Around the year 2000, applied economics was in the midst of what later came to be called a “credibility revolution”: a shift away from heavily model-dependent, structural claims toward research designs that made causal identification as transparent as possible. The backdrop was decades of debate over whether econometric estimates in labor, education, and development economics were credible as causal statements. In the 1970s and 1980s, simultaneous-equations models and elaborate functional-form assumptions often carried the burden of identification. Yet economists increasingly confronted a practical problem: policy questions—Does schooling raise earnings? Do minimum wages reduce employment? Do immigration inflows hurt native wages?—hinged on causal counterfactuals that were hard to infer from observational data. Joshua Angrist, David Card, and Guido Imbens became central figures in a new approach: treat causality as an identification problem, look for quasi-random variation in the world (“natural experiments”), and use econometric tools—especially instrumental variables—to link that variation to credible causal effects.
The core ideas crystallized through a series of influential papers in the 1990s that shaped the research culture economists brought into the new millennium. Card’s 1990 study of the Mariel Boatlift, which exploited the sudden arrival of Cuban immigrants in Miami in 1980, exemplified an event-based comparison design. Card and Alan Krueger’s minimum wage work (notably their 1994 paper and 1995 book) used differences in state policies as a quasi-experiment to challenge a canonical prediction of competitive labor markets. Angrist and Krueger’s 1991 paper used quarter of birth—interacting compulsory schooling laws with school-starting-age rules—as an instrument for education. Angrist’s work on Vietnam-era draft lotteries (early 1990s) showed how literal randomization could be embedded in administrative processes. Imbens, often collaborating with Angrist, supplied the conceptual and mathematical foundations for what instrumental variables actually identifies when treatment effects vary across individuals. Their 1994 Econometrica paper, “Identification and Estimation of Local Average Treatment Effects,” and subsequent work reframed IV not as a mechanical fix for endogeneity, but as a design-based estimator tied to a specific causal estimand.
At the heart of the credibility revolution is the potential outcomes framework. For each unit (i), define potential outcomes (Y_i(1)) and (Y_i(0)) corresponding to treated and untreated states. The causal effect is (Y_i(1)-Y_i(0)), but only one outcome is observed: (Y_i = D_i Y_i(1) + (1-D_i)Y_i(0)), where (D_i\in{0,1}) is treatment status. The central identification problem is that (D_i) is typically correlated with unobserved determinants of (Y_i), so the naïve difference in means does not equal a causal effect. Natural experiments aim to isolate variation in (D_i) that is “as good as random,” conditional on design assumptions, so that treated and control groups are comparable in the relevant counterfactual sense.
Instrumental variables became the flagship method. An instrument (Z_i) affects treatment (D_i) but does not directly affect the outcome except through (D_i). In the simplest linear setup, one writes $$ Y_i = \alpha + \beta D_i + \varepsilon_i, $$ with endogeneity (E[\varepsilon_i \mid D_i]\neq 0). An instrument satisfies (i) relevance (Cov(Z_i,D_i)\neq 0) and (ii) exclusion (E[\varepsilon_i \mid Z_i]=0). The population IV estimand is $$ \beta_{IV}=\frac{Cov(Z_i,Y_i)}{Cov(Z_i,D_i)}. $$ But the credibility revolution’s key advance was to clarify what (\beta_{IV}) means when treatment effects are heterogeneous. Under the additional “monotonicity” condition—no “defiers,” meaning the instrument does not push some people systematically into treatment while pushing others out—Angrist and Imbens showed that IV identifies a Local Average Treatment Effect (LATE): the average causal effect for “compliers,” those whose treatment status is changed by the instrument. In potential-outcomes terms, if (D_i(z)) is the treatment that would be chosen under instrument value (z), compliers satisfy (D_i(1)=1) and (D_i(0)=0). LATE is $$ E!\left[Y_i(1)-Y_i(0)\mid D_i(1)>D_i(0)\right]. $$ This interpretation transformed how economists read IV results: not as a universal “return to schooling,” but as the return for people induced to get more schooling by a specific policy lever (such as compulsory attendance interacting with birth timing).
By 2000, these ideas had reshaped applied labor economics and were rapidly diffusing to public, development, and political economy. The impact was methodological and sociological. Methodologically, economists emphasized transparent identification strategies—difference-in-differences, regression discontinuity, randomized controlled trials when feasible, and IV with clear institutional narratives. Sociologically, norms of evidence changed: journals and seminars increasingly demanded explicit discussion of counterfactuals, threats to identification, balance checks, placebo tests, and robustness to alternative specifications. The “reduced-form” relationships (E[Y\mid Z]) and (E[D\mid Z]) were treated as first-class objects, not mere stepping-stones to structural parameters.
Criticism accompanied the revolution. One critique was external validity: LATE is local, and different instruments identify different complier groups. Another was that the design emphasis could crowd out theory, leaving economists with credible estimates of narrow effects but limited guidance for welfare analysis or general equilibrium responses. A further concern was that “natural experiments” are not truly random; institutional stories can be incomplete, and subtle selection mechanisms can remain. In IV specifically, weak instruments—small (Cov(Z,D))—can bias estimates and inflate uncertainty; by 2000, econometricians were already developing diagnostics and robust inference methods to address this. There were also philosophical disputes about whether economics should prioritize design-based partial effects or structurally interpretable parameters capable of simulating counterfactual policies.
The legacy today is visible across empirical economics. The credibility revolution helped standardize the language of causal inference—identification, assumptions, estimands, and research design—bringing economics closer to modern statistical causal analysis while retaining its policy focus. It also accelerated econometric developments in inference under clustering, heteroskedasticity, many instruments, and finite-sample robustness—pragmatic tools needed for credible applied work. Importantly, the Angrist–Imbens LATE framework continues to discipline interpretation: researchers are pushed to ask “whose effect?” and “which margin of change?” rather than claiming a single policy-invariant parameter.
In the line of continuity, this movement did not emerge from nowhere. It built on earlier econometrics—Cowles Commission structural modeling, the identification debates of the 1940s–1960s, and the IV tradition dating to early 20th-century statistics—while reacting against overconfidence in specification-driven causality. It also connected to the Rubin causal model in statistics and to program evaluation methods developed in the 1970s and 1980s. After 2000, the revolution fed into adjacent waves: the rapid expansion of regression discontinuity designs, the scaling of randomized evaluations in development economics, and later the integration of machine learning with causal inference (“double/debiased” methods and causal forests). Yet the intellectual signature associated with Angrist, Card, and Imbens remains the same: start from a concrete source of quasi-random variation, state assumptions plainly, and let the design—not the elegance of the model—carry the causal claim.