Loading field map
The formal econometric pursuit of causal inference originated with the Cowles Commission's Structural Econometrics program in the 1940s and 1950s. This paradigm, championed by figures like Haavelmo, Koopmans, and Hood, established the probability approach to econometrics, treating causality as a property of deep, theory-derived simultaneous equations models. Identification was secured through exclusion restrictions and other theoretical assumptions, with the goal of estimating structural parameters that were invariant to policy changes. This framework dominated for decades, facing early criticism from the Measurement without Theory stance associated with Burns, Mitchell, and the NBER, which advocated for descriptive empirical cycles over premature theoretical formalization.
A profound shift began in the 1970s with the articulation of the Rubin Causal Model (or Potential Outcomes Framework) by Donald Rubin, building on foundational work by Neyman. This introduced a design-based approach to causality, separating the definition of causal effects (via potential outcomes) from the methods to estimate them. This paradigm gained decisive influence in applied microeconomics through the Design-Based Approach (or Credibility Revolution), led by Angrist, Imbens, and others. It prioritized research designs like randomized experiments, instrumental variables, regression discontinuity, and difference-in-differences that aimed to mimic experimental conditions, often explicitly avoiding heavy reliance on structural economic theory.
The late 20th and early 21st centuries saw the maturation and extension of these core paradigms. Microeconometrics, as a broad school, integrated both design-based and structural methods for cross-sectional and panel data, focusing on individual- and firm-level causality. Simultaneously, Bayesian Econometrics developed a comprehensive framework for causal inference, formally incorporating prior information and offering a unified approach to estimation and uncertainty within both structural and potential outcomes models. Recent syntheses seek to bridge the historical divide, while new frontiers explore the integration of causal inference with machine learning for high-dimensional settings and the formal development of Partial Identification methods, which quantify the bounds on causal effects when point identification is not possible.