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 on April 23, and its replacement, statalist.org is already up and running.


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

Re: st: Gllamm: Convergence not achieved: try with more quadrature points


From   Kjetil Arne Van Der Wel <kjetil.wel@sam.hio.no>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: Gllamm: Convergence not achieved: try with more quadrature points
Date   Thu, 23 Sep 2010 09:39:44 +0200

Thank you very much for your valuable comments, Steve and Stas!
Best,
Kjetil

On 22.09.2010 15:54, Steve Samuels wrote:
--

Kjetil-

To Stas's comments, I would only add:

In my post, I used the technical sampling term "strata": units which
together constitute the population. In your data, countries are
formally strata.

To get an idea of what the probability weighted estimates will look
like, set up the survey design approach:

svyset _n [pweight= wt1], strata(country)
svy: logit [your model, which can include country-level indicators and
covariates]


Standard errors will be valid for the individual-level covariates. The
validity is "design validity", based on the possibly false assumption
that observations within country were sampled independently, and not
in multiple-stages or clusters. For true design validity, you would
need to add information about the sample design within each country
(strata, first-stage clusters or "primary sampling units")

Another possibility: try HLM, which will do multi-level models with
survey data and give design-based estimates of standard errors.

Steve

Steven J. Samuels
sjsamuels@gmail.com


On Wed, Sep 22, 2010 at 9:03 AM, Stas Kolenikov<skolenik@gmail.com>  wrote:
2010/9/22 Kjetil Arne Van Der Wel<kjetil.wel@sam.hio.no>:
Many thanks for this. Sorry for describing the analysis so poorly. I did
actually run the command with pweight (I tried both methods), but that did
not change the situation. Also, each country within the survey provides a
representative sample of its population. Together, the countries constitutes
EU-25, and hence, do not represent a greater *population' of countries.
Furthermore, the countries are not merely administrative borders, but
political entities employing different policies that have profound impact on
the lives of their inhabitants. Because of this, I feel it is correct to
treat them as 'units' in which individuals (and regions) are nested (i.e.
appropriate entities of variation that can be described with variables)
rather than 'strata' (which I take to mean as constructed by the researcher,
such as social class). Sorry if I completely misunderstood you, Steve and
Stas.

There are several schools of thought in survey statistics. The
hard-core design based paradigm states that uncertainty in the data
and statistics based on them is exclusively due to sampling, and
inference based on the probability spaces generated by randomization
mechanism. Both Steve and I tend to gravitate towards this end of
specturm. On the other end of it, there is model-based inference that
postulates an underlying model for the data generating process, like a
multilevel model with fixed or random effects for strata and PSUs.
http://www.citeulike.org/user/ctacmo/article/850244 comes more or less
from this school, and it is a natural place to gravitate for
multilevel people. There is no "truth", there are different
approximations to reality, and depending on the objectives of your
analysis and the tools available to you, one or the other or something
in between might be more appropriate.

Running this command:

gllamm depvar indepvar1-4, i(country) pweight(wt) link(logit)
fam(binom) nip(20) adapt

leads to the error message, while omitting the pweight works fine. (By the
way; depvar and all indepvars are binary individual level variables, except
age. in the next model, continous counry level variables will be added).

N(i)=188 000
N(j)=25

That's a sizeable data set. Gllamm might be running into numerical
difficulties with it, I am afraid. How much variation is there in your
weights? If there is substantial variation, then the few (a couple
thousand, in your case) observations dominate the estimation
procedure, and -gllamm- just cannot find a good place for the
remaining ones in the model. If that is the case, you would want to
scale and possibly trim some weights to make the model more stable.

--
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/


*
*   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/

--
Kjetil van der Wel
Stipendiat/PhD student
Gruppe for inkluderende velferd (GIV) / Research Group for Inclusive Social Welfare Policies,
Avdeling for samfunnsfag (SAM) / Faculty of social sciences
Høgskolen i Oslo / Oslo University College

Rom: SG304 (Stensberggata 29)
tlf/phone: +47 22 45 35 76 / +47 906 41 606
http://www.hio.no/giv

Postadresse/ postal address:
SAM
Høgskolen i Oslo
Postboks 4, St. Olavs plass
0130 Oslo


*
*   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