Tracing the Arc of Causal Inference in Economics

 The provided sources identify the 18th and 19th centuries as the foundational "Philosophy" era of causal inference in economics, setting the stage for three centuries of methodological evolution.

18th Century: David Hume and the Problem of Induction

The sources credit David Hume with initiating the "modern problem of causation". His contributions, specifically in A Treatise of Human Nature (1739) and An Enquiry Concerning Human Understanding (1748), focused on a fundamental skeptical challenge known as the Problem of Induction. Hume questioned the empirical basis of causality, asking: "what do we actually observe when we say that one thing causes another?". The sources note that this philosophical question remains without a "fully satisfying answer" even in the modern era.

19th Century: John Stuart Mill and Methods of Inquiry

The timeline progresses into the 19th century with John Stuart Mill, specifically citing his 1843 "Methods of Inquiry". While the text does not elaborate on Mill's specific methods, it positions his work as the next major philosophical milestone following Hume’s skepticism.

The Larger Context in Economics

In the broader context of causal inference, these philosophical roots are the starting point for a lineage that later transitioned into various methodological frameworks:

• Experiments: Moving from theory to practical application with figures like Neyman (1923) and Fisher (1935).
• Structural and Predictive Causality: Developing formal models such as Haavelmo’s probability approach (1944) and the Lucas critique (1976).
• The Credibility Revolution: Addressing the "credibility crisis" in the late 20th century through the work of Leamer, Rubin, and Angrist.
• Modern Methods: Integrating Directed Acyclic Graphs (DAGs) and Machine Learning into causal analysis.

The sources suggest that the philosophical inquiries of Hume and Mill established a "recurring tension" that persists in economics today: the choice between Design vs. Structure (identifying effects without knowing mechanisms) and Local vs. General (the ability of estimates to generalize).

In the early 20th century, the study of causal inference in economics transitioned from philosophical inquiry into the "Experiments" era, which provided the mathematical and statistical foundations for identifying causal effects.

According to the timeline provided in the sources, this period is defined by two landmark contributions:

• 1923: Neyman and Potential Outcomes: Jerzy Neyman introduced the concept of potential outcomes, a framework that remains central to causal inference today. This approach allows researchers to conceptualize what would have happened to the same unit under different treatment conditions.
• 1935: Fisher and Randomization: Ronald Fisher formalised the role of randomization. By randomly assigning treatments, researchers could ensure that groups were comparable, thereby isolating the causal effect of a specific variable.

The Larger Context in Economics

In the broader history of economic methodology, the "Experiments" era represents a shift from the skepticism of the 18th and 19th centuries (Hume and Mill) toward a more practical, design-based approach to science.

1. Addressing the "Design vs. Structure" Tension: This era prioritizes "Design"—the use of experimental controls and randomization—to identify effects. This often sits in tension with "Structure," which seeks to understand the underlying mechanisms of why something happens. The sources note that a recurring question for this approach is: "Can we identify effects without mechanisms?".
2. Foundation for the "Credibility Revolution": The work of Neyman and Fisher laid the groundwork for the later "Credibility Revolution" (1970s–1990s). During that later period, economists like Rubin (1974) refined the causal model and Angrist (1996) developed the LATE (Local Average Treatment Effect) framework, both of which are deeply rooted in the early 20th-century experimental logic.
3. Generalization Challenges: The sources highlight a second recurring tension relevant to this era: "Local vs. General." While experiments (and the potential outcomes framework) are powerful for identifying effects in a specific context, there is a persistent debate over whether these estimates generalize to broader populations or different settings.

While these early 20th-century developments were revolutionary, the sources characterize the history of causal inference as an "unfinished" journey that continues to evolve through structural modeling, the credibility revolution, and modern machine learning.

In the mid-20th century, the study of causal inference in economics entered the "Structural" era, a period characterized by the development of formal mathematical and probabilistic models to explain the underlying mechanisms of economic behavior.

According to the provided sources, this era is defined by several pivotal milestones:

• 1944: Haavelmo and the Probability Approach: Trygve Haavelmo pioneered the probability approach to econometrics, which provided a formal framework for treating economic models as systems of simultaneous equations.
• 1969: Granger and Predictive Causality: Clive Granger introduced predictive causality (now known as Granger causality), a method for determining whether one time series is useful in forecasting another, which added a temporal dimension to causal analysis.
• 1976: Lucas and the Policy Critique: Robert Lucas famously argued that historical relationships between economic variables might change if government policy changes because individuals adjust their expectations. This policy critique highlighted the danger of relying on "structure" that is not grounded in fundamental behavioral parameters.
• 1979: Heckman and the Selection Model: James Heckman developed the selection model to address bias arising from non-random samples, providing a structural way to account for why certain data points are observed while others are not.

The Larger Context in Economics

Within the broader evolution of the field, the Structural era represents a specific philosophical and methodological stance:

1. The "Structure" in Design vs. Structure: This era directly addresses the "recurring tension" of whether we can identify effects without understanding mechanisms. Unlike the preceding "Experiments" era (Neyman and Fisher), which focused on the design of trials to isolate effects, the Structural era sought to model the mechanisms—the "how" and "why" behind economic phenomena.
2. Addressing Generalizability: This period also speaks to the tension of Local vs. General. By attempting to model the fundamental "structure" of the economy, researchers in this era aimed to create estimates that could generalize across different policies and environments, rather than just being valid for a specific experimental group.
3. Bridge to the Credibility Revolution: The sources position this era between the early experimentalists and the "Credibility Revolution" of the late 20th century. While structural modeling provided deep insights into mechanisms, the later revolution (led by figures like Leamer, Rubin, and Angrist) would eventually challenge the "credibility" of these complex structural assumptions, pushing the field back toward design-based approaches.

Ultimately, the Structural era was an ambitious attempt to provide the "satisfying answer" to David Hume’s 18th-century skepticism by moving beyond mere observation to a deep, model-based understanding of causal relationships.

In the late 20th century, the Credibility Revolution emerged as a critical methodological shift within the history of causal inference in economics, primarily aimed at addressing the reliability of empirical findings.

According to the sources, this era is defined by three landmark contributions:

• 1974: Rubin and the Causal Model: Donald Rubin introduced a formal causal model (often referred to as the Rubin Causal Model), which built upon the potential outcomes framework to provide a rigorous mathematical basis for identifying causal effects in non-experimental data.
• 1983: Leamer and the "Credibility Crisis": Edward Leamer published a highly influential critique of econometric practices, highlighting a "credibility crisis". He argued that many empirical results were fragile and highly dependent on specific, often arbitrary, modeling choices made by researchers.
• 1996: Angrist and the LATE Framework: Joshua Angrist developed the LATE (Local Average Treatment Effect) framework, which provided a clear interpretation for causal estimates derived from instrumental variables, acknowledging that these effects are often "local" to a specific sub-population affected by the instrument.

The Larger Context in Economics

The Credibility Revolution is positioned as one of "at least two methodological revolutions" in a three-century-long timeline that remains "unfinished". Within the broader evolution of the field, this era represents a pivotal response to the "Recurring Tensions" of causal inference:

1. Design vs. Structure: This era signaled a move away from the complex, assumption-heavy "Structural" models of the mid-20th century. Instead, it prioritized "Design"—emphasizing research designs (like natural experiments) that could identify causal effects even if the underlying behavioral mechanisms were not fully modeled. This directly addresses the question: "Can we identify effects without mechanisms?".
2. Local vs. General: The Credibility Revolution brought the tension of "Do estimates generalize?" to the forefront. While the LATE framework (1996) offered a way to identify credible causal effects, it also forced economists to confront the fact that these effects are often "Local" and may not easily generalize to broader populations or different contexts.

By focusing on transparency and robust research designs, the Credibility Revolution sought to provide a more "satisfying answer" to the fundamental skepticism first raised by David Hume in the 18th century regarding what we truly observe when we claim one thing causes another.

In the 21st century, the field of causal inference has entered its "Modern" era, which the sources characterize as an integration of computer science, graphical modeling, and machine learning into economic analysis.

According to the provided timeline, this era is defined by three major milestones:

• 2000: Pearl and DAGs: Judea Pearl introduced Directed Acyclic Graphs (DAGs) and do-calculus. This provided a new mathematical and visual language to represent causal assumptions and rigorously determine whether a causal effect can be identified from available data.
• 2018: Athey and Causal Forests: Susan Athey developed causal forests, a machine learning approach based on random forests. This method is specifically designed to estimate heterogeneous treatment effects, allowing researchers to understand how causal impacts vary across different types of individuals or environments.
• 2018: Double Machine Learning (DML): This period also marked the rise of DML (ML + causality), a framework that uses machine learning to better control for complex, high-dimensional variables that might bias causal estimates.

The Larger Context in Economics

The Modern era is the latest chapter in a journey that "spans three centuries and at least two methodological revolutions". Within the broader evolution of the field, these modern developments address the "Recurring Tensions" identified in the sources:

1. Design vs. Structure: Modern methods like DAGs and DML offer a way to bridge this tension. While the Credibility Revolution (Late 20th Century) prioritized design, the Modern era uses graphical structures and machine learning to bring back a level of "structure" that is more flexible and data-driven than the simultaneous equations of the mid-20th century.
2. Local vs. General: By focusing on tools like causal forests that identify varying effects across populations, the Modern era directly tackles the question: "Do estimates generalize?". These tools move beyond the single "Local" estimates of the late 20th century toward a more nuanced understanding of how effects might apply more generally or change in different contexts.
3. An "Unfinished" Journey: The sources emphasize that despite the sophistication of 21st-century machine learning, the field remains "in important ways, unfinished". The journey from Hume’s 1739 skepticism to modern algorithms shows that while our tools have become more powerful, the fundamental philosophical challenge of what we "actually observe" when we claim causality still lacks a "fully satisfying answer".

Based on the sources and our conversation history, the "Recurring Tensions" represent the fundamental, unresolved debates that have persisted throughout the three-century evolution of causal inference in economics. These tensions highlight the trade-offs researchers face when choosing a methodological approach.

The sources identify two primary recurring tensions:

1. Design vs. Structure: Can we identify effects without mechanisms?

This tension centers on whether a researcher should focus on the "Design" of a study (how data is generated) or the Structure of the underlying economic system (the theoretical "why" behind an effect).

• Design-focused eras: In the Experiments (1920s-30s) and Credibility Revolution (1970s-90s) eras, the priority was on isolating a specific effect through randomization or natural experiments. This approach often identifies that something happened without necessarily explaining the behavioral mechanisms.
• Structure-focused eras: The Structural era (mid-20th century) prioritized modeling the internal logic and mechanisms of the economy (e.g., Haavelmo’s probability approach or Lucas’s policy critique).
• Modern Synthesis: The Modern era (21st century) uses tools like DAGs and Machine Learning to attempt to reconcile these two, using data-driven structures to inform better design.

2. Local vs. General: Do estimates generalize?

This tension addresses the external validity of causal findings—whether a result found in one specific context can be applied to other populations or policies.

• The "Local" Challenge: The Credibility Revolution (specifically Angrist’s 1996 LATE framework) acknowledged that many credible designs only identify effects for a specific sub-group (a "local" effect).
• The "General" Goal: The Structural era aimed for more generalizable "structural parameters" that would remain stable even if policies changed.
• Modern Advancements: Current methods, such as Susan Athey’s causal forests (2018), use machine learning to map out heterogeneous treatment effects, providing a more nuanced way to see how "local" findings might generalize across different environments.

The Larger Context: An Unfinished Journey

These tensions exist within a larger historical context that begins with David Hume’s 1739 skepticism regarding the "problem of induction". The sources emphasize that because we cannot "actually observe" one thing causing another, these tensions remain "unfinished". Despite two major methodological revolutions and the rise of modern algorithms, there is still "no fully satisfying answer" to the core philosophical challenges of causality, making these recurring tensions the central drivers of ongoing innovation in the field.