Kaplan Meier Survival Analysis - Censoring

[simon metin]

Hey guys and girls,

i am doing a project in a hospital and am looking at the survival of renal cancer patients on a particular treatment. Now, i am going to need to do the kaplan meier analysis to get some survival curves and this is my first time doing this. i have separated all my data into what i need them to be and have calculated the days each person survived (or until they were last seen in clinic if they were still alive). i am slightly confused about the process of censoring my data. do i censor the ones who are still alive? if so, do i ascribe them a number 0 or a 1? i know this is a fundamental part of the analysis but i have never done stats before in my life. i dont want to get this essential component wrong!!

best,

simon

[Clyde Schechter]

At the conceptual level, you have a last follow-up time for each person in your data set. If that time represents their time of death, it is an uncensored observation. If that variable represents the last point at which they were known to be alive, but weren't dead yet, then you have a censored survival time. That is, you know that they survived at least up to that point, but you don't know when, after that point, they died or will die. That is what censored means.

As for how you code death vs censoring, you can do it pretty much any way you like. You need to read the manual section on the -stset- command, and you will see that it is quite flexible about the ways in which it lets you represent death (failure) vs censoring. You just have to specify the -failure()- option to -stset- in a way that accurately reflects how you code it in your data. Do read also the -sts graph- command, which is how you will get your K-M plots.

[Roman Mostazir]

In addition to Clyde's valuable advise, a little more to add if this is your first time and never have taken any stat analysis. Think about the analysis time for risk exposure. Did all patients start to be in risk at the same time? do you have delayed entry? Is it only KM curves you want? KM curves are un-adjusted. Do you think you have covariates that may affect the survival? In that case, you need to choose a model based on your data. Cox-regression, exponential, Weibull, Gompertz are the options for adjusted curves. Also how time is recorded is an important one. If they are discrete you need to choose a discrete time-to-event model otherwise, continuous. As per Clyde's advise read the -st-manual first and -sts graph- . If you need to fit a model, then you also will need -sts curve- for adjusted curves.