Microeconometrics

Microeconometrics is the branch of econometrics that studies economic behaviour at the level of individual units—people, households, firms, and the transactions between them—using data observed on those units.

9 frameworks, from 1900 to today · Econometrics · Economics

still a working view ended · each leads to its entry below

Overview

Microeconometrics is the branch of econometrics that studies economic behaviour at the level of individual units—people, households, firms, and the transactions between them—using data observed on those units123. Its central problem is that such data are almost never generated by controlled experiments4. A household's spending, a worker's wage, a firm's decision to enter a market: each is shaped by many forces at once, and the analyst sees only the outcome, not the counterfactual of what would have happened otherwise. Microeconometrics is the accumulated toolkit for separating one influence from another in that setting, for deciding whether a relationship is real or an artifact of selection and chance, and for predicting how individuals will respond to a change in prices, policies, or incentives. Its history is a sequence of attempts to make credible statements about individual behaviour from data that resist easy interpretation.

Regression and the Problem of Selection

The oldest tool is ordinary least squares (OLS), the method of fitting a line or surface to data by minimising the sum of squared deviations5. Its statistical foundations were laid in the early nineteenth century by Legendre and Gauss, and by the early twentieth century it was the workhorse of applied economics. OLS answers a precise question: given observations on an outcome and a set of explanatory variables, what linear combination of those variables best predicts the outcome in the sample? What it does not automatically answer is whether the fitted coefficient measures a causal effect. If the variables that determine both the outcome and the explanatory variable are unobserved—ability in a wage equation, motivation in a schooling equation—the coefficient is contaminated by selection. This gap between prediction and causation is the thread that runs through the whole history of the field.

Simultaneous Equations and the Cowles Commission

In the 1940s and 1950s, a group of economists and statisticians centred at the Cowles Commission for Research in Economics at the University of Chicago confronted a related problem: economic quantities are determined together. The quantity of a good and its price are set jointly by supply and demand, so a regression of quantity on price does not recover either curve6. The Cowles researchers—among them Trygve Haavelmo, Tjalling Koopmans, and later Lawrence Klein—developed the simultaneous equations framework: write the economy as a system of equations, identify which parameters can be recovered from the data given restrictions on the system, and estimate them by methods such as two-stage least squares and limited-information maximum likelihood. Haavelmo's 1944 monograph on the probability approach in econometrics supplied the statistical foundation, treating economic data as realisations of a probability model. The Cowles program made explicit that identification—whether a parameter is recoverable at all—is a separate question from estimation, and that insight remains foundational.

Limited Dependent Variables

Through the 1960s and 1970s, microeconometricians turned to outcomes that are not continuous: whether a person works, which mode of transport they choose, how many children a household has. OLS applied to a binary outcome produces fitted values outside the unit interval and coefficients that are hard to interpret. The response was a family of limited-dependent-variable models. The probit and logit models, whose statistical lineage runs back to the nineteenth century, were adapted to economic choice; Daniel McFadden's work in the 1970s on discrete choice, culminating in the conditional logit model and its extensions, gave economists a rigorous way to model choices among alternatives with different attributes, and earned him a share of the 2000 Nobel Prize. James Tobin's 1958 model of household expenditure on durable goods introduced what became known as the Tobit model for censored outcomes7. These models made it possible to estimate behavioural relationships on the kinds of data microeconomists actually had, and they forced attention onto the assumptions—about the distribution of unobservables—that such estimates require.

Panel Data and Unobserved Heterogeneity

A parallel development addressed the selection problem directly. When the same units are observed repeatedly—a panel—the analyst can control for characteristics that are fixed over time but unobserved. The fixed-effects estimator, which differences out unit-specific constants, and the random-effects estimator, which treats those constants as draws from a distribution, were formalised in the 1960s and 1970s, with major contributions from Pietro Balestra, Marc Nerlove, and Gary Chamberlain, among others. The fixed-effects approach made a strong claim: by comparing a unit to itself over time, it removes any time-invariant confounder, however unmeasurable. Its cost is that it discards all variation between units and cannot identify the effect of anything that does not change. The random-effects approach is more efficient but requires that the unobserved effect be uncorrelated with the regressors—an assumption often implausible. The tension between these two strategies, and the development of tests and extensions (dynamic panels, nonlinear panels, and later interactive fixed effects), has been a continuing theme.

Structural Modelling of Individual Behaviour

From the 1970s onward, a distinct strand sought to estimate the deep parameters of economic models of individual decision-making—preferences, discount rates, the costs of adjustment—rather than reduced-form relationships. Dynamic discrete choice models, in which agents choose among discrete alternatives over time anticipating future consequences, were developed and applied to problems such as retirement, fertility, and machine replacement. The nested fixed-point algorithm, associated with John Rust's work in the late 1980s, made such models computationally feasible. Structural microeconometric modelling aims at counterfactual prediction: if the estimated parameters describe preferences and constraints, the model can be used to simulate the effects of policies never observed. Its critics note that the results depend heavily on the assumed model, and that identification often rests on functional-form assumptions that the data cannot check.

