[R] survreg with measurement uncertainties
therneau at mayo.edu
Wed Jun 12 15:51:00 CEST 2013
I will assume that you are talking about uncertainty in the response. Then one simple way
to fit the model is to use case weights that are proprional to 1/variance, along with
+cluster(id) in the model statement to get a correct variance for this case. In linear
models this would be called the "White" or "Horvitz-Thompsen" or "GEE working
independence" variance estimate, depending on which literature you happen to be reading
(economics, survey sampling, or biostat).
Now if you are talking about errors in the predictor variables, that is a much harder problem.
On 06/12/2013 05:00 AM, Kyle Penner wrote:
> I have some measurements that I am trying to fit a model to. I also
> have uncertainties for these measurements. Some of the measurements
> are not well detected, so I'd like to use a limit instead of the
> actual measurement. (I am always dealing with upper limits, i.e. left
> censored data.)
> I have successfully run survreg using the combination of well detected
> measurements and limits, but I would like to include the measurement
> uncertainty (for the well detected measurements) in the fitting. As
> far as I can tell, survreg doesn't support this. Does anyone have a
> suggestion for how to accomplish this?
More information about the R-help