Multivariate Analysis of Variance

#1
Hi all

I am currently preparing my research for ethics submission but trying to establish an accurate sample size using G*Power.

Apologies in advance if I do not explain this accurately and will try to simplify it as much as possible.

I have three groups that will be determined based on test scores. Let's say G1, G2, and G3.

In the experiment I am going to be comparing each group on several different measures (6 and 1 confounding variable).

Will this require a MANOVA?

As one of the tests have small effect sizes, I am concerned I am going to need to recruit an unachievable sample size. However, I have read where a MANOVA analysis is concerned the effects sizes should be adjusted. Is this correct?

Thank you in advance.
 

Karabiner

TS Contributor
#2
I have three groups that will be determined based on test scores. Let's say G1, G2, and G3.
Is this really necessary? Grouping might be arbitrary and non-generalizable, and it will reduce statistical power.
Why not use the test score as predicor variable?
Will this require a MANOVA?
This depends on your research goal, and how the 6 dependent variables are related to
each other.
If they jointly operationalize one theoretical construct, and they do not correlate extremely
high (= redundancy), and not very low (= weak association), and you just want to say
something about that construct, then MANOVA would be an option.
As one of the tests have small effect sizes,
What do you mean by this? That you consider the 6 variables seperately, and for one
variable you expect only a small effect?
However, I have read where a MANOVA analysis is concerned the effects sizes should be adjusted.
Is this correct?
I'm afraid that I do not quite understand what you mean here.

With kind regards

Karabiner
 
#3
Is this really necessary? Grouping might be arbitrary and non-generalizable, and it will reduce statistical power.
Why not use the test score as predicor variable?

This depends on your research goal, and how the 6 dependent variables are related to
each other.
If they jointly operationalize one theoretical construct, and they do not correlate extremely
high (= redundancy), and not very low (= weak association), and you just want to say
something about that construct, then MANOVA would be an option.

What do you mean by this? That you consider the 6 variables seperately, and for one
variable you expect only a small effect?

I'm afraid that I do not quite understand what you mean here.

With kind regards

Karabiner
Hi Karabiner - Thank you for getting back to me.

Why not use the test score as predicor variable?
The experiment is using clinical samples. The three groups will be control, at-risk and clinical. The idea would be to examine the significant differences between the groups.

This depends on your research goal, and how the 6 dependent variables are related to
each other.
If they jointly operationalize one theoretical construct, and they do not correlate extremely
high (= redundancy), and not very low (= weak association), and you just want to say
something about that construct, then MANOVA would be an option.
One of the measures being used has never been looked at in the sample I am using although it is possible a relationship will exist although at the moment it is theoretical. The primary aim of the study is to examine relationships between clinical assessment and each of the tasks.

What do you mean by this? That you consider the 6 variables seperately, and for one
variable you expect only a small effect?
Yes.

I'm afraid that I do not quite understand what you mean here.
This was more to do with calculating power in my study. I think I have this bit sorted. It depends on the family of tests that are used z, t, f etc. However for each family of tests the effect sizes are adjusted in accordance with the test being used.