Hello Yuanyuan Zhu. My advice is to read the nice article by Francis Huang (see link below) and ask yourself if one of the other analytic approaches he suggests might be more suitable. HTH.

Article MANOVA: A Procedure Whose Time Has Passed?

Bruce Weaver has a good point here and it's one that I generally agree with. However in this case I think that Bruce is using an elephant gun to shoot a gnat

Let's think about this problem for a minute. We know manova can be treated as a multivariate regression. In this case there's only one IV. Multiplying the regression coefficient by the only 1 IV gives only one linearly independent DV which is what the Box test said. In other words this situation is impossible and hence this experiment is severely messed up and even Bruce's normally good suggestion can't save it. This can't happen because a constant multiplied by a vector only changes that vector's length. So one must ask where did you get this data. If this is actually real data please explain the experiment. Best wishes David Booth

Hello Yuanyuan,

Allow me to take a different tack from what Professors Weaver and Booth have offered:

First, don't use the composite scores!

The reason is, all the information in those scores is carried by their constituent subscores. So, if you're comparing two or more groups on all of the subscores, this will be redundant with comparing those groups on the composite scores.

Second, if your IV is categorical (e.g., a grouping variable), consider using discriminant analysis instead of manova. Internally, the same algorithms are applied, but most D.A. programs offer specific information about the weights (both unstandardized and standardized) applied to the measures to optimally separate the group centroids in space. It is this information which can be helpful in gauging how the groups differ, if at all, rather than having to resort to a bucketful of univariate tests, which is the advice given in Huang article.

Finally, if the IV is continuous, then essentially you have a single OLS regression model to evaluate.

Good luck with your work.

Just for the heck of it I'll take issue with my colleague Prof. Morse. First,we agree that the composite scores cannot lead to independent DVs. This would reduce the number of DVs. that's true. But how do the Box M test results fit in here. This test said there's is no more than one covariance matrix that is nonsingular. Let's now cut to the chase and as I said produce the data and the paper it came from so we can clearly resolve this hideous mess. One other possibility still seems to exist. That's the possibility that this is a linear algebra take home test. In any case produce the data and this issue can be finally resolved.. Hidden data is no data at all. Best wishes to all and to all a good night. David Booth

Yuanyuan Zhu, please tell us more about the IV. E.g., is it categorical or quantitative? Thanks.