Choosing decision trees or Markov models
Use a decision tree for a bounded sequence and a Markov model when repeated events and time in health states drive the answer.
Fit
When it matters, and when it does not
It matters when
- You can describe each relevant pathway as a finite sequence over a suitable horizon.
- Events do not recur in ways that make the tree unwieldy.
It matters less when
- Choose a Markov state-transition model when recurring events or long-term state occupancy matter. Read: Choosing an economic evaluation
- If individual history drives risk, consider expanded states, semi-Markov modeling or individual simulation. Read: Choosing an economic evaluation
Before you start
What to check in your study
- Decision and horizonChoose the strategies, perspective and duration needed to capture important consequences.
- Relevant historyList recurring events and whether risk depends on duration or earlier events.
- Transitions and valuesIdentify event probabilities, costs, utilities and uncertainty for the intended setting.
- The model retains the history needed for future risks.How to check: Use states that distinguish relevant histories; test alternative structures.
- Probabilities use compatible time intervals.How to check: Check that outgoing transition probabilities sum to one for each state.
- The same perspective and horizon apply to each strategy.How to check: Include relevant consequences and test parameter and structural uncertainty.
Plain language
How it works
An economic decision tree branches through choices and events to final costs and outcomes. It is different from a classification tree used for prediction. A cohort Markov model repeatedly moves a population between health states, adding costs and outcomes over cycles.
Beck and Pauker develop a Markov prognosis model; Sonnenberg and Beck explain when repeated events make trees difficult. Markov does not mean that transition risks must be constant over calendar time. The memory limit concerns history absent from the states. Tunnel states can track duration; other state structures or individual simulation can retain more history. Both approaches need validation and sensitivity analyses.
Worked example
A worked example
- Setting
- A model starts with 100 people in a well state.
- Question
- Can a short tree capture repeated periods of illness?
- Approach
- Each cycle, well people have a 0.8 chance of staying well and a 0.2 chance of becoming ill. Ill people stay ill.
- Result
- After one cycle, 80 of 100 remain well, or 80 in 100. After two cycles, 64 remain well and 36 are ill.
- What it means
- The second-cycle well count is 100 times 0.8 times 0.8. Allowing recovery would require another transition and change these counts.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
A cohort Markov model assumes future transitions depend on the current modeled state and time, not unrecorded history. Add states or another model when duration or prior events matter. A simpler tree is valid only if its horizon captures relevant consequences.
- Applying an annual probability each month.Instead: Convert risks using a justified rate or time model.
- Letting past events matter without recording them.Instead: Expand the states or choose a model that retains history.
- Calling the more detailed model more credible.Instead: Validate structure and inputs against the decision and evidence.
Inside Mallard
How Mallard uses this
Mallard includes this comparison in Learn search and the Atlas. Plans do not link to this comparison automatically.
concept.choosing_tree_markov, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- The Markov process in medical prognosis (opens in a new tab) · Med Decis Making · doi 10.1177/0272989X8300300403
- Markov models in medical decision making: a practical guide (opens in a new tab) · Med Decis Making · doi 10.1177/0272989X9301300409