SCM Prior: Structural Causal Models

• SCM priors are frameworks that explicitly define causal structures by specifying graphical, parameter, functional, and latent constraints.
• They facilitate robust Bayesian inference, synthetic data generation, and out-of-distribution generalization by integrating domain knowledge.
• Advanced techniques like the Causal Knowledge Hierarchy combine expert, data-driven, and literature-derived priors to optimize causal discovery and model selection.

A Structural Causal Model (SCM) prior specifies assumptions about the underlying causal structure, parameters, and variables governing a system. SCM priors are used as foundational components in causal learning, Bayesian inference, robust prediction, and synthetic data generation. By encoding domain knowledge or structural constraints, they play a critical role in disambiguating causal discovery, improving learning efficiency, regularizing inference, and enabling out-of-distribution generalization.

1. Mathematical Foundation of Structural Causal Model Priors

An SCM is formalized as a tuple $M = \langle U, V, F, P(U) \rangle$, where $U = \{U_1,...,U_k\}$ are exogenous variables (background factors), $V = \{X_1,...,X_n\}$ are endogenous variables, $F = \{f_1,...,f_n\}$ are structural functions (each $X_i = f_i(PA_i, U_i)$ where $PA_i \subset V \setminus \{X_i\}$), and $P(U)$ is the joint distribution of exogenous variables. The causal relationships among $V$ are encoded as a directed acyclic graph (DAG) $G$ (Adib et al., 2022).

An SCM prior can be specified in several layers:

• Graphical prior: Specifies allowable, required, or prohibited edges (causal relationships) in $G$.
• Parameter prior: Regularizes the structural parameters (e.g., edge weights, noise variances) of the SCM.
• Functional prior: Constrains the form of the structural functions $U = \{U_1,...,U_k\}$0 (e.g., linear, nonlinear, neural).
• Latent variable prior: In generative settings with unobserved variables, places a prior directly over possible realizations.

This formal structure enables both hard constraints—such as whitelisting certain edges or fixing functional forms—and soft constraints, as in Bayesian settings with heavy-tailed or sparsity-inducing distributions (Subramanian et al., 2022).

2. SCM Priors in Bayesian Latent Causal Inference

In scenarios where high-level causal variables, parameters, and structure are all unobserved, SCM priors enable tractable Bayesian inference from low-level data (e.g., images, high-dimensional vectors). In the latent SCM framework for linear Gaussian additive-noise models, the prior decomposes as follows (Subramanian et al., 2022):

1. Prior over DAG structure $U = \{U_1,...,U_k\}$1: $U = \{U_1,...,U_k\}$2 is parameterized by a permutation matrix $U = \{U_1,...,U_k\}$3 (node ordering) and strictly lower-triangular weight matrix $U = \{U_1,...,U_k\}$4, with adjacency $U = \{U_1,...,U_k\}$5. The prior $U = \{U_1,...,U_k\}$6 is uniform over all $U = \{U_1,...,U_k\}$7 permutations, while $U = \{U_1,...,U_k\}$8 assigns nonzero probability only to acyclic graphs: $U = \{U_1,...,U_k\}$9.
2. Prior over structural parameters $V = \{X_1,...,X_n\}$0 given $V = \{X_1,...,X_n\}$1: For nonzero $V = \{X_1,...,X_n\}$2, a horseshoe prior is used: $V = \{X_1,...,X_n\}$3, $V = \{X_1,...,X_n\}$4. The global scale $V = \{X_1,...,X_n\}$5 controls sparsity. Noise variances $V = \{X_1,...,X_n\}$6 are regularized with a log-Gaussian: $V = \{X_1,...,X_n\}$7.
3. Prior over latent causal variables $V = \{X_1,...,X_n\}$8 given $V = \{X_1,...,X_n\}$9: Each sample $F = \{f_1,...,f_n\}$0 is generated by the structural equations, i.e., $F = \{f_1,...,f_n\}$1, with $F = \{f_1,...,f_n\}$2.

Combined, the joint prior is $F = \{f_1,...,f_n\}$3. This enters variational inference through a KL penalty in the ELBO, governing joint inference over structure, parameters, and variables (Subramanian et al., 2022).

