1 Foundations

Causal discovery is the task of learning cause-and-effect structure from data. Rather than asking only whether two variables move together, it seeks to identify which variables may influence others and under what conditions those influences hold. The field draws on statistics, computer science, philosophy of science, and domain knowledge to distinguish direct causal relations from indirect associations.

1.1 Definition and scope

Causal discovery focuses on uncovering causal links, causal direction, and sometimes the broader network of influences among variables. Its scope includes learning from observational data, controlled experiments, or a mixture of both. In practice, the aim is often to recover a causal graph or a comparable representation that supports explanation, prediction under intervention, and scientific understanding.

1.2 Correlation versus causation

Correlation describes statistical association, whereas causation implies that changing one variable can produce a change in another. Two variables may be correlated because one causes the other, because both are affected by a third factor, or because the relationship is indirect. Causal discovery methods are designed to separate these possibilities more carefully than ordinary association analysis.

1.3 Causal inference and causal discovery

Causal inference estimates the effect of a specified cause on an outcome, usually with a model or design already in mind. Causal discovery is broader in that it attempts to infer the structure itself when the causal relations are not known in advance. The two areas are closely connected: discovery can suggest candidate models, while inference can quantify the effects implied by those models.

1.4 Counterfactual reasoning

Counterfactual reasoning asks what would have happened under a different condition, such as whether an outcome would still occur if a cause were absent. This kind of reasoning is central to many causal frameworks because it captures notions of hypothetical intervention and alternative histories. In causal discovery, counterfactual concepts help formalize what it means for one variable to be a genuine cause of another.

2 Historical development

Causal discovery emerged from multiple intellectual traditions rather than from a single origin. Its modern form was shaped by philosophical work on causation, by statistical methods for studying dependence, and later by computational approaches that made larger and more complex models feasible.

2.1 Early philosophical roots

Questions about causation have long occupied philosophy, especially in discussions of explanation, necessity, and the limits of observation. Early thinkers explored how causes might be distinguished from mere regularities in experience. These debates provided conceptual foundations for later formal approaches to causal structure.

2.2 Development in statistics

Statistics contributed methods for studying dependence, regression, hypothesis testing, and experimental design. Over time, researchers recognized that standard statistical association did not by itself identify causal direction. Work on graphical models, conditional independence, and structural equations helped formalize causal questions in a way that could be addressed mathematically.

2.3 Influence of artificial intelligence and machine learning

Artificial intelligence and machine learning expanded the scale and automation of causal discovery. Search algorithms, optimization methods, and probabilistic modeling made it possible to analyze larger variable sets and more intricate dependencies. These fields also encouraged the use of data-driven techniques that infer candidate causal structures with limited human specification.

3 Causal models

Causal discovery relies on formal representations of how variables affect one another. These models provide the language in which assumptions are stated, algorithms are defined, and results are interpreted.

3.1 Structural causal models

Structural causal models represent variables through equations that specify how each is generated from its causes and an exogenous error term. They are useful because they can encode both causal relations and the effect of interventions. In this framework, changing a variable means altering the equation that produces it.

3.2 Directed acyclic graphs

Directed acyclic graphs are diagrams in which arrows indicate direct causal influence and cycles are absent. They offer an intuitive and mathematically tractable way to describe causal structure. Such graphs are especially common in discovery methods because they support tests of conditional independence and other structural properties.

3.3 Potential outcomes framework

The potential outcomes framework describes causal effects by comparing what would happen under different treatment assignments. Each unit is associated with possible results under alternative interventions, though only one is observed in practice. This perspective is widely used in causal inference and also informs discovery when the goal is to understand treatment relationships among variables.

3.4 Causal Bayesian networks

Causal Bayesian networks combine probabilistic graphical models with causal interpretation. They represent variables as nodes with directed edges and attach probability distributions that reflect dependencies. When coupled with causal assumptions, they allow reasoning about both observation and intervention.

4 Assumptions and principles

Any causal discovery method depends on assumptions that connect data patterns to causal structure. These principles are not always guaranteed in real data, but they make the problem mathematically identifiable and algorithmically tractable.

