Reading focus
Syllabus: What Would Justify a Causal Claim?
Compare the descriptive, predictive and causal questions that could be asked of one dataset.
Why you’re reading it. Separate a useful forecast from a claim about changing an outcome.
A question to consider. For the association between study-group attendance and exam performance, write one question of each kind.
Why this reading
Hernán, Hsu and Healy distinguish describing observations, predicting outcomes and asking what would happen under an intervention. Their examples show why a successful predictive analysis can leave a causal question unanswered.
Before you read
A counterfactual describes an alternative to what occurred. A causal contrast compares outcomes under different interventions for a specified population and period. The same data and statistical technique can serve different scientific questions.
Read for
- Track what changes in the required assumptions when a predictive question becomes a causal one.
- Use the smoking and birthweight example to examine why a useful predictor can be an inappropriate adjustment variable.
Consider afterwards
- A model predicts which students will leave a course. What additional question must be answered before using one of its predictors as a target for an intervention?
Editorial context
This is a methodological and educational argument. Its distinction among tasks does not establish the causal structure of a particular problem, and subject knowledge can be mistaken. The discussion of artificial intelligence reflects the examples and debates of its publication period; it is not a current assessment of particular systems. Accurate prediction and causal identification require different kinds of justification.