A large observational dataset can reveal associations, but it does not automatically answer what would happen under an intervention or in a counterfactual world. Causal reasoning separates these questions into three levels: association, intervention, and counterfactual. section 10.1 section 10.2.1
The causal hierarchy theorem states that a level-k question requires level-k assumptions. In general, data from a lower level cannot identify a query at a higher level without additional modeling assumptions. More data alone does not bridge that gap; even a deep learning model needs to encode suitable assumptions to answer higher-level questions reliably. section 10.1.4 section 10.2.1
The hierarchy distinguishes the data and questions involved:
- Association: observational data records statistical relationships among variables. section 10.1.3
- Intervention: experimental data, such as a randomized A/B test, records outcomes under an imposed change. section 10.1.3
- Counterfactual: reasoning compares possible outcomes for the same case; such data may be generated by rerunning a deterministic process under changed conditions. section 10.1.3
Causal identification asks whether assumptions and available lower-level data suffice to infer a higher-level query. In practice, identify the query’s level first, then check whether the causal model supplies the assumptions needed to connect it to the data. section 10.2.1 section 10.1.4
