Dear all, I would have some questions on the coxph function for survival analysis, which I use with frailty terms. My model is: mdcox<-coxph(Surv(time,censor)~ gender + age + frailty(area, dist='gauss'), data) I have a very large proportion of censored observations. - If I understand correctly, the function mdcox$frail will return the random effect estimated for each group on the scale of the linear predictor? - Similarly, the variance of the frailties is the variance of the terms on the scale of the linear predictor? - A p-value is provided for this variance. Is that possible to obtain a 95% CI for the variance of the random effect instead of a p-value? - I would like to make predictions for different combinations of covariate values (but with an average value of 0 for the gaussian frailty term). I saw on the list that it is impossible to obtain a mean or a median survival time for each combinaison of covariates? Would that be possible to predict a probability of survival at the end of my follow up period, even if most observations are censored? How could I do that within the survival package? - Is there somewhere an explanation of the different possibilities of predictions offered by the predict.coxph function: lp, risk, expected, terms? Is that possible to use it with a model including frailties? In that case, should I include in the new dataset with observations to predict at, a column corresponding to the frailty term (containing the code of the group for which a prediction is desired)? Thank you so much for your help for solving some or all of these questions. Best regards, Basile Chaix French National Institute of Health and Medical Research [[alternative HTML version deleted]]
On Wed, 7 Sep 2005, Basile Chaix wrote:> Dear all, > I would have some questions on the coxph function for survival analysis, > which I use with frailty terms. > > My model is: > mdcox<-coxph(Surv(time,censor)~ gender + age + frailty(area, dist='gauss'), > data) > I have a very large proportion of censored observations. > > - If I understand correctly, the function mdcox$frail will return the random > effect estimated for each group on the scale of the linear predictor?Yes> - Similarly, the variance of the frailties is the variance of the terms on > the scale of the linear predictor?Yes> - A p-value is provided for this variance. Is that possible to obtain a 95% > CI for the variance of the random effect instead of a p-value?I don't think anyone knows how to do this. Personally, I'm not really convinced of the usefulness of these frailty models and I don't know how well their properties are known. I wouldn't use them except when I was actually interested in the variance components, and I haven't worked on any problems like that, so I haven't investigated the issue. [I don't write the survival package, I just port it] -thomas Thomas Lumley Assoc. Professor, Biostatistics tlumley at u.washington.edu University of Washington, Seattle