Summary Causal Graphs

• Summary causal graphs are graph-based abstractions that consolidate time-indexed variables into macro-nodes representing entire time series or variable clusters.
• They depict causal connections using directed and bidirected edges to indicate any influence or latent confounding across unknown temporal offsets.
• SCGs facilitate causal effect estimation in complex systems where detailed temporal data is unavailable, supporting robust analysis with limited data granularity.

A summary causal graph is a graph-based abstraction that represents potential causal relationships between variables—most commonly time series—while omitting exact temporal details such as lag times and, frequently, allowing for cycles and confounding. Unlike fully specified dynamic causal graphs (such as full-time acyclic directed mixed graphs or FTCGs), each node (often called a macro-node) in a summary causal graph (SCG) denotes an entire time series or variable cluster, and edges reflect the existence of some causal connection at some (potentially unknown) temporal offset. This structure is designed for settings where constructing or obtaining a granular, temporally-indexed causal model is infeasible, and only aggregate knowledge of causal interactions is available or can be detected.

1. Fundamental Concepts and Formal Definition

Summary causal graphs are obtained by “collapsing” all temporally-indexed variants of variables from a fully specified dynamic graph into single macro nodes. The resulting edge set includes a directed edge $X \rightarrow Y$ if there exists any lagged or contemporaneous edge from a component of $X$ to a component of $Y$ in the underlying fine-grained (micro) graph. Analogously, a bidirected edge $X \leftrightarrow Y$ denotes possible latent confounding—that is, cases where $X$ and $Y$ have a shared (possibly hidden) cause at some temporal point.

Formally, from a full-time causal graph ($\mathbb{V}, \mathbb{E}$), the associated SCG has:

• Nodes: The set of time series or clusters $\mathbb{S} = \{ X, Y, \ldots \}$, each summarizing all time-indexed variables $X_{t}$.
• Directed Edges ($\mathbb{E}^{s1}$): $X \rightarrow Y$ if $$\exists t', t : X_{t'} \rightarrow Y_{t} \in \mathbb{E}$$.
• Bidirected Edges ($\mathbb{E}^{s2}$): $X \leftrightarrow Y$ if $X_{t'}$ and $Y_{t}$ have unmeasured shared causes.

This abstraction allows multiple (possibly infinitely many) full-time graphs to correspond to the same SCG.

2. Rationale and Practical Motivation

SCGs are constructed for contexts where:

• The precise temporal dynamics of causal influence are unknown, too complex, or measured at too coarse a resolution.
• Domain experts can provide only high-level assertions (“X affects Y”) without temporal specificity.
• Data availability is limited to aggregate or cross-sectional representations, as often occurs in register-based epidemiological studies or system monitoring applications.
• Feedback or cyclic effects exist between variables (e.g., diseases, economic indicators), violating classical requirements of acyclic temporal models.

SCGs help reduce cognitive load when reasoning about highly interconnected systems by summarizing essential causal pathways without overcommitting to unverified temporal details.

3. Identifiability and Causal Effect Estimation

A central question for SCGs is the identifiability of causal effects: under what circumstances can the effect of interventions be uniquely expressed and estimated from observational data, given only the SCG? This involves determining the existence of adjustment sets, application of the front-door/back-door criteria, and construction of do-free formulas (i.e., formulas not involving interventions).

Key results include:

• Adjustment via Common Sets: A set of variables qualifies as a “common adjustment set” if it blocks all backdoor paths in every compatible underlying micro-level graph, while not conditioning on descendants of interventions. Sound and complete graphical criteria now exist for multiple interventions and time series (2506.14534).
• Front-Door Criterion Extensions: When adjustment is infeasible (e.g., due to latent confounders), sufficient conditions for total effect identifiability are available by extending the front-door criterion to SCGs; these conditions involve intercepting all directed paths from $X$ to $Y$ via a mediator and blocking all backdoor paths into the mediator or effect (2406.05805).
• Do-Calculus and Macro Effects: For macro-level queries (effects of interventions applied to entire clusters), d-separation and do-calculus are both sound and complete tools in the SCG (2407.07934), even with cycles and latent confounding.
• Adjustment Algorithmics: Efficient, pseudo-linear algorithms have been developed for checking if a given causal effect is identifiable by adjustment in an SCG, crucial for high-dimensional systems (2506.14534, 2506.14862).

