Emma Slaymaker <[email protected]> noticed that -estat gof- and
certain -predict- options are not allowed after -svy: poisson-, but are
allowed after -poisson- with -pweight-s and the -cluster()- option:
> I have a question regarding the goodness of fit measures available for
> poisson models based on survey data.
>
> I've tried three commands to fit the same poisson model and find that
> Stata is not consistent in the post-estimation options offered after
> these three different commands. I can't find an explanation for the
> difference in the manuals (though perhaps I've missed something).
>
> I started by fitting a poisson model to the survey data using
> -svy:poisson-. After this command Stata does not allow -estat gof- or
> the prediction of any residuals (anscombe, deviance etc).
>
> My data are stratified, which is why I'm using -svy:poisson- and not
> -poisson- with the cluster and weight options. However, if I ignore the
> strata and use -poisson- with the weight and cluster options then Stata
> allows -estat gof-. If I use -glm-, again with the weight and cluster
> options and ignoring the stratification, I can obtain the anscombe,
> pearson and deviance residuals.
>
> Is there is something about the stratification of the data that makes
> it difficult, or inappropriate, to calculate the residuals and the
> goodness of fit test? I'm under the impression that weights and
> clusters are more of a problem in this respect.
>
> Is it appropriate to calculate residuals/gof following glm or poisson
> with the weight and cluster options? Is -svy:poisson- too strict, or
> are -glm- and -poisson- too lenient?
>
> I've not found anything in the literature so would be grateful for any
> help.
Stata is a little more strict about what is available by way of postestimation
with survey estimation results. This is mostly due to the fact that the
distribution is not generally known for most of these statistics or if they
are even appropriate when computed from survey data.
Consider -estat gof- after -poisson-. To our knowledge, the distribution of
the Pearson gof statistic from clustered data is not known (and similarly for
stratified or any other kind of data not from an IID design).
-poisson- and -glm- are too lenient in this regard. This is generally true
for most of the standard estimation commands; however, you will notice from
the survey manual that the corresponding survey estimation commands restrict
the available postestimation statistics to those known to be valid for use
with survey data.
--Jeff
[email protected]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