4.1 Causal Markov condition

The causal Markov condition states that each variable is conditionally independent of its non-effects given its direct causes. This principle links graph structure to observable statistical independencies. It is one of the main bridges between a causal model and the data used to estimate it.

4.2 Faithfulness and stability

Faithfulness assumes that the conditional independencies seen in the data arise from the causal structure itself rather than from accidental numerical cancellation. Stability refers to the idea that small perturbations in the data or model should not completely alter the inferred structure. Together, these ideas help justify why detected independencies can be treated as meaningful clues about causation.

4.3 Causal sufficiency

Causal sufficiency means that all common causes of the measured variables are themselves measured. When this assumption fails, hidden variables can create spurious associations that mislead discovery procedures. Many methods therefore state explicitly whether they require causal sufficiency or can accommodate latent confounding.

4.4 Interventions and invariance

Interventions change a system by altering one part of the causal mechanism while leaving other mechanisms intact. Invariance refers to the expectation that certain causal relationships remain stable across conditions or environments. These ideas are useful both for identifying causal direction and for checking whether a proposed model generalizes beyond the data on which it was built.

5 Methods of causal discovery

Causal discovery methods can be grouped by the kind of information they extract from data and the assumptions they use. Some rely on patterns of conditional independence, others on scoring candidate graphs, and others on functional restrictions that distinguish directionality.

5.1 Constraint-based methods

Constraint-based methods infer causal structure by testing which independence relations are present in the data. They then use those results to narrow the set of graphs consistent with the observed distribution.

5.1.1 Conditional independence testing

Conditional independence testing asks whether two variables remain unrelated after conditioning on a third set of variables. If so, the test can suggest that a direct causal link is absent or that the variables share a particular graphical relationship. The reliability of these tests depends on sample size, noise, and the correctness of model assumptions.

5.1.2 PC algorithm

The PC algorithm is a well-known procedure that begins with a fully connected graph and removes edges based on conditional independence tests. It then orients some edges using logical rules that preserve consistency with the observed independencies. The result is usually a partially directed graph representing an equivalence class of plausible causal models.

5.1.3 FCI algorithm

The FCI algorithm extends constraint-based discovery to settings with possible hidden confounders. It aims to recover a more cautious graphical representation when not all common causes are observed. Because it addresses latent variables, the output is often less specific than that of methods designed for fully measured systems.

5.2 Score-based methods

Score-based methods search over candidate graphs and evaluate each one with a scoring criterion. The preferred graph is the one that best balances goodness of fit with model simplicity.

5.2.1 Bayesian scoring approaches

Bayesian scoring approaches compare graphs using posterior probability or related quantities. They incorporate prior beliefs about structure and update them with evidence from the data. This framework is attractive because it naturally handles uncertainty and model comparison.

Greedy equivalence search explores the space of graph equivalence classes by making local improvements to a score. It moves step by step toward a better model, rather than examining every possibility. This makes it practical for problems where exhaustive search would be infeasible.

5.2.3 Regularization and model selection

Regularization adds a penalty for complexity so that overly dense graphs are discouraged. Model selection criteria help choose among candidates by weighing explanatory power against parsimony. These tools are especially important in high-dimensional settings where many structures can fit the data nearly equally well.

5.3 Functional causal model methods

Functional causal model methods use assumptions about the form of the causal mechanism to identify direction more directly. They often exploit asymmetries in noise, nonlinearity, or distributional shape.

5.3.1 Additive noise models

Additive noise models assume that each effect is a function of its causes plus an independent noise term. Under suitable conditions, this structure can make causal direction identifiable. These models are valued because they are simple, interpretable, and often effective in practice.

5.3.2 LiNGAM

LiNGAM, short for linear non-Gaussian acyclic model, uses linear relationships and non-Gaussian noise to infer causal ordering. The non-Gaussian assumption gives directionality that ordinary linear models may lack. It has become a standard example of how distributional assumptions can resolve ambiguity.

5.3.3 Nonlinear and non-Gaussian approaches

More general methods extend beyond linearity and Gaussian error terms. By exploiting nonlinear relations or richer noise structures, they can sometimes identify causal effects in situations where simpler models fail. These approaches are flexible but may be harder to estimate and interpret.

