Probable error stset

[Francisco Pelegrin]

Hi,

I'm trying to perform a survival analysis on relapse-free survival after rectal cancer surgery. When I stset the data, the message is as follows:

gen TSLR = int((Frecaigud - Fcir )/30.4375)
(51 missing values generated)

stset TSLR, failure(RecaidaSiNo == 1)
failure event: RecaidaSiNo == 1
obs. time interval: (0, TSLR]
exit on or before: failure
------------------------------------------------------------------------------
63 total observations
51 event time missing (TSLR>=.) PROBABLE ERROR
1 observation ends on or before enter()
------------------------------------------------------------------------------
11 observations remaining, representing
10 failures in single-record/single-failure data
116 total analysis time at risk and under observation
at risk from t = 0
earliest observed entry t = 0
last observed exit t = 22

I've read previous questions on the matter, but I cannot solve the problem on my dataset. I'd appreciate if anyone could help me to find the error.

Thank you,

[Clyde Schechter]

Well, it isn't clear whether there is really an error or not. But it does seem clear that your data, as they stand, are not suitable for the analysis you want to do.

Notice that Stata says "PROBABLE ERROR." It is not necessarily wrong to have missing values for the time variable: sometimes that happens. But it is unusual, because typically when the time of failure is unkown, we can instead use a time when last observed and consider the observation censored. Anyway, when the time variable is missing, the observation cannot contribute to the survival analysis, so it is excluded. Since you only have 63 observations to begin with, when you lose 51 you are left with only 12, which, frankly, isn't enough to get meaningful results. On top of that, you lose another one more observation where, it seems, the calculated failure time is 0 or negative (also not allowed).

So the question is, why do you have so many observations with missing values for TSLR. Given how it was computed this means that Frecaigud or Fcir (or both) is missing in these observations. So you need to check why so many observations have missing values on these variables. I have no idea what these variables represent--their names don't convey any particular meaning to me. But the fact that your analysis time is calculated as their difference normalized by 30.4375 suggests to me that perhaps Frecaigud is the date of death/recurrence, and Fcir is the date the person entered the study, and the normalization is to convert the unit of time from days to months.

If I have correctly guessed what these variables are, I'm guessing that the missing variable is most often Frecaigud. Now, if it is missing because you never saw the patient again after they entered the study and you have no information about what happened to him/her, then your data are correct but simply not usable for the study--you don't have enough information to do anything meaningful.

But I wonder if perhaps the missing values of Frecaigud are for patients who are still alive and relapse free: they haven't yet "achieved" the failure event, so we don't know when, if ever, they will. It is not correct to represent those with missing values in the analysis time variable. Instead, this variable should contain the date at which the patient was last evaluated and found to still be alive and relapse-free. And, in these observations, RecaidaSiNo must be set to 0. This is how Stata will recognize these as censored observations and will include them in the analysis.

As for the one observation where Frecaigud precedes Fcir, that is clearly an error: if the patient already died or relapsed, then that patient should not be in the study. More likely is that one of those two variables is incorrectly entered into the data set.

[Francisco Pelegrin]

Thank you Clyde. It is exactly as you described. Problem solved.