[R] Survival Analysis Daily Time-Varying Covariate but Event Time Unknown

[David Winsemius]

On Sep 16, 2010, at 4:43 PM, David Winsemius wrote:

>
> On Sep 16, 2010, at 12:14 PM, smm7aa wrote:
>
>>
>> Help!
>>
>> I am unsure if I can analyze data from the following experiment.
>>
>> Fish were placed in a tank at (t=0)
>> Measurements of Carbon Dioxide were taken each day for 120 days  
>> (t=0,...120)
>> A few fish were then randomly pulled out of the tank at different  
>> days,
>> killed and examined for the presence of a disease
>> T= time of examination in days from start (i.e. 85th day), E = 0/1  
>> for
>> nonevent/event
>>
>> My problem has been linking all the Carbon Dioxide measurements up  
>> to the
>> day of examination and trying to create a survival object.
>>
>> I have considered interval censoring with right censored for fish  
>> without
>> disease and then left censored for fish with the disease, but i  
>> really
>> cannot structure the data or intuitively figure out how to  
>> incorporate the
>> daily Carbon Dioxide values up until day of examination.
>>
>> The end goal to to predict an event based on Carbon Dioxide levels
>
> I think the goal should be restated as estimation of the proportion  
> of disease in the population as a function of time and CO2  
> concentration. I think Poisson regression would be sensible analysis  
> framework. I don't think you need to consider censoring unless your  
> repeated sampling has removed a substantial proportion of the  
> starting population.
>
> ?glm  # with family="poisson"
>
> Poisson regression is a proportional hazards framework that is  
> suitable for grouped data such as you have. You do need to ask  
> whether recovery is possible from a diseased state and what sort of  
> analysis you will apply to individuals who died during hte study  
> period, but those are domain questions, as much as statistical  
> questions.
>

As a further note: There is a nice paper by Atkinson and colleagues at  
the Mayo Clinic with R/S code for analyses
"Poisson models for person-years and expected rates", Elizabeth J.  
Atkinson, Cynthia S. Crowson,  Rachel A. Pedersen,  Terry M. Therneau.  
Technical Report #81

http://mayoresearch.mayo.edu/mayo/research/biostat/upload/81.pdf

You will be shown how to set up the time intervals (and  population at  
risk if that happens to change materially) as offsets.


-- 
David Winsemius, MD
West Hartford, CT