Significant Multivariate Effects but not Univariate

[Marissa4usa]

---

Members don't see this ad.

Hi all,
I am a bit confused about an analysis I am running (well technically it's for students in my TA class).
So, they have two DV's (both are continous variables) and one caterical IV with two levels. I told them to run a MANOVA.

The output says that there is significant multivariate effect, however, for all four multivariate tests everything is exactly the same, except for "value". The significance is .013 (i.e. significant).

However, when I then check for the univariate effects, nothing is significant.

I am now wondering whether I should just let them run two t-tests or whether I messed up. Was a MANOVA the right choice?

Any ideas?

 

[cara susanna]

It's possible for that to happen. IIRC, it just means that the effect for one IV is only significant in the context of the other IVs.

 

[loveoforganic]

Would that indicate an augmenting effect between at least a couple of the variables on one another, or does this just represent a sample size too small to make significant the effect size of the individual variables, which becomes significant when the effects are summed, looking multivariately?

 

[aequitasveritas]

i second Cara.

You're looking at an interaction with no main effects. Add post hocs and assess further.

 

[Marissa4usa]

What if SPSS won't let me do post-hocs? Remember, the IV only has 2 levels.

 

[hopefuldoc97239]

The problem is not a mathematical problem, but the fact that after years of undergrad training in trying to visualize problems in univariate terms it's hard to switch back into multivariate thinking.

If multivariate is significant but univariate is not... that is your reason why. You are really dealing with a multivariate phenomenon.

The opposite is when you find significant univariate effects without significant multivariate effects... means that the phenomenon is expressed at those individual variables.

It's like if you found that men are taller on average than women, but don't have longer individual bones. Just because the bones aren't significantly longer individually doesn't mean they aren't taller overall.

 

[aequitasveritas]

What if SPSS won't let me do post-hocs? Remember, the IV only has 2 levels.

My mistake. You could use logistic regression, no? They are not seperate groups but in this way you could assess the direction of difference between the the levels of the IV.

 

[cara susanna]

Couldn't you do a t test?

 

[Marissa4usa]

My mistake. You could use logistic regression, no? They are not seperate groups but in this way you could assess the direction of difference between the the levels of the IV.

I would consider that if this wasn't for the undergraduate Social Psych class I am ta'ing. They were already freaking about the fact that they had to run a MANOVA. Try teaching them logistic regression when all they ever want to run are t-tests😀. Essentially, I just want to make sure I'm not telling them complete bogus information.

Couldn't you do a t test?

well, yeah. Essentially the two univariate tests are t-tests. But I wanted to teach them how to do it "the right way" 😉.

It's like if you found that men are taller on average than women, but don't have longer individual bones. Just because the bones aren't significantly longer individually doesn't mean they aren't taller overall.

Thanks, that was a really helpful example.