5.4 Hybrid methods

Hybrid methods combine ideas from more than one discovery strategy. They are designed to capture the strengths of different approaches while reducing their individual weaknesses.

5.4.1 Combining constraints and scores

Some algorithms first use independence tests to narrow the search space and then apply scoring to rank candidate structures. This combination can improve efficiency and robustness. It also allows the method to use both logical constraints and quantitative fit measures.

5.4.2 Bootstrap and ensemble techniques

Bootstrap and ensemble techniques repeat discovery over resampled data or across multiple models. They help assess which edges are stable and which may be sample-specific artifacts. Such approaches are often used to summarize uncertainty in complex analyses.

6 Data and experimental design

The quality of causal discovery depends heavily on the data source and on how the study is designed. Observational and experimental settings present different opportunities and limitations, and both can be affected by missingness, measurement noise, and temporal structure.

6.1 Observational data

Observational data are collected without deliberate intervention by the analyst. They are common in naturalistic studies, administrative records, and many scientific databases. Because they reflect preexisting conditions rather than randomized changes, they usually require stronger assumptions for causal discovery.

6.2 Experimental data and interventions

Experimental data arise when variables are manipulated under controlled conditions. Interventions can reveal causal direction more directly than passive observation because they alter the system in a planned way. Even so, experiments may be limited by ethical, practical, or cost constraints.

6.3 Missing data and measurement error

Missing values and measurement error can distort the dependencies on which causal discovery relies. If not handled carefully, they may create false links, hide real ones, or change the apparent direction of influence. Methods often need preprocessing, imputation, or explicit error modeling to remain reliable.

6.4 Time series and longitudinal data

Time series and longitudinal data track variables across repeated observations. Temporal ordering can supply valuable clues about causation, since earlier events may influence later ones. However, delayed effects, feedback, and changing conditions can make the analysis more complex than in cross-sectional data.

7 Algorithms and computational issues

Many causal discovery problems are computationally demanding because the number of possible graphs grows rapidly with the number of variables. Practical methods therefore rely on efficient search, approximation, and careful uncertainty handling.

7.1 Graph search strategies

Graph search strategies explore candidate structures using heuristics, local moves, or pruning rules. They try to avoid the combinatorial explosion of checking every graph. The choice of strategy affects both runtime and the likelihood of finding a good solution.

7.2 Scalability and high-dimensional data

High-dimensional data pose special difficulties because the number of variables may be large relative to the sample size. In such settings, methods must reduce complexity through sparsity assumptions, feature screening, or restricted search spaces. Scalability is often a deciding factor in whether a method can be used on real-world problems.

7.3 Computational complexity

Causal discovery can be computationally hard, especially when many graph structures are compatible with the same data. Some subproblems are known to grow rapidly with the number of variables, making exact search impractical. As a result, many algorithms favor approximate solutions that are faster but not guaranteed to be optimal.

7.4 Uncertainty estimation

Uncertainty estimation quantifies how confident one should be in an inferred edge, orientation, or model class. It can be based on resampling, Bayesian posterior probabilities, or asymptotic approximations. Reporting uncertainty is important because causal conclusions are often sensitive to sampling variation and modeling choices.

8 Evaluation and validation

Because causal discovery aims to recover structure that is not directly observed, evaluation is challenging. Researchers therefore use simulation, benchmarks, sensitivity checks, and external validation to assess performance.

8.1 Simulation studies

Simulation studies generate synthetic data from a known causal model and then test whether an algorithm can recover it. They are useful for comparing methods under controlled conditions. Their limitation is that simulated settings may not fully represent the complexity of real data.

8.2 Ground truth benchmarks

Ground truth benchmarks use datasets or systems in which the causal structure is at least partly known. These may come from well-studied scientific domains, curated tasks, or controlled experiments. Benchmarks provide a common reference for comparing methods, though true causal ground truth is often incomplete.

8.3 Sensitivity analysis

Sensitivity analysis examines how results change when assumptions, parameters, or preprocessing choices are altered. It helps identify fragile conclusions and can reveal which parts of a model are most dependent on unverified conditions. This is especially useful when data quality is uneven or hidden confounding is possible.

