Residual Sojourn Time After State Survival

Why survival to a landmark changes the waiting-time question

Suppose a person enters a clinical state, such as living with an illness, and is still in that state at a later landmark time. The question “how much longer will this person remain here?” is not generally answered by the original distribution of time spent in the state. The person has already spent time there without exiting. That survival is information about the sojourn, and the remaining-time distribution must be conditioned on it.

The distinction matters whenever the chance of leaving a state depends on how long the current sojourn has lasted. A person who has remained in a state for a long time may have a different near-term transition risk from someone just entering it. This tutorial derives that conditional distribution, including its destination-specific form.

The landmark \(a\) is measured from entry into the current sojourn, not automatically from the study’s calendar-time origin. The conditioning event \(T_i>a\) means that the current visit to state \(i\) has lasted beyond \(a\); it does not describe a person who has already left and later returned.

Deriving the residual distribution

For \(r\geq 0\), the event \(R_a>r\), given \(T_i>a\), is exactly the event that the sojourn lasts beyond \(a+r\), given that it lasts beyond \(a\). Therefore,

This is the residual survival function. Subtracting it from one gives the conditional cumulative distribution function:

The numerator in the last expression is the probability of exiting between elapsed times \(a\) and \(a+r\). Dividing by \(S_i(a)\) conditions on having reached the landmark without exiting. If the sojourn has a continuous density \(f_i(t)\), the residual density is \(f_i(a+r)/S_i(a)\) for \(r>0\). These formulas also allow distributions with atoms when probabilities are interpreted using the stated strict survival event and the corresponding interval events.

When exit can lead to several destinations, the same conditioning applies to the transition kernel. Write \(Q_{ij}(t)=P(T_i\leq t,J=j)\), the probability of exiting state \(i\) by elapsed time \(t\) and entering destination \(j\). For a sojourn that has survived to \(a\),

Summing over possible destinations gives the total residual distribution, because the events of exiting to distinct destinations are disjoint. If exit occurs eventually with probability one, the probability that the eventual destination is \(j\), conditional on surviving to \(a\), is \([Q_{ij}(\infty)-Q_{ij}(a)]/S_i(a)\). Conditioning on survival may therefore change not only the expected remaining time but also the distribution of the eventual destination.

What happens to the hazard?

For a continuous sojourn distribution, define the sojourn hazard at elapsed duration \(t\) as the instantaneous exit rate among individuals who have remained in the state until \(t\):

At residual time \(u\) after surviving to landmark \(a\), the individual has accumulated total sojourn duration \(a+u\). Thus the residual hazard, at residual duration \(u\), equals \(h_i(a+u)\), not \(h_i(u)\). The accumulated time in the state is still relevant; conditioning has not reset it.

This identity follows directly from the conditional probability defining the hazard: being at residual time \(u\) means surviving in the original sojourn to \(a+u\), and an exit in the next short interval is an exit at original duration \(a+u\). In a model with destination-specific cause hazards, each destination’s hazard is evaluated at that same accumulated duration.

A useful special case is the exponential distribution, for which \(S_i(t)=e^{-\lambda t}\), with constant rate \(\lambda>0\). Then

The residual survival is identical to the fresh sojourn survival. This is the memoryless property. For a general continuous proper distribution, equality of residual and fresh distributions at every landmark requires the memoryless property and characterizes the exponential distribution. Equality at one particular landmark alone does not establish that the whole distribution is exponential.

Destination-specific residual risks

In a multi-state setting, an exit from state \(i\) can have different destinations. The residual waiting-time calculation must retain which destination occurs; summing over destinations too early can hide changes in destination composition after landmark survival. The following example uses a simple kernel to make that distinction explicit.

Survival selection and unknown sojourn age

Landmark conditioning can change the composition of a group even when the model contains latent types. For example, people with longer typical sojourns are more likely to remain in the state until a later landmark. The next calculation shows how that selection changes residual survival.

If the current sojourn age is unknown, one cannot generally substitute the calendar landmark into \(S_i(a+r)/S_i(a)\). Let \(A\) denote the elapsed age of the current sojourn at the landmark. If the distribution of \(A\) among individuals known to be in state \(i\) is available, then the residual survival is a mixture of age-specific residual survivals:

This expression presumes that, conditional on age and the model information used, the sojourn distribution is \(S_i\). If people in the state have different unobserved types, the relevant conditional distribution must also account for those types. The age mixture and the type mixture are distinct sources of selection; neither is resolved by simply treating everyone as a new entrant at the landmark.

Assumptions, checks, and alternatives

The residual formulas are probability identities once the sojourn distribution is specified. Their model-based use nevertheless depends on defining a coherent sojourn and conditioning event. In particular:

• The time origin is entry into the current visit to state \(i\), and \(T_i\) measures time to its exit.
• The landmark is an elapsed sojourn duration \(a\), and \(S_i(a)>0\); otherwise the conditional distribution is undefined.
• The survival or transition kernel used applies to the population and conditioning information of interest. If relevant types or covariates differ, their distribution among landmark survivors must be represented.
• For destination-specific calculations, the kernel \(Q_{ij}\) must describe the joint event of exit time and destination; destination independence must not be assumed unless justified.

A practical check is to verify that the residual survival at \(r=0\) is 1, is nonincreasing in \(r\), and lies between 0 and 1. Destination-specific probabilities should be nonnegative and sum to the total residual exit probability. If a model’s residual survival increases with \(r\), exceeds 1, or fails these sum checks, there is likely an algebra, coding, or kernel-definition error. Also check whether \(a\) is duration since entry or calendar time: confusing the two is a common source of incorrect predictions.

The main alternative is to use a fresh sojourn distribution \(S_i(r)\) at the landmark. That is correct when the landmark individual is a new entrant, or when the sojourn distribution is memoryless in the relevant model. It is generally wrong for a person who has already survived a nonzero duration under a duration-dependent model. A state-transition probability over a calendar-time interval answers a different question: it may account for multiple transitions and paths, whereas the residual formula concerns the remaining time in one current sojourn. Use the quantity matching the question.

The residual distribution is limited by the quality and scope of the assumed sojourn model. It does not by itself adjust for an unknown age distribution, unmeasured heterogeneity, or a change in the process at the landmark. In applied work, the landmark population and the meaning of state entry should be stated explicitly. A fitted sojourn distribution that ignores selection among survivors may produce a mathematically valid conditional calculation for the wrong target population.