Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down at the end of May, and its replacement, statalist.org is already up and running.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Generalized lineal models with survey data


From   Stas Kolenikov <skolenik@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Generalized lineal models with survey data
Date   Tue, 27 Jul 2010 18:02:00 +0100

No, you don't have any problems with the degrees of freedom, which is
#PSUs - #strata = 837-4 = 833, and is reported as such. So I tend to
believe in Steven's story about empirical underidentification of the
overdispersion parameter: the likelihood is so flat in alpha that the
curvature (inverse of the variance) of the likelihood wrt this
parameter cannot be estimated with numeric accuracy that Stata would
find acceptable to report. And yes, this is an indication that
overdispersion is not such a great problem: coniditioning on
covariates and taking weights into account seems to make your data
approximately OK.

As for the general convergence problems, they may be caused by the
scale of weights. Note that your log pseudo-likelihood has 8 digits
before the decimal point, and typically Stata wants to optimize things
down to 7 or so digits after the decimal point, that is, you need to
have about 15 reliable digits to declare convergence. That's too much
to ask for, as 15 digits is the accuracy limit of the -datatype-
double. In this situation (and in this situation only), it would be OK
to relax the convergence criteria by specifying something like
-ltolerance(1e-3)- instead of the default 1e-7; or rescale the weights
so that they sum up to say sample size rather than the population
size.

On Tue, Jul 27, 2010 at 5:30 PM, Paolina Medina
<carmencitamedina@gmail.com> wrote:
> Thank you both, very much.
> So this almost zero alpha, without a confidence interval can be taken
> to indicate that there is no overdispersion in the model?
> Here is my svyset statement and the complete output..
> I am using 52 regressors (including the constant), i really dont know
> how many are the design degrees of freedom... But in fact whenever i
> take any of these regressors i get a lot of troubles with convergence
> in the survey results (not concave or backed up) and i have to throw
> away many other regressors to get convergence again.
> Do you know anything i can do to fix this?

-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
*   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/


© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index