Dear R-helpers, excuse me if this is not exclusively an R-related question. I have data from a nested design, both temporally and spatially, and the reponse variable of interest is left-censored. That is, only values > "some treshold" are available, otherwise "LOW" is reported. Are there ways of building a linear model with both fixed and random effects, when the response variable is censored? Can the tobit model be modified to do this? Does anyone have experience with this type of dataset? Help is much appreciated, Remko Duursma ^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~'^'~,_,~' Remko Duursma, Ph.D. student Forest Biometrics Lab / Idaho Stable Isotope Lab University of Idaho, Moscow, ID, U.S.A.
mixed-effects models for left-censored data?
3 messages · Remko Duursma, Thomas Lumley, A.J. Rossini
On Wed, 11 Jun 2003, Remko Duursma wrote:
Dear R-helpers, excuse me if this is not exclusively an R-related question. I have data from a nested design, both temporally and spatially, and the reponse variable of interest is left-censored. That is, only values > "some treshold" are available, otherwise "LOW" is reported. Are there ways of building a linear model with both fixed and random effects, when the response variable is censored? Can the tobit model be modified to do this? Does anyone have experience with this type of dataset?
For a random intercept model you could use survreg() and frailty() in the survival package. In general the random effects tobit model will be quite hard to fit, involving a numerical integration whose dimension is the number of random effects. Some sort of EM algorithm might work. There is a paper by Pettit in Biometrics some time ago on censored linear mixed models -- I don't have the reference with me. -thomas
Thomas Lumley <tlumley at u.washington.edu> writes:
On Wed, 11 Jun 2003, Remko Duursma wrote: For a random intercept model you could use survreg() and frailty() in the survival package. In general the random effects tobit model will be quite hard to fit, involving a numerical integration whose dimension is the number of random effects. Some sort of EM algorithm might work.
One huge catch with that approach is heterscedasticity, which seems to pop its head up all too often with limit-of-detection assay data.
There is a paper by Pettit in Biometrics some time ago on censored linear mixed models -- I don't have the reference with me.
There is also a paper by a fellow named Jim Hughes, in Biometrics (late 90s?), on this exact topic -- he used single imputation, whereas he mentioned later (private communication) that a multiple imputation approach would be better. The S-PLUS code (it isn't pretty) is somewhere on his WWW page, buried deep in the U Washington Biostatistcs WWW site. At least it used to be. best, -tony
A.J. Rossini / rossini at u.washington.edu / rossini at scharp.org
Biomedical/Health Informatics and Biostatistics, University of Washington.
Biostatistics, HVTN/SCHARP, Fred Hutchinson Cancer Research Center.
FHCRC: 206-667-7025 (fax=4812)|Voicemail is pretty sketchy/use Email
CONFIDENTIALITY NOTICE: This e-mail message and any attachments ... {{dropped}}