in interaction, can i just do ANOVA on each DV but not do MANOVA?

[@StatisticsofDOOM]

Yes!

[@amogh2101]

This is great! Am I correct in assuming that MANOVA can also be applied to self-report inventories where they have sub-constructs? If so, is there an example/resource that I can look for?

[@StatisticsofDOOM]

Mostly, I don't recommend MANOVA much at all - it doesn't always necessarily do what you expect it to. What about normal repeated measures? Or using factor analysis to create a composite score?

[@tingchanghui]

Thank you for this really clear explanation and demonstration. I've a question about the IVs of femininity and masculinity. As participants are not randomly allocated into low and high fem and masc groups, would these IVs be covariates instead, and would a MANCOVA be more appropriate? Thanks in advance for clarifying!

[@StatisticsofDOOM]

It would just be a quasi experimental design, rather than experimental design - since these are our IVs, we want to use MANOVA. If they were only meant for adjustment/control, then you could use MANCOVA because the covariates are often only used to adjust for factors. You still have to have IVs!

[@juliemilovanovic2679]

Thank you for this video, helps a lot. I have one question though, I tried that on my dataset and when doing the manova I get an error message saying: residuals have rank 4 < 5. What does this mean? I find a way to get around it by including tol=0 like this summary(res.man, test='Wilks',tol=0). I can't find what it means and if going around it means that my results are not reliable.

[@StatisticsofDOOM]

More than likely that implies that two of your DVs are so highly correlated that it can't deal. Try running a correlation table on them to see.

[@juliemilovanovic2679]

@@StatisticsofDOOM Thanks for the answer and advice. I did that and the highest correlation is 0.5 so not that correlated. I think the problem is that my data is not normal. The kurtosis and skewness was ok but when I did a Shapiro test it shows that the data is not normal. Is there a statistical test to explore differences between my groups that I can do since I have dependant variables and a non normal dataset? (I don't have a background in statistics so I struggle a bit)

[@StatisticsofDOOM]

@@juliemilovanovic2679 If the DV isn't normal and samples are small, you can try a Mann-Whitney test - honestly, I would recommend just running the ANOVAs on each DV separately. I'm not sure I think MANOVA is very useful.

[@DulalAlauddin]

Again Amazing. A BIG Thanks, I will definitely put your name in my wordings of thanks in our article. But i still don't understand why people use barcharts for displaying group means, since bars start from zero, it suggest that there are people that score zero. boxplots with means (and error bars) make more sense to me and display more information about the distributions. Rick Scavetta, pointed this out in one of his fantastic courses on ggplot2 on DataCamp.

[@StatisticsofDOOM]

That's a good point - I usually try to get real papers to show the actual range of the data (i.e. 1 to 7 should start at 1 and go to 7 so as not the distort the data). I think it's probably a historical precedent thing - also think that dot plots with error bars are easier to read personally.

[@sannebast]

Thanks

[@StatisticsofDOOM]

Appreciate your support!

[@sannebast]

Thanks

[@StatisticsofDOOM]

Appreciate the kind gesture!