I am having issues including "season" as a predictor variable in my survival analysis using cox proportional hazard models. I work with sage-grouse and have reason to believe their survival varies by season. I provided a sample data set below.


BirdID <- c(1,1,1,1,2,2,2,2,3,3,3)
Event <- c(0,0,0,0,0,0,0,1,0,0,1)
Start <- c(0,13,26,39,0,13,26,39,0,13,26)
Stop <- c(13,26,39,52,13,26,39,48,13,26,35)
Avg.Move <- c(2,5,6,9,15,9,16,25,1,3,18)
Season <- c("spring", "summer", "fall", "winter", "spring", "summer", 
"fall", "winter", "spring", "summer", "fall")

C <- cbind(BirdID, Start, Stop, Event, Avg.Move, Season)
T <- as.data.frame(C)
T$Start <- as.numeric(as.character(T$Start))
T$Stop <- as.numeric(as.character(T$Stop))
T$Event <- as.numeric(as.character(T$Event))
T$Avg.Move <- as.numeric(as.character(T$Avg.Move))

I have formatted my data so that each row is a different season because I have time-dependent variables (e.g. Avg.Move = average distance moved per season). I'm guessing that this may be part of my problem. Anyway, when I try to run basic coxph models with different combinations of my predictors (Season and Avg.Move) I continually get errors.

L <- coxph(Surv(Start, Stop, Event) ~ Season, data=T)
## Error in fitter(X, Y, strats, offset, init, control, weights = weights, : 
routine failed due to numeric overflow.This should never happen.  Please 
contact the author.

X <- coxph(Surv(Start, Stop, Event) ~ Avg.Move+Season, data=T)
## Warning messages:
1: In fitter(X, Y, strats, offset, init, control, weights = weights,  :
Loglik converged before variable  1,3,4 ; beta may be infinite. 
2: In coxph(Surv(Start, Stop, Event) ~ Avg.Move + Season, data = T) :
X matrix deemed to be singular; variable 2

I've tried coding Season as numeric (1,2,3,4) and binary (10000, 0100, 0010, 0001) but nothing has helped.

Thanks in advance,


  • $\begingroup$ Hello, I have the same problem. did you ever figure this out? $\endgroup$
    – nasim
    Jul 21, 2017 at 16:57
  • $\begingroup$ I guess you could try this: 1) table(T$season)to is if it is singular; 2) sample size is only 11 obs, too small, try to get more obs and run it again. $\endgroup$
    – Joey Zhou
    Sep 19, 2018 at 9:42

2 Answers 2


You only have three individuals and two events. Thus, you only have two partial likelihood terms on the log-likelihood scale. Thus, you are very limited by your data. In particular, you cannot estimate a dummy variable with four-levels as DJA mention's as you will have at-least two levels without any events. Thus, you can set the coefficient for these levels to minus infinity and get a hazard of zero (which is what you observe in the data but likely not what you expect).

The code below stress that you cannot estimate much from the data

da <- data.frame(
  BirdID   = c(1,1,1,1,2,2,2,2,3,3,3),
  Event    = c(0,0,0,0,0,0,0,1,0,0,1),
  Start    = c(0,13,26,39,0,13,26,39,0,13,26),
  Stop     = c(13,26,39,52,13,26,39,48,13,26,35),
  Avg.Move = c(2,5,6,9,15,9,16,25,1,3,18),
  Season   = c("spring", "summer", "fall", "winter", "spring", "summer", 
               "fall", "winter", "spring", "summer", "fall"))

# you only have three birds and two events
#R [1] 3
#R [1] 2

# you have hard time getting a good estimate of the non-parametric baseline
fit <- coxph(Surv(Start, Stop, Event) ~ 1 + cluster(BirdID), data = da)
plot(survfit(fit)) # plot survival curve

enter image description here

# You are still limited if you assume a constant intercept
fit <- survreg(Surv(Stop - Start, Event) ~ 1, dist = "exponential", data = da)
#R Call:
#R survreg(formula = Surv(Stop - Start, Event) ~ 1, data = da, dist = "exponential")
#R             Value Std. Error    z       p
#R (Intercept) 4.212      0.707 5.96 2.6e-09
#R Scale fixed at 1 
#R Exponential distribution
#R Loglik(model)= -10.4   Loglik(intercept only)= -10.4
#R Number of Newton-Raphson Iterations: 5 
#R n= 11 

I believe the reason you are having an issue is because 2/4 levels of your factor "Season" have no events. Only winter and fall have a single event occur during those seasons. All other seasons haven 0 events occur. That is why you are not getting convergence.

See: Cliff AB (https://stats.stackexchange.com/users/76981/cliff-ab), Dealing with no events in one treatment group - survival analysis, URL (version: 2015-05-19): https://stats.stackexchange.com/q/153070


Your Answer

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

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