Should I be using a MANOVA?

#1
Hi all,

I apologise in advance if I explain this poorly - this is all somewhat new to me.
I'm currently doing a small piece of classroom research looking at the effects of two different types of teacher feedback. It's very small (n30).

I was recently told that to keep the number of tests to a minimum I should be using MANOVA, but I'm not entirely convinced this is the case.

Do I need multivariate analysis when:

1) I have three groups (2 different treatments, and 1 control), each of which received a different treatment over 4 sessions.

2) I am looking to compare these groups on two outcome variables, but I am not interested in the relationship between these two variables (only how the three groups fare in each).

If a MANOVA is appropriate, my understanding is that there is no non-parametric MANOVA. I was told that if I just ranked my data, and then used a parametric test, this would effectively be the same as a non-parametric MANOVA. Does this make sense?

Many thanks in advance for any suggestions.
 
#2
Hi, you need a MANOVA instead of an ANOVA if you have more than 1 outcome variables. Even if you are not interested in the relationship between the two outcomes, a MANOVA can be more sensitive / can have more power compared to 2 separate ANOVAs. But especially if you have only two outcomes and you say you are not interested in the relationship between them, I think two ANOVAs are OK for your purposes.

"...each of which received a different treatment over 4 sessions..."

Keep in mind: As soon as you have repeated measurements of the same participants, you have to use a mixed ANOVA, since this is not anylonger an independent design.

You need non-parametric ANOVAS only if parametric assumptions (such as normality of the outcome, homogeneity of variances...) are not met. If they are met, you should use parametric tests since they often have more power compared to non-parametric ones.
 

Karabiner

TS Contributor
#3
You need non-parametric ANOVAS only if parametric assumptions (such as normality of the outcome, homogeneity of variances...) are not met.
Some additions:

ANOVAs (or other GLM procedures, such as regressions) do not assume a normal distribution
of the response variable, only normality of the residuals (and if the sample is large enough,
> 30 or > 50 or so, even this assumption is negligible).

The only non-parametric ANOVA is the Kruskal-Wallis H-test, which is analogous to a oneway
ANOVA, but cannot handle multifactorial designs, including mixed designs.

With kind regards

K.
 
#4
ANOVAs (or other GLM procedures, such as regressions) do not assume a normal distribution
of the response variable, only normality of the residuals
Hi Carabiner, for GLM's/ANCOVAS I agree, but in case of an ANOVA I think normality of residuals and normality of the outcome is virtually the same - the difference is only a constant: The group mean
 
#6
The language used by some people and some statistical packages is confusing here. However, what you need to account for is the dependence of the data.

There are several ways to do this, but probably the most flexible is to use a multilevel model. In SAS this would be PROC MIXED.
 
#7
Thanks for your replies all - very much appreciated.

So should I take this:

The only non-parametric ANOVA is the Kruskal-Wallis H-test, which is analogous to a oneway
ANOVA, but cannot handle multifactorial designs, including mixed designs.
to mean that I'm basically out of luck when it comes to non-parametric tests for my situation (comparing results for three groups over four treatments, as well as within each group over four treatments)?