Causal Inference

Causal inference is a framework of statistical methods and principles aimed at identifying genuine cause-and-effect relationships between variables, rather than mere correlations. While two variables may move together in data, causal inference asks whether changing one variable actually produces a change in another. This distinction is critical: a model that learns correlations may fail catastrophically when deployed in environments where the underlying causal structure differs from training data, making causal reasoning increasingly important in machine learning.

The field draws on two major formal frameworks. The potential outcomes framework, developed by Neyman and later expanded by Rubin, reasons about what would have happened to a unit under different treatments — a counterfactual perspective. Judea Pearl's structural causal models and do-calculus provide a complementary graphical approach, using directed acyclic graphs (DAGs) to encode assumptions about causal structure and derive identifiable causal quantities from observational data. Together, these frameworks give practitioners rigorous tools for asking and answering causal questions.

In practice, causal inference relies on techniques such as randomized controlled trials (the gold standard), instrumental variables, regression discontinuity designs, difference-in-differences, and propensity score matching. Each method addresses the fundamental challenge of confounding — the presence of variables that influence both the treatment and the outcome, creating spurious associations. When randomization is impossible, these quasi-experimental methods attempt to approximate the conditions of a controlled experiment using observational data.

Causal inference has become increasingly central to machine learning research, particularly in areas like algorithmic fairness, domain generalization, and reinforcement learning. Models trained purely on observational data can inherit and amplify spurious correlations, leading to poor generalization and harmful decisions. By incorporating causal structure, researchers aim to build systems that are more robust, interpretable, and capable of reasoning about interventions — not just predicting outcomes under the status quo. As ML systems are deployed in high-stakes domains like healthcare, economics, and policy, the ability to distinguish causation from correlation is no longer optional.