[R] bootstrap in time dependent Cox model?

[Ehsan Karim]

Thanks Prof. Terry for the response.

To answer the second comment, the weights I am considering are inverse
probability of treatment weights (IPTW). To perform bootstrapping, my
initial thought would be to
- select patients (sampling from unique id with replacement) and
- then include all the multiple observations of the selected patients
(include all rows with the selected id) to make a bootstrap sample.
- Then create new weights based on this bootstrap sample (I mean,
based on covariate history of the selected patients) and
- then run the stated weighted Cox model with this bootstrap sample.
This should give me point estimate (of treatment effect or HR) from
one sample. Similarly I would repeat for all bootstrap samples. My
question is: does this bootstrap strategy sound reasonable? I am
guessing boot() can be used here once the above is put in a function.

Any suggestions/references will be highly appreciated.

cheers,

Ehsan

On Tue, Feb 21, 2012 at 11:11, Ehsan Karim <wildscop at hotmail.com> wrote:
>
>
>> Subject: Re: bootstrap in time dependent Cox model?
>> From: therneau at mayo.edu
>> To: wildscop at hotmail.com
>> CC: r-help at r-project.org
>> Date: Tue, 21 Feb 2012 07:06:16 -0600
>>
>> Two comments.
>> I've not found bootstrapping worthwhile for Cox models. If one has
>> 10-20 events per covariate and no outrageous coefficients (risks of >10
>> fold) the standard asymptotics are very good. With an multiple events
>> AG model there is the additional consideration that no one subject is
>> responsible for too many of the events. (Years ago I was working on a
>> study of recurrent syncope in Long QT syndrome. Most subjects had 0
>> events, some 1, and one gentleman 29. The multi-event analysis could be
>> summarized as "don't look like him.")
>>
>> I like to think of bootstrapping in two stages. First, the "best"
>> answer which is what I would get if I had enough money to repeat the
>> study 100 times. Then set up the bootstrap to resemble that. In your
>> case, what would the pattern of weights be in the hypothetical repeat?
>> For instance, if your weights are sampling weights from a structured
>> design, the repeat might have the same set of weights and creating a
>> bootstrap strategy to replicate it will be more challenging.
>>
>> Terry T.
>>
>> ---- begin included message ----
>>
>> I am wondering how to perform a bootstrap in R for the weighted time
>> dependent Cox model? (Andersen?Gill format, with multiple observations
>> from each patients) to obtain the bootstrap standard error of the
>> treatment effect.
>>
>>