Re: st: clustering in quantile regressions with sampling weights

[Stas Kolenikov <skolenik@gmail.com>]

Michael refers to the likelihood-based approach to survey data
analysis, in which conditioning on the design variables makes the
probability weights superfluous (note that conditioning is a very
vague term; adding the design variables to the regression may not
suffice, as you'd have to add their powers and interactions for an
approximation to effective conditioning). Unfortunately, I don't think
this will work out with quantile regression, as this is not a
likelihood-based method. In my experience, the bootstrap methods seem
to work well, although formal proofs are convoluted and rely on rather
complicated conditions on the designs and sampling probabilities, and
small sample performance is basically unknown (Austin and I are
running some simulations to try to identify the limits of
applicability, but we don't have any particularly strong findings yet
to report). See my description of the survey bootstrap in SJ 10 (3)
and accompanying package.

On Fri, Oct 22, 2010 at 12:44 AM, Michael N. Mitchell wrote:
>  I know this is a longshot, but do you know and have the variables involved
> in creating the sampling weights (e.g., gender, age, race, etc...). If you
> have the variables that were used to create the sampling weights, then you
> could include those variables as predictors in your model to account for
> them.

-- 
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/