Re: st: Nonparametric Methods for Longitudinal Data

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

On Mon, Feb 11, 2013 at 7:16 AM, Thomas Herold <thomasherold@gmx.net> wrote:

> The problem is that it has recently been discovered that the depression
> score we are working with can only be interpreted as ordinal data. What´s
> more, the resulting variable is far from being normally distributed - the
> data is just not suitable for parametric analysis.
>

These are two separate issues. Parametric analysis based on the
assumption of the predictor being metric may be a very good
approximation even if, strictly speaking, the measure doesn't satisfy
the necessary assumptions of interval measurement. Also the crucial
assumption is that errors are Gaussian, not that the dependent
variable itself is; it can be markedly non-normal. Thus a model that
assumes Gaussian errors and a linear regression may be pretty
reasonable despite the issues you mention.

However, screening data tend to be highly non-normal, often to the
point that normal-theory based models such as linear mixed models
break down. There are a number of ways to solve this problem but
without knowing more about your data it's impossible to know.

Questions you need to address:

-Are there missing observations? If so, what is the nature of the
missingness by variable? Do you have only time-varying missingness or
are the time-constant variables missing? If the latter, the analysis
has just become WAY more complex.
-What kind of analysis do you want, conditional on subjects or some
kind of population-averaged?

If they are as I would suspect, you may benefit from a model such as a
gamma Generalized Estimating Equation (GEE), or you might well be able
to trick xtmepoisson into doing the analysis even if the data aren't
integer-valued.

*
*   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/