Multi-State Models: When One Time-to-Event Endpoint Hides the Clinical Path
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
A time-to-event analysis often asks one clean question: how long until progression, hospitalization, relapse, or death? Clinical courses are rarely that tidy. Patients move through intermediate states, and a treatment may change each move differently.
Two treatments can produce similar 12-month survival while creating very different journeys. One may prevent progression but offer little benefit after progression. Another may not delay progression but substantially extend life afterward. A single Kaplan–Meier curve can be correct and still compress away the mechanism clinicians care about.
From One Finish Line to a Map
A multi-state model represents the clinical course as a set of states and the allowed transitions between them. In a simple oncology example, every patient begins alive without progression. They may progress, die before recorded progression, or die after progression. Death is absorbing; the other states are transient.
The clean metaphor
Ordinary survival analysis records when the journey ends. A multi-state model keeps the route map: which station a patient reached, when they arrived, and which exits remained available.
Competing-risks analysis is a special, simpler map: one starting state with several mutually exclusive first events. Multi-state models extend that logic to intermediate events and later transitions. That distinction matters when the intermediate state is itself clinically meaningful.
Interactive pathway explorer
One survival number can hide three transitions
Start 100 patients in the stable state. The control pathway is fixed at 6% stable-to-progression, 1% stable-to-death, and 7% progression-to-death per month. Change the treatment pathway and watch the 12-month state distribution move.
Control at 12 months
Treatment at 12 months
What the endpoint headline misses
Treatment changes 12-month survival by +0.4 percentage points, but changes the probability of being alive without progression by +12.2 points. Those are different clinical stories produced by different transitions.
Teaching model only: expected counts from fixed monthly transition probabilities, with no recovery, covariates, censoring, or uncertainty. It illustrates why pathways matter; it is not an estimator or trial-design calculator.
Choose the Quantity Before the Model
“We used a multi-state model” does not identify the scientific question. The same state diagram can support several answers:
| Question | Useful quantity | Plain-language meaning |
|---|---|---|
| Where are patients at 12 months? | State occupation probability | Probability of occupying each state at that time |
| How quickly do patients move? | Transition-specific hazard | Instantaneous transition rate among those currently at risk for that move |
| What is the chance of a later move? | Transition probability | Probability of reaching a destination state over a stated horizon |
| How much healthy time is preserved? | Restricted mean time in state | Expected time spent in each state up to a fixed horizon |
The Aalen–Johansen estimator is the multi-state analogue of familiar nonparametric survival estimators and is commonly used for transition or state-occupation probabilities. Transition-specific Cox models can examine associations with each move. Neither method makes a vague question precise after the fact.
The Post-Progression Hazard Is Not Automatically Causal
Suppose a randomized treatment reduces progression. The patients who nevertheless progress under treatment may be a selected, higher-risk subset. Comparing death after progression between trial arms conditions on a post-randomization event affected by treatment. Randomization no longer guarantees that those two progressed groups are exchangeable.
A transition-specific treatment coefficient can describe how the observed pathways differ. Calling it the causal effect of treatment “among progressors” requires additional assumptions and usually a more explicit causal framework. Multi-state decomposition is not mediation analysis in disguise.
Reviewer red flag
The paper moves from a randomized total treatment effect to causal language about a post-randomization transition without discussing selection into the origin state.
Four Design Choices That Can Change the Story
- State definitions: progression confirmed on imaging, first symptomatic deterioration, and treatment change are not interchangeable. States should be clinically interpretable and reproducibly measured.
- Allowed transitions: a progressive illness–death model forbids recovery. That may fit metastatic progression and badly misrepresent relapsing conditions or repeated hospital care.
- Clock choice: the next transition may depend on time since study entry, time since entering the current state, or both. A simple Markov assumption says the present state carries the relevant history; that assumption should not be smuggled in as software default.
- Observation process: scan schedules, visit frequency, interval censoring, missing transition times, and unequal surveillance can alter when—or whether—an intermediate state is recorded.
A Reviewer Checklist for Multi-State Claims
Question and map
- Is the target quantity named?
- Are every state and allowed transition prespecified?
- Does the map match the disease process?
Time and measurement
- Is the time scale clear for every transition?
- Were intermediate states sought equally?
- Are same-day and interval-censored events handled explicitly?
Assumptions and support
- Is independent censoring plausible?
- Is the Markov or semi-Markov choice justified?
- Are sparse transitions reported rather than overmodeled?
Interpretation
- Are hazards separated from probabilities?
- Are descriptive transitions kept distinct from causal effects?
- Does uncertainty accompany every pathway claim?
When a Multi-State Model Earns Its Complexity
Use one when intermediate events change prognosis, remain clinically important, and are measured well enough to support a pathway analysis. Good examples include remission and relapse, discharge and readmission, transplant and graft failure, or disability states that can improve and worsen.
Do not add states merely because the dataset contains timestamps. Every extra transition divides the events, multiplies modeling decisions, and creates another opportunity for sparse estimates. If the clinical question is genuinely time to one terminal event, a simpler analysis may be more honest.
Why This Matters for Aqrab
Multi-state papers often look rigorous because the diagram is elaborate. The critique still begins with ordinary questions: Which clinical decision does the map serve? Which transitions were observable? Which assumptions connect transition estimates to the conclusion?
Use Aqrab Try to pressure-test whether a manuscript's state diagram matches its stated estimand and whether descriptive pathway contrasts have been promoted into causal claims. The useful review is not “use a more advanced model.” It is “show that every extra state earns its place.”
Methods Anchors
Putter, Fiocco, and Geskus provide a practical tutorial connecting competing risks, cumulative incidence, transition probabilities, and multi-state models. Andersen and Keiding introduce event-history analysis through transition intensities and multi-state structures. Le-Rademacher and colleagues show how state probabilities and time in state can complement standard survival analyses in cancer trials while separating multi-state and time-dependent Cox questions.
Keep reading
Don't stop at one method.
Good methods judgment comes from contrast. Read the neighboring guides, see where the assumptions diverge, and avoid treating every observational problem like it needs the same hammer.
Recurrent Events in Clinical Trials: When Time to First Event Hides Disease Burden
A practical guide to recurrent-event analysis in clinical trials. Learn when time to first event discards patient burden, how death changes the estimand, and what reviewers should demand.
Nonproportional Hazards: When One Hazard Ratio Pretends the Treatment Effect Never Changes
A practical guide to nonproportional hazards for clinical researchers. Covers delayed effects, crossing curves, waning benefit, why one hazard ratio can mislead, and what reviewers should demand instead.
The Hazard-Ratio Autopsy: When Delayed Effects Make One Number Misleading
A practical trial autopsy for non-proportional hazards. Read the curves, risk sets, landmarks, and restricted mean survival time before translating one hazard ratio into a constant treatment effect.