Loading field map
Microeconometrics is the branch of econometrics concerned with the behavior of individual economic agents—households, firms, workers, consumers—and with the data that record their choices and outcomes. Where macroeconometrics studies aggregate time series such as national output or inflation, microeconometrics works with data on many distinct units observed at a point in time or over a short panel. Its central task is to estimate the causal effects of policies, incentives, and constraints on individual behavior, and to do so credibly when the data were not generated by a controlled experiment.
The field is defined less by a single method than by a shared set of problems. How does an additional year of schooling affect earnings? How does a job-training program affect employment? How does a price change alter a household's consumption? These questions share a common difficulty: the individuals being compared are not randomly assigned to treatment and control groups. Those who choose more schooling, enroll in training, or buy at a higher price differ systematically from those who do not. Microeconometrics is, in large part, the science of making credible comparisons despite this self-selection.
The foundational challenge of microeconometrics is the missing counterfactual. To know the causal effect of a treatment—schooling, a program, a price—on an outcome, one would ideally observe the same individual both with and without the treatment. Since only one of these states is ever observed, the analyst must construct a comparison group whose outcomes approximate what the treated would have experienced absent treatment.
The difficulty is that treated and untreated individuals typically differ in ways that also affect the outcome. A worker who completes a training program may be more motivated than one who does not; a student who attends a selective college may have higher ability than one who does not. If these pre-existing differences are not accounted for, the estimated effect conflates the treatment's impact with the effect of the underlying differences. This problem—selection bias—is the central obstacle that microeconometric methods are designed to overcome.
A second pervasive issue is endogeneity: the explanatory variable of interest is correlated with unobserved factors that influence the outcome. This can arise from omitted variables, from measurement error, or from simultaneity, where the outcome and the explanatory variable influence each other. The presence of endogeneity means that ordinary regression coefficients do not have a causal interpretation, no matter how many control variables are added, because unobserved confounders may remain.
Microeconometrics emerged as a distinct enterprise in the mid-twentieth century, but its intellectual roots lie earlier. Agricultural economists in the 1920s and 1930s studied farmers' responses to prices using cross-sectional survey data, and labor economists analyzed wage determination with individual-level data. These early efforts were largely descriptive, using regression methods to summarize associations.
The modern field took shape in the 1960s and 1970s as economists began to confront selection bias explicitly. A landmark development was the Heckman selection model, introduced in the 1970s, which treated sample selection—such as the fact that we only observe wages for people who choose to work—as a problem to be modeled rather than ignored. Around the same period, the "Lucas critique" of econometric policy evaluation, though aimed at macroeconomics, reinforced a broader skepticism about using reduced-form correlations for policy prediction.
The 1980s and 1990s brought a decisive shift toward what is now called the design-based or quasi-experimental approach. Rather than attempting to model the full structure of behavior, this tradition seeks settings where treatment assignment is "as good as random" conditional on observable variables, or where a natural experiment provides exogenous variation. The key insight was that credible causal inference requires a defensible source of variation, not just elaborate statistical machinery.
This shift was accompanied by a methodological pluralism. Some researchers continued to develop structural models—explicit economic theories of behavior estimated from data—while others pursued reduced-form estimates of treatment effects. The tension between these approaches remains one of the field's defining features.
Ordinary least squares regression remains the workhorse of microeconometrics. The researcher specifies an outcome as a linear function of explanatory variables, estimates the coefficients, and interprets them as partial associations. When the explanatory variable of interest is randomly assigned, or when all confounders are observed and correctly specified, regression coefficients estimate causal effects.
The limitations are well understood. If unobserved confounders exist, regression fails. If the true relationship is nonlinear, a linear specification misleads. If the treatment effect varies across individuals, regression estimates an average that may not correspond to any particular person's experience. These limitations motivated the development of more sophisticated methods, but regression remains the baseline against which other approaches are judged.
One response to selection bias is to make the selection-on-observables assumption: conditional on a set of observed covariates, treatment assignment is independent of potential outcomes. Under this assumption, comparing treated and untreated individuals with the same covariate values yields a valid causal effect.
Matching methods implement this idea directly by pairing each treated unit with one or more untreated units that have similar covariate values. Propensity score methods, introduced in the 1980s, reduce the dimensionality of the problem by matching on the probability of treatment given covariates rather than on the covariates themselves. These methods are intuitive and transparent, but their validity rests entirely on the assumption that no unobserved confounders remain. This assumption is often implausible, and the methods provide no direct test of it.
The instrumental variables (IV) approach addresses endogeneity by finding a variable—the instrument—that affects the treatment but has no direct effect on the outcome except through the treatment. If such a variable exists, it can be used to isolate the exogenous variation in the treatment and estimate its causal effect.
The classic example is schooling and earnings. Because ability affects both, the estimated return to schooling is biased. An instrument such as quarter of birth—which affects schooling through compulsory schooling laws but is plausibly unrelated to ability—can break the endogeneity. IV methods are powerful when a credible instrument exists, but good instruments are rare. A weak instrument—one that explains little variation in the treatment—produces imprecise and potentially misleading estimates. An invalid instrument—one that affects the outcome through other channels—produces biased estimates that may be worse than no correction at all.
