Advantages

• MANOVA means multiple ANOVA - multiple dependent variables to be analyzed simultaneously.
• Repeated measures is a subset of MANOVA.
• If you use multiple one way ANOVAS to try to do this, you will raise the probability of a Type I error too high. MANOVA controls the experiment-wide error rate. Lots of ANOVAS raise power, but that is spurious because the Type I error rate goes up as well.
• MANOVA is a "gateway". Look at the multivariate F. If it is significant, then it allows you to look at the individual univariate analyses (rather than with ANOVA, having to run a separate post hoc test.)
• You can use it with assorted dependent variables, or with repeated measures. This is important because you'd usually use it when you can't collapse the measures into a few factors because they're all different.
• When there is multicollinearity (you can make a linear combination out of the dependent variables), MANOVA may detect combined differences not found in the univariate tests.
• You gain power over separate ANOVAs.
• The more DVs you have, the more you need MANOVA.

Limitations

• The number of people in the smallest cell should be larger than the total number of dependent variables.
• It can be very sensitive to outliers, for small N.
• It assumes a linear relationship (some sort of correlation) between the dependent variables.
• MANOVA won't give you the interaction effects between the main effect and the repeated factor.

Trade-offs

Assumptions

• ANOVA assumptions - normality, independence, HOV, plus:
• linearity/multicollinearity of dependent variables.
• MANOVA does not have the compound symmetry requirement that the one-factor repeated measures ANOVA model requires.

When is it good or bad to have variables correlate?

• We don't want high correlation among independent variables. That means they are redundant.
  • If the IVs are highly correlated, when you do crosstabs, you will see empty cells. You need at least two people in each cell. The cells ought to be roughly balanced.
  • You could add related variables and make a new factor, but the new factor has to have real meaning. Else choose one or the other.
• We don't want high correlation among dependent variables either.
  • If you have fairly low correlations among dependent variables (e.g., knowledge of stat and attitudes toward stat correlate 0.3) then you are OK.
  • If you have two highly correlated DVs (e.g., patterns of use and reasons for use correlate 0.95), then the errors in the data analysis are more highly compounded because you are measuring the same people.
• Is there a correlation between the independent and dependent variable? This is what your major analysis ought to tell you. That is real information.

Multivariate Tests in SPSS

• These tests are all a bit different.
• The multivariate F ratio is not just additive - that's why.
• Use Wilks or Pillais. They have good power and are most immune to violations of assumptions.
• Roys is the most powerful if all assumptions are met.
• Pillais is better if the homogeneity of covariance is not met.
• If the F ratio is significant, then SPSS will go on to calculate the univariate results for you.