---
title: 'Staged Tree Models: Theory & Applications'
url: https://www.emergentmind.com/topics/staged-tree-models
type: topic
---

# Staged Tree Models: Theory & Applications

Searching arXiv for recent and foundational work on staged tree models to ground the article in cited literature.
arXiv search query: "staged tree models context-specific independence balanced staged trees chain event graphs"
Staged tree models are probabilistic graphical models for discrete processes and multivariate categorical data in which a rooted event or probability tree is equipped with a partition of its internal vertices into stages. Vertices in the same stage share the same outgoing edge labels, equivalently the same local conditional distribution for the next variable. This construction yields a representation of a full joint distribution through root-to-leaf path probabilities while allowing equality constraints that depend on history, so staged trees generalize Bayesian networks from symmetric conditional independence to context-specific, partial, and local forms of non-symmetric dependence. They are also used to represent asymmetric event unfoldings, structural zeros, structural missing values, and non-regular sample spaces [2010.15515, 2108.01994, 2307.16301, 2006.15857].

## 1. Formal construction and parametrization

A staged tree is built on a rooted directed tree whose non-leaf vertices are often called situations and whose outgoing edges from a situation form a floret. For an \(\mathbf X\)-compatible staged tree with ordered variables \(\mathbf X=(X_i)_{i\in[p]}\), the vertex set is
\[
V=\{v_0\}\cup\bigcup_{i\in[p]}\mathbb X_{[i]},
\]
and edges connect prefixes to one-step extensions. The edge-labeling takes the form
\[
\theta(v,\mathbf x_{[i]})=(\kappa(v),x_i),
\]
where \(\kappa\) is the coloring function. Two vertices are in the same stage precisely when their sets of outgoing edge labels coincide:
\[
\theta(E(v))=\theta(E(w)).
\]
Thus stage membership is an equality constraint on conditional distributions at different contexts [2203.04390, 2108.01994].

The probabilistic parametrization is multiplicative along paths. A generic parameter space for a staged tree is
\[
\Theta_T=\Big\{\mathbf y\in\mathbb R^{|\theta(E)|}:\ \forall e\in E,\ y_{\theta(e)}\in(0,1),\ \sum_{e\in E(v)}y_{\theta(e)}=1\Big\},
\]
and the model map is
\[
\phi_T:\Theta_T\to \Delta^\circ_{|\mathbf l_T|-1},\qquad \mathbf y\mapsto \Big(\prod_{e\in E(\lambda(l))}y_{\theta(e)}\Big)_{l\in\mathbf l_T}.
\]
Equivalently, if \(\lambda=(v_0,v_1,\dots,v_k)\) is a root-to-leaf path, then its probability is
\[
\theta_\lambda=\prod_{i=1}^k\theta_{v_{i-1},v_i}.
\]
In this sense, staged trees retain the chain-rule factorization of discrete probability models, but replace fixed parent-set structure by stage constraints on a tree of partial histories [2203.04390, 2405.18306].

An equivalent formulation starts from a probability tree with local normalization
\[
\sum_{j=1}^{\kappa_i}\theta_{ij}=1
\]
at each inner vertex and path probability
\[
p_\theta(x)=\prod_{i=1}^k \prod_{j=1}^{\kappa_i} \theta_{ij}^{\alpha_{ij}(x)}.
\]
A staged tree is then a probability tree together with an equivalence relation on inner vertices such that vertices in the same stage have the same outgoing labels. This framing is particularly useful when deriving exponential-family structure directly from the graph [2010.15515].

## 2. Independence semantics and relation to Bayesian networks

The defining semantic feature of staged trees is that they encode equality of local conditional laws only where the corresponding histories are staged together. Consequently, they represent ordinary conditional independence together with more granular forms of dependence symmetry. The literature distinguishes context-specific independence, partial conditional independence, and local conditional independence. In a CStree, for example, a stage corresponding to a context \(X_C=x_C\) induces statements of the form
\[
X_k \perp X_{[k-1]\setminus C}\mid X_C=x_C,
\]
while the broader staged-tree literature also considers equalities that hold only on subsets of a parent domain or only between specific pairs of contexts, without a global conditional-independence interpretation [2210.11521, 2108.01994].

