AI Glossary / Causal inference and experimentation / Intermediate
Causation
Definition
Causation is a cause-and-effect relationship in which one event, condition or intervention contributes to producing another. In statistical analysis, a claim of causation requires more than observing that two quantities occur together or that one can predict the other.
What it means
A causal claim says something about how an outcome would differ under a relevant change. It need not mean that the cause is the only influence, that the outcome always follows, or that the effect is identical for everyone.
For example, changing a process step may reduce the probability of an error without eliminating every error. Other conditions can still matter. Conversely, two measures may move together because of a common cause, reverse direction of influence or a shared trend.
The everyday words cause and causation cover broad questions. In an evaluation, make the intended claim precise by naming the action, comparison, outcome, population and period.
How it works
Evidence for causation comes from a combination of design, assumptions and observations. Random assignment can make intervention groups comparable in expectation. Other designs use credible natural variation or explicit adjustment strategies, each with requirements that need scrutiny.
Investigate alternative explanations, including confounding, selection, measurement error and changes occurring at the same time. Consider whether the proposed cause precedes the effect and whether the intervention is clearly defined. Quantify the effect and uncertainty where the design supports that calculation. Mechanistic understanding can strengthen interpretation, but a convincing story alone is not a substitute for evidence that distinguishes it from competing explanations.
Three practical examples
- A controlled product experiment studies whether a revised form causes more completed applications.
- An engineering test evaluates whether a changed component reduces a measured failure probability under specified conditions.
- A program evaluation distinguishes the effect of a new workflow from a seasonal change that happened at the same time.
Think of it this way
Think of a row of switches and indicator lights. A light may be associated with a switch position because the switch controls it, or because both respond to another part of the circuit. Merely watching them together does not reveal the wiring. The analogy explains alternative causal structures; real social and technical systems are often less directly inspectable than a circuit.
Worked example
Suppose a team introduces a simplified online form, and completion rises from 40% in one month to 50% in the next. The observed increase is 10 percentage points. At the same time, an advertising campaign attracts visitors who are more ready to apply.
The before-and-after comparison alone cannot determine how much of the increase the form caused. In a separate hypothetical randomized comparison during the same period, the old form has a 47% completion rate and the new form 50%. The difference is 3 percentage points, with sampling uncertainty still to be assessed.
These invented numbers illustrate why contemporaneous alternatives matter. They do not imply that randomization always produces a smaller effect or that the campaign necessarily explains the remaining change. The causal claim must follow the actual design and evidence.
Illustrative case study
Fictional case: Hawthorn Operations celebrates a decline in delivery complaints after introducing a new dispatch rule. A review finds that order volume also fell and a problematic supplier stopped shipping during the same period.
The team revises its report to describe the observed decline without assigning all of it to the rule. It then designs a stronger comparison across comparable operating periods and considers a controlled rollout where practical. This is an example of improving the evidence behind a causal claim. It does not conclude that the dispatch rule had no effect, only that the initial observation could not isolate it.
Common misconceptions and mistakes
After does not automatically mean because of. A causal relationship need not be deterministic, and several causes can contribute to one outcome. Statistical significance does not remove confounding or establish that an effect is operationally important.
Do not assume that a variable with high predictive importance is a good intervention target. Do not overcorrect by claiming that observational evidence can never support causation; credible designs and assumptions can sometimes identify causal effects. The strength and scope of the claim should match the evidence rather than the confidence of its wording.
Business use
Clear causal language improves decisions and reporting. Organizations often need to know whether a program, feature or policy produced a result before expanding it. Unsupported attribution can direct resources toward ineffective actions or obscure the actual source of an improvement.
Use different wording for observation and intervention evidence. “Completion increased after the change” describes timing; “the change increased completion” requires stronger support. Document comparison groups, simultaneous changes and measurement consistency. Report effect size and uncertainty where justified. If evidence is incomplete, identify the next study that would most improve the decision instead of turning a plausible explanation into a settled conclusion.
People also ask
Must a cause always produce the effect?
No. Many causal effects change probabilities rather than guarantee outcomes.
Can an outcome have several causes?
Yes. Causes can interact or contribute through different pathways.
Does a large correlation prove causation?
No. Even a strong association can arise from common causes, selection or other mechanisms.
Check your understanding
Sales rise immediately after a new dashboard is installed. What can be concluded from timing alone?
Answer: Sales rose after installation. Timing alone does not establish that the dashboard caused the increase; other changes and the way the dashboard was adopted must be considered.
Related terms
- Causality
- Correlation
- Causal inference
- Confounding
- Effect size
- intervention
- counterfactual
- randomised experiment
- temporal order
Learn next
Learn correlation versus causation and the role of a comparison group next. Take a before-and-after business claim and list competing explanations. Then identify a design that could distinguish the proposed intervention effect from those alternatives without overstating what the current data show.
More in Causal inference and experimentation
- 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...
- Causal positivity - Causal positivity is the requirement that the treatment choices needed for a causal comparison have positive probability within the relevant covariate...
- Character N-gram F-score - Character N-gram F-score evaluates generated text by combining precision and recall of matching character sequences against reference text.
- Citation precision - Citation precision measures how well the citations an AI system supplies support the statements to which they are attached, under a specified evaluation rule.
- Citation recall - Citation recall measures how much of an answer's evaluated content is supported by the citations provided, under a defined claim or statement-level protocol.
Sources
Australian Bureau of Statistics, Statistical terms and concepts glossary, Causation: https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/statistical-terms-and-concepts-glossary Miguel A. Hernán and James M. Robins, Causal Inference: What If: https://miguelhernan.org/whatifbook Official glossary definition read and author source located. Illustrations are original; full methodological and editorial review remain pending.
Previous: Causality Next: Cause-effect diagram