8.4 Out-of-sample validation

Out-of-sample validation tests whether a learned causal model performs well on new data. A structure that generalizes may be more credible than one that only fits the original sample. Validation is particularly valuable when the intended use involves prediction under new conditions or interventions.

9 Applications

Causal discovery is used wherever scientists and analysts want to move beyond association toward mechanism. Its applications span laboratory science, health research, social analysis, and computational systems.

9.1 Biology and genomics

In biology and genomics, causal discovery can help identify regulatory relationships among genes, proteins, and cellular processes. It is often used to propose networks that explain how biological signals propagate. These models may guide laboratory experiments by suggesting promising intervention targets.

9.2 Medicine and epidemiology

In medicine and epidemiology, causal discovery supports the study of risk factors, treatment pathways, and disease progression. It can help organize complex clinical data into interpretable structures. The field is particularly useful when randomized experiments are limited or when multiple interacting variables influence outcomes.

9.3 Economics and finance

In economics and finance, causal discovery may be applied to markets, policy variables, risk factors, and macroeconomic indicators. Analysts use it to explore how changes in one quantity might affect others over time. The methods can aid forecasting and scenario analysis, though results depend strongly on modeling assumptions.

9.4 Social and behavioral sciences

Social and behavioral sciences use causal discovery to study relationships among attitudes, actions, institutions, and environments. These settings often involve many interdependent variables and imperfect measurement. Discovery methods can help generate hypotheses about social mechanisms and inform follow-up research.

9.5 Computer science and artificial intelligence

In computer science and artificial intelligence, causal discovery is used for representation learning, system diagnosis, recommendation, and robust decision-making. It also helps address questions of fairness, transferability, and distribution change. More broadly, causal structure is increasingly seen as useful for building systems that reason beyond pattern matching.

10 Challenges and limitations

Despite major progress, causal discovery remains difficult in realistic settings. Many of its core assumptions are only approximately true, and this can limit the certainty of any inferred structure.

10.1 Hidden confounders

Unobserved common causes can induce associations that mimic direct causal effects. When such variables are missing from the dataset, algorithms may infer incorrect edges or directions. Methods that account for latent confounding are therefore important but often less definitive.

10.2 Feedback loops and cycles

Some systems contain feedback loops, where a variable influences another that later affects the first. This creates cycles that are not compatible with all acyclic graph-based methods. Dynamic models and time-indexed representations can sometimes handle these settings better than static ones.

10.3 Nonstationarity and distribution shift

Nonstationarity occurs when the data-generating process changes over time or across environments. Distribution shift can make a model learned in one setting perform poorly in another. Causal methods sometimes use such changes as clues to structure, but they can also undermine standard assumptions.

10.4 Identifiability issues

Different causal graphs can produce the same observational distribution. When that happens, the true structure cannot be uniquely recovered from the available data alone. Identifiability depends on assumptions, data richness, and sometimes experimental input.

10.5 Interpretability and reproducibility

A causal model should be understandable enough for experts to inspect and debate. If the method is too opaque, its conclusions may be hard to trust or apply. Reproducibility is also essential, since causal claims should remain stable across datasets, analytical choices, and independent studies.

Causal discovery is closely linked to several neighboring areas that share methods, concepts, or applications. These fields often overlap in practice, even when their goals are not identical.

11.1 Causal inference

Causal inference estimates the consequences of interventions or treatments. It often uses assumptions supplied by causal discovery or preexisting theory. The two areas complement each other by connecting structure learning with effect estimation.

11.2 Graphical models

Graphical models represent variables and their dependencies through network structures. They provide the mathematical machinery for many causal discovery algorithms. Their tools include conditional independence, factorization, and message passing.

11.3 Experimental design

Experimental design concerns how to structure studies so that causal effects can be identified efficiently and reliably. It includes randomization, control groups, and intervention planning. Good design reduces ambiguity and strengthens the evidence used in discovery.

11.4 Machine learning for causality

Machine learning for causality refers to methods that use predictive modeling, optimization, and representation learning for causal tasks. It includes algorithms for discovery, effect estimation, and robustness analysis. This area has helped bring scalable computation and data-driven search into causal research.