[R] imputation of sub-threshold values

[A.J. Rossini]

Thomas Lumley <tlumley at u.washington.edu> writes:

> On Mon, 15 Mar 2004, Jonathan Williams wrote:
>
>> Is there a good way in R to impute values which exist,
>> but are less than the detection level for an assay?
>
>
> If there were a good way to do it, it would probably be implementable or
> implemented in R.
>
>
> If you can persuade the people measuring the values to give you the
> numbers (assuming they are just below `limit of detection' rather
> than genuine non-detects) you will reduce the need for imputation. This is
> often the most powerful technique -- analytical chemists are trained not
> to give out numbers less than some multiple of the measurement error, but
> they can often be persuaded that statisticians can be trusted with these
> numbers.
>
> If you think your data are close to one of the parametric survival models
> in the survival package, you can analyse the data as left-censored,
> without imputation. You could impute from the fitted model, if you need
> imputations.
>
> Otherwise you may be stuck with some ad hoc imputation with a sensitivity
> analysis to see how the results depend on the imputed value.

Jim Hughes, in Thomas' department, had written a paper on this for
Biometrics and implemented an approach for this problem in S-PLUS.  It
used to be on Jim's WWW site.  Not sure where that is at UW.  It was
mostly in the context of repeated assays, but probably could be
trimmed down or reimplemented.

I'm not claiming it's a good way, either.  But it does what is
requested (and similar to Thomas' comment, I'd suggest being careful
with this approach).

best,
-tony

-- 
rossini at u.washington.edu            http://www.analytics.washington.edu/ 
Biomedical and Health Informatics   University of Washington
Biostatistics, SCHARP/HVTN          Fred Hutchinson Cancer Research Center
UW (Tu/Th/F): 206-616-7630 FAX=206-543-3461 | Voicemail is unreliable
FHCRC  (M/W): 206-667-7025 FAX=206-667-4812 | use Email

CONFIDENTIALITY NOTICE: This e-mail message and any attachme...{{dropped}}