Decision analysis and decision trees
- Concept
When the best choice depends on chances and on what each outcome is worth, a decision tree lays the options out and adds them up so they can be compared.
What it is
Decision analysis models a choice explicitly: options start at a decision node, uncertain events branch at chance nodes with probabilities, and each path ends in an outcome with a value (utility). Each option's expected value is the probability-weighted sum of its outcomes; the highest wins, and sensitivity analysis checks how firm that is.
Key facts
- Folding back (averaging out): at each chance node multiply each branch's value by its probability and add; at a decision node keep the best option.
- Utilities can be elicited from the patient, for example with a standard gamble: the risk of death a person would accept to avoid a state sets its value.
- QALYs multiply years in a state by its utility; the QALY page shows the arithmetic and how costs are added.
- Sensitivity analysis: one-way (one input), two-way (two at once) or probabilistic (Monte Carlo draws from each input's range; the share of runs an option wins).
- NNT against NNH: made-up drug with NNT 25 for a stroke and NNH 50 for a major bleed prevents 2 strokes per bleed caused; whether that is worth it depends on how bad each event is, which utilities supply.
- When events recur over years, a Markov model (health states, fixed cycles, no memory of the past) replaces a simple tree.
Worked example
- Made-up choice: operate now (A) or watch and wait (B). Utilities: well 1.0, well after surgery 0.95, death 0.
- A: operative death 2%, so 98% end at 0.95. EV(A) = 0.98 × 0.95 + 0.02 × 0 = 0.931.
- B: 70% settle alone (1.0); 30% worsen and need emergency surgery, where 90% recover (0.9) and 10% die (0).
- Fold back the emergency branch first: 0.9 × 0.9 + 0.1 × 0 = 0.81.
- EV(B) = 0.70 × 1.0 + 0.30 × 0.81 = 0.700 + 0.243 = 0.943.
- Base case: B wins by 0.012, a small margin, so the inputs must be tested.
- One-way sensitivity on p, the chance of worsening: EV(B) = 1 − 0.19 × p, which equals 0.931 when p ≈ 0.36.
- Reading: if this patient's chance of worsening is above about 36%, operating now wins; below it, waiting wins.
Three decision tools that are often confused
| Feature | Decision tree | Markov model | Threshold model |
|---|---|---|---|
| Question | Which option has the highest expected value for one choice | The same, when events recur over a long time | At what probability of disease to test or treat |
| Time | Short horizon, one-off events | Repeated cycles of fixed length, often lifelong | One moment of decision |
| Built from | Probabilities and utilities on branches | Health states and transition probabilities | Benefit and harm of treatment, plus test accuracy |
| Output | Expected value (utility) per option | Expected QALYs, often costs, per strategy | A test threshold and a treatment threshold |
| Main limit | Branches multiply when reality is complex | No memory: the next state depends only on the current one | Rests on group data and on probabilities and utilities being independent |
In practice
- At the bedside, a small tree built with the patient makes their values explicit in a complex choice and supports shared decisions.
- Probabilities come from trials or cohorts, adjusted for the patient in front of you; where there are no data, from expert opinion, labelled as such.
- Guideline panels and health technology assessment use decision models, often with costs added, to compare strategies.
- Some patients prefer to see the chances of each outcome (a risk analysis) rather than accept the option with the highest expected utility.
Common pitfalls
- Putting relative risk reductions on the branches: the tree needs absolute probabilities for this patient.
- Reporting a model without its inputs, ranges and sources, so nobody can check it.
- Trusting a winning margin without a sensitivity analysis.
- Modelling options that lie outside the standard of care.
Common traps
- The option with the highest expected value is best for every patient: it is best on average for the inputs used; a patient with different risks or values may do better with the other option
- Utilities and probabilities are the same thing: a probability says how likely an outcome is; a utility says how much it is worth, from 0 (death) to 1 (full health)
- Any winning margin settles the question: when expected values are close, only a sensitivity analysis shows whether the choice holds or flips with plausible inputs
- A decision model produces new evidence: it only recombines existing data and judgements, so its answer is as good as its inputs
- If the NNT is smaller than the NNH, treat: counts ignore how bad each outcome is; a rare harm can outweigh a common benefit once outcomes are weighted by utilities
Test yourself
Think of the answer, then tap the question to see it.
The branches leaving a chance node carry probabilities 0.6 and 0.3. What is wrong?
Probabilities leaving one chance node must add up to 1: a 0.1 branch is missing or the numbers are wrong.
Option A has an 80% chance of an outcome worth 0.9 and a 20% chance of one worth 0.5. What is its expected value?
0.8 × 0.9 + 0.2 × 0.5 = 0.72 + 0.10 = 0.82.
Two options differ by 0.005 in expected utility. What should be done before recommending one?
A sensitivity analysis: vary the uncertain probabilities and utilities; if plausible values flip the choice, it is a close call to talk through with the patient.
A drug has NNT 20 for a stroke and NNH 40 for a major bleed. Per 40 patients treated, what happens, and what else is needed to decide?
About 2 strokes prevented and 1 bleed caused; the decision also needs how bad a bleed is compared with a stroke for this patient (the utilities).
Why is a simple decision tree a poor fit for a chronic disease with relapses over decades?
Recurring events over long horizons make the branches multiply; a Markov model with health states and fixed cycles handles them better.
Elsewhere in DurgiPedia
Connections
Educational content on epidemiological and statistical methods; the examples use made-up numbers to teach the calculation. Not a clinical or legal guide: current Colombian rules are in the official sources.
Sources: Tips for teachers of evidence-based medicine: making sense of decision analysis using a decision tree (Journal of General Internal Medicine, 2009); Medical decision analysis for personalized oncology at the patient bedside (Tumori, 2026); Decision analysis in cardiac surgery: a scoping review and methodological primer (European Journal of Cardio-Thoracic Surgery, 2024); The threshold model revisited (Journal of Evaluation in Clinical Practice, 2019).
Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake
Representative model, for educational purposes only.
