AI Glossary / Causal inference and experimentation / Intermediate
Causal graphs
Definition
Causal graphs are diagrams that represent assumptions about causal relationships among variables. In a common form, a directed acyclic graph uses arrows from causes to effects to help reason about interventions, confounding, selection and which variables may need to be measured or adjusted for.
What it means
A causal graph is more than a picture of correlations. An arrow proposes a causal relationship, while a missing arrow also carries an assumption about the absence of a direct relationship within the model. The graph is therefore a compact statement of substantive beliefs that can be examined and challenged.
Graphs can help distinguish a common cause from a mediator or a collider. Those roles matter because adjusting for every available variable can create bias rather than remove it.
The diagram alone does not prove its assumptions. Domain knowledge, study design, measurement quality and empirical evidence are needed to justify the structure and the resulting conclusions.
How it works
List variables relevant to the causal question and establish their time order where possible. Draw arrows for plausible direct causal relationships at the chosen level of detail, including unmeasured common causes when needed. A directed acyclic graph excludes directed feedback loops within its representation; time-indexed variables can represent many dynamic processes.
Use graphical rules to inspect paths between intervention and outcome. A back-door path can carry confounding, while conditioning on a collider can open a previously blocked path. Determine whether the desired effect is identifiable under the graph and available data. Only then choose a suitable estimation method and examine sensitivity to uncertain assumptions.
Three practical examples
- A training evaluation maps how prior experience influences both program participation and later performance.
- A product team distinguishes variables measured before a feature launch from behaviors caused by the feature.
- A researcher identifies how selection into a dataset could create misleading associations between otherwise unrelated variables.
Think of it this way
Think of a map of proposed information about cause and effect, with routes that can transmit misleading associations as well as intended effects. The map helps plan an analysis, but drawing a road does not prove it exists. Unlike a geographic map, its arrows and omissions encode assumptions that may be difficult to verify directly from observational data.
Worked example
Suppose experience E affects whether a worker joins training T and also affects later performance Y. Training may itself affect performance. The graph contains E→T, E→Y and T→Y. The path T←E→Y is a back-door path connecting training and performance through a common cause.
If the graph and other identification assumptions are credible, adjusting for E can block that path when estimating the total effect of training. Now add skill gain M with T→M→Y. Skill gain is a mediator on a causal pathway.
Adjusting for M would change the analysis away from a straightforward total-effect question and can introduce further complications. This example shows why covariates should be selected by their causal roles, not simply by whether they predict Y or are available in a dataset. It does not establish that experience is the only confounder in an actual training study.
Illustrative case study
Fictional case: Rowan Retail wants to estimate whether a new service workflow changes customer retention. Analysts initially propose controlling for every logged variable, including satisfaction measured after the workflow change.
The team builds a causal graph and recognizes that satisfaction may lie on the pathway through which the workflow affects retention. It distinguishes the total-effect question from a possible mediation question and identifies missing pre-change service-quality information. The graph makes disagreements visible and guides additional data collection. It does not automatically settle those disagreements or transform the observational records into a randomized experiment.
Common misconceptions and mistakes
An arrow is not merely a positive correlation, and it does not specify an effect size. A missing arrow is an assumption, not an omission without consequences. Causal direction cannot generally be recovered from association alone without additional assumptions.
Do not adjust for all variables indiscriminately. Mediators and colliders have different roles from confounders. A graph that excludes cycles does not imply that real-world feedback never exists; it may require a time-expanded representation. Finally, identification under a graph and accurate estimation from finite data are separate problems.
Business use
Causal graphs can improve business evaluations by making assumptions explicit before modeling begins. They are useful for cross-functional discussions because operational experts may know why people receive a program or why records are missing, while analysts know the statistical implications.
Keep the graph tied to one clearly stated intervention question. Record uncertain arrows and plausible alternatives. Use it to decide which measurements are needed and whether the available data can support the intended claim. A polished diagram should not be used as evidence of certainty; its value lies in transparent reasoning, reviewable assumptions and a clearer path from business question to research design.
People also ask
Does every graph describe causality?
No. Many graphs represent associations, computation or other relationships.
Can a causal graph prove an effect from data alone?
No. Its causal interpretation relies on assumptions and evidence beyond the drawing.
Must every variable be observed?
No. Unmeasured variables can be represented, although they may prevent identification of a desired effect.
Check your understanding
A variable is caused by the intervention and helps transmit its effect to the outcome. Should it automatically be adjusted for when estimating the total effect?
Answer: No. It may be a mediator. Conditioning on it can remove part of the pathway of interest and requires a different causal question and additional assumptions.
Related terms
- Directed acyclic graph
- Confounder
- Collider
- Causal inference
- Identification (information)
- Structural causal model
- mediator
- back-door criterion
- identification
Learn next
Learn the three basic graph patterns: common cause, causal chain and collider. For each, examine what conditioning does to associations. Then draw alternative plausible graphs for a real evaluation question and identify which conclusions depend on disputed assumptions.
More in Causal inference and experimentation
- Brier score - The Brier score measures the quality of probability predictions by averaging squared differences between predicted probabilities and observed outcomes.
- Calinski-Harabasz Index - The Calinski-Harabasz Index is an internal clustering score that compares separation between cluster centers with variation inside clusters, adjusted for...
- Causal estimand - A causal estimand is the precisely defined causal quantity that an analysis aims to learn.
- Causal language modeling - Causal language modeling is a training task in which a language model predicts the next token using only the preceding context.
- Causal link - In partial-order AI planning, a causal link records that one action establishes a condition needed by another action.
- Causal network - A causal network is a model that represents variables and their causal relationships, commonly using a directed graph together with probability...
Sources
Judea Pearl, Causal Diagrams for Empirical Research: https://bayes.cs.ucla.edu/R218-B.pdf Author repository record: https://escholarship.org/uc/item/6gv9n38C Primary paper search excerpts consulted. The training example is original; full methodological and editorial review remain pending.
Previous: Causal estimand Next: Causal inference