When a customer clicks after seeing an advertisement, attribution asks whether that exposure caused the click, whether the customer would have clicked anyway, or whether several messages acted cumulatively. It assigns a causal source and corresponding credit or blame to a particular outcome. section 1.3.2
Attribution differs from prediction or correlation because it depends on counterfactual reasoning: would the outcome have occurred if the suspected cause had been absent or changed? Causal inference turns this question into a query answered with observational or experimental data. section 1.0 section 1.6
A practical attribution analysis separates several questions:
- Identify the outcome and candidate cause, such as subscriber churn and an absent recommendation. section 1.0
- Estimate whether the cause was sufficient: would it alone have been enough to produce the outcome? section 8.3.2
- Estimate whether the cause was necessary: would the outcome have failed to occur without it? section 8.3.2
- Account for abnormality: unusual events may receive more blame than ordinary events with similar causal roles. section 8.3.2
Root-cause analysis
A network crashes after Lazlo accesses a faulty server. Attribution does not stop at observing his access: it asks whether that access was sufficient, whether it was necessary, and whether it was abnormal. Actual causal strength combines abnormality with probabilities of causation, treating attribution as a trade-off between an abnormal necessary cause and a normal sufficient cause. section 8.3.2
Use the figure to inspect how abnormality and probabilities of causation combine into an attribution measure. section 8.3.2
