What Is Causal Inference?

causal-inferencestatisticsrcta-b-testingdata-science

Abstraction: Causal inference methods from RCTs to Pearl's causal graphs explained

Key points:

Judea Pearl's "ladder of causation" has three rungs: association (statistics/ML), intervention (A/B tests/RCTs), and counterfactuals (causal models that answer "what if?"); most ML operates only at rung 1
Three preconditions for causation: correlation, temporal precedence, and absence of confounding variables; meeting all three still doesn't prove causation
A/B tests can only estimate impact of incremental product changes; they cannot answer why customers churn or discover qualitatively new product directions — "local maxima" problem
Causal graphs enable reasoning about forks (confounders), colliders (spurious correlations when conditioning on a common effect), and information flow; back-door and front-door adjustments let you remove confounder effects even when confounders are unobservable
Omitted variable bias (OVB) is a key hazard in regression-based causal inference; sensitivity analysis on controls is a standard check; instrumental variables (e.g., John Snow's water company example) can identify causal effects when direct measurement of confounders is impossible

Connections: Judea Pearl · Causal Inference · Randomized Controlled Trials · Statistics