R help - Dec 2003 - Design functions after Multiple Imputation

I am a new user of R for Windows, enthusiast about the many functions
of the Design and Hmisc libraries.
I combined the results of a Cox regression model after multiple imputation
(of missing values in some covariates).
Now I got my vector of coefficients (and of standard errors).
My question is: How could I use directly that vector to run programs such
as 'nomogram', 'calibrate', 'validate.cph' which, in
contrast, call for
the saved results form 'cph' ?
I did not use 'aregImpute' for multiple imputation. However, even if
I did it, 'fit.mult.impute'  seems not to allow specifying the option
'surv=TRUE' (essential to get a nomogram) or 'x=TRUE, y=TRUE'
(which
are essential for 'calibrate' and 'validate.cph'. Therefore, I
dont't see
how I could get a nomogram or run other Design functions after
'aregImpute'.

thank you so much in advance
Umberto

_________________________________________________________________
MSN Extra Storage! Hotmail all'ennesima potenza. Provalo!

On Mon, 08 Dec 2003 21:23:40 +0000
"Umberto Maggiore" <umberto_maggiore at hotmail.com> wrote:
> I am a new user of R for Windows, enthusiast about the many functions
> of the Design and Hmisc libraries.
> I combined the results of a Cox regression model after multiple
> imputation(of missing values in some covariates).
> Now I got my vector of coefficients (and of standard errors).
> My question is: How could I use directly that vector to run programs
> such as 'nomogram', 'calibrate', 'validate.cph'
which, in contrast, call
> for the saved results form 'cph' ?
> I did not use 'aregImpute' for multiple imputation. However, even
if
> I did it, 'fit.mult.impute'  seems not to allow specifying the
option
> 'surv=TRUE' (essential to get a nomogram) or 'x=TRUE,
y=TRUE' (which
> are essential for 'calibrate' and 'validate.cph'.
Therefore, I dont't
> see how I could get a nomogram or run other Design functions after
> 'aregImpute'.
> 
> thank you so much in advance
> Umberto
Good questions Umberto.  The Design package handles multiple imputation
and model validation, but currently not at the same time.  But model
descriptions such as those provided by summary.Design, nomogram.Design,
contrast.Design are fully operational after the variance-covariance matrix
is corrected for multiple imputation by fit.mult.impute.  There is one
caveat.  When the nomogram not only has predicted relative hazards but
absolute survival estimates (probabilities, quantiles, restricted mean
life), fit.mult.impute gets the baseline survival estimates from surv=T
from the first imputation.  We need to extend that to average baseline
survival estimates over all multiple imputations.

You should have had no difficulty using nomogram after fit.mult.impute
(after aregImpute); only fully trust the relative hazard estimates from
the resulting nomogram.

I have put out a note on the IMPUTE e-mail list (see
http://hesweb1.med.virginia.edu/biostat/rms for subscription information)
about how to develop a reasonable algorithm for simultaneous multiple
imputation and bootstrap model validation/calibration.  So far no takers.

Frank

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