Advanced modules¶

Four optional, self-study modules that take the course beyond ISLP: causal inference with randomised experiments, explaining black-box models with Shapley values, distribution-free uncertainty with conformal prediction, and the exponential-family machinery that unifies GLMs and penalized splines.

4 modules 300 slides 25 + 8 exercises with solutions 4 companion notebooks optional self-study

Each module is built exactly like a course deck — the same colour-coded callout boxes, every exercise followed immediately by its worked solution, a closing summary block and an optional appendix — and each is paired with a companion notebook whose printed numbers match the slides seed-for-seed (np.random.default_rng(2024)). The modules sit outside the twelve-lecture plan: nothing in the course or its exams depends on them.

The modules¶

	Module	What it covers	Slides	Deck	Notebook
A1	Randomised Controlled Trials	Potential outcomes, the selection-bias decomposition, why randomisation works, regression adjustment with HC2 errors, power and sample size, ITT vs per-protocol, the peeking problem	71	Open	Rendered
A2	Explainable AI with Shapley Values	Cooperative games and the four axioms, the marginal value function, exact Shapley by enumeration, Monte-Carlo sampling, waterfalls and global importance, the correlated-feature and retrain pitfalls	73	Open	Rendered
A3	Conformal Prediction	Exchangeability, split conformal and the finite-sample quantile, marginal vs conditional coverage, locally weighted scores and CQR, classification prediction sets, the OLS stress test	74	Open	Rendered
A4	GLMs and Splines	The exponential family, Poisson regression on `Bikeshare`, deviance and LRTs, overdispersion (quasi-Poisson, negative binomial), penalized splines and effective df, a count GAM	82	Open	Rendered

Tip

Everything is implemented from scratch — exact Shapley values, split conformal, P-splines — so the notebooks need nothing beyond the course environment. The slides name the production tools (shap, MAPIE, mgcv-style GAM software) once the mechanics are understood.

What each module assumes¶

Each module bridges back to the course chapters it extends, and assumes them:

Module	Builds on
A1	Ch 0 (standard errors, tests) · Ch 3 (regression) · Ch 5 (simulation) · Ch 13 (multiple testing — the peeking problem)
A2	Ch 0b (counting and the 2ᵖ cost) · Ch 8 (boosting, variable importance) · Ch 10 (black-box models)
A3	Ch 3 (prediction intervals) · Ch 5 (train/validation splits)
A4	Ch 3–4 (linear and logistic regression) · Ch 5–6 (CV, AIC) · Ch 7 (splines and GAMs)

How they are built¶

The sources live in Chapters/Advanced/ — one folder per deck with its LaTeX source, its make_figures.py (every figure is computed from the course datasets or a seeded simulation) and the compiled PDF; the notebooks ship with stored outputs and run on Colab with one click, resolving data exactly like the course labs. From the repository root, make advanced rebuilds any deck whose source changed. The folder also carries the distilled style guides (STYLE_DECK.md, STYLE_NOTEBOOK.md) used to author them — the starting point for adding a module A5.

All advanced notebooks¶

Where to go next¶

Lecture slides — the twelve course decks these modules extend.
Lab notebooks — the course labs, built to the same pattern.
The course at a glance — where the twelve-lecture plan ends and these begin.
Repository layout — where everything lives.