4
$\begingroup$

My application is not a traditional survival analysis scenario. However, I believe survival analysis methods, e.g., Cox regression, can be a possible solution. In particular, my dataset contains two kinds of data:

  1. Left-censored: For example, I know a patient has been dead at time t8. But the exact time point that the patient died is unknown. Also, in the nature of the problem, it is impossible to know.
  2. Right-censored: Similar in the traditional case. For example, I know a patient was alive from t0-t6. But it is unknown when the patient died after t6.

In both case, the exact "death point" is unknown. My data only contains these two cases.

I plan to use widely used survival analysis packages (e.g. lifelines) to solve this problem:

  1. Left-censored: For the example above, I will label it as "duration: t8, event:1". (note that t8 is NOT the exact time that the patient dies)
  2. Right-censored: For the example above, I will label it as "duration: t6, event 0".

My questions:

  1. Does using Cox regression make any sense?
  2. How does these two types of data impact the final model? Say, under-estimate/over-estimate the baseline hazard?
  3. Are these any other models can better handle this case, instead of Cox regression?
Improve this question
$\endgroup$

2 Answers 2

Reset to default
1
$\begingroup$

For left censored data, they are excluded from follow-up at the point of censoring and coded as a non-event. That is in fact the definition of censoring. Also, you need to know what "time 0" is. So if you don't understand that, you can't include right censored data. You will get different results if the time under observation goes from 0 to 6 as compared to 12 to 18. The last consideration is that you can't just sample observations on the basis of being failures. It will inflate the incidence and bias the hazard ratios. Cox regression works, but I think you need a clearer understanding of how censoring is encountered. A practical description of the data is warranted.

Improve this answer
$\endgroup$
1
$\begingroup$

EDIT: see AdamO's answer as well.

unfortunately that solution won't work. One can't encode left-censorship and right censorship together like that. Lifelines currently (as of v0.14.1) doesn't have an interval censorship model, so you'll have to use something like the solution here: https://stats.stackexchange.com/a/198748/11867

Improve this answer
$\endgroup$
3
  • $\begingroup$ @Cam.Davidson.Pilson Do you think the Cox PH in lifelines can be modified in a stochastic way? Feed a mini-batch and set max iteration = 1, get the gradients and update betas iteratively. Do you think it is possible? $\endgroup$
    – Munichong
    May 10, 2018 at 15:07
  • $\begingroup$ That is not quite correct. You can in fact combine left and right censoring. $\endgroup$
    – AdamO
    Feb 25, 2019 at 16:09
  • $\begingroup$ @AdamO Yes I agree with you now. I was too hasty. $\endgroup$
    – Cam.Davidson.Pilon
    Feb 25, 2019 at 17:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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