The R Book

Models with Censoring

Censoring occurs when we do not know the time of death for all of the individuals. This comes about principally because some individuals outlive the experiment, while others leave the experiment before they die. We know when we last saw them alive, but we have no way of knowing their age at death. These individuals contribute something to our knowledge of the survivor function, but nothing to our knowledge of the age at death. Another reason for censoring occurs when individuals are lost from the study: they may be killed in accidents, they may emigrate, or they may lose their identity tags.

In general, then, our survival data may be a mixture of times at death and times after which we have no more information on the individual. We deal with this by setting up an extra vector called the censoring indicator to distinguish between the two kinds of numbers. If a time really is a time to death, then the censoring indicator takes the value 1. If a time is just the last time we saw an individual alive, then the censoring indicator is set to 0. Thus, if we had the time data T and censoring indicator W on seven individuals,

T  4  7  8  8  12  15  22
W  1  1  0  1  1  0  1

this would mean that five of the times were times at death while in two cases, one at time 8 and another at time 15, individuals were seen alive but never seen again.

With repeated sampling in survivorship studies, it is usual for the degree of censoring to decline as the study progresses. Early on, many of ...