Classical ML models answer predictive questions but fail at causal ones — they cannot reliably estimate what will happen if you intervene. This piece introduces causal inference as a distinct discipline, explaining the fundamental problem (counterfactuals are unobservable), the potential outcomes framework (ATE, ATT, CATE), and the three key assumptions (ignorability, positivity, SUTVA). It also covers Structural Causal Models and DAGs from Judea Pearl's work, explaining confounders, mediators, and colliders, and why 'controlling for everything' can introduce rather than remove bias. The post is Part 1 of a three-part series; Parts 2 and 3 will cover classical methodologies and ML-based causal methods respectively.
Table of contents
Prediction vs CausationThe Fundamental Problem of Causal InferenceThe Potential Outcomes FrameworkStructural Causal Models and DAGsTwo Frameworks, One PracticeWhy this matters to you?Coming up in Part 2…638 Impressions