I am trying to run a MANOVA, but the Box's M and Levene's tests are significant. Does this mean I can not run a MANOVA?
Is it OK to continue if I choose to report Pillai's trace in stead of Wilks' Lambda under these circumstnces?
My sample size is 300, and I have 3 groups, each consisting of 100 participants. So is it OK to ignore Levene's test?
Thank you.
Box's M: If group sizes are over 30, then the MANOVA is robust against violations of homogeneity of variance-covariance matrices assumption.
Levene's Test: if homogeneity of variance can't be assumed for one (or more) dependent variables, then use an alpha level stricter than .05 (ie: use .001) when you evaluate the univariate ANOVAs.
[From "SPSS for the Health and Behavioural Sciences" (Allen & Bennett, 2008)]
What are you measuring? Depending on what you are measuring Levene's test being significant matters more or less. Because you have large and equal Ns, reporting Pillai's trace is likely sufficient.
That being said, if your data is greatly homoscedatic and you are concerned this is altering your outcomes you may want to use a non-parametric test. This article may help if you can access it.
https://www.researchgate.net/publication/247638643_Bias_in_the_Analysis_of_Repeated-Measures_Designs_Some_Alternative_Approaches
Article Bias in the Analysis of Repeated-Measures Designs: Some Alte...
Thank you.
I am measuring the frequency and intensity of negative and positive cognitions in bipolar patients, depressed patients and healthy controls. (N = 100 in each group, and quite a lot of scales are being used to compare these groups)
Is it OK to continue using MANOVA if Levene's test is significant?
Which variables are failing Levene's test? If they're all failing that's much more problematic than a single variable.
Though it seems like all of your variables are going to be susceptible to drastic individual differences, and as such a non-parametric test might be best.
Yes, most of the variables are significant in Levene's test. Is there a non-parametric test as an alternative for MANOVA?
Are you able to find the article I linked in my first response? If not you can search Google and research gate. Off the top of my head I want to say Friedman's rank order test might do it, but I am away from my computer and unable to verify that presently.
Since you have equal group sizes MANOVA is more robust to heterogenity of variances.You can report Pillai's trace statistic in this case instead of Wilk's lambda. I don't think a nonparametric test is necessary.
Box's M: If group sizes are over 30, then the MANOVA is robust against violations of homogeneity of variance-covariance matrices assumption.
Levene's Test: if homogeneity of variance can't be assumed for one (or more) dependent variables, then use an alpha level stricter than .05 (ie: use .001) when you evaluate the univariate ANOVAs.
[From "SPSS for the Health and Behavioural Sciences" (Allen & Bennett, 2008)]
- Badges
- Science topic