Skip to content

Quantile regression (rq) and complex samples

2 messages · James Shaw, Thomas Lumley

#
I am new to R and am interested in using the program to fit quantile
regression models to data collected from a multi-stage probability
sample of the US population.  The quantile regression package, rq, can
accommodate person weights.  However, it is not clear to me that
boot.rq is appropriate for use with multi-stage samples (i.e., is
capable of sampling primary sampling units instead of survey
respondents).  I would like to apply Rao's rescaling bootstrap
procedure and poststratify the weights to population control totals in
each bootstrap replicate.  I know how to do all of this in Stata but
have not yet seen any means of doing so in R.  I  presume I could do
what is needed using batch processing but was hoping that there might
be a way to pass the rq parameter estimates to a package that performs
resampling variance estimation in order to simplify the task.  Any
programming suggestions or directions to informational resources would
be greatly appreciated.

Jim
1 day later
#
You could use the survey package to run the bootstrapping, if you mean
the Rao & Wu bootstrap that samples n-1 of n PSUs in each replicate.

Set up a survey design object with bootstrap replicate weights: use
svrepdesign() if you already have replicate weights, use svydesign()
and then as.svrepdesign() to get R to make the replicate weights for
you.

Then use withReplicates() to run rq() for each set of replicate
weights and compute the variance.

       -thomas
On Thu, Jan 27, 2011 at 11:18 AM, James Shaw <shawjw at gmail.com> wrote: