Background and Theory

This page aims to quickly cover some basic survival analysis.


This page is under construction.

Survival Analysis Setup

Suppose we are interested in a particular time-to-event distribution (e.g., time until death, time until disease occurrence or recovery, or time until failure in a mechanical system). The phrase “time-to-event” can mean nearly any positive quantity being measured, not necessarily time.

The time-to-event distribution is completely determined by its survival function \(S(t) = \Pr(X > t)\), where \(X\) is the time-to-event (a positive random variable), and \(t\) is a positive time. The survival function might be of interest itself since it answers questions like, “what is the probability that a cancer patient will survive at least five more years?” or “what is the probability that this machine part won’t break in the next six months?”

Suppose we have \(n\) individuals with independent and identically distributed event \(X_1, \ldots, X_n\) with survival function \(S\). If we could observe \(X_1, \ldots, X_n\), then one obvious candidate for estimating \(S(t)\) is the empirical survival function:

\[\widehat{S}(t) = \frac{1}{n} \sum_{i=1}^n \mathbf{1}_{\{X_i > t\}}.\]

This is just the number of individuals observed to survive past time \(t\) divided by the sample size \(n\). (Here \(\mathbf{1}_A\) denotes the indicator of an event \(A\).) This unbiased, consistent, and asymptotically normal estimator is nevertheless not suitable for many situations of interest in survival analysis. The main problem is that it relies on all the event times \(X_1,\ldots,X_n\) being observable.

Censoring and Truncation

It is common in practice that not all the times \(X_1, \ldots, X_n\) are observed. We list some possible examples.

  • A clinical trial investigating a disease treatment might end before a patient recovers from the disease, in which case the patient’s true recovery time is not known.
  • A patient whose time until death from cancer is being monitored might die from a different disease, so the patient’s death-from-cancer time will not be known.
  • In engineering, a reliability experiment might be stopped after a predetermined number of parts fail. A part that is still operational after this time will not have its failure time observed.

In these cases, the true time-to-event is not known, but a lower bound for it is available. This situation is called right-censoring.

Another source of incomplete information is left-truncation (also known as delayed entry), in which an individual’s time-to-event may only be observed if it exceeds a certain time. Some examples:

  • Patients for a certain disease might only be observed after the disease has been diagnosed. A patient who died from the disease before diagnosis is unknown to investigators.
  • If we are observing the age at death of residents in a retirement home, then we do not observe the ages at death of individuals who died before becoming residents.
  • If we are measuring the diameters of particles with a microscope, then only particles large enough to be detected by the microscope will be observed.

There are other types of censoring and truncation, but for now we will focus on right-censoring and left-truncation, the most common variants.

When there is right-censoring or left-truncation, the empirical survival function is not a suitable estimator of the survival function \(S\) because the event times \(X_1, \ldots, X_n\) are not known. A popular alternative is the Kaplan-Meier estimator, which we will define below.

Let us first formalize the sampling situation with right-censoring and left-truncation. Suppose as before that we have \(n\) individuals with with independent and identically distributed times-to-event \(X_1, \ldots, X_n\) with survival function \(S\). Moreover, suppose we observe a sample \((L_1, X_1^\prime, \delta_1), \ldots, (L_n, X_n^\prime, \delta_n)\), where \(L_i\) is the entry time of the \(i\)-th individual, \(X_i^\prime\) is the observed final time for the \(i\)-th individual (so \(X_i^\prime > L_i\)), and \(\delta_i\) is zero or one according to whether \(X_i^\prime < X_i\) (the \(i\)-th individual is censored) or \(X_i^\prime = X_i\) (the \(i\)-th individual’s time-to-event is observed).

The Counting Process Formulation

We define two stochastic processes associated with this sample: the event counting process

\[N(t) = \sum_{i=1}^n \mathbf{1}_{\{X_i^\prime \leq t, \delta_i = 1\}},\]

and the at-risk process

\[Y(t) = \sum_{i=1}^n \mathbf{1}_{\{L_i < t \leq X_i^\prime\}}.\]

The event counting process counts the number of events that have been observed to occur up to a certain time, and the at-risk process counts the number of individuals who are “at risk” (those who are already being observed but have not yet been censored or experienced the event of interest) at a certain time.

The trajectories (i.e., sample paths) of the event counting process are right-continuous with left limits (commonly called càdlàg). In fact, the event counting process is non-decreasing and piecewise constant with at most \(n\) jumps (the jumps happen at observed event times). Let \(0 < T_1 < T_2 < \cdots\) denote the ordered jump times. The size of the jump at time \(T_j\) is

\[\Delta N(T_j) = N(T_j) - N(T_j-),\]

where for \(t > 0\),

\[N(t-) = \lim_{s \uparrow t} N(s)\]

is a limit from the left. In other words, \(\Delta N(T_j)\) is the number of observed events at time \(T_j\). For convenience, also define \(T_0 = 0\).

Stochastic integrals with respect to \(N\) can be easily computed as sums:

\[\int_0^t H(s) \, dN(s) = \sum_{j : T_j \leq t} H(T_j) \Delta N(T_j).\]

The Kaplan-Meier Estimator

We will now give a derivation of the Kaplan-Meier survival function estimator. This will be an estimator of the survival function \(S(t) = \Pr(X > t)\), where \(X\) is the time-to-event, based on the right-censored and left-truncated sample \((L_1, X_1^\prime, \delta_1), \ldots, (L_n, X_n^\prime, \delta_n)\) described above.

First, observe that if we have times \(s < t\), then

\[\begin{split}S(t) &= \Pr(X > t) \\ &= \Pr(X > s) \Pr(X > t \mid X > s) \\ &= S(s) \Pr(X > t \mid X > s).\end{split}\]

Thus, it suffices to estimate the conditional probability \(\lambda(s, t) = \Pr(X > t \mid X > s)\) of surviving the time interval \((s, t]\) given survival up to time \(s\).


Finish this section.