Does causation require certainty?

No. A causal factor can change probabilities without determining every outcome.

Is correlation enough to estimate a causal probability change?

Not alone. A study design or additional causal assumptions are needed.

Does a population effect establish individual attribution?

Not generally. Individual counterfactual attribution can require information and assumptions beyond an average intervention comparison.

AI Glossary / Statistics and probability / Advanced

Probabilistic causation

Definition

Probabilistic causation refers to approaches that describe causal relationships in terms of probabilities rather than requiring a cause to guarantee its effect. A cause may change the probability of an outcome without producing that outcome on every occasion. The term covers several philosophical and statistical approaches, not one universally accepted probability formula. A central challenge is distinguishing a causal change in probability from an observed association that arises through common causes, selection or other features of how data are generated.

What it means

A factor can matter causally even when many exposed cases do not experience the outcome and some unexposed cases do. Conversely, a factor can predict an outcome without causing it. Simple probability raising, such as P(E|C) being larger than P(E|not C), is therefore not sufficient on its own. Modern causal models distinguish observing C from intervening to set C. Philosophical accounts differ in how they characterize the relationship, relevant background conditions and individual causal events, so terminology should be tied to the intended question.

How it works

Define the possible intervention or causal factor, the outcome, population and time ordering. Specify what probability contrast is intended and what assumptions connect observed data to it. Randomized experiments can identify certain intervention effects under their design assumptions. Observational analysis requires a justified causal structure and conditions for controlling confounding or using other identification strategies. In intervention notation, a population probability effect can compare P(E|do(C=1)) with P(E|do(C=0)). This comparison is different from simply conditioning on observed C. Attribution for a particular individual event can require additional counterfactual assumptions beyond identifying a population-average probability change.

Three practical examples

Experiment interpretation: explain why an intervention that increases success probability does not guarantee success for every participant.
Failure analysis: distinguish a warning indicator that predicts failure from a mechanism that changes failure probability when altered.
AI decision support: separate predictive features from candidate actions, preventing a model's association from being treated as evidence that manipulating the feature will improve outcomes.

Think of it this way

Think of a factor that changes the chances of a result rather than acting like an infallible switch. To ask whether it causes the change, compare what would happen under relevant interventions, not merely the cases where the factor happened to be present. A clue and a cause can point in the same direction while playing very different roles.

Worked example

Consider an original artificial causal model with a background condition Z present in half the population. An indicator C occurs with probability 0.8 when Z is present and 0.2 otherwise. Outcome E has probability 0.6 when Z is present and 0.1 otherwise, and C has no causal effect on E in this stipulated model. C and E are conditionally independent given Z. Observing C gives P(Z=1|C=1)=0.8, so P(E|C=1)=0.8 times 0.6 + 0.2 times 0.1 = 0.50. Without C, the corresponding probability is 0.2 times 0.6 + 0.8 times 0.1 = 0.20. Yet intervening on C leaves Z unchanged, so either intervention gives outcome probability 0.5 times 0.6 + 0.5 times 0.1 = 0.35. The observed 30-percentage-point association is not causal here. By contrast, a separately justified intervention that changes an outcome probability from 0.20 to 0.30 has a 10-percentage-point effect without guaranteeing the outcome.

Illustrative case study

In a fictional case, Alder Support finds that cases receiving specialist attention have more unresolved outcomes. Managers initially infer that specialist involvement causes poor results. Analysts note that difficult cases are preferentially escalated, making case difficulty a plausible common cause of both attention and failure. They define an intervention question and design a suitable evaluation rather than using the raw association to remove specialist support. The report distinguishes predictive evidence from the assumptions required for causal conclusions.

Common misconceptions and mistakes

A probabilistic cause need not be necessary or sufficient for its effect. Observed probability raising is not automatically causation, and lack of an unadjusted association does not exclude a causal relationship hidden by opposing pathways or confounding. More data can estimate an association precisely without identifying a causal effect. A causal effect on a population probability does not by itself reveal whether a factor caused one particular person's outcome. Causal diagrams also do not prove their own arrows; they encode assumptions that need substantive justification and, where possible, empirical scrutiny.

Business use

Businesses benefit from this distinction whenever predictions are used to choose actions. A useful risk marker may be difficult or pointless to manipulate, while a causal intervention can improve outcomes even if it is not a strong observational predictor. State the decision and the intervention contrast before analyzing data. Report effect sizes, uncertainty and identification assumptions. For AI systems, avoid translating feature importance directly into recommendations for changing those features. Evidence about actions should come from a design or causal model capable of addressing what happens when the action is taken.

Check your understanding

An indicator predicts failures but changing the indicator leaves failure probability unchanged in the stipulated causal model. Is its predictive association evidence of an effect of manipulating it?

Answer: No. Prediction and intervention are different questions.

Related terms

Causal inference
Confounding
Conditional probability
Causal graphs
intervention
counterfactual
potential outcome
causal graph

Learn next

Learn the distinction between conditioning and intervention, then examine how experiments, causal assumptions and sensitivity analysis support claims about probability changes.

Sensitivity analysis

Sources

Stanford Encyclopedia of Philosophy, Probabilistic Causation: https://plato.stanford.edu/entries/causation-probabilistic/ Pearl, The Foundations of Causal Inference: https://www.cs.ucla.edu/~kaoru/r355-corrections.pdf

Previous: Probabilistic automaton Next: Probabilistic circuit