Re: st: R-squared measures proposed by Cameron and Windmeijer (1996) in stata

[Nick Cox <njcoxstata@gmail.com>]

I'll take (2).

You need to think about what it would mean in your field to have
perfect predictions. It means that the data provide a complete
description _and_ the model captures the generating process exactly.
Or it would mean that you were using a model with too many parameters.
In most observational fields even large datasets provide only partial
data and the model is at best a caricature of the underlying process.

Imagine for example that the response variable was the number of
children born so far to a set of women. We can get variables like age,
occupation, religious affiliation, etc. that may have a variety of
direct or more likely indirect effects on the response, but it seems
likely that there will always be a lot of unexplained variation
arising from individual details.

Literature in your field will tell you what kind of fit is considered
strong enough to be publishable.

Nick

On Fri, Mar 25, 2011 at 4:46 AM, Francisco Rowe <frowe@ucn.cl> wrote:

> I have two questions regarding the estimation of a Negative Binomial Regression Model with quadratic variance (NBRM2).
>
> 1) How could the R-squared for NBR2 based on deviances proposed in the paper below be implemented using STATA? Does it require any changes if robust standard errors are used for the estimation?
> Cameron, C and Windmeijer, F 1996, 'R-squared measures for count data regression models with applications to health-care utilization', Journal of Business & Economic Statistics, vol. 14, no. 2, pp. 209-20.
>
> 2) In general -for what I have seen-, why measures of fit in form of (Pseudo) R-squared are relatively low (<0.3)? Are there any special conditions that cause this?
*
*   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/