causal-inference

Causal Inference

Causal inference is the discipline of estimating cause-and-effect relationships from data — answering "what would have happened?" rather than "what is correlated with what?" This skill covers the identification strategies that make causal claims defensible (DAGs, assumptions, estimation methods) and the computational tools (CausalML, DoWhy) that implement them. It appears alongside analytics and ML engineering whenever the goal is decision-making, not prediction.

Core Concepts

DAGs and the Language of Causation

A Directed Acyclic Graph (DAG) is the formal representation of causal assumptions. Nodes are variables; directed edges represent direct causal effects. Three fundamental structures:

• Chain: X → M → Y. M is a mediator. Controlling for M blocks the path (bad if you want total effect; intentional for direct effect).
• Fork (Common Cause): X ← C → Y. C is a confounder. Controlling for C blocks the backdoor path and removes confounding.
• Collider: X → C ← Y. Controlling for C (or conditioning on it in sampling) opens a spurious path — this is selection bias. Never control for a collider.

Backdoor criterion: a set Z blocks all backdoor paths from treatment T to outcome Y (paths with an arrow into T) and contains no descendants of T. If such Z exists and is observed, the causal effect is identified by adjustment: P(Y | do(T)) = Σ_z P(Y | T, Z=z) P(Z=z).

Do-calculus (Pearl): three rules for transforming expressions with do() operators into observational distributions. In practice, use the backdoor/frontdoor criteria rather than applying do-calculus rules manually. DoWhy handles identification automatically given a graph.

Identification gotcha: unmeasured confounders break identification from observational data. Always draw the full DAG before choosing an estimator — the choice of method is determined by the graph structure and which variables are observed.

DoWhy

DoWhy formalizes the four-step process: model → identify → estimate → refute.