The difference-in-differences (DiD) method exploits panel data or repeated cross-sections to compare changes over time between a treated group and an untreated comparison group. The identifying assumption is that, absent treatment, the two groups would have followed parallel trends. Under this assumption, the difference in the treated group's change over time minus the comparison group's change isolates the treatment effect.
DiD has become enormously popular because it requires only the parallel-trends assumption, which is often more plausible than the assumptions of matching or IV. It has been applied to minimum wage changes, health insurance expansions, and many other policy reforms. Recent methodological work has clarified that DiD estimates are weighted averages of group-specific effects and that the method can be biased when treatment timing varies across units and effects evolve over time. These refinements have led to new estimators that are robust to such complications.
The regression discontinuity (RD) design exploits a threshold rule: units on one side of a cutoff receive treatment, while those on the other side do not. If the cutoff is arbitrary and units cannot precisely manipulate their position relative to it, then units just on either side of the threshold are nearly identical, and the discontinuity in outcomes at the cutoff estimates the treatment effect.
RD designs arise naturally in many settings: scholarship eligibility based on test scores, program eligibility based on income thresholds, and electoral outcomes based on vote margins. The method's appeal is its transparency—the identifying assumption is local and often quite credible. Its limitation is that the estimated effect applies only to units near the threshold, which may not generalize to the broader population.
A distinct tradition estimates explicit economic models of individual behavior. Rather than seeking a reduced-form treatment effect, the structural approach specifies a utility function, a budget constraint, and an optimization problem, then estimates the parameters of that model from data. The goal is not just to measure an effect but to understand the mechanism and to simulate counterfactual policies.
Structural models are common in industrial organization, where researchers estimate demand systems and firm behavior to evaluate merger policy or market power. They are also used in labor economics to model job search and in public economics to analyze tax responses. The advantage is that structural estimates can answer "what if" questions that reduced-form estimates cannot, such as the effect of a policy that has never been observed. The cost is that the results depend on the model's assumptions, which are often strong and difficult to verify. If the model is misspecified, the estimates may be badly wrong even if they fit the data well.
The most influential development of recent decades has been the rise of randomized controlled trials (RCTs) and the broader quasi-experimental movement. RCTs, long standard in medicine, were increasingly adopted in development economics and social policy beginning in the 1990s. By randomly assigning treatment, an RCT eliminates selection bias by construction, making causal inference straightforward.
The quasi-experimental movement extends the logic of the experiment to observational data. Its practitioners search for settings where nature, policy, or institutional rules approximate random assignment, and they design their empirical strategy around that source of variation. This approach has been enormously influential, reshaping how empirical work is conducted and evaluated. Its critics argue that it has gone too far, privileging internal validity—the credibility of the estimate in a particular setting—over external validity, the generalizability of findings to other contexts. They also note that many important questions do not admit of experimental or quasi-experimental answers.
The field today is characterized by methodological eclecticism and a healthy debate about the proper balance between internal and external validity, between reduced-form and structural approaches, and between design-based and model-based inference.
Several developments define the current landscape. Machine learning methods have entered the toolkit, used for high-dimensional control variables, for estimating heterogeneous treatment effects, and for improving the precision of matching and weighting estimators. These methods are powerful but raise their own questions about inference and interpretability.
The credibility revolution has raised the bar for empirical work. Journals now demand that authors justify their identification strategy explicitly, discuss threats to validity, and conduct robustness checks. This has improved the quality of applied work but has also led to a certain narrowness, as researchers gravitate toward questions that admit clean identification rather than questions that matter most.
The debate between structural and reduced-form approaches has become more nuanced. Many researchers now combine elements of both, using reduced-form estimates to discipline structural models or using structural models to interpret reduced-form findings. The old antagonism has softened into a recognition that the two approaches answer different questions and that the choice depends on the research question.
A further development is the increasing attention to external validity and generalizability. As the field has accumulated credible estimates from many specific settings, questions have arisen about how to aggregate findings, whether effects estimated in one context transfer to another, and how to design research programs that produce both credible and useful knowledge.
The field also faces ongoing challenges. Publication bias—the tendency for journals to publish statistically significant results—distorts the accumulated evidence. Replication and data-sharing norms are improving but remain uneven. And the increasing complexity of methods has raised concerns about transparency and the reproducibility of empirical findings.
Beneath the methodological debates lie enduring questions that define the field's purpose. How can we learn about cause and effect from data that were not generated by experiment? What assumptions are necessary for causal inference, and how can we assess their plausibility? How should we balance the desire for credible estimates against the need for generalizable knowledge? How can we use economic theory to interpret empirical findings without letting unverifiable assumptions drive the conclusions?
These questions have no final answers. The field advances by developing new methods that relax old assumptions, by finding new sources of variation that make causal inference more credible, and by refining our understanding of when existing methods work and when they fail. What remains constant is the commitment to using data to understand how individuals behave and how policies affect them—a task that requires both technical skill and careful judgment about what the data can and cannot support.