Cox regression model convergence issues

[Pareesa Kassam]

Hi there!

I am fitting a Cox proportional hazards model to a subgroup of my data but I keep receiving the following error: flat region resulting in a missing likelihood

A snapshot of the structure of my data is below:

* Example generated by -dataex-. For more info, type help dataex
clear
input long participantid byte visit str8 subgroup float exposure byte outcome float(covariate time)
1 1 "group1" 0 0  .8414373    0
1 2 "group1" 0 0  .8414373   28
1 3 "group1" 0 0  .8414373   84
1 4 "group1" 0 0  .8414373  147
1 5 "group1" 0 0  .8414373  231
2 1 "group1" 1 0  .6503111    0
2 2 "group1" 1 0  .6503111   28
2 4 "group1" 1 0  .6503111  165
2 5 "group1" 1 0  .6503111  257
2 6 "group1" 1 0  .6503111  348
2 7 "group1" 1 0  .6503111  428
3 1 "group1" 0 0 .56734204    0
3 2 "group1" 1 0 .56734204  114
3 3 "group1" 1 0 .56734204  212
4 1 "group1" 0 0  .8987436    0
4 2 "group1" 1 0  .8987436   70
4 3 "group1" 1 0  .8987436   96
4 4 "group1" 1 0  .8987436  131
4 5 "group1" 1 0  .8987436  174
4 6 "group1" 1 0  .8987436  517
4 7 "group1" 1 0  .8987436 1237
5 1 "group1" 1 0  .3669614    0
5 2 "group1" 1 0  .3669614   35
5 3 "group1" 1 0  .3669614   63
5 4 "group1" 1 0  .3669614   91
5 5 "group1" 1 0  .3669614  117
5 6 "group1" 1 0  .3669614  222
5 7 "group1" 1 0  .3669614  336
end

My code for the model is:

stset time, failure(outcome==1) id(participantid) sc(30.4)

stcox exposure covariate if subgroup=="group2"

I have tried different methods to handle tied failures (efron, exactm, exactp), as well as increasing the number of iterations but I keep running into the same issue. Any advice of other options to try to help the model converge would be much appreciated.

Thank you!

[Paul Dickman]

Welcome to the forum! Thanks for reading the FAQ and presenting a data example.

Unfortunately, the example doesn't help much, because it contains no failures/events.

My guess is you have very few events. Is it possible to post a copy of your output from -stset- for the data to which you are hoping to fit the model? Output from -stcox- would also be useful. Later you might be asked to post the output with additional display options.

As an aside, you'll probably find the observations with time 0 are redundant (-stset- will set _st to 0 for those observations and you'll get a note in the log).

There are various maximization options (see the help). Using the -difficult- option would be a start. You can also specify different initial values (the -from()- option). For now, showing the output from -stset- will help.

Note to other forum users: Please don't hesitate to respond.

[Pareesa Kassam]

Hi there, thank you for your response! I apologize for my delayed reply.

Below is the output from stset:

Survival-time data settings

ID variable: participantid
Failure event: outcome==1
Observed time interval: (time[_n-1], time]
Exit on or before: failure
Time for analysis: time/30.4

--------------------------------------------------------------------------
1,140 total observations
250 observations end on or before enter()
60 observations begin on or after (first) failure
--------------------------------------------------------------------------
830 observations remaining, representing
197 subjects
4 failures in single-failure-per-subject data
1,860.559 total analysis time at risk and under observation
At risk from t = 0
Earliest observed entry t = 0
Last observed exit t = 40.69079

Unfortunately, there are very few events/failures but I am still hoping to document the hazard ratios nonetheless. Thank you for your help!