AI Glossary / Statistics and probability / Advanced
Staged tree (mathematics)
Definition
A staged tree is a probabilistic event tree whose nonterminal situations are grouped into stages when they share the same conditional distribution over corresponding next events. Stages are often shown with matching colors. The tree represents possible event histories, while stage membership expresses equality constraints on selected transition probabilities. This allows detailed context-specific relationships to be represented. A staged tree is different from a decision tree used simply to split features and predict a label.
What it means
Two different histories can lead to the same distribution for what happens next. A staged tree records that equivalence without pretending the histories themselves are identical. It can also represent processes in which different branches have different possible next events. The stage structure determines which probabilities are shared and can reduce the number of free parameters. Sharing a stage concerns the immediate next-event distribution; it does not necessarily mean that every later development has the same distribution.
How it works
Build an event tree in which each root-to-leaf path represents a possible history. Assign outgoing edges conditional probabilities that sum to one at each nonterminal situation. Identify situations whose corresponding outgoing distributions are constrained to match and place them in the same stage, with a clearly defined correspondence between edge labels. Estimate shared probabilities from suitable data or elicit them under stated assumptions. Calculate a path probability by multiplying its edge probabilities, and calculate broader event probabilities by summing relevant paths. Evaluate whether the proposed stage constraints fit the application and evidence.
Three practical examples
Example 1: A service process models alternative routes through enquiry handling while sharing a next-step distribution across selected histories. Example 2: A reliability analysis represents different failure and repair paths with some common transition patterns. Example 3: An AI workflow model distinguishes branches where different follow-up actions are possible. These examples require a meaningful ordering of events and justified equality constraints. A visually appealing color grouping alone does not establish that the corresponding probabilities should be equal.
Think of it this way
Think of several routes through a building that arrive at different desks, each offering the same immediate choice in the same proportions. Those desks can share a stage even though visitors took different routes to reach them. The rest of the building need not be identical beyond each desk. This distinction separates a shared next-step distribution from a stronger claim that entire future subtrees are equivalent.
Worked example
Consider an event tree whose first branch is contact channel A with probability 0.4 or channel B with probability 0.6. At either channel, the next event is resolved with probability 0.7 or unresolved with probability 0.3. Put the two channel situations in the same stage because their corresponding next-event distributions are equal. The four path probabilities are 0.4 times 0.7 = 0.28, 0.4 times 0.3 = 0.12, 0.6 times 0.7 = 0.42 and 0.6 times 0.3 = 0.18. They sum to one, and the marginal resolution probability is 0.28 + 0.42 = 0.70. The model needs two free probabilities: channel A and the shared resolution rate. Without that shared stage, separate channel-specific resolution rates would give three free probabilities. This equality is a model assumption, not a conclusion established by the arithmetic.
Illustrative case study
In this fictional case study, Harbor Workflow maps a customer enquiry process. Analysts initially propose a shared resolution stage for two channels because their procedures seem similar. They inspect adequate channel-specific data, compare a shared-rate model with a more flexible alternative and record uncertainty. One later branch has a different escalation pathway, so it is not merged merely because its first transition looks similar. Managers receive a model that makes probability-sharing assumptions visible and keeps the event histories understandable.
Common misconceptions and mistakes
Common mistakes include equating a stage with a time step, assuming all nodes at the same depth share a stage, and grouping situations solely because their observed proportions happen to match. Some definitions impose additional restrictions, such as matching edge labels or the same depth, so the convention should be stated. A stage is also weaker than a position in a chain event graph: positions require appropriate equivalence of the future staged subtrees. Probabilistic structure alone does not justify a causal intervention interpretation.
Business use
Businesses can use staged trees to discuss branching processes and make selective probability-sharing assumptions explicit. Shared stages can pool information across contexts, but inappropriate pooling can hide meaningful differences. Document the event ordering, stage definitions, missing-data treatment and estimation method. More detailed trees can grow rapidly, while sparse branches provide weak evidence. A useful model balances interpretability and fidelity, and its probabilities should be validated against relevant new process data before being used for planning or service commitments.
People also ask
Is every event tree a staged tree with nontrivial sharing?
No; stage groupings may be trivial.
Does sharing a stage mean the same past history?
No.
Must all later outcomes be identical?
No; the stage describes the corresponding immediate transition distribution.
Is it the same as a chain event graph?
No; a chain event graph can compress an associated staged tree through a stronger equivalence of situations.
Does the diagram alone identify causal effects?
No.
Check your understanding
What is the probability of channel B followed by resolution?
Answer: 0.42.
Why are the two channel situations in the same stage?
Answer: Their next-event probability vectors are both (0.7,0.3) with corresponding labels.
How many free probabilities does this simple shared-stage model use?
Answer: Two.
Does matching the supplied numbers prove that the shared-stage assumption is true in a real service?
Answer: No.
Related terms
- Conditional probability
- Context-specific independence
- Related terms include event tree
- floret
- chain event graph
- position and probabilistic graphical model
- while path probabilities follow ordinary conditional probability multiplication
Learn next
Learn next how stage constraints can encode context-specific relationships that are awkward to display in a conventional Bayesian network. Then examine the stronger position equivalence used to construct chain event graphs. Study model selection and uncertainty so that estimated similarities are not confused with known equality of probabilities.
More in Statistics and probability
- Spillover (experiment) - A spillover in an experiment occurs when a unit's outcome is affected by treatment assigned to another unit.
- Squared ranks test - The squared ranks test, often called the Conover squared ranks test, is a rank-based procedure for comparing dispersion across independent groups.
- Stable model - A stable model is a formally defined interpretation of a logic program with default negation.
- Standard deviation - Standard deviation is a measure of spread obtained by taking the square root of variance.
- Standard error - The standard error is the standard deviation of an estimator's sampling distribution.
- Standard error of differences between means - The standard error of differences between means is the standard deviation of the sampling distribution of a difference between estimated means.
Sources
Sources: Shenvi and Smith, Staged Tree Algorithm, https://proceedings.mlr.press/v138/shenvi20a/shenvi20a.pdf; Leonelli and Varando, staged-tree research, https://proceedings.mlr.press/v206/leonelli23a/leonelli23a.pdf; Görgen and Smith, A Differential Approach to Causality in Staged Trees, https://proceedings.mlr.press/v52/goergen16.html. These primary papers explain stages, probability structure and additional causal considerations. The channel example is original and independently checked; the case is fictional.
Previous: Staged training Next: Staging (data)