Competing risks with multiple failures per subject not supported by stcrreg

[Ginny Natale]

I am trying to estimate time to onset of Alzheimer's with a competing risk of death. what are the hazards of onset of AD or death? I have multiple failures per subject.

agey = agey in years
idauniq = id variable
bagey = baseline age, origin
iadrd = failure Alzheimer's ==1
die = did respondent die

stset age, id(idauniq) failure(iadrd==1) o(bagey) exit(DIE==1)

id: idauniq
failure event: iadrd == 1
obs. time interval: (agey[_n-1], agey]
exit on or before: DIE==1
t for analysis: (time-origin)
origin: time bagey

------------------------------------------------------------------------------
138,510 total observations
18,296 observations end on or before enter()
33,373 observations begin on or after exit
------------------------------------------------------------------------------
86,841 observations remaining, representing
15,643 subjects
8,646 failures in multiple-failure-per-subject data
172,249.08 total analysis time at risk and under observation
at risk from t = 0
earliest observed entry t = 0
last observed exit t = 23


stcrreg bagey, compete(DIE==1)
data with multiple failures per subject not supported by stcrreg

What commands do i need to do competing risks of failure(iadrd==1 | DIE==1)?

[Nicola Carta]

Hi Ginny,
Have you found a solution to your problem? I am currently dealing with the same situation!

I need to account for multiple failures in the presence of competing risks.

Thank you!

[Tiago Pereira]

I am not sure if I understand the problem.
Have you tried stpm2?

https://journals.sagepub.com/doi/pdf...867X1301300209

[Barna Szabo]

I am quite sure this comes way too late for the original question, but someone may still find it useful.

What you are looking for is the Ghosh-Lin model (2002), which is an extension of the Cox proportional hazards model to accomodate competing/terminal events (similar to the Fine-Gray model), and also multiple failures per subject (as in the Lin-Wei-Yang-Ying model).

I have uploaded my implementation to github (the current version requires Stata 16.1 or newer). Run the following command in Stata to access it:

net from "https://szabobarnaki.github.io/ghoshlin"

The ghoshlin command expands the currently stset data (and copies it into a new frame), adding weighted pseudo-observations after terminal events to modify the risk set compensating for the informative censoring. After this dataset modification, using the Stata built-in stcox command will provide results for the Ghosh-Lin multiple-failure-terminal-events model. All postestimation commands of stcox work as well.

The ghoshlin command works both with unimputed and multiple imputed data.

You can also limit the maximal allowed number of failures per patient (to match the data collection procedures). The special case of max 1 event per subject coincides with the Fine-Gray model.

Comparing with stcrreg with single failure per subject, the coefficients obtained are almost identical, and the whole dataset preparation and estimation by stcox together takes much less time.

NOTE: This is my first published command. If you encounter any issues, or the help file is not clear enough, do not hesitate to ask me.