In these models, the uniform prior over permutations avoids bias toward any ordering, the horseshoe prior encourages sparse, interpretable graphs, and log-Gaussian priors on noise variances prevent degenerate solutions. These design choices support both identifiability and generalization out-of-distribution.

3. Encoding and Aggregating Priors: The Causal Knowledge Hierarchy

The Causal Knowledge Hierarchy (CKH) framework introduces a methodical abstraction for encoding and combining SCM priors derived from diverse information sources (Adib et al., 2022). Prior knowledge is organized into three levels:

• CK_E: Expert opinion
• CK_D: Data-driven structure learning algorithms
• CK_L: Literature mining (published research, text mining)

Each tier receives a reliability weight ($F = \{f_1,...,f_n\}$4, $F = \{f_1,...,f_n\}$5, $F = \{f_1,...,f_n\}$6), typically $F = \{f_1,...,f_n\}$7 and $F = \{f_1,...,f_n\}$8. For each unordered pair $F = \{f_1,...,f_n\}$9, each tier produces edge-confidences $X_i = f_i(PA_i, U_i)$0, combining voting frequencies, inter-rater agreement (e.g., Fleiss' $X_i = f_i(PA_i, U_i)$1), and tier weight.

The target SCM is constructed by convexly combining these sources, enforcing hard constraints for strong expert priors and soft penalties/bonuses for weaker, ambiguous priors during structure search. The Bayesian-style objective maximizes $X_i = f_i(PA_i, U_i)$2, integrating priors at multiple reliability levels. This aggregation yields robust graph identification even with conflicting or partially incorrect priors, as demonstrated by empirical true positive rates $X_i = f_i(PA_i, U_i)$3 under adversarial perturbations (Adib et al., 2022).

Summary Table: CKH Edge Confidence Computation

Tier	Evidence Type	Weight ($X_i = f_i(PA_i, U_i)$4)	Confidence Calculation
CK_E	Expert opinion	$X_i = f_i(PA_i, U_i)$5	Vote aggregation $X_i = f_i(PA_i, U_i)$6 IRR $X_i = f_i(PA_i, U_i)$7
CK_D	Data-driven	$X_i = f_i(PA_i, U_i)$8	Score-based or constraint-based $X_i = f_i(PA_i, U_i)$9
CK_L	Literature/NLP	$PA_i \subset V \setminus \{X_i\}$0	NLP-extracted edges $PA_i \subset V \setminus \{X_i\}$1 IRR $PA_i \subset V \setminus \{X_i\}$2

The CKH methodology allows for principled robustness and scalability, while failure modes include over-dominance by incorrect priors in a tier or subjectivity in tier weighting.

4. Role of SCM Priors in Robust Model Selection and Hypothesis Testing

SCM priors serve as constraints in model selection, promoting generalization and robustness to covariate shifts by enforcing domain-invariant dependencies. The Causal Assurance Metric (CAM) leverages an SCM prior $PA_i \subset V \setminus \{X_i\}$3 (possibly incomplete) to rank models by a composite score: $PA_i \subset V \setminus \{X_i\}$4, where $PA_i \subset V \setminus \{X_i\}$5 quantifies log-likelihood under the SCM and $PA_i \subset V \setminus \{X_i\}$6 is standard loss (e.g., MSE) (Kyono et al., 2019).

A candidate model is favored if its predictions adhere to the conditional dependencies of $PA_i \subset V \setminus \{X_i\}$7, even under domain or distribution shift. Empirical assessments demonstrate that selection by CAM reduces out-of-distribution error by $PA_i \subset V \setminus \{X_i\}$8–$PA_i \subset V \setminus \{X_i\}$9 compared to pure statistical ranking, and maintains robustness even as the available known prior edges decrease to $P(U)$0–$P(U)$1 of the full graph (Kyono et al., 2019). Failure modes occur when the causal graph is seriously mis-specified, dominates with incorrect edges, or confounders violate acyclicity assumptions.

SCM priors can also serve as hypotheses in comparative settings. Structural priors are encoded as binary adjacency matrices $P(U)$2, where $P(U)$3 is the adjacency (allowable causal edges) and $P(U)$4 marks exogenous nodes (Jiang et al., 2022). Models are compared based on out-of-distribution reconstruction error, using non-i.i.d. train/test splits along specific dimensions—a hypothesis test preferring the structure with lowest OOD loss across splits (Jiang et al., 2022). This methodology enables statistical selection and synthesis of causally compliant synthetic data.

5. Structural Causal Priors in Time Series Forecasting

Structural causal priors are extended to temporal domains via the Time-Series Structural Causal Model (TS-SCM) framework (Chen et al., 5 Jun 2026). TS-SCM defines a directed causal graph $P(U)$5, with nodes representing univariate time series $P(U)$6. Nodes are partitioned into root (exogenous) and affected (endogenous) variables, with edges governing lagged, weighted, and possibly nonlinear causal effects.

The update equation for each affected node is

$P(U)$7

where $P(U)$8 aggregates parent messages with histories, lags, edge transforms, and noise. Lags can be fixed or dynamic, and distributional drift is accommodated by resampling model and noise parameters across intervals. This structure enables the generation of synthetic forecasting problems exhibiting cross-variable dependencies, feedback, and nonstationarity, mirroring real-world scenarios (Chen et al., 5 Jun 2026).

In practice, TS-SCM-generated episodes are used to pretrain forecasting models (e.g., Trio) by supplying synthetic problems that transfer structural knowledge to the model. Empirical results show 10–20% reductions in forecasting error on real-world benchmarks when using TS-SCM pretraining relative to less structured synthetic priors. Ablations confirm that TS-SCM covers diverse temporal structures—lags, feedback, regime shifts—improving zero-shot transfer and out-of-distribution resilience (Chen et al., 5 Jun 2026).

6. Practical Construction and Integration of SCM Priors

The specification and integration of SCM priors in applications involve:

• Encoding the prior: Directly constructing adjacency matrices, parameter priors, or higher-level constraints from expert knowledge, structure learning algorithms, or computational extraction from literature/databases (Adib et al., 2022).
• Aggregating multiple priors: Convex combination of sources via CKH, with empirical or justified assignment of reliability weights.
• Hard and soft constraints: Tiers or confidence levels dictate whether an edge is strictly enforced or weighs as a soft penalty/bonus in search or inference (Adib et al., 2022).
• Bayesian or hybrid inference: SCM priors enter directly into the ELBO in Bayesian settings, or penalize score-based search/inference procedures in non-Bayesian algorithms (Subramanian et al., 2022).
• Synthetic data generation and hypothesis selection: SCM priors serve as blueprints for generating synthetic observations, with structural plausibility imposed through the model's generative mechanism (Jiang et al., 2022, Chen et al., 5 Jun 2026).

These methods are robust to small- to medium-scale datasets and can be adapted to time series, image, and high-dimensional modalities, given algorithmic and computational scalability.

7. Implications, Limitations, and Open Challenges

SCM priors provide a principled mechanism for encoding domain knowledge, mitigating identifiability problems inherent in purely observational causal discovery, and regularizing model capacity in high-dimensional or small-sample regimes. Critical success factors include the correctness and informativeness of the prior, the robustness of aggregation schemes (as in CKH), and careful regularization when only partial prior information is available.

Limitations arise from possible dominance of erroneous priors, subjectivity in tier weighting, unknown functional forms or noise distributions, and scalability as the number of variable pairs grows combinatorially (Adib et al., 2022). Automatic extraction of high-quality priors from literature remains an open problem. Extensions to continuous-time and dynamic SCMs, posterior integration over causality graphs, and differentiable causal regularizers at training time are active research directions.

In summary, SCM priors—spanning structure, parameters, and latent variables—form an essential substrate for causal discovery, model selection, robust inference, and synthetic data generation across a range of domains and data modalities (Subramanian et al., 2022, Adib et al., 2022, Jiang et al., 2022, Kyono et al., 2019, Chen et al., 5 Jun 2026).