Re: st: Poisson regression with score/scale as DV

[Nick Cox <njcoxstata@gmail.com>]

That's Austin Nichols' presentation.

The folklore is that -poisson- works well even if the usual
assumptions are not well satisfied. Furthermore, as is customary, the
exact form of marginal distribution of the response is not part of the
assumptions.

So, I would try -poisson-. But since your response (you say DV)  is
also bounded, I would consider dividing by 27 and using -glm,
f(binomial)-. Why stick to one model when you can try two?

Nick

On Tue, Apr 3, 2012 at 8:58 AM, Clinton Thompson
<clintonjthompson@gmail.com> wrote:

> Based on Austin Nichol's July 2010 presentation on use of Poisson for
> non-negative skewed variables
> (http://www.stata.com/meeting/boston10/boston10_nichols.pdf) as well
> as Bill Gould's Stata blog post about the same
> (http://blog.stata.com/2011/08/22/use-poisson-rather-than-regress-tell-a-friend/),
> I'm considering Poisson for a problem I have, albeit the dependent
> variable is different enough from their examples that I'm unsure
> whether Poisson is entirely appropriate.  In my problem, the DV is a
> "score" assigned to each subject based on their responses to several
> component (ordinal scaled) questions.  This variable is bound between
> zero and 27 and the distribution of responses is decidedly non-normal
> (pile-up of responses at the zero value).  Any thoughts on whether
> Poisson is still a good candidate or should I be considering other
> approaches?  And if Poisson is not the right way to go, any advice on
> how to model this DV?
>
> I'm using Stata/SE 11.2 for Windows.
>

*
*   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/