We would like to present a study reporting the results of an orthopaedic implant in patients with a follow up over ten years. There is no problem with that. But when it comes to the survival (meaning: the patient is still carrying the implant because it did not get loose nor infected) we have a methodological doubt: Should we include all patients in the Kaplan Meier estimation?
On one side we would say yes, since there might be some patients that did not reach the 10 years follow up because the implant failed before that limit. But, on the other side, there will be tons of patients with very short follow ups (3 months, 4 months, 5 months...) and we are afraid that introducing a lot of censored cases will make the estimate less accurate.
It helped me to better understand the Kaplan-Meier model and its limitations. Although I don't know how to do it in R yet, I guess that I should use a multiple decrement model when competing risk are present.
But this is not the case. Luckily, only one patient in the whole series had passed away for non-implant related reasons . The problem in our series is that we have around 100 patients over 10 years of follow up... but more than 300 with less than that. So, if we use the 400 in the survival analysis, we will be introducing tons of censored data (many of them with less than 12 months of follow up) and -as far as I understand- they will have a multiplicatory effect on the events that have actually occurred.
So we do not know what will reflect the reality with the most accuracy: to leave out of the analysis all the patients with less than 10 years follow up and not getting too many censored cases or to include everyone and overestimate the risk of failure.
Computes an estimate of a survival curve for censored data using either the Kaplan-Meier or the Fleming-Harrington method or computes the predicted survivor function. For competing risks data it computes the cumulative incidence curve. This calls the survival package’s survfit.formula function. Attributes of the event time variable are saved (label and units of measurement).
For competing risks the second argument for Surv should be the event state variable, and it should be a factor variable with the first factor level denoting right-censored observations.