Ask Mallard
Analysis method

Multistate models

4 min read

Multistate models follow people through defined health states. They separate the speed of each transition from the chance of being in a state.

Fit

When to use it, and when not to

Use it when

  • An intermediate event changes the later course, such as relapse before death.
  • You need probabilities of being alive in different health states over time.

Not when

Before you start

What you need

Decide these first
  1. States and allowed transitionsMake states distinct and state whether recovery or repeat entry is possible.
  2. Entry, transition and last-contact timesA state first detected at a visit may have begun earlier; account for interval censoring.
  3. The probability or transition contrast of interesttargetcompetingEventsSpecify the starting state, time origin and follow-up horizon.
It rests on these
  1. MCAR excludes dependence on observed or missing values. MAR allows dependence on observed values. MNAR retains dependence on missing values after conditioning.How to check: Relate the assumption to the unobserved outcomes and the history actually recorded.
  2. The chosen model captures the relevant past.How to check: A Markov model uses the current state and time. A semi-Markov model can also use time since entering that state.
  3. Transition recording and censoring support the chosen estimator.How to check: Check visit timing, delayed entry and independent censoring given the modeled history.
  4. Name missing completely at random (MCAR), missing at random (MAR) or missing not at random (MNAR).How to check: If MAR is plausible, consider history-based censoring weights. If MNAR is plausible, vary unobserved transition risks in a sensitivity analysis.

Plain language

How it works

Draw states before choosing a model. Arrows mark allowed transitions; an absorbing state has no outgoing arrow. A transition hazard describes the instantaneous rate among people in its source state. State probabilities combine paths through the system. De Wreede and colleagues developed mstate tools for transition models, probability estimates and their standard errors. Beyersmann and colleagues modeled infection during intensive-care stays rather than treating infection as present at admission.

Relapse is an intermediate state; death is absorbing
An initial state leads to an intermediate state or an absorbing endpoint. The intermediate state also leads to that endpoint. There is no recovery in this chosen model, and the endpoint has no outgoing arrow.Two paths to an absorbing endpointInitial stateIntermediate stateAbsorbing endpointArrows are allowed transitions, not probabilities. Other settings may need recovery or repeat transitions;state that choice before fitting.
Conceptual diagram. It shows the structure of the idea, not data from any study.
Choose a probability estimator for the question

Aalen-Johansen combines estimated transition hazards into probabilities under appropriate assumptions. Markov transition predictions and probabilities in the original starting cohort need different history assumptions. Report those assumptions and intervals suited to the probability estimator. A Cox coefficient for one transition is not the probability of that transition or state.

Worked example

A worked example

Invented for illustration
Setting
All 100 patients start alive without relapse. Death is absorbing, so no one leaves that state.
Question
Where are patients expected to be after two intervals?
Approach
In each interval, well patients stay well with probability 0.7, relapse with 0.2, or die with 0.1. Relapsed patients remain relapsed with 0.75; otherwise they die.
Result
After two intervals, expected counts are 49 well, 29 relapsed and 22 dead: 22 in 100 dead. All state probabilities sum to 1.
What it means
Assume fixed transition probabilities and complete follow-up. Death can occur with or without relapse; the relapse-to-death probability alone is not the overall death risk.

In practice

Recent studies PubMed search · not vetted

Searching PubMed…

Watch out

Common mistakes

The one to remember

A transition hazard is not a probability of occupying a state. When death prevents the event, one minus Kaplan-Meier overstates cumulative incidence. Define every allowed transition and justify any assumption that future risk depends only on the current state.

  • Reading a transition hazard as a state probability.Instead: Combine all allowed paths to calculate state probabilities from the specified starting state and time.
  • Censoring competing deaths in a risk curve.Instead: Use cumulative incidence for competing outcomes. One minus Kaplan-Meier overstates incidence when death prevents the event.
  • Applying a Markov assumption without checking the disease course.Instead: Consider time since entry and earlier events. Use a model and probability estimator appropriate to that history.

Inside Mallard

How Mallard uses this

Plans do not link to this method yet. Mallard has no dedicated multistate sizing form.

No Mallard code example points to this page.
Atlas node analysis.multistate, the id a plan cue records.

Related Learn pages

Related designs and methods

Evidence

Sources & further reading

  1. The mstate package for estimation and prediction in non- and semi-parametric multi-state and competing risks models (opens in a new tab) · Comput Methods Programs Biomed · doi 10.1016/j.cmpb.2010.01.001
  2. Transmission-associated nosocomial infections: prolongation of intensive care unit stay and risk factor analysis using multistate models (opens in a new tab) · Am J Infect Control · doi 10.1016/j.ajic.2007.06.007
De Wreede and colleagues support transition-model estimation, probability prediction and estimator-specific standard errors. Beyersmann and colleagues support treating acquired infection as a time-dependent state. The discrete transition calculation is invented; it is not an estimate from either paper. Each record was verified by PMID against PubMed on 2026-11-06, matching title, authors, year and DOI. Record verification does not establish support for every use. No independent professional or methodological approval is claimed.
Last source review: 2026-11-06
AI-generated educational content
Mallard uses reputable methodological and reporting sources, but this material may contain errors or omit important context. Verify important methodological decisions with the cited sources and, when appropriate, a qualified research methodologist or statistician.