Learn Causal Inference with Phoebe

Leader track 🪜 · six sessions, no code

For whoever approves the spend and reads the lift number. Which rung a claim is on, what to audit, what to refuse to sign, and the four options when the question turns out to be unanswerable.

🪜 Leader session 1 · start here

Three questions, three rungs

Association, intervention, counterfactual. A rung-two question cannot be answered with rung-one data, which is why a dashboard cannot tell you whether the email worked.

▶ Start here 🔬 Leader session 2

The claim audit

Five questions for any causal claim that crosses your desk, and the vocabulary that signals an observational study wearing experimental clothes.

🚫 Leader session 3

What not to adjust for

Mediators, colliders and the pre-treatment collider that looks safe because it was measured first. Each one priced on the same data.

🔎 Leader session 4

Identifiable or not

Whether the question can be answered from this data at all, checkable on paper in ten minutes, and the four honest options when the answer is no.

👥 Leader session 5

Not everyone responds

Average effects hide who they help and who they cost. Why a positive average can still be the wrong policy.

📋 Leader session 6

Commissioning and reading a causal claim

What to require in the brief, above the method, and what to refuse to sign.

the board version

Practitioner track 🔧 · ten sessions, one defensible claim

One intervention, Kestrelmark's re-engagement email, taken from a blank page to a claim you can defend in review. Session 10 scores it against a truth the engine computes rather than estimates.

✏️ Practitioner session 1 · start here

Draw the DAG before touching the data

An arrow is a claim and a missing arrow is a stronger one. Kestrelmark's naive lift is 10.54 against a truth of 6.23, and the gap has two names.

▶ Start here 🔀 Practitioner session 2

Confounder, collider, mediator

Three shapes, three fates. Conditioning closes one path, removes real effect from another, and creates association out of nothing in the third.

🚪 Practitioner session 3

Identification before estimation

The back-door criterion stated properly, its two conditions, and what to do when no valid adjustment set exists in your data.

🚪 Practitioner session 4

The front door

When an unmeasured confounder shuts the back door and a mediator rescues the estimate. Three conditions, and why Kestrelmark fails the first.

🔄 Practitioner session 5

Simpson's paradox

The same numbers, two opposite conclusions, and a graph rather than a table deciding which one to act on.

🌗 Practitioner session 6

Counterfactuals

Rung three: would this particular customer have churned anyway? Two worlds at once, and what a fully specified model buys you.

📈 Practitioner session 7

CATE and uplift

The average effect is a summary of a distribution. S-, T- and X-learners, and the documented weakness of each.

🎯 Practitioner session 8

Qini curves and who to treat

Targeting by predicted uplift against targeting at random, and the customers a campaign actively costs you.

⚖️ Practitioner session 9

Sensitivity analysis

How strong would an unmeasured confounder have to be to overturn this? The honest answer to the objection you cannot rule out.

🎛️ Practitioner session 10

The adjustment bench

Six adjustment sets scored against a truth known by construction. The right one leaves noise; every column in the table misses by 5.25.

🔒 the capstone

start here foundations hands-on advanced and capstone

Choose your path 🗺️

Three doors into the same discipline. Both tracks start at their own session 1.

💼 I approve the spend and read the lift L1 → L2 → L3 → L6 the ladder, the audit, the covariate list, the brief

🔧 I build the model P1 → P2 → P3 → P9 → P10 the graph, the three shapes, identification, the objection, then score it

🎯 Everything, in order L1 → L4 → P1 → P4 → P5 → P6 → P7 → P10 from a lift number nobody can defend to an adjustment set you can

Your passport 🎫

Clear all three quiz questions in a session and it gets stamped. Sixteen stamps across both tracks, and they live in this browser only.

What is real here 🧾

A course about defensible claims should be exact about its own.

Claim	Status
The 5,000 Kestrelmark customers	Generated from a structural causal model written out in the source file, one fixed seed, identical for every learner. Every coefficient is visible
The true effect of 6.23	Computed, not estimated: the model is run twice with the email forced on for everyone and then forced off, and the mean spend differenced. That is the do-operator by brute force
The engine agrees with the algebra	Worked out by hand the effect should be the mean direct effect of 5.16 plus 3.0 times the change in open rate, which is 6.22. The simulation gives 6.23. A simulation checked only against itself would prove nothing
Every estimate on the bench	A real least-squares fit of spend on the email plus the ticked covariates, solved in the page. The coefficient on the email is the estimate
The ladder: 10.54, 6.56, 5.26, 5.48, 2.59, 0.98	Arithmetic on that data. Nothing on this bench is modelled and no result is written down in advance
Noise is not bias, and the bench separates them	The correct set's residual of +0.33 changes sign across five seeds and averages +0.06. Every contaminated set keeps its sign every time
The bystander is genuinely harmless	Adding the device variable changes the estimate by nothing to two decimal places, which is exactly why a covariate list cannot be read by eye
No estimator is taught anywhere	Propensity scores, inverse probability weighting, doubly robust methods, difference-in-differences and synthetic control all belong to the experimentation course, which is named on every page that reaches for one
Kestrelmark	A fictional subscription retailer. The confounding is ordinary; the company is not real

The knowledge map 🧠

Both tracks at a glance - hover a session to spotlight its concepts, click any node to jump in.