- 1Causal deniabilityA large study of coffee drinkers reported a 5% lower risk of dying, stated plainly that it could not establish cause and effect, and then suggested people drink coffee for their health. Naming that move, and the three conditions for demonstrating causality.8 min
- 2The road not takenA causal effect is a difference between two outcomes for the same person, and only one of them ever happens. Potential outcomes and the fundamental problem of causal inference, worked through a population whose true average treatment effect is known to be 1.5.9 min
- 3Confounder control and drawing your assumptionsIce cream sales and violent crime rise together, and no amount of staring at the data will settle why — a dataset carries no memory of what produced it. A causal diagram is where you write the story down before estimation starts, built on forks, pipes, and colliders.7 min
- 4Good paths and bad pathsBetween any two variables in a causal diagram there are usually several routes, and identification is the discipline of knowing which ones carry the causal effect and which carry confounding. Path tracing, minimum sufficient adjustment sets, d-separation, and bad controls.5 min
- 5Closing backdoor pathsA simulated dataset where the true effect is 1.5 by construction: the naive comparison returns 2.07, holding the confounder constant returns 1.47, and the specification that adds a collider moves the estimate away from the truth. More control, worse answer.9 min
- 6Instrument-based approachesRandomization destroys confounding by construction, because nothing upstream can influence an assignment made by a coin. When a coin is not available, the alternative is variation in the world that behaves like one — here, an HPV vaccination program where eligibility turned on a birth date.9 min