Re: st: Sampling weights in gllamm

[Stas Kolenikov <skolenik@gmail.com>]

On Thu, Mar 18, 2010 at 1:47 AM, Alexis Le Nestour
<alexis.lenestour@gmail.com> wrote:
> I'm working on a dataset collected in Senegal including 4513 individuals living in 505 households nested in 93 farming organizations.
> We use a two stage stratified sampling procedure. A fixed number of farming organizations were sampled in 3 rural communities, and the probability to be picked up for a farming organization depends on its size (number of households). Then, we randomly chose 5 households in each farming organization (few of them were oversampled to increase the sample size). Finally, we interviewed all the household members.

Since you have a nearly equal probability of selection design, you can
probably get a pretty good start with -xtmelogit-. It does not support
the weights, but I would argue you don't need them that badly. At the
very least you would get very good starting values for -gllamm-. Of
course an accurate variance estimator would have to take clustering on
PSUs and SSUs into account, but as long as these will be modeled
explicitly, you may not need any other corrections.

> On this dataset I estimate a three level random intercept logistic model to assess the probability of seeking treatment. My levels are individuals, households and farming organizations.
> I'd like to introduce the sampling weights in my estimate but:
> - I read the paper written by Chantala et al. on this issue and I downloaded the pwgils command but unfortunately it's designed for a two level model.
> - I don't know exactly how to use the sampling weights for the individuals. Since we interviewed all the individuals in a household the probability of inclusion on someone equals the probability of inclusion of his/her household. Therefore I'm wondering if I can only use the weights at the household and farming organization levels, even if my analysis is at the individual level.

You are correct that since there was no sampling at individual levels,
there are no separate weights. You would still need to feed something
to -gllamm-, but you can simply put a vector of ones for your level-1
weights.

> - I don't know which rescaling method I should use.

There is simply no good answer to that, period. I have not been
convinced by the existing literature that one of the methods is always
better than others. It looks like the only reasonable way to decide is
to run a small simulation with your typical sample size and structure
of the sample to see which method gives the most accurate estimates. A
simulation with -gllamm- will come out... hm... what's a good way to
put that... computationally costly.

You might want to echo your questions on multilevel@jiscmail.ac.uk
list. Be prepared for a swarm of not-so-relevant answers suggesting to
switch to another software though. Basically you want to hear what a
guy by the name of Cam MacIntosh would have to say :)).

-- 
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/