4. Graphical Criteria and Algorithms

The existence of cycles, overlapping descendants, and hidden confounding introduces new complexity to identifiability in SCGs. Recent research has established:

• Forbidden Sets and Descendant Cones: Adjustment sets must be disjoint from the “cone of descendants” (the union of all possible descendants of interventions across all consistent micro-graphs). If a collider-free backdoor path exists within this cone, identifiability fails.
• Common Backdoor Set Algorithm: Algorithms efficiently compute cones of descendants and check for such dangerous paths, enabling practitioners to decide identifiability on large graphs without examining all micro-level realisations (2506.14862).
• Role of Consistency Through Time: If the causal structure does not change over time (stationarity), identifiability conditions often become easier to check and are more likely to hold.

5. Applications and Implications

SCGs have been deployed for:

• Root Cause Analysis in IT and Engineering Systems: Estimating the effect of component interventions in complex infrastructures when only coarse-grained system diagrams are available (2306.16958).
• Epidemiology and Medicine: Assessing the direct or total effect of exposures on outcomes, even where detailed temporal data is unavailable or cycles (bidirectional feedback) occur (e.g., kidney function and hypertension, disease spread models) (2310.14691, 2407.07934, 2406.05805).
• Finance and Economics: Analyzing policy or intervention effects in macroeconomic models with feedbacks or incomplete data (2310.14691).

These applications benefit from explicit adjustment guidelines, robustness to unmeasured confounders, and algorithmic efficiency in effect identification even in high-dimensional and partially specified settings.

6. Limitations and Open Challenges

Despite their utility, SCGs present certain challenges:

• Ambiguity and Overlap: Many different micro-level graphs may be consistent with a given SCG—this can both protect against overconfident inference and make adjustment sets overly large or estimates high-variance.
• Cycles and Non-identifiability: The presence of cycles means that certain effects (especially micro-level or path-specific effects) can be non-identifiable, even without classical “hedges”.
• Expressiveness: SCGs are best suited for macro-level (cluster- or time series-based) queries. Micro-level queries, or those requiring precise temporal information, may be non-identifiable or require stronger assumptions (2407.07934).
• Algorithmic Complexity in Edge Cases: While most identifiability checks are efficient, worst-case complexity may grow for highly entangled graphs or multifactorial interventions.

7. Future Directions

Ongoing research targets:

• Generalizing Do-Calculus and Front-Door Criteria: Extending complete identification criteria to SCGs with additional types of interventions or partial knowledge.
• Refining Adjustment Set Minimality: Developing algorithms to deliver the smallest possible adjustment sets, improving estimator variance and interpretability.
• Integration with Discovery and ML Pipelines: Fusing SCG-based reasoning with estimation, structure discovery, and large-scale data science workflows in dynamic and heterogeneous environments.
• Informative Summarization: Embedding SCGs and their adjustment criteria into tools (e.g., graph databases, data analysis platforms) to facilitate robust causal queries without sacrificing interpretability (2412.13965).

---

Summary Table: Key Aspects of Summary Causal Graphs

Aspect	Details
Node semantics	Macro-level (time series or clusters), not single variables
Edge semantics	Aggregated over all lags; may involve cycles or confounding
Identifiability tools	Adjustment sets, front-door, do-calculus, macro d-separation
Key applications	Epidemiology, IT, finance, large-scale dynamic systems
Algorithmic status	Efficient, complete algorithms for many identifiability tasks
Limitations	Not all effects identifiable; large adjustment sets possible

Summary causal graphs provide a flexible, theoretically grounded, and computationally practical means for representing and reasoning about causal effects in complex, dynamic systems where granular causal information is unavailable or incomplete. Recent work has established sound and complete algorithms for effect identifiability—especially via adjustment and generalized “backdoor” methods—making SCGs a central tool for robust, large-scale causal inference.