Why Add Recovery?
The basic illness-death structure distinguishes being alive without illness, being alive with illness, and being dead. It allows illness onset and death, but it treats illness as a one-way move: once a person enters the illness state, the only permitted next state is death. That may be appropriate when illness is irreversible by definition, but not when patients can recover and later become ill again.
Adding a transition from illness back to the illness-free state changes which histories the model can represent. A person may now recover, remain alive without illness for a time, and experience another illness episode. The expanded diagram still has only three states; the return transition means a state can be visited more than once.
A transition is an allowed change between states. It does not say how often the change occurs, how long a person stays in a state, or what causes the change. Those questions require additional modelling assumptions or data. In particular, adding recovery to a diagram does not by itself assert that the process is Markov.
The Expanded Structure and Its Paths
The return edge \(1\to0\) allows a path to revisit state \(0\), and from there to enter state \(1\) again. For example, \(0\to1\to0\to1\to2\) describes illness, recovery, a second illness, and death while ill. The first and second visits to state \(1\) are separate illness episodes under the state definition above. By contrast, \(0\to1\to2\) describes one illness episode ending in death, with no recovery transition recorded.
The diagram describes allowable successive states, not every detail of the underlying biology. An edge \(0\to2\) means death occurs without an intervening illness state in the recorded history. It need not claim that no symptoms or unrecorded disease existed. Likewise, a recorded \(1\to0\) transition requires an operational definition of recovery: for example, absence of active illness according to a specified clinical criterion. If illness remains active but improves in severity, that change is not recovery under this three-state definition.
| Transition | Interpretation | What it does not establish |
|---|---|---|
| \(0\to1\) | Illness begins while the person is alive and illness-free. | The cause of illness or its future duration. |
| \(1\to0\) | Active illness ends and the person is alive without active illness. | That the person will never become ill again. |
| \(0\to2\) | Death occurs from the illness-free state. | That illness was biologically impossible or absent but unrecorded. |
| \(1\to2\) | Death occurs while the person is in the illness state. | That illness caused the death. |
Worked Example: Reading a History With Two Illness Episodes
Consider a hypothetical patient observed from study entry. The recorded state sequence is \(0\to1\to0\to1\to2\), with transitions at times 2, 5, 8, and 11 months after entry. Interpret the path and count its transitions.
At entry the patient is alive and illness-free. At month 2, the transition \(0\to1\) marks the first illness onset. At month 5, \(1\to0\) marks recovery. At month 8, \(0\to1\) marks a second illness onset. At month 11, \(1\to2\) marks death while ill.
There are two illness episodes because the patient enters state \(1\) twice. The transition counts are \(N_{01}=2\), \(N_{10}=1\), \(N_{12}=1\), and \(N_{02}=0\), where \(N_{ij}\) is the number of observed transitions from state \(i\) to state \(j\) in this history. Death is the final state, so the path has no transitions after month 11.
A Transition-Balance Check
A useful way to check a history is to account for each entry to and exit from the illness state. Let \(X(t)\) denote the patient’s state at time \(t\), and let \(N_{ij}(t)\) count observed \(i\to j\) transitions by time \(t\). Each \(0\to1\) transition adds one to the indicator that the patient is in state \(1\). Each \(1\to0\) or \(1\to2\) transition subtracts one. No other allowed transition changes that indicator.
Summing those changes over the observed history gives the identity below. It holds for a path with the stated transitions, whether or not the process is Markov.
Here \(\mathbf{1}\{\cdot\}\) is 1 when its condition is true and 0 otherwise. The left side is the change in illness-state occupancy: it is 1 if the patient starts outside state \(1\) and ends in state \(1\), \(-1\) if the patient starts in state \(1\) and ends outside it, and 0 if the patient starts and ends on the same side. The right side records entries to illness minus exits through recovery or death.
For the path in the worked example, the patient starts and ends outside state \(1\), so the left side is \(0-0=0\). The right side is \(2-1-1=0\), confirming the accounting. This check can expose omitted or duplicated transitions in a coded history, but it cannot verify that the clinical state definitions themselves are appropriate.
Worked Example: Checking State Occupancy Against Transition Counts
In a hypothetical follow-up record, a patient starts in state \(0\) and has \(N_{01}=3\), \(N_{10}=2\), and \(N_{12}=1\) by the end of observation. Use the balance identity to determine the final state.