Bayesian networks appear as a special case. A discrete DAG model factorizes as
\[
f(x_1,\dots,x_p)=\prod_{k=1}^p f(x_k \mid x_{\mathrm{pa}(k)}),
\]
and every discrete Bayesian network can be translated into a staged tree representation by coloring each vertex according to the relevant parent configuration. In this construction, if \(G\) is a DAG and \(T_G\) its staged-tree representation, then
\[
\mathcal M_G=\mathcal M_{T_G}.
\]
The converse does not generally hold: staged trees can encode non-symmetric independence statements with no exact DAG representation. This strict extension is one of the central reasons they are used in settings where dependence varies by context or the unfolding process is asymmetric [2105.05907, 2203.04390, 2010.15515].

To summarize the symmetric part of a staged-tree model, several papers define a minimal DAG \(G_T\). Its edges record where changing one earlier variable can change a stage label later in the tree. The resulting DAG is minimal with the property that the staged-tree model is contained in the corresponding DAG model, and d-separation in \(G_T\) captures exactly the symmetric conditional independences implied by the staged tree. Because staged trees usually encode more than symmetric conditional independence, a second summary device, the asymmetry-labeled directed acyclic graph, augments the minimal DAG with edge labels from
\[
\mathcal L^A=\{\text{context},\text{partial},\text{context/partial},\text{local},\text{total}\}.
\]
This allows one to distinguish ordinary DAG-type dependence from context-specific, partial, and local asymmetries in a compact graph [2108.01994].

## 3. Exponential-family, toric, and decomposable structure

From a statistical viewpoint, probability trees form regular exponential families, and staged trees are curved exponential families obtained by imposing stage equalities. In one explicit construction, the sufficient statistic is
\[
T_{ij}=\alpha_{ij},
\]
the natural parameters are
\[
\eta_{ij}(\theta)=\log\!\left(\theta_{ij}\frac{N_{ij}(\theta)}{N_i(\theta)}\right),
\]
and the cumulant-generating function is
\[
\psi(\theta)=-\log(N_1(\theta)).
\]
Stage constraints are linear in the original edge parameters but may become nonlinear in \(\eta\)-coordinates, which is why staged tree models are typically curved rather than regular exponential families. A staged tree is regular if and only if for every pair of same-stage vertices \(v_i,v_s\) and each corresponding edge,
\[
N_{ij}N_{s\kappa} = N_{sj}N_{i\kappa}.
\]
All balanced staged trees are regular, and simple staged trees are automatically balanced, hence regular. This regularity result underlies the asymptotic validity of BIC for staged trees [2010.15515].

Balancedness is the central combinatorial property linking staged trees to decomposable graphical models. If \(v\) is a node and \(t(v)\) its interpolating polynomial, then two same-stage vertices \(v,w\) form a balanced pair when
\[
t(v_i)t(w_j)=t(w_i)t(v_j)\qquad\text{for all }i\neq j.
\]
A staged tree is balanced if every same-stage pair is balanced. For a DAG \(G\), this property exactly characterizes decomposability:
\[
 \mathcal M(G)\text{ is decomposable}
 \iff G\text{ is a perfect DAG}
 \iff T_G\text{ is balanced}.
\]
Balanced staged trees are therefore the staged-tree analogue of decomposable discrete graphical models [2105.05907].

The algebraic-statistical consequences are substantial. Balanced staged trees are toric in the original leaf coordinates, and their prime ideals have strong commutative-algebra properties such as quadratic Gröbner bases, Koszulness, normality, and the Cohen–Macaulay property. The toric viewpoint also extends beyond the balanced case: the class of staged tree models with toric structure is larger if one allows a linear change of coordinates, with the subtree-inclusion property providing a broad non-balanced toric class [2107.04516].

A further refinement is the theory of decomposable context-specific models. In that setting, the relevant objects are CStrees and, more specifically, balanced CStrees. A decomposable context-specific model is defined to be a balanced CStree model, and its defining ideal satisfies a context-specific Hammersley–Clifford theorem:
\[
\ker(\psi_T)=I_{\mathrm{Sat}(T)},
\]
where \(\mathrm{Sat}(T)\) is the set of saturated context-specific independence statements holding in the tree. These statements generate the model ideal, furnish a Markov basis, and can be organized through a finite collection of minimal context DAGs and, after directed moralization, a finite collection of perfect DAGs. This generalizes decomposable graphical-model theory from ordinary conditional independence to context-specific independence [2210.11521].

## 4. Chain event graphs, positions, and equivalence classes

Stages record equality of one-step transition distributions, but they do not exhaust the downstream symmetry of a staged tree. For this reason the literature defines positions: two internal vertices are in the same position when their rooted subtrees are identical. Positions are coarser than stages, and coalescing vertices in the same position yields a chain event graph. The chain event graph is statistically equivalent to the staged tree but can be much more compact, because it merges not only local stage symmetries but whole future developments [2203.04390, 2004.06459].

