Thank you for the informations, especially for the cumhaz tip. I'll explain a bit more of my experimental design: I have 4 different replicate populations of drosophila, from which I take samples to infect, in 5 independent replicates (tubes) of 10 individuals each. That makes 200 individuals per treatment (50 per pop), and if I treat it as a whole like in coxph(Surv(Time,Status)~Treatment) or coxph(Surv(Time,Status)~strat(Treatment) some of the variability associated with the population and replicate would go unconsidered, and the variance would be wrongly estimated. That's my interest in using (nested) random models. For now, I would only like to plot the survival curves of each treatment, with the "correct" s.e., and I'm having some troubles with that using the standard R functions. In the future I would like to know if a) what are the different survival curves for each treatment b) the variability between populations (i.e. if the different populations are significantly different between each other, for the same treatment) c) if the variability within populations for a given treatment is different from other treatments. d) all this removing the error associated with the replicates. Thank you again for your help
On Sun, Jan 16, 2011 at 16:44, David Winsemius <dwinsemius at comcast.net> wrote:
On Jan 16, 2011, at 5:07 AM, Nelson Martins wrote:
I am a relative newbie to survival analysis and R in general, but would like to use the coxme package to analyse some data I currently have. The data is relative to survival times of drosophila melanogaster populations to infection with pathogens, and has the variables: Time, Status, Treatment (4 treatments + 2 controls) Population Replicate ?and I'm currently using the following call mixed<- coxme(formula = Surv(Time, Status) ~ strata(Treatment) + (1 | Population/Replicate),x=T,y=T) The treatments have very different mortality profiles, that's why I'm using stratification. I have several problems: 1 - I'm able to get the results, and to compare the different treatments using glht (e.g. - glht(mixed,mcp(Treatment="Tukey"))), only if I don't stratify the treatments
My understanding is that strata are used when you _don't_ want estimates for that factor (at least in ordinary coxph() modeling.)
2 - I'm ?unable to plot the results as in other cox models (cph,coxph) using survplot/survfit. The other models don't allow specifying random variables (Populations and Replicates), which is of great importance to me.
My understanding is that _this_ is the situation in which one would use strata (for factors like Replicates) within regular coxph() models. I'm not clear what you mean by "Population" and so cannot comment on how it might be handled.
3 - One other problem I'm having is in plotting the hazard curves (i.e. the instantaneous risk at each day). Is it possible to do it using these functions, hopefully with the associated standard errors?
(Sorry, I don't have sufficient understanding of the operation of coxme to make informed comments. Terry Therneau visits R-help but not generally on the weekends. In ordinary survfit use, there is a "cumhaz" option to the fun argument.))
Thank you for your time and help Best regards Nelson Martins
David Winsemius, MD West Hartford, CT