Hi Stas and Steve and the rest of statalist,
I spoke with a statistician yesterday (that did assist with some of the making of the original pweigts) and agreed that my first level should be individuals, level 2-census tract (codeupm), and level 3 if i wanted it, state.
And that the L1 weights should be the original individual pweights
And that the L2 census tract weights should be the average of the pweights for each specific census tract
And L3 state given a constant weight of 1
So i constructed the glamm pweight as follows and yet, I still cannot get a basic random-intercept only model to converge...i also tried including state...thoughts? Thanks again for all your help, I really appreciate it!
*MLM-level pweights
generate pwadulsr1=adul_sr
*cenus tract average pw adul_sr
by codeupm, sort: egen pwadulsr2=mean(adul_sr)
generate pwadulsr3=1
gllamm bmi2 if diettag==1 & exwt==1, i(codeupm) pweight(pwadulsr) adapt nip(20)
Running adaptive quadrature
Convergence not achieved: try with more quadrature points
gllamm bmi2 if diettag==1 & exwt==1, i(codeupm) pweight(pwadulsr) adapt nip(15) cluster(state)
Running adaptive quadrature
Convergence not achieved: try with more quadrature points
___________________________________________
Rebecca M. Kanter
PhD Candidate
Johns Hopkins Bloomberg School of Public Health
Department of International Health
Center for Human Nutrition
________________________________________
From: [email protected] [[email protected]] On Behalf Of Stas Kolenikov [[email protected]]
Sent: Saturday, July 18, 2009 12:53 PM
To: [email protected]
Subject: Re: st: gllamm with pweights
Can you contact the data provider to get the probability weights for
the higher level units? Most weight scaling procedures look to me like
at attempt to guess the color of somebody's dress on a black-and-white
picture.
Also, I understand that this is interesting substantively, but I feel
uneasy specifying the random effects at the level higher than the PSU
level. I've seen that done though if those higher levels correspond to
strata (although typically strata are understood as fixed rather than
random effects). You said the design is stratified; where was
stratification applied? My guess would be that your strata are the
state by urban/rural cells, which are your urstate factors. It looks
to me that at least nominally the model for the design would be to
have the higher level random effects correspond to these strata. But
in your application, as Steve S noted, they look kinda weird, and
states may indeed be conceptually better units to think of. As I said,
you would probably want to use -geq(rural)- or something like that to
get the main effect of the urban/rural location.
On Fri, Jul 17, 2009 at 11:38 AM, Kanter, Rebecca<[email protected]> wrote:
> Hi Steve and list,
>
> The original survey design is a multi-stage stratified design. The PSU is essentially the equivalent of a U.S. census tract (the probability that one of these tracts was selected was proportional to the number of households within it and the number of tracts selected corresponded to the sample size in the strata within the state) ...from which households are selected (with probability proportional to size). For each census tract selected six "blocks" are selected with probability proportional to the number of houses in each block; within each chosen block 6 households are selected via systematic random sampling and then individuals within the household via simple random sampling.
--
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/
