MANOVA question

[lazure]

---

Members don't see this ad.

Hi,
I am doing MANOVAs followed by Roy-Bargmann Stepdown analyses to check which dependent variables contribute unique variance to the observed effects. In a few instances, the multivariate effect for the set of dependent variables is significant, all stepdown effects are significant but only one of the univariate effects is significant. Tabachnic and Fidell refer to these situations as "difficult" to interpret and don't offer much more guidance than that. How do I report these results?
Thanks....

 

[lazure]

If you could answer both questions, that'd be great 🙂

 

[phd2006]

I'm a little confused; how many times have I said that when it comes to stats? You did a MANOVA, followed by the R-B stepdown analysis -- I'm assuming because you found sig effects for the DVs. Now, the stepdown analysis shows each of the DVs to be sig yet you are concerned because these effects were not shown at the univariate level (for all IVs and DVs?). Did I read that correctly? If so, it sounds like you know which #'s you need to report, or you couldn't have given this much info. Now interpretation....whether it's MANOVA, Multiple Regression, ANOVA etc., when something is not sig at the univariate level but is at the multivariate level, sounds like an "additive effect" is occuring. That is -- something by itself is not sig but when you add in other info, whether IVs or DVs, the original variable that was not sig becomes significant.....potentially speaks to a moderator (although Baron and Kenney would disagree, Kramer might agree) or a suppressor effect. See Cohen, Cohen, West, Aiken (2003), they expand on T & F.

Of course, you need to also consider sample size (total and per group), # of IVs and # of DVs. You have given no info as to the 'fit' of your model. In the R-B stepdown analysis, each time you add a DV to the next sequential ANOVA, you are removing explained variance...may be the DV you are concerned about not being sig at the univariate level is sig at the multivariate level as the model is now 'over fit.'