Causality
Just wandered in? Here is the course in ten short stops. Each one is a picture that moves, and something you get to grab. Together they tell one story: from a simple question about 100 patients to a survival curve you can defend.
Each button opens a lesson at the right step. Play with it for a few minutes, then come back here with your browser's Back button. The colours follow the course route: question, identification, model, estimation, uncertainty, interpretation.
Treated patients did 2.59 better. The true effect is 2. Where did the extra come from?
What moves: the cohort sorts itself, first by who got treated, then by outcome, then within severity, until the gap shrinks toward the truth. Tap anyone to meet them.
Some gaps cannot be closed by a better estimator. This is the one to see with your own eyes.
What moves: untick exchangeability and slide κ. Only the hollow outcomes nobody observed move, and the true effect moves with them. The data stay exactly the same.
A registry device that does nothing at all can look lifesaving. It depends on one innocent row of the protocol: when follow-up starts.
What moves: twenty patients on one clock. Move time zero and watch waiting time turn into immortal time, while the rate ratio drops far below 1.
On the day he joins the waiting list, which plan is Mr. Ortiz on? Both. So we copy him.
What moves: one timeline splits into two copies, one per plan. Let his real life run month by month and watch a copy get cut the moment it stops fitting its plan.
“Influence function” sounds abstract. It is how hard one extra patient tips a beam.
What moves: drag one patient along a beam holding 100 others. The balance point slides by exactly the patient's share of the mass times their distance from it.
AIPW is not a black box. It is a prediction plus a weighted correction, and you can watch every patient add theirs.
What moves: each patient carries two sticks, a predicted effect and a residual drawn as thick as its weight. Play adds them tip to tail into the estimate.
An interval is a promise about studies you did not run. Here you can run two hundred of them.
What moves: each ball is one whole study, pushed sideways by its patients, carrying its own 95% interval. The AIPW pile centres on the truth; the plug-in's pile centres at 2.64.
You can never rule one out. You can say how strong it would need to be to change your answer.
What moves: drag a hypothetical confounder across a map of its two strengths and watch whether a risk ratio of 1.50 survives. The next step turns the map into one number, the E-value.
You have drawn a hundred Kaplan–Meier curves. Here is what the estimator does with the people who leave.
What moves: every patient holds an equal share of mass. At a death it falls out; at a censoring it is handed to everyone still at risk to the right, and their dots grow.
Everything comes together: a device registry, an emulated trial, and an answer you could put in front of a reviewer.
What moves: the estimates arrive one by one on a forest plot: the unadjusted difference points the wrong way, the plug-in, AIPW and TMLE move toward the truth, and then the truth itself appears.
Twenty-two lessons walk the same road slowly, one idea at a time, with predictions to make and checks to try. You can also dive into any lesson from the course map.