Causal Inference

Causal inference is the part of statistics that tries to answer not just what is associated with what but what would happen if we intervened.

13 frameworks, from 1923 to today · Statistics · Mathematics

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

Overview

Causal inference is the part of statistics that tries to answer not just what is associated with what but what would happen if we intervened1. Does a drug save lives, or do healthier people simply happen to take it? Would a minimum wage raise unemployment, or do states that raise it differ from those that don't in ways that matter? These are causal questions, and the central difficulty is that we never observe the same unit both treated and untreated. Causal inference builds the concepts, designs, and estimators that let researchers reason about such counterfactuals from data—and, just as importantly, state honestly what the data cannot tell them.

From Correlation to Cause: The Prehistory

For most of the nineteenth and early twentieth centuries, statistics was wary of causation. The dominant tradition, associated with Karl Pearson and the biometricians, treated correlation as the proper scientific object and causation as metaphysics2. John Stuart Mill's methods of agreement