Cutting the depressed sample down to 70 might help the robustness aspect, but you're also likely to lose a fair bit of information by doing so, and reduce the degree to which you can generalise the results to other depressed people. I might suggest only doing this if the variance-covariance hetereogeneity is a problem. Check via Levene's test and Box's m (there are others that I don't recall offhand, sorry!)
Another technique you could utilise if you don't feel you can justify the assumptions of MANOVA is multinomial logistic regression, an unfairly ignored technique in psychology
By the way - I must say I'm wondering (I'm a psychology student/researcher myself) - what is the theoretical bent behind your research? As far as I know the association between disorders s/a schizophrenia/bipolar/MDD and common personality tests has been researched pretty extensively already?