2
$\begingroup$

Background

I am still learning survival analysis, and try to understand why we cannot use linear or logistic regression on it. This question helped me a lot.

Question

When I am reading some tutorials another question pops in my mind: suppose we have a poor experiment design and the observation window is very short, and we have a lot of right censored cases, let's say it is 70% right censored vs 30% died in the observation window.

In such a case, would survival analysis still applicable and "preferred approach" comparing to logistic regression?

The reason I think logistic regression will do well is that, it is a nice binary classification problem. But in survival analysis, the response variable would have "low resolution" because of the right censored cases. And should predicting a contentious number is "harder" than predicting a binary label?

$\endgroup$
6
$\begingroup$

There is absolutely no advantage to using e.g. logistic regression in the case you describe, if you actually have the individual records with survival and censoring times. The type of situation you describe happens all the time in practice (even lower percentage of records with an observed are common) and this is not necessarily bad experimental design, but sometimes simply unavoidable (e.g. randomized cardiovascular outcome trials).

And there are a lot of downsides. If there are losses to follow-up prior to the end of the planned observation window (or it was a window of different length for different individuals as in event-driven clinical trials), then survival analysis is the appropriate way to account for these differences between records. Doing so is particularly critical, if some variables of interest in your model are also associated with different censoring patterns. When you use binomial data methods, you do not really have a way (or requires ad-hoc fixes that are at best approximately valid) to predict what would happen during observation periods of a different length.

You can actually think of logistic regression and other binomial data methods as a survival analysis that throws away all information about exactly when events happened and that also assumes that there is no censoring prior to the end of the planned observation window. When viewed this way, it is sort of clear that the first part primarily entails some loss of information (not so bad, if you have an exponential event time distribution with a really low rate, extremely bad if all records have an event), while the second assumption brings in a substantial risk of bias when it is not true (alternatively it is also enough to assume that all records need to have the exact same distribution of censoring times).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.