🔗 Causal Inference

Causal Inference#

Drawing cause-and-effect conclusions from data, not merely associations.

Important

✨ AI-generated content. This page was written with the assistance of an AI language model and is provided as a learning aid. Despite careful review, it may still contain mistakes, omissions, or out-of-date information. Whether you are new to the topic, a team lead, or a senior practitioner, treat it as a starting point rather than an authoritative reference: read it critically and independently verify anything you act on (code, commands, figures, and factual claims) against official documentation and primary sources before relying on it.

What it is#

Causal inference is the discipline of establishing whether one variable actually causes a change in another, rather than merely correlating with it. The classic trap: ice-cream sales and drownings rise together, but neither causes the other — hot weather drives both. Causal inference is the toolkit for separating causation from association.

Why it matters#

Decisions hinge on it. Did the campaign cause the sales lift, or would sales have risen anyway? Did the treatment cause recovery, or was it natural? Did the policy cause the change, or did something else? Acting on a spurious correlation wastes money and harms people.

The gold standard and its alternatives#

A randomised controlled trial (RCT) assigns treatment and control at random, which balances confounders and makes the comparison causal. When randomisation is impossible (ethics, cost), observational methods approximate it: regression adjusts for measured confounders; propensity-score matching pairs treated and untreated units with similar covariates; difference-in-differences compares before/after trends against an untreated group; instrumental variables exploit a variable that moves treatment but not the outcome directly; and regression discontinuity compares units just above and below a treatment cutoff.

The counterfactual core#

Underneath sits the potential-outcomes (Rubin) model: each unit has a treated outcome \(Y_1\) and an untreated outcome \(Y_0\), and the causal effect is \(E[Y_1 - Y_0]\). The fundamental problem is that we never observe both for the same unit — which is exactly why we need RCTs, DiD and the rest. Directed acyclic graphs (DAGs) complement this by mapping confounders and mediators visually.

The pitfalls#

The enemies of causal claims are confounding (a third variable driving both), selection bias (groups that differ systematically), reverse causality (the outcome also influences the cause), and unobserved variables that can’t be adjusted for if they’re never measured.

Theme: Causal Inference & Uplift · All terminology

Hint

Mind map — connected ideas

Causal Impact · Treatment Effect · Regression Coefficient · A/B Testing · Counterfactual Explanations · Frequentist

Hint

More in Causal Inference & Uplift

AUUC (Area Under the Uplift Curve) · Causal Effect · Causal Impact · Causal ML (Causal Machine Learning) · Causal Trees · Cumulative Incremental Gain (CIG) · Cumulative Uplift · Incremental Conversions · Incremental Gain · Incremental Recovery Rate (IRR) · Incremental Revenue · Incremental Sales · Qini Coefficient · Qini Curve