Potential Outcomes and Local Average Treatment Effects

A different response to the selection problem came from the statistics of randomised experiments. The potential-outcomes framework, associated with Jerzy Neyman's 1923 work on agricultural experiments and formalised for observational settings by Donald Rubin in the 1970s, defines a causal effect as the difference between an outcome under treatment and the same unit's outcome under control—only one of which is ever observed. This framing made precise what a causal claim requires. Its extension to economics came through the instrumental-variables tradition. In the 1990s, Joshua Angrist, Guido Imbens, and others showed that when a binary instrument shifts treatment for some units but not others, the instrumental-variables estimand can be interpreted as a local average treatment effect: the average effect for the "compliers," those whose treatment status the instrument changes. This result, published in the mid-1990s, gave a behavioural interpretation to a technique that had previously been justified mainly by algebraic conditions, and it linked the experimental and observational traditions8.

Natural Experiments and the Credibility Revolution

The 1990s and 2000s saw what is often called the credibility revolution: a shift in applied microeconomics toward research designs in which the source of variation is transparent and plausibly exogenous. The strategy is to find a natural experiment—a policy change, a lottery, a discontinuity in eligibility rules—that mimics random assignment. Landmark studies used differences in compulsory schooling laws, the Vietnam-era draft lottery, and geographic and temporal variation in policy to estimate returns to schooling, the effects of military service, and the consequences of minimum wages. Difference-in-differences, regression discontinuity, and event-study designs became standard491011. The appeal was that the assumptions required for identification were stated in plain terms and could be assessed by the reader; the cost was that the estimates were often local to a particular population or period, and that the search for clean variation could steer research toward questions that happened to have good instruments. The debate between design-based and model-based approaches—between the credibility of a narrow estimate and the reach of a structural model—remains live.

Machine Learning Enters the Toolkit

From the 2010s, machine learning methods—regularised regression, decision trees, ensemble methods, neural networks—began to enter microeconometrics. Their initial appeal was prediction: in high-dimensional settings with many potential controls, algorithms can select and combine variables in ways that classical methods handle poorly. But the deeper contribution has been to causal estimation. Methods such as double/debiased machine learning, developed by Victor Chernozhukov and coauthors in the 2010s, use flexible algorithms to control for confounders while retaining valid inference on the parameter of interest. Causal forests, introduced by Susan Athey and Stefan Wager, estimate heterogeneous treatment effects—how an effect varies across individuals—rather than a single average. These tools extend the potential-outcomes tradition rather than replace it, and they raise new questions about interpretability, overfitting, and the distinction between prediction and explanation.

The Present

Microeconometrics today is plural. Design-based causal inference—difference-in-differences, regression discontinuity, instrumental variables, and the potential-outcomes framework—dominates much applied work, especially in labour and development economics. Panel methods and limited-dependent-variable models remain standard, now often combined with flexible machine-learning estimators. Structural modelling continues in industrial organisation, public finance, and elsewhere, where counterfactual policy questions demand a model of behaviour. The open questions are familiar in new form: how to make credible causal claims when clean natural experiments are unavailable; how to combine the transparency of design-based estimates with the reach of structural models; how to use machine learning without sacrificing interpretability or valid inference; and how to handle the increasingly rich, high-dimensional, and unstructured data—administrative records, text, and transactions—that now describe individual behaviour at a scale earlier generations could not have imagined.

References

  1. Pravin K. Trivedi, Wikipedia
  2. Elena Manresa, Wikipedia
  3. Stéphane Bonhomme, Wikipedia
  4. Econometrics (Methods), Wikipedia
  5. Ordinary least squares, Wikipedia
  6. Simultaneous equations model, Wikipedia
  7. Tobit model, Wikipedia
  8. Local average treatment effect, Wikipedia
  9. Difference in differences, Wikipedia
  10. Regression discontinuity design, Wikipedia
  11. Event study, Wikipedia

The frameworks of Microeconometrics

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1900-Present

    Ordinary Least Squares Regression

    What linear combination of variables best predicts an outcome?

    Fly to its tower in Nousfarer

  2. 1944-1980

    Simultaneous equations model

    Which parameters of a system of equations can be recovered from data?

    Fly to its tower in Nousfarer

  3. 1958-Present

    Limited dependent variable

    How do we model choices among discrete or censored outcomes?

    Fly to its tower in Nousfarer

  4. 1960-Present

    Panel data

    How can repeated observations remove unobserved differences between units?

    Fly to its tower in Nousfarer

  5. 1970-Present

    Structural estimation

    What are the deep parameters of individual decision-making?

    Fly to its tower in Nousfarer

  6. 1974-Present

    Rubin causal model

    What does a causal effect mean when only one outcome is observed?

    Fly to its tower in Nousfarer

  7. 1990-Present

    Credibility revolution

    Which research designs give transparent and credible causal estimates?

    Fly to its tower in Nousfarer

  8. 1994-Present
  9. 2010-Present

    Machine Learning for Causal Inference

    How can flexible prediction algorithms improve causal estimates?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Ordinary Least Squares RegressionSimultaneous equations modelLimited dependent variablePanel dataStructural estimationRubin causal modelCredibility revolutionLocal Average Treatment Effects and I…Machine Learning for Causal Inference

Beyond the field

Where Microeconometrics's ideas came from in other fields, and what they led to there: the strongest 12 of 47 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Microeconometrics is a planet in Nousfarer

Its city has a tower for each of these 9 frameworks. Land, walk in, and climb one concept at a time.

Land on Microeconometrics

Related fields

Fields its ideas are linked to

More in Econometrics

See where Microeconometrics sits on the Map of Human Knowledge