MANOVA, finally
So, after three posts of
we have arrived at MANOVA. If you skipped those three posts, feel shame at trying to take shortcuts, go back and read them.
Before we dive into coding, let’s take a look at some basic background on MANOVA.
The difference between ANOVA and MANOVA is simple
- With ANOVA you have one dependent variable
With MANOVA you have multiple dependent variables
How does that work? Think back to what you know about multiple correlation
In correlation, you are looking at the relationship between two variables, X and Y. You predict changes in X from changes in Y
Y = bX
In multiple correlation you are looking at the relationship between Y and MULTIPLE X variables.
You have an equation something like
Predicted Y = b0X0 + b1X1 + b2X2 + b3X3
And you are looking at how the Y variable changes in relation to the PREDICTED Y. Notice that predicted Y is a sum of all of your variables, each of which is multiplied by a regression coefficient.
The correlation between these predicted Ys and the actual Y is your multiple R and the multiple R-squared in ANOVA or regression is the square of the multiple R.
The multiple R-squared answers the question – how much of the variance in the dependent variable can be explained by variance in the independent variable (s) ?
In the case of ANOVA, this variance is in group membership, so we are testing the null hypothesis that the mean of group1 = the mean of group 2 all the way to group N
With MANOVA, you have multiple variables on the Y side of the equation
The variable you are predicting/ explaining in this case is also a weighted sum
Dependent = w1Y1 + w2Y2 + w3Y3
Our null hypothesis is that the mean of this weighted combination is equal for groups 1, 2 and all the way up to group N
Instead of looking at a multiple R-squared in this case, we look at two other statistics, Wilk’s lambda and Pillai’s trace
- Assumptions of MANOVA
- Independent, randomly sampled observations
- Variables follow a multivariate normal distribution
- Homoscedasticity – population covariances for the dependent groups are equal
- Relationship of dependent variables is linear (because notice you made the dependent into a linear equation)
Also note that in the case of a repeated measures ANOVA certainly assumption 1 and possibly assumption 3 are violated
When you have conducted your MANOVA the first thing you should look at is the Multivariate tests – Wilk’s lambda, Pillai’s trace . Rejecting the null hypothesis that the model does not explain the difference in the VECTOR of means then leads you to examine the second logical question, which of these dependent variables differs ? So , if you don’ t have a significant, lambda, trace, etc. STOP. If you do, move on and check out the univariate F-tests. If your F is significant, go on to post hoc tests.
ETA-squared is the variance accounted for IN THE LINEAR COMBINATION OF THE DEPENDENT VARIABLES by the model.
Mertler and Vannata said it well.
“When the IV has only two categories, the F test for Pillai’s Trace, Wilks’ Lambda, and Hotelling’s Trace will be identical. When the IV has three or more categories, the F test for these three statistics will differ slightly but will maintain consistent significance or nonsignificance. Although these test statistics may vary only slightly, Wilks’ Lambda is the most commonly reported MANOVA statistic. Pillai’s Trace is used when homogeneity of variance-covariance is in question. If two or more IVs are included in the analysis, factor interaction must be evaluated before main effects. “
