Left censored responses in mixed effects models

[Bert Gunter]

Dear R Fellow-Travellers:

What is your recommended way of dealing with a left-censored response
(non-detects) in (linear Gaussian) mixed effects models?

Specifics: Response is a numeric positive measurement (of volume, actually);
but when it falls below some unknown and slightly random value (depending on
how the sample is prepared and measured), it cannot be measured and is
recorded as 0.

There is some statistical literature on this, but I was unable to find
anything that appeared to me to implement a strategy in any R package. If it
matters, I am less interested in inference than in removing possible bias in
estimation.

Feel free to respond off-list if you feel that this would not be of general
interest.

Cheers,

Bert Gunter
Genentech

______________________________________________
R-help@... mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

[Don MacQueen]

I assume you've looked at the NADA package(?) While I don't believe
it goes as far as dealing the mixed effects models, it might give you
a starting point, and possibly some additional references.

-Don

At 9:08 AM -0700 5/12/08, Bert Gunter wrote:

>Dear R Fellow-Travellers:
>
>What is your recommended way of dealing with a left-censored response
>(non-detects) in (linear Gaussian) mixed effects models?
>
>Specifics: Response is a numeric positive measurement (of volume, actually);
>but when it falls below some unknown and slightly random value (depending on
>how the sample is prepared and measured), it cannot be measured and is
>recorded as 0.
>
>There is some statistical literature on this, but I was unable to find
>anything that appeared to me to implement a strategy in any R package. If it
>matters, I am less interested in inference than in removing possible bias in
>estimation.
>
>Feel free to respond off-list if you feel that this would not be of general
>interest.
>
>Cheers,
>
>Bert Gunter
>Genentech
>
>______________________________________________
>R-help@... mailing list
>https://stat.ethz.ch/mailman/listinfo/r-help
>PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
>and provide commented, minimal, self-contained, reproducible code.

--
--------------------------------------
Don MacQueen
Environmental Protection Department
Lawrence Livermore National Laboratory
Livermore, CA, USA
925-423-1062

______________________________________________
R-help@... mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

[Gregor Gorjanc-2]

Bert Gunter <gunter.berton <at> gene.com> writes:
> Dear R Fellow-Travellers:
>
> What is your recommended way of dealing with a left-censored response
> (non-detects) in (linear Gaussian) mixed effects models?

Your description of the data calls for a tobit model

http://en.wikipedia.org/wiki/Tobit_model

I think you need to take a look in the survival package.

Regards, Gregor

______________________________________________
R-help@... mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

[giovanni parrinello]

Dear Bert,
a solution is the 'package'
censre3 by  Hughes JP
Reference: Hughes JP: Mixed effects models with censored data with
application to HIV RNA levels.  Biometrics, 55:625-629, 1999.

Giovanni

Don MacQueen ha scritto:

> I assume you've looked at the NADA package(?) While I don't believe it
> goes as far as dealing the mixed effects models, it might give you a
> starting point, and possibly some additional references.
>
> -Don
>
> At 9:08 AM -0700 5/12/08, Bert Gunter wrote:
>> Dear R Fellow-Travellers:
>>
>> What is your recommended way of dealing with a left-censored response
>> (non-detects) in (linear Gaussian) mixed effects models?
>>
>> Specifics: Response is a numeric positive measurement (of volume,
>> actually);
>> but when it falls below some unknown and slightly random value
>> (depending on
>> how the sample is prepared and measured), it cannot be measured and is
>> recorded as 0.
>>
>> There is some statistical literature on this, but I was unable to find
>> anything that appeared to me to implement a strategy in any R
>> package. If it
>> matters, I am less interested in inference than in removing possible
>> bias in
>> estimation.
>>
>> Feel free to respond off-list if you feel that this would not be of
>> general
>> interest.
>>
>> Cheers,
>>
>> Bert Gunter
>> Genentech
>>
>> ______________________________________________
>> R-help@... mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-help
>> PLEASE do read the posting guide
>> http://www.R-project.org/posting-guide.html
>> and provide commented, minimal, self-contained, reproducible code.
>
>

--
dr. Giovanni Parrinello
Department of Biotecnologies
Medical Statistics Unit
University of Brescia
Viale Europa, 11 25123 Brescia
email: parrinel@...
Phone: [phone]
Fax: [phone]

______________________________________________
R-help@... mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

[giovanni parrinello]

Dear Bert,
you can also translate in nlme, as I'm trying to do, the approach of
Thiébaut and *Gadda( *Mixed models for longitudinal left-censored
repeated measures. Computer Methods and Programs in Biomedicine 74
<http://www.informatik.uni-trier.de/%7Eley/db/journals/cmpb/cmpb74.html#ThiebautJ04>(3):
<javascript:void(0)>(2004)) written in nlmixed(SAS)
Giovanni
Bert Gunter ha scritto:

> Dear R Fellow-Travellers:
>
> What is your recommended way of dealing with a left-censored response
> (non-detects) in (linear Gaussian) mixed effects models?
>
> Specifics: Response is a numeric positive measurement (of volume, actually);
> but when it falls below some unknown and slightly random value (depending on
> how the sample is prepared and measured), it cannot be measured and is
> recorded as 0.
>
> There is some statistical literature on this, but I was unable to find
> anything that appeared to me to implement a strategy in any R package. If it
> matters, I am less interested in inference than in removing possible bias in
> estimation.
>
> Feel free to respond off-list if you feel that this would not be of general
> interest.
>
> Cheers,
>
> Bert Gunter
> Genentech
>
> ______________________________________________
> R-help@... mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
>
>  

--
dr. Giovanni Parrinello
Department of Biotecnologies
Medical Statistics Unit
University of Brescia
Viale Europa, 11 25123 Brescia
email: parrinel@...
Phone: [phone]
Fax: [phone]

______________________________________________
R-help@... mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

[Richard Cotton]

Bert Gunter wrote:

What is your recommended way of dealing with a left-censored response
(non-detects) in (linear Gaussian) mixed effects models?

Specifics: Response is a numeric positive measurement (of volume, actually);
but when it falls below some unknown and slightly random value (depending on
how the sample is prepared and measured), it cannot be measured and is
recorded as 0.

There is some statistical literature on this, but I was unable to find
anything that appeared to me to implement a strategy in any R package. If it
matters, I am less interested in inference than in removing possible bias in
estimation.

The simplest way is to substitute some number for the nondetects, typically half the limit of detection.  This method is usually fine if you have less than about 10% nondetects, but can lead to big biases with larger numbers of nondetects.

Denis Helsel has written fairly extensively on this topic, and is the author of the book "Nondetects and data analysis", which the NADA R package is based upon.  He discusses it here:
http://www.practicalstats.com/nada/nadafiles/files/NADAforR_Examples.pdf

Possibly the best option is to use OpenBUGS to run an MCMC model.  It has a nice interface with R using the R2WinBUGS package.  Be warned though, this option does require some understanding of Bayesian stats.