# [R] Understanding of survfit.formula output

Terry Therneau therneau at mayo.edu
Fri May 7 15:21:40 CEST 2010

You have not given enough information to reproduce your problem, so it
is difficult to say.  Most of the time results such as you display will
be due to a data error.  It is possible to get a crossing result,
however.
subject     1  2   3   4  5
-----------------------------
Death      10  -   -   -  4
Relapse     2  -   4   -  3
Last FU    10  6   6   11 4

The Kaplan-Meier curves are
time   relapse                 death
2      4/5                     1
3      4/5*3/4                 1
4      4/5* 3/4 * 2/3=.4      2/3
10     .4                     2/3 * 1/2 = .3333

Subject 1 has a relapse early when all 5 are still at risk, dropping the
curve by .2 units.  Their death is late when only 2 are at risk,
dropping the curve by 1/2.

This crossing anomaly usually only happens near the end of a
Kaplan-Meier, when the confidence intervals are as wide as a river.
Terry Therneau

------------ begin included message --------------------------

At year 5, in group C2 I have one more patient with an event when
looking at
DFS (13) than when looking at relapse (12). However, the probability is
higher when looking at DFS (0.23) than relapse (0.18), which I cannot
understand as I have one more event.