How to code event variable for survival analysis model to predict employee attrition

Question

I have a dataset consisting of 6000ish employees and am trying to predict rapid employee turnover (someone who quits within 180 days) with a survival analysis model. Roughly 3000 have quit within 180 days and the other half either have not or haven't been employed long enough to know at the time the data was pulled. My problem is figuring out how to deal with the event variable. In most cases I've seen, the even is usually binary (0 = still employed, 1 = censored), but my case is a bit different I assume. I have to account for people in the data who have yet to quit but who haven't worked 180 days yet, those who have quit within 180 days, and those who quit after 180 days.

What I'd like to know is whether my event should be like this:

0: still employed / quit, but did not quit before 180 days

1: quit within 6 months at the time the data was pulled (event occurred)

2: employed for fewer than 180 days at the time the data was pulled, but is but censored because we don't know if they will quit within 180 days.

or like this:

0: still employed / quit, but did not quit before 180 days

1 & 2 combined as same variable

Or should I code the event some other way?

I apologize if this question seems trivial but I've searched other articles on here such as the one below, but cannot seem to find a suitable answer. Any help would be greatly appreciated!

Survival analysis: How to handle data where I know the event occurred, but I don't know when it occurred?

Answer 1

I have to account for people in the data who have yet to quit but who haven't worked 180 days yet, those who have quit within 180 days, and those who quit after 180 days.

If you know the actual time-to-quitting for those who quit within 180 days and the employment length to date of those who worked less than 180 days but haven't yet quit, then there should be no problem with a simple 0/1 coding of non-quit/quit. You would have two different types of right censoring, some cases with an "administrative censoring" at the 180 days (if that's your longest time of interest) and others censored due to follow up for less than 180 days.

That combination of types of right-censoring is fine provided that censoring isn't informative. Standard survival analysis then uses information from each individual so long as the individual has information to provide. If you don't care about times after 180 days, then you don't need to incorporate further information from individuals who did last 180 days.

If you don't know the actual quitting or last-observation times for those whose observation times didn't reach 180 days, then you have interval censoring and should proceed in a way that can deal with that type of censoring. The icenReg package cited in the thread to which you link is one tool for such analysis. To code the event/censoring times you would need to follow the instructions for whichever tool you chose.