Interventions
A causal question asks: if I set $x$ (a treatment), how does $y$ change?
Seeing that people who take a vitamin have fewer colds is an association. Forcing a random half of people to take the vitamin is an intervention. Causal effects are about the second kind of question.
Treatments can be binary (drug / no drug) or continuous (dose). Outcomes likewise. Randomized experiments assign treatment by chance to break confounding. Observational studies need stronger assumptions / adjustments.
A/B test
Show half of users button colour A and half colour B at random. Difference in click rate estimates the causal effect of colour — because other factors balance out on average. If users self-select the colour, confounding sneaks back in.
Key takeaways
- Intervention ≠ passive observation.
- Randomization is the gold standard.
- Estimate effects, not only associations.