dandar
Reputation
Top tag
Next privilege 125 Rep.
Vote down
1 11
Impact
~3k people reached

• 0 posts edited

# 69 Actions

 Mar 22 awarded Tumbleweed Mar 15 asked Missing survival probability estimates and times using R using an Anderson-Gill model for recurrent events Mar 15 comment Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model To follow on from above: we might fit dynamic covariates to help understand the true nature of the dependence of the outcomes on the outcome history, but for understanding the effect of a non-dynamic covariate (say the effect of treatment in a clinical trial) we would not want the non-dynamic covariate effect underestimated, nor would we want to risk introducing bias into the estimate, and so we would fit an AG model without dynamic covariates - strictly not an AG model but a "rate" model as proposed by Lin et al 2000. Mar 15 comment Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model In terms of your general comments @Theodor on the AG model, I am not saying you are wrong since I am no expert, but I believe time-varying covariates can be used which depend on the outcome history since this ensures conditional independence of all the recurrent events within a subject. I believe this is in the spirit of the original model proposed by Andersen and Gill, 1982. Even though there are reasons why we would not wish to include these in the model (i.e. they "steal strength" from non-dynamic covariates, and in designed experiments they are post-randmoisation) they are "allowed". Mar 15 comment Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model Even though @Theodor your equation $S_i(t|s) = \exp \left( -\int_s^t \lambda_0(u) \exp(\beta'x_i(u)) du \right)$ does not use the product form as mine does in the question (not important) is your equation not similar in principle to equation (1) - i.e. path of the covariate vector through time from $T=0$ to $T=t$ is not taken into account? Does this makes sense? Should I be using something like my equation (2)? Furthermore I see your R code call to "survfit" does not use the "id" statment. Is it true that using the "id" statement gives equation (2) and not using it gives (1) or your equation? Mar 15 revised Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model Updating R code due to errors in previous code Mar 15 revised Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model Clarifying how the survival probabilities should be calculated Mar 14 comment Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model Thanks @Theodor for your answer - I really appreciate this but will need some time to digest your comments. For now I have updated my original question to clarify the recurrent event model I am using. I will get back to you ASAP. Mar 14 revised Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model Explained in more detail the model I am using Mar 12 comment Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model Yes where I wish to calculate $S[t|x(t)]:=P[T>t|x(t)]$ which is the probability of the time of the next event being greater than $t$. Thus as $t$ varies the covariate vector also varies, but nonetheless I we should still end up with a "survival" curve that decreases towards zero as $t$ increases - just like we do when we have covariates that do not vary with time? Mar 11 comment Using R to calculate survival probabilities with time-varying covariates using an Andersen-Gill model Thanks for your comment @Theodor, I have no terminal event other than censoring, and yes as you describe I have recurrent events at t1, t2, t3, .. for a subject. What I wish to do is to calculate the survival probability at time $T=t$ conditional on the time varying covariates. So say \$t2< t