Re: st: Nonparametric Methods for Longitudinal Data

["JVerkuilen (Gmail)" <jvverkuilen@gmail.com>]

On Wed, Feb 13, 2013 at 6:08 AM, Thomas Herold <thomasherold@gmx.net> wrote:
> Thanks a lot to all of you. That really sounds quite promising and will
> probably keep me busy over the next week as many of the proposed methods are
> completely new to me.
>
> Regarding the questions asked by Jay (thanks for your detailed answer):
>
> -There are missing observations and the dropout rate is quite high. However,
> I do have all the time-constant information (gender, marital status etc.).
> Here´s a little overview (please ignore the numbering; effectively, there
> are 5 waves):
> N (1st wave)=117
> N (3rd wave)=91
> N (5th wave)=68
> N (6th wave)=58
> N (7th wave)=36 (with 9 obs belonging to the 1st, 18 to the 2nd and 9 to the
> 3rd group).
>

That sounds fairly ugly. I think you have a huge missing data problem
that needs to be addressed first. Unfortunately GEE based models don't
cope with missingness as well as a mixed model, and if there is
time-invariant missingness (e.g., gender being missing) then it's time
for MI.

This analysis sounds quite complex. As a consultant I'd charge a lot for it. ;)


> - If I am not completely mistaken I am looking for a PA model because I am
> mainly interested in the effects of the 2 treatments vs. the control group.

Probably but that doesn't mean you wouldn't want to use a conditional
model and then calculate the relevant marginal response at the end of
the day.

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