[R] fit.mult.impute and quantile regression

Frank E Harrell Jr f.harrell at vanderbilt.edu
Tue Jun 15 12:53:03 CEST 2004

roger koenker wrote:
> Having not tried this, it is dangerous to speculate, but it appears to 
> me that there
> would be no problem passing rq arguments (crucially, only tau, the 
> specification
> of the quantile of interest) to fit.mult.impute, since the call to the 
> "fitter" procedure
> includes a ... argument.  The real question would seem to be:  are the 
> assumptions
> underlying the imputation procedure consistent with the rq fitting, that 
> is are they
> assuming something stronger than that the tauth conditional quantile 
> function of
> y is linear in x?   There seem to be quite a variety of options for the 
> imputation
> in transcan, maybe Frank could advise on this?

You are right Roger, fit.mult.impute is generic.  It just assumes that 
impute.transcan can find the imputations to insert, to complete the 
dataframe.  Note that transcan is no longer the function to use for 
multiple imputations.  Use aregImpute.  It is flexible as long as 
additivity holds.

> url:    www.econ.uiuc.edu/~roger            Roger Koenker
> email    rkoenker at uiuc.edu            Department of Economics
> vox:     217-333-4558                University of Illinois
> fax:       217-244-6678                Champaign, IL 61820
> On Jun 15, 2004, at 11:52 AM, <david_foreman at doctors.org.uk> wrote:
>> I have a largish dataset (1025) with around .15 of the data missing at 
>> random overall, but more like .25 in the dependent variable.  I am 
>> interested in modelling the data using quantile regression, but do not 
>> know how to do this with multiply imputed data (which is what the 
>> dataset seems to need).  The original plan was to use qr (or whatever) 
>> from the quantreg package as the 'fitter' argument in Design's 
>> fit.mult.impute, but it is not clear whether this would work, 
>> especially as fit.mult.impute seems only to work with the default 
>> settings of its 'fitter' arguments, which rather defeats the purpose 
>> of quantile regression.  Help!!

Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University

More information about the R-help mailing list