Are your response variables distributed...
...linearly
...unimodally
Or will you work with (dis)similarity matrices of your response variables?