How to do survival analysis in SAS with interval censoring and random effects

[Tarek]

Hi,

I am interested in using survival analysis methods on seed germination data. I have just started reading Allison's Survival analysis in SAS. Accelerated failure time analysis (using PROC LIFEREG) and Cox proportional hazards models using PROC PHREG seem promising, but I have a feeling that neither will do what I need and I may have to resort to GENMOD, GLIMMIX, or NLMIXED. The various experiments I am working on have 2-4 treatments each (e.g. 3 maternal temperatures, 4 seed germination temps, 5 genotypes, 3 seed cold stratification durations), and I'd like to know how the treatments affect the mean or median time for a seed to germinate.

I have data on seeds on petri dishes that were censused on day 1, 3, 5, 8, and 14. These are interval censored, with a seed that germinated on d8 really germinated between day 6 and day 8. LIFEREG can handle interval censoring. But seeds on the same petri dish are not independent, so I need to be able to include the dish as a random effect (often called "frailty").

Is there an approach that allows interval censoring as well as random effects? Is there a way to do this in NLMIXED or GLIMMIX? IF so , how should I modify the code below to include random effects?

PROC LIFEREG DATA=survGeno2;
CLASS geno mat_temp;
MODEL time*status(0)=geno | mat_temp;
RUN;

thanks!

 

[hlsmith]

Trying to remember why you can't just add and control for the clusters in model if you feel they contribute toward the outcome. Also, I think people may usually invert the generated graphs for similar types of question, so instead of descending stairs the graph will look like ascending stairs.

 

[hlsmith]

Bump - please don't create duplicate posts. Doing so can lead to too many abandoned threads and confusion if people end up helping you on two different threads. In addition, in the future other site visitors may not get the full question or responses in one localized place, since components of the question/response may be scatter throughout the forum.

My vague question or statement was "why can you just introduce the clusters via a categorical covariate?". I don't know if that would be appropriate or not/

 

[Tarek]

That's a good point about spreading out my question not helping future visitors. I was worried that a survival analysis question would be better suited to readers of "biostats" than SAS since I'm now open to using other software packages.

As for your question, there too many petri dishes to use as a categorical covariate instead of a random effect. There might be 500+ petri dishes in some of the datasets I'm working on. This is because the petri dish was really the experimental unit, not the individual seed. I had to recode my data as if each seed was independent (even though there were 12 seeds/dish) in order to try survival analyses. Even though I can only independently apply a treatment to a whole dish not to individual seeds on a dish, other researchers seem to not regard this as pseudoreplication as long as a petri random effect is included (e.g. Onofri et al. 2011 "the cure model:an improved way to describe seed germination?" Weed research 51:516-24 DOI: 10.1111/j.1365-3180.2011.00870.x). Also, petri dish passes the "replacability" test for a random effect:an individual petri dish can be replaced with another without reducing my ability to draw conclusions form the experiment.


So any survival analysis experts out there know how to deal with grouped-time (interval censored) survival data with random effects?