A simple staged tree is one in which stages and positions coincide:
\[
\theta(E(v))=\theta(E(w))\Rightarrow T_v=T_w.
\]
This subclass is especially useful for graphical interpretation, because every stage becomes a position in the coalesced graph. The resulting chain event graph can then be read directly for non-symmetric conditional independences, while remaining smaller and less cluttered than the underlying tree [2203.04390].

The passage from staged tree to chain event graph is lossless. A general iterative backward algorithm starts from the staged tree, merges leaves into a sink, and then recursively merges same-stage vertices whose already-collapsed futures are isomorphic. The resulting chain event graph is uniquely defined by the staged tree, and the mapping between staged trees and chain event graphs is bijective, so no information is lost by the transformation [2006.15857].

Statistical equivalence among staged trees is described by two operators: swap and resize. The swap operator reorders a twin subtree without changing the represented model, whereas the resize operator contracts or expands saturated or conditionally saturated subtrees. Two staged trees are statistically equivalent if and only if one can be obtained from the other by a finite composition of swaps and resizes. This equivalence theory motivates score equivalence. A Bayesian Dirichlet score based on path uniformity and mass conservation, BDepu, was introduced specifically to satisfy
\[
\text{BDepu}(\mathcal S)=\text{BDepu}(\mathcal S')
\]
for statistically equivalent staged trees [2206.15322].

## 5. Structural learning, scoring, and scalable estimation

Learning a staged tree typically involves three tasks: selecting a variable ordering or event-tree structure, learning the staging at each depth, and estimating stage probabilities. For stratified staged trees, the `stagedtrees` R package provides saturated and independence starting models through `full(...)` and `indep(...)`, conversion of Bayesian networks via `as_sevt`, coalescence to chain event graphs through `ceg()`, and several score-based and clustering-based structure-learning routines. The score-based algorithms include `stages_hc`, `stages_bhc`, `stages_fbhc`, and `stages_bhcr`; the clustering-based algorithms include `stages_bj`, `stages_hclust`, and `stages_kmeans`. By default, hill-climbing routines optimize the negative BIC [2004.06459].

Several later papers refine this basic search paradigm. For simple staged trees, direct structure-learning algorithms include the marginal and total procedures for fixed orderings, together with greedy marginal and exhaustive-order variants when the variable order is unknown. For sparse staged trees, a scalable \(k\)-parents strategy first learns a Bayesian network with at most \(k\) parents per node, converts it to the equivalent staged tree, and then applies backward hill-climbing over stage merges. This reduces the search space while preserving the ability to introduce asymmetric structure beyond the Bayesian-network model [2203.04390, 2206.06970].

Scalability has also been pursued through a posteriori restrictions on admissible merges. Mean-posterior clustering introduces a totally ordered hyperstage on binary event trees, orders situations by mean posterior probabilities, and then restricts agglomerative hierarchical clustering to consecutive merges. In a hyperset of size \(N\), the unrestricted first step considers \({N\choose 2}\) candidate merges and the full run can require
\[
\sum_{n=2}^{N} {n \choose 2} = \frac{N^3-N}{6},
\]
whereas the ordered restriction yields at most
\[
\frac{N(N-1)}{2}
\]
candidate merges over the full run. The reported effect is a reduction from cubic to quadratic scaling [2211.07228].

A different scalable route treats stage recovery as hierarchical clustering on the probability simplex. Starting from the saturated tree, one estimates local conditional distributions, computes dissimilarity matrices within each depth, performs hierarchical clustering, and cuts the dendrogram by a model score such as BIC. The distances studied include Total Variation, Hellinger, Fisher, Jensen–Shannon, Kaniadakis, and Jeffreys/total KL; the linkages include average, complete, McQuitty, and Ward.D2. The reported simulation conclusion is that Total Variation with Ward.D2 often achieves staged trees with strong relative BIC, Hamming-distance recovery, and much lower computational cost than Backward Hill Climbing [2603.15568].

