Discussion: View Thread

This message has been cross posted to the following eGroups: Biopharmaceutical Section and Statistical Consulting Section .
-------------------------------------------

Dear all,

    I just wonder if there is a general way to simualte time-to-event (survival) data.

    Currently, I generate survival data from exponential distribution then compare them with a max time. All the times greater than the max time would be considered as censored. But it seems still not mimic the reality close enough. For instance, it does not account for some subjects censored before the max time. Do you have any suggestion to improve the simulation? Or is there a general way to simualute survival data.

    Thank you.

-------------------------------------------
Shihua Wen
 -------------------------------------------

Here's a paper that discusses simulating time-to-event data, including with time-dependent covariates:

Sylvestre MP and Abrahamowicz M. Comparison of algorithms to generate event times conditional on time-dependent covariates. Statistics in Medicine 2008; 27 (14): 2618-2634.

-------------------------------------------
Robert Podolsky
Georgia Health Sciences University
-------------------------------------------

Hi Shihua,

I was just going through my literature and came across this paper about simulation and then saw your message about simulation survival data. This paper discussed the general issues for simulation and also has a section talking about how to simulate survival data in particular.

The design of simulation studies in medical statistics, STATISTICS IN MEDICINE
Statist. Med. 2006; 25:4279-4292


Hope it will help you.

 

-------------------------------------------
[Caiyan] [Li]

-------------------------------------------

 
Very helpful discussion. Thank you all.

-------------------------------------------
Shihua Wen
 -------------------------------------------

I would like to make one last comment about modeling the censor time distribution.  If the plan is to simulate a clinical trial with a time-to-event end point such as in an oncology trial  I do not think an exponential distribution would be appropriate for censoring.  Patients will more likely be censored when the time is small rather than large.  That is because patient drop out is more likely to occur early on (either because the patient thinks the treatment is ineffective or unpleasant).  Also patients enroll throughout the trial and those entering late would be more likely to be censored at the end of the trial than those who enter early.  So if a survival model is used for censoring it should have a decreasing hazard rather than a constant ( exponential) or increasing or bathtub rate (as is common with many survival models). 

-------------------------------------------
Michael Chernick
Director of Biostatistical Services
Lankenau Institute for Medical Research
-------------------------------------------







Original Message:
Sent: 01-11-2012 13:30
From: Shihua Wen
Subject: simualte time-to-event (survival) data

 
Very helpful discussion. Thank you all.

-------------------------------------------
Shihua Wen
 -------------------------------------------