Bootstrapping instable for iv_2SRIF

[jguski]

Hello again!
Thank you very much for closing the two issues so quickly, this is very helpful.
I've actually just run into another obstacle related to bootstrapping, a very obscure one this time:

If the event rate is very low in the data (~1 % or similar), it may happen that the bootstrapped data sample does not contain any observation for which the event occurred, at least in one of the groups. This does not seem to be a problem for some of the estimators (e.g., AIPTW seems to work well), but others crash (I observed the problem for iv_2SRIF). The error I get for bootstrap iteration X is: task X failed - "need at least two non-NA values to interpolate".

Maybe it would be possible to enable bootstrap sampling stratified by event as a new feature, or fix this problem further downstream?

[RobinDenz1]

Thanks again for your input. This is not an bug per se. It does make sense that the function throws an error if there are no events (I am actually suprised that the AIPTW method doesn't). Stratified bootstrapping sounds promising. I don't know a lot about the statistical implications though. I am currently focused on another package called simDAG and have other responsibilities that I have to attend to, so I don't really have the time to look into this right now.

If you have the time and knowledge I am lacking, you are highly welcome to file a pull request with the desired changes, though. Also, I am curious: Are you performing some sort of simulation study? Or are you analyzing empirical data?

[jguski]

Hello,
Thanks for your reply. It would of course be no problem to change the code and file a pull request. However, I am also not sure at this point if a stratified bootstrap would be statistically sound. I will try to find out and see if I can share some sources with you.

I am working on empirical data (real world clinical data, to be more precise). However, the outcome is a rare event in general (Alzheimer's disease), and especially one of my groups has a very limited follow-up horizon.

[jguski]

So I did some research and came across the boot library, which implements the methods from this book and seems to be the go-to library for bootstrapping in R. They actually have a method censboot, which deals specifically with censored data. Stratified bootstrapping does not seem to be a problem, because it is supported by the method and even done in the examples, and I cannot find any discussion of potential problematic implications.

However, they provide different "simulation types" there. As far as I understand, what is currently done in adjustedCurves is their option "ordinary", which is just straightforward sampling with replacement from the data. For this method, we have to assume random censoring though, which is not realistic for many application (including mine). With non-random censoring, we may end up with bootstrap samples with censoring and survivor distributions that are very different from the original data. Thus, there are also methods available there which do not only resample observations from the original data, but also use models to generate new failure and censoring times for the bootstrapped samples.

So there would be two possibilities: Either I file a pull request in which I only change the code to enable stratified sampling with replacement, or I try to integrate the censboot method. However, this will probably be more challenging and take more time.

Best
Jannis

[RobinDenz1]

Thank you for looking into this. I am still a little on the fence about this though. Do these different bootstrapping methods require different analysis or pooling strategies? I know about the boot package, but had decided not to use it because of the added complexity. I would prefer to keep it that way.

Maybe you can fork this repository and just do what works best for you in that forked version first. I would like to put some thought into this before merging a pull request.

[jguski]

Okay, that sounds like a good idea. It would be great if you could keep me posted if you decide to change or not to change anything in your code in this regard.

[RobinDenz1]

Sure, I will keep you updated! Will leave this issue open till then.