The patient starts outside state \(1\), so \(\mathbf{1}\{X(0)=1\}=0\). The net change in the illness-state indicator is \(3-2-1=0\). Therefore \(\mathbf{1}\{X(t)=1\}=0\): the patient ends outside the illness state.
The transition counts alone leave two possible final states, \(0\) or \(2\). However, the record includes one \(1\to2\) transition, and state \(2\) is absorbing. After that transition, no return to state \(0\) is allowed. Thus the patient ends in state \(2\), dead. The balance confirms the illness-state accounting; the absorbing-state rule resolves which non-ill state is the endpoint.
What the Diagram Assumes—and What It Leaves Open
Before using this structure, specify how states are assigned and what counts as a transition. The interpretation depends on several conditions:
- States are mutually exclusive at a given time. A patient is classified as either alive without active illness, alive with active illness, or dead.
- The categories are sufficiently complete for the question. A patient whose illness status is unknown should not be silently classified as recovered.
- Death is absorbing. A recorded death cannot be followed by recovery or a new illness onset.
- The listed edges match the intended clinical process. If illness can begin and resolve between assessments, the observed record may miss transitions that the underlying process made.
- Entry and observation rules are clear. The initial state and the time follow-up begins determine which episodes are visible. An episode already underway at entry is not an observed \(0\to1\) onset.
A state diagram alone does not specify transition rates, waiting-time distributions, or whether the future depends only on the current state. Even if two patients are both currently in state \(0\), one may have recovered recently and the other may never have been ill. Whether that difference matters for their future transition behaviour is a separate modelling question. The Markov property is not implied by drawing the return edge.
Observation schedules matter too. If visits are months apart, a patient may move \(0\to1\to0\) between visits and appear to remain in state \(0\). Conversely, a change in recorded symptoms may be mistaken for a transition if measurement rules are inconsistent. When exact transition times are unavailable, the observed sequence may not reveal the full path. The diagram specifies which paths are allowed in the model; it cannot recover changes that the data do not record.
Worked Example: Finding a Structural Inconsistency
Suppose a hypothetical data file records the sequence \(0\to1\to2\to0\) for one patient. Determine whether this is a valid path under the illness-death model with recovery.
The first two transitions are allowed: illness can begin through \(0\to1\), and a person who is ill can die through \(1\to2\). The final transition \(2\to0\) is not allowed because state \(2\) is absorbing. The recorded sequence is therefore inconsistent with the model.
The appropriate response is to check the source record, state coding, event dates, and patient identifier. One should not silently delete the death transition or reinterpret it as recovery. If the record is correct as entered, then either the data are erroneous or the state definitions and model are not describing the recorded process.
How Recovery Changes the Modelling Choice
The original three-state illness-death structure without recovery is simpler when a first illness event permanently defines the patient’s subsequent state until death. Adding \(1\to0\) is necessary when recovery is a meaningful observed state change and later illness can occur. Omitting recovery then forces the model to treat a recovered patient as still ill, which can distort state histories and any summaries based on time spent in each state.
A different choice is to focus only on the first illness onset and then follow death. That may suit a question about time to first illness, but it does not represent recovery or a second onset as transitions in the state process. These structures answer different questions; the more elaborate diagram is not automatically preferable. Choose states and edges to match the scientific question and the available measurements.
The main practical diagnostics are to inspect observed transition sequences for forbidden moves, verify that recovery is defined consistently, and compare recorded state occupancy with transition counts. Also ask whether the observation schedule can miss recovery-and-recurrence paths. A diagram that fits the visible data may still be a poor representation if important transitions are routinely unobserved.
Check Your Understanding
Answer each question using the state definitions \(0=\) alive without active illness, \(1=\) alive with active illness, and \(2=\) dead.
- How many illness episodes are represented by \(0\to1\to0\to1\to0\)?
- Is \(0\to2\) an allowed transition, and what does it mean in the recorded history?
- A patient starts in state \(0\), with \(N_{01}=2\), \(N_{10}=1\), and \(N_{12}=0\). What state must the patient occupy at the end?
- Does adding the recovery edge establish that the process is Markov? Explain briefly.
Solutions. (1) There are two episodes, one for each entry into state \(1\). (2) Yes; it records death without an intervening illness-state transition. (3) The occupancy balance gives a net increase of \(2-1-0=1\) in the illness indicator. Starting in state \(0\), the patient therefore ends in state \(1\). (4) No. The edge specifies an allowed transition; the Markov property is a separate assumption about how the future relates to the past given the current state.