I intend to perform simulation based on a kaplan-meier estimation of survival on a dataset. I planned to use uniform simulation over [0,1] and the percentile function that i obtain from the kaplan-meier estimate.

Nevertheless, the fact that the kaplan-meier estimate S(t) is piecewise constant, and the use of the "classical" percentile function Q(p) = inf { t | S(t) ≤ 1-p } bring me to simulate only over a set of different time points. To add a more flexible approach, i would like to use weighted percentiles : with pj < p < pj+1, and pj, pj+1 well defined percentiles (i.e with S(Q(pj)) = pj),

Q*(p) = [(p-pj)/(pj+1-pj)]*Q(pj) + [(pj+1-p)/(pj+1-pj)]*Q(pj+1)

The fact is that it is equivalent to use a modification of the kaplan-meier estimate to construct our percentile function, where all the depression in the stairs representing this piecewise-constant estimator are linked by a straight line. So here is my question : does such an estimator exist in the litterature, and why don't we use it, as the hypothesis of constant survival seems in most practical case a bit far-fetched ?

thanks for reading

  • $\begingroup$ i will add latex formulation, i just need to understand how to include it and i have an urging task to do besides, during this time i've set this question back to the original shape so that it's a bit readible $\endgroup$
    – user310928
    Commented Feb 10, 2021 at 17:33

1 Answer 1


The Kaplan-Meier estimate makes jumps where there are observed event times.

Suppose there is an event on Day 150 and the next time of event is Day 200.
The K-M estimate of probability of surviving beyond Day 151 is the same as that of surviving beyond Day 199 because there is no data to tell me that they should be different. I did not observe anybody having an event in any of those days between 151 and 199. Maybe the truth is that no events can happen on those days for some scenarios.

If you really have a good feeling for what the curve should look like and there is not much data, then parametric survival analysis may work better in those scenarios. Another possible alternative is using piecewise constant hazard functions. Those will give you an estimated survival curve that is continuously decreasing.

  • $\begingroup$ The parametric modeling also overcomes a typical problem when trying to sample from Kaplan-Meier or Cox models: if the last data point in time isn't an event, then you can't sample from survival quantiles lower than the lowest reached survival fraction. (+1) $\endgroup$
    – EdM
    Commented Feb 11, 2021 at 17:25

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.