2
$\begingroup$

I have survival data on patients, coming from a clinic's datawarehouse. I want to do a survival analysis. The timeframe starts on the day a patient gets a certain examination, and ends 730 days (two years) later.

The data comes in with timestamps.

patient examination.date date.of.death date.of.last.contact comment
A 01.01.2000 NA 01.01.2010 Alive
B 01.02.2000 NA 01.03.2000 Lost to followup
C 01.03.2000 01.04.2000 01.04.2000 Died during study
D 01.04.2000 01.04.2010 01.04.2010 Died after study

So far, I have transformed the data to have the length of actual survival times, and an event column.

patient surv.time event comment
A NA FALSE Alive
B NA FALSE Lost to followup
C 31 TRUE Died during study
D 3653 FALSE Died after study

Surv seems to take no argument for a time of last contact, so I am quite certain that I should at least make a correction for patient B, who was lost to followup, and enter 28 there (the duration between examination and last contact). Patient C is also quite clear: died during the study, and the duration until death is entered.

But my question is about the other two patients:

  • should I enter 730 (the study duration) for patients like A, who lived beyond the end of the study?
  • should I enter 730 for patients like D, who died after the study timeframe? Note that I have already set the event column to FALSE for these cases.

I read some examples on how to use the survival package, but they used very simplified cases with already-prepared datasets.

$\endgroup$

1 Answer 1

2
$\begingroup$

Patient A should have time = 730 and event = FALSE. They were censored at the end of the study. EDIT: EdM makes a good point that this depends on whether Patient A was actually studied for a longer period of time, like Patient D apparently was. If A was followed after the original 730 days, then we have to make the same decision for A as we make for D below, either to censor at 730 days or at the time when we actually stopped observing them.

Patient D's data is very unusual. In typical studies there are three possibilities:

  1. The patient died during the study (event = TRUE, with event time).
  2. The patient was lost to follow-up (event = FALSE, at last observed time).
  3. The patient was alive at the end of the study (event = FALSE, last observed time is the end of the study).

It seems from your description that you kept observing patient D long after the "study" ended, which doesn't make sense. If the study is over, how are you still collecting data? But, given that you apparently have this data I think there are two ways to proceed.

  1. Enter the data as though the study really did continue until they died, time = 3653 and event = TRUE

  2. Enter the data as though they were censored at the end of the study, time = 730 and event = FALSE

If you're using a Cox proportional hazards model and patient D is the only one who is alive and uncensored after time 730 then the results will be identical for the two methods. If you're using a parametric regression or estimating survival curves then I think the results will depend on what you choose.

$\endgroup$
1
  • 1
    $\begingroup$ I agree in principle (+1), although the way to handle Patient A depends on the decision made about Patient D. It seems that Patient A was followed for 10 years. So if Patient D is handled according to your first suggestion, the censoring time for Patient A should be at 3653 days, through last follow up. If Patient D is handled according to your second suggestion, then censoring Patient A at 730 days makes sense. Also, for completeness, Patient B should have a censoring time at 29 days (if these are dd.mm.yyyy date formats with 2000 being a leap year). $\endgroup$
    – EdM
    May 6, 2021 at 17:30

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.