RE: st: Non-normal alternative to MANOVA

["Nick Cox" <n.j.cox@durham.ac.uk>]

This is a frequently asked question without an official FAQ. (It's not
StataCorp house style to say "No, we don't do it" unless there is also
an explanation of why it is a very bad idea.) 

See, for example,  

http://www.stata.com/statalist/archive/2010-05/msg01339.html

and several replies. 

Nick 
n.j.cox@durham.ac.uk 

marietherese.kelly@flinders.edu.au

Thanks for the replies.
I've had a look at transforming the data or using lnskew0, the problem
is there
are 15 variables, each measured at 3 times periods (baseline, 7 weeks
and 3
mothns after treatment) and it's pretty difficult to find a
transformation that
works for all of them. The data were skewed at baseline and it only got
worse
after treatment. The confounder numint (number of interventions) is a
major
issue so I can't ignore it. With 3 repeated measurements I wasn't sure
if you
could use lnskew0 separately on all 3 of them if I'm interested in
changes from
baseline?

I'll have a look at ranking the data - is there a good reference
anywhere on how
to do that? I did it once in the past but can't remember the details.

> I have been asked to analyse a small dataset from a randomized
controlled
trial,
> with 24 subjects (12 in each group, control and intervention). There
are 3
sets
> of continuous outcome measures (b1, b2, b3) taken at baseline, 7 weeks
and 3
> months, and 2 variables  which need to be controlled for  - level of
impairment
> impair (binary) and number of interventions numint (continuous
integer).
>
> If the data were normal I would have done something like this:
>
> manova b1 b2 b3=group impair numint, continuous(numint)
>
> and then do various manova tests.
>
> Unfortunately the outcome data b1-3 are highly skewed, so I can't use
manova.
> What would be a good alternative?
>

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/