Median Overall Survival and Right Censoring


I’m currently working on a retrospective analysis of survival from a very rare cancer (N = 23) and I have a question about what is the most appropriate way calculating the median overall survival in a retrospective study.

In my data, some patients had complete response to treatment and never recurred and consequently their follow-up is longer (~ 10 years). My question is if I should right-censor these patients on a estipulated follow-up maximum time (example, 5 years) or if I should use the complete time of follow-up to calculate the median overall survival. This problem does not happen in clinical trials because the end of follow-up will be the end of the study and patients will have similar time. However, for observational data, I have not found any discussion on that.

In my project we are not investigating correlates of mortality but describing outcomes and survival of patients with this specific rare cancer.

Thanks in advance,


i think you should use complete follow up which in any case would be right censored. maybe you want to do some influence analysis though?

N=23 is probably not sufficient for estimating the median from the highest information variable one could measure (a continuous variable with little measurement error and no censoring). For small samples one typically has to make some assumptions (e.g., exponential or Weibull distribution), compute the median from the maximum likelihood estimates of that distribution’s parameters, and hope for the best.

1 Like

Thank you Paul and Frank.

Paul, my initial approach was using the complete follow-up as you mentioned. Theoretically, the complete follow-up would give me the answer of how patients with this disease survive in the long-term. Some doubts that arise to me is the fact that cancer is theoretically curable and therefore, after fiver (or some more) years you wouldn’t expect the survival to be related to the disease anymore. That would not be true for chronic diseases as diabetes, congestive heart failure, etc. That’s the reason why I suggested ending follow-up (right censoring) in five years. However, I haven’t found a discussion about this and in routine clinical practice, oncologists sometimes present to the patients estimates based on trials (which are medium term - 3-5yrs).

Frank, this is true. The estimate (calculated with Kaplan Meyer) is very variable in our case (low N) - adding two patients with 120 months of OS and alive, changes the median from 28 to 62 months. And that’s one of the reasons for this question. However, I never used any of the models you commented and don’t know which one should I use. It’s a cancer, aggressive, but diagnosed early stages so with an not low but not high expected survival. Which one you suggest to use?

1 Like

there is a literature on cure models although i don’t know it well eg: The N is an issue though as Frank says