Dear Sameer,
As I have understand, you want to do following:
1. Generate true survival time and censored time.
2. Then to identify some individuals lost to follow-up.
3. Finally calculating bias.
According to your algorithm, you are taking the small survival time as
min(event_times, censored_times). In this situation your estimate of the
parameter is highly affected by the parameter of the censoring
mechanism.There are two independent processes generating survival time but
you are estimating one(right?). If you take lamda_e=.5 and lamda_c=2 then
you will get overestimated value and if u take lambda_c=0.1 then you will
get underestimated value of the parameter.
In this situation I would prefer only consider right censoring mechanism .
Just to generate all data by giving one parameter and then choose a suitable
end point for follow-up. Those survival time who exceed the cut point would
be considered as censored observation as real life situation. In this
situation bias estimation is more logical.
For considering lost to follow-up you could use a percentage and generate
dummy variable and then those who gets 1 will be considered as censored.
This is my view. Please let me know yours. I will be happy to discuss with
you about it.
Regards,
Kishor Kumar Das
On Wed, Sep 28, 2011 at 8:22 AM, Sameer Nizarali Parpia <[log in to unmask]>wrote:
> Hi AllStaters
>
> I need some help on a survival simulation I am conducting.
>
> I generate survival data using an exponential distribution for event and
> censoring times
>
> event_times=exp(lamda_e)
> censored_times=exp(lambda_c)
>
> Then the minimum between the two is used and event or censoring is
> dependent on which is the minimum. This is considered the truth (similar to
> one group in a trial of time to event outcome)
>
> I then generate another variable for loss to follow-up times using
> lost_times=exp(lamba_l).
>
> I then use imputaion methods to predict whether the lost time was either a
> event or censored time. Once I have completed this, I would like to compare
> this predicted survival function to the "truth" and I am wondering what
> would be the best measure of bias??? Quantiles is not possible because the
> truth data may not have exact 0.25, 0.5 and 0.75 times.
>
> Any help wiill be appreciated. Thanks
>
> sameer
>
> You may leave the list at any time by sending the command
>
> SIGNOFF allstat
>
> to [log in to unmask], leaving the subject line blank.
>
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|