Learning from incomplete data requires modifying both likelihoods and search procedures. Under MCAR or MAR, the incomplete-data likelihood is
\[
L(\theta\mid D)\propto \prod_{x\in D}\sum_{\lambda\in\Lambda_x}\theta_\lambda,
\]
where \(\Lambda_x\) is the set of root-to-leaf paths compatible with the observed part of \(x\). The literature proposes omit, first-missing, and stage-average pseudo-likelihoods, together with EM and structural EM algorithms. In the E-step, expected path counts take the form
\[
n_\lambda^{(t)}= \sum_{x:\lambda\in\Lambda_x} \frac{\theta_\lambda^{(t-1)}}{\sum_{\lambda'\in\Lambda_x}\theta_{\lambda'}^{(t-1)}},
\]
and the M-step reduces to complete-data staged-tree estimation with expected counts. This extends staged-tree structure learning beyond the fully observed setting [2405.18306].

## 6. Causal, predictive, longitudinal, and applied developments

Staged trees have become a major vehicle for causal modeling with categorical variables because they can represent asymmetric and context-specific causal relationships that standard DAGs cannot encode directly. One causal-discovery framework learns a stage structure for a fixed ordering by backward hill-climbing or \(k\)-means clustering, then searches over orderings by dynamic programming using a decomposable score such as BIC. To compare staged trees causally, the context interventional discrepancy is defined by
\[
CID(T,S)=\sum_{i\in[p]} CID_i(T,S),
\]
where each \(CID_i(T,S)\) counts contexts in which interventional distributions are wrongly inferred. This construction is explicitly tailored to context-specific interventions and asymmetric dependence [2106.04416].

A fully Bayesian causal-inference framework has also been proposed. In that formulation, Dirichlet priors are placed on stage probabilities, structural priors on stage partitions are induced through product partition models and a distance-based prior, and posterior inference proceeds by a split-and-merge Markov chain Monte Carlo algorithm. Posterior samples of staged trees are then used to estimate causal quantities such as
\[
\text{ATE} = \mathbb{E}[Y(1)-Y(0)],
\qquad
\widehat{\text{ATE}} = \frac{1}{M}\sum_{m=1}^M \text{ATE}^{(m)}.
\]
This approach is designed to propagate both parametric and structural uncertainty into causal effect estimates [2511.03399].

Interventional staged tree models extend the framework still further. Soft interventions are represented by partitioning labels into intervention-regime labels and ordinary conditional-probability labels, leading to models that live in products of simplices. For DAG-based interventional models, the factorization
\[
f^{(I)}(x) = \prod_{j\in I} f^{(I)}(x_j\mid x_{\operatorname{pa}(j)}) \prod_{j\notin I} f(x_j\mid x_{\operatorname{pa}(j)})
\]
is realized by an interventional staged tree, and balancedness again characterizes toricity of the defining ideal [2012.03593].

Predictive modeling is another important application. Staged tree classifiers learn the joint model \(p(c,\mathbf x)\) for a class variable and features, with the class usually placed at the root. The naive staged tree classifier is defined so that each level has exactly \(|\mathbb C|\) stages, giving the same parameter count as naive Bayes while strictly enlarging the model class. In simulation studies, staged tree classifiers were reported to be competitive with state-of-the-art classifiers, and the naive staged tree was able to capture parity/XOR-like structure that naive Bayes could not [2012.13798].

For discrete longitudinal data, staged trees exploit the natural temporal order but face severe graph growth and sparsity. Two remedies have been proposed: longitudinal staged trees with Markov assumptions, which use a DAG to impose initial conditional independences before staged-tree refinement, and marginal longitudinal staged trees, which model selected margins such as \((Z,X_t,Y_t)\) rather than the full joint distribution. These constructions trade flexibility against tractability and interpretability, especially when repeated observations create many zero-count or weakly supported contexts [2401.04297].

Applied work has emphasized domains where asymmetric dependence and constrained sample spaces are intrinsic. In health data, staged trees have been used to analyze invasive fungal infections, where they revealed non-observed or impossible combinations such as patients who do not enter ICU cannot be intubated, and they produced richer dependence summaries than logistic regression or Bayesian networks [2307.16301]. In transport-service evaluation, bootstrap-based order aggregation, bootstrap stage aggregation, cross-validation, and what-if sensitivity analysis have been developed for staged trees and their ALDAG summaries, allowing robustness assessment and intervention-style scenario analysis in large survey datasets [2401.01812].

Across these developments, a common theme persists: staged tree models preserve full joint-distribution semantics while replacing globally symmetric dependence assumptions by stage-wise equality constraints that can depend on context, path, and event structure. This places them between Bayesian networks and richer event-based graphical models, with algebraic, statistical, and algorithmic theories that now extend from decomposable and toric subclasses to learning, scoring, interventions, causality, longitudinal analysis, and domain-specific applications [2210.11521].

Source: https://www.emergentmind.com/topics/staged-tree-models