In traditional causal discovery, the goal is typically to find the "best" Directed Acyclic Graph (DAG) that explains a dataset. However, relying on a single point estimate is often misleading. Observational data is frequently noisy, limited, or mathematically compatible with multiple conflicting causal structures. As Henao and Winther note, outputting a single graph obscures the degree of confidence in individual edges, leading to overconfident conclusions in high-stakes domains like medicine or policy-making.
The emerging research direction of causal uncertainty quantification shifts the focus from finding one optimal model to estimating a full posterior distribution over possible causal structures and effects.
The Problem of Structural Ambiguity
Causal discovery faces a fundamental challenge: identifiability. Often, multiple graphs belong to the same Markov equivalence class, meaning they imply the same statistical dependencies and are indistinguishable from observational data alone. Annadani et al. argue that point-estimate optimization methods fail to account for this epistemic uncertainty, ignoring the existence of multiple plausible models that fit the data equally well.
By adopting a Bayesian approach, researchers can represent this ambiguity directly. For instance, Dhir et al. demonstrate that amortized Bayesian inference models can correctly assign equal probability to competing causal directions when the data is unidentifiable. This prevents the model from "collapsing" to a single, potentially incorrect, point estimate.
Sampling the Space of Graphs
Estimating a posterior distribution over DAGs is computationally difficult because the space of possible graphs grows exponentially with the number of variables. Recent work has introduced several mechanisms to make this tractable:
- Variational Inference: Cundy et al. introduced BCD Nets, which use continuous relaxations of permutations to estimate distributions over causal graphs. Similarly, Thompson et al. propose ProDAG, which projects continuous distributions onto the space of sparse, acyclic matrices to provide well-calibrated confidence intervals for edge predictions.
- Generative Flow Networks: Deleu et al. developed DAG-GFlowNet, treating graph construction as a sequential decision process. This allows for efficient sampling of diverse, high-scoring, and strictly acyclic graphs, providing a representative set of candidates rather than a single "best" solution.
- Differentiable Sampling: To speed up discovery, Charpentier et al. introduced DP-DAG, a framework for fast, differentiable sampling that guarantees acyclicity at every training step, enabling edge-specific confidence scores.
Uncertainty in Causal Effects
Quantifying uncertainty is not just about the graph structure; it is also critical for estimating Conditional Average Treatment Effects (CATE). In high-stakes environments, knowing that a treatment works is as important as knowing how sure the model is about the result.
Frauen et al. argue that standard models often prioritize numerical accuracy over decision reliability. They suggest that integrating uncertainty estimation is essential for identifying cases where a model is unsure about a treatment's sign (e.g., whether a drug will help or harm), preventing dangerous real-world decisions.
Furthermore, when a causal query is non-identifiable, Xia and Pan show that optimization-based neural causal models can identify a range of possible outcomes. This provides a more robust representation of uncertainty than traditional methods that might output a single, confident, but incorrect value.
From Discovery to Active Learning
A major benefit of quantifying causal uncertainty is the ability to perform Active Bayesian Causal Inference (ABCI). As Knoll et al. describe, by marginalizing over the posterior distribution of both structures and functional relationships, systems can identify the most informative experiments to perform.
This creates a feedback loop: the model identifies which parts of the causal web are most ambiguous (Hägele et al.), and the researcher performs targeted interventions to resolve that specific uncertainty. This data-efficient process ensures that scientific discovery is guided by the limits of current knowledge rather than trial and error.
Key Foundational Work
Bayesian Sparse Factor Models and DAGs Inference and Comparison (2009) — contributed 2 concept elements to this direction.
Bayesian Discovery of Linear Acyclic Causal Models (2009) — contributed 2 concept elements to this direction.