Causal Discovery and Causal Inference

Overview

Correlation quantifies association. Causation explains mechanism. Every scientist knows the gap between these two concepts, yet much of data-driven discovery operates entirely in the correlational regime: fit a model, find a pattern, report a relationship. The field of causal inference provides the formal machinery to bridge this gap, answering questions that no amount of observational data analysis alone can settle. Does a drug reduce mortality, or do healthier patients simply choose to take it? Does a gene expression change drive tumor growth, or is it a downstream consequence of a shared regulatory signal?

This chapter builds the complete causal reasoning toolkit. We start with structural causal models (SCMs), the mathematical language that makes causal claims precise enough to test. We then turn to causal discovery algorithms that learn causal graph structure directly from data, moving beyond the hand-drawn DAGs that dominate textbook examples. With a graph in hand, we estimate treatment effects using modern semiparametric methods (double machine learning, targeted learning) that leverage the predictive power of ML while preserving valid statistical inference. Finally, we assemble a full pipeline: from raw observational data through graph discovery, effect estimation, and rigorous refutation testing.

The ideas here connect directly to the causal reasoning foundations introduced in Chapter 4, where we first encountered the do-calculus and causal graph queries. They build on the exploratory discovery techniques of Chapter 25, which provide the initial patterns that causal analysis can elevate from associations to mechanisms. The anomaly detection methods of Chapter 30 tell us what is unusual; causal inference tells us why. Looking forward, causal reasoning powers the social and economic systems of Chapter 52 and feeds into the AI scientist architectures of Chapter 53, where autonomous agents must reason about interventions.