Loading field map
Imagine two economists looking at the same data on class size and student test scores. One builds a model from economic theory: parental choice, school funding rules, peer effects. The other searches for a natural experiment: a district that changed class sizes arbitrarily, perhaps due to a funding formula cutoff. Each path embodies a different vision of what makes an empirical finding credible. This tension—between theory-driven structural modeling and design-driven quasi-experimental identification—runs through the entire history of econometrics, shaping every major framework in the field.
In the 1930s and early 1940s, much empirical macroeconomics consisted of descriptive pattern-finding. Researchers compiled business-cycle indicators, computed correlations, and plotted turning points without an explicit economic model. This approach, later labeled "Measurement without Theory" by Tjalling Koopmans and others, treated data as self-sufficient: if the numbers showed a cycle, one could describe it. The National Bureau of Economic Research (NBER) under Wesley Mitchell epitomized this style, producing detailed chronologies of expansions and contractions. Its strength was transparency—the numbers spoke directly. But critics argued that without a probabilistic model, one could not distinguish signal from noise, let alone test hypotheses or predict under new policies.
The decisive break came in 1944 with Trygve Haavelmo's "The Probability Approach in Econometrics." Haavelmo insisted that economic data should be viewed as realizations of a stochastic process, not as deterministic patterns. He argued that any empirical claim presupposes a probability model—whether the researcher admits it or not. By making that model explicit, the econometrician could use formal statistical inference: estimation, hypothesis testing, and prediction. The Cowles Commission, led by Koopmans and including Haavelmo, transformed this insight into a research program. They specified systems of simultaneous equations derived from economic theory, then estimated parameters using methods like limited-information maximum likelihood. The Probability Approach replaced the pattern-finding tradition by imposing three changes: (1) theory must specify the probability distribution of the data; (2) identification—whether parameters can be uniquely recovered—became a prerequisite; (3) inference was framed as a statistical decision problem.
The Probability Approach crystallized into what is now called Structural Econometrics. Its hallmark is the recovery of deep, policy-invariant parameters—preference elasticities, production function coefficients—from a model that explicitly represents optimizing agents and market equilibrium. The Cowles Commission's simultaneous-equation models were the template. A structural approach typically relies on strong assumptions (functional forms, exclusion restrictions, rational expectations) to achieve identification. The payoff is counterfactual analysis: if you know the structural parameters, you can simulate the effect of a tax change or a tariff, even if those policies were never observed. Structural methods dominated macroeconometrics through the 1960s and remain central in industrial organization, labor economics, and macroeconomics. They are the direct descendants of the Probability Approach, preserving its commitment to theory-driven model specification.
By the early 1970s, the structural consensus had frayed. Critics pointed to three weaknesses: (1) aggregate models failed to forecast the stagflation of the 1970s; (2) the Bayesian school challenged the frequentist foundations of inference; (3) micro data revealed the limitations of aggregate models for studying individual behavior. Three new frameworks emerged in response.
Rather than impose theory-derived equations, time-series econometricians let the data speak through autoregressive and moving-average processes. George Box and Gwilym Jenkins's 1970 book provided tools for univariate forecasting that ignored economic theory entirely. More dramatically, Christopher Sims's 1980 "Macroeconomics and Reality" argued that structural models imposed incredible restrictions. He proposed vector autoregressions (VARs), which treat every variable as endogenous and identify shocks through timing assumptions rather than theory. Time-Series Econometrics thus reacted against the Probability Approach by narrowing the role of economic theory: one could either predict without theory (Box-Jenkins) or trace out empirical impulse responses (VARs) without structural priors. Over time, the framework's identification shortcomings—the same criticism it leveled at structural models—led to hybrid approaches like structural VARs, but the core idea of data-driven dynamics persists.
Arnold Zellner's 1971 "An Introduction to Bayesian Inference in Econometrics" brought Bayesian methods into full view. The Bayesian framework competed directly with the frequentist tradition of the Probability Approach. Instead of treating parameters as fixed unknowns to be estimated, Bayesians treat them as random variables with prior distributions that are updated by data via Bayes' theorem. This shift changed the standard of acceptable evidence: Bayesian credibility intervals are direct probability statements about parameters, not long-run frequencies. For decades, Bayesian econometrics remained a minority school, limited by computational hurdles. But with Markov chain Monte Carlo methods in the 1990s, it became practical—and began to absorb into structural and microeconometric practice. Today Bayesian methods coexist with frequentist approaches, and many applied papers report both p-values and posterior probabilities.
The rise of micro data—household surveys, firm panels, program evaluations—demanded methods tailored to individual-level behavior. Microeconometrics derived from the Probability Approach but adapted its toolkit to discrete choices, sample selection, and duration models. Daniel McFadden's conditional logit (1974) and James Heckman's sample-selection model (1979) were landmarks. Microeconometrics preserved the structural tradition's goal of recovering behavioral parameters, but it also adopted design-based insights from statistics (randomization, matching) that eventually split the subfield. This internal division—structural vs. design-based—mirrors the broader tension between theory-driven and data-driven identification.
In the 1990s, a growing number of researchers argued that structural assumptions were too often untestable and that identification should rest on observable features of the research design. Design-Based Econometrics reacted against the Probability Approach's reliance on model-based identification. Its champions—Joshua Angrist, Guido Imbens, Alan Krueger—borrowed from statistics and epidemiology: instrumental variables, regression discontinuity, difference-in-differences. The flagship paper was Angrist and Imbens's 1991 "Identification and Estimation of Local Average Treatment Effects," which showed that instrumental variables identify a specific local effect without parametric assumptions. The framework transformed applied microeconomics by demanding that researchers justify their identification strategy from institutional details—a sharp contrast with the structural tradition's emphasis on theoretical derivation. Design-based methods are now standard in labor economics, public economics, development economics, and political science.
Today, no single framework dominates. Structural macroeconomists continue to estimate dynamic stochastic general equilibrium (DSGE) models using Bayesian methods. Applied microeconomists choose between structural and design-based approaches depending on the question: design-based methods excel at credibly estimating average treatment effects for a specific population; structural methods allow counterfactual policy simulations. Time-series econometricians use VARs to assess monetary policy shocks, while Bayesian econometrics provides a unifying language for handling model uncertainty and shrinkage. The leading frameworks agree on several fundamentals: inference requires a probabilistic model (a legacy of Haavelmo); replication and transparency are paramount; and identification must be defended, not asserted. They disagree on how much untestable theory should be accepted. Design-based researchers see strong assumptions as a last resort; structural researchers see them as the price of answering causal questions that transcend the data. This disagreement is productive: it forces clarity, and many best-practice studies now combine elements—using structural insights to motivate a design or using a design to calibrate a structural parameter. The history of econometrics is not a story of one school defeating another, but of successive frameworks sharpening the question: what makes an empirical claim credible?
Econometrics remains a field in which methodological pluralism is not a weakness but a defining feature. Each framework reveals something the others overlook: measurement without theory shows what data would say if left to itself; the probability approach insists on formal inference; structural models enable counterfactuals; time-series methods capture dynamics unrestricted by theory; Bayesianism quantifies uncertainty in a natural way; microeconometrics handles individual heterogeneity; and design-based methods demand credible identification. No single framework can do everything, and the best applied work increasingly borrows from several.