Reproducing the Literature

A gallery of classic causal-inference and causal-RL results, each reproduced end-to-end with causalrl. Every case below is a passing test in tests/test_literature_classics.py that asserts the textbook result — run them with pytest tests/test_literature_classics.py.

Causal-inference classics

Case	What it shows	In causalrl
Simpson's paradox (kidney stones)	Adjusting for the confounder reverses the naive sign	estimate_effect (back-door) vs the raw association
Front-door (smoking → tar → cancer)	A confounded effect is still identifiable through a mediator	identify_effect / estimate_effect
Pearl's napkin	Identifiable despite latent confounding	is_identifiable_effect → True
Instrumental variable	Not point-identified, but boundable	is_identifiable_effect → False; manski_bounds
Bow arc	The simplest non-identifiable confounded effect	is_identifiable_effect → False
Transportability (LA → NYC covariate shift)	Re-weight a source effect to a new population	identify_transport

Difficult RL problems where causal beats associational

These are classic decision problems where a confounding-aware (causal) agent provably outperforms an associational one on the same data:

Problem	Associational RL	Causal RL
MABUC — bandit with unobserved confounders	NaiveThompsonSampling (ignores the context)	CausalThompsonSampling conditions on the "intuition" and earns more
Greedy Casino — counterfactual decision-making	best fixed arm do(X=a) ≈ 0.37	acting on the counterfactual (intuition) ≈ 0.80
Hard exploration — sparse far goal	flat Q-learning rarely reaches it	a causal prerequisite curriculum (curriculum_q_learning) does

The point each makes: associational policies condition on what they see, while causal policies reason about what they do (Layer 2) and what they would have done (Layer 3). Under unobserved confounding the gap is not subtle — the counterfactual "follow your intuition" policy more than doubles the best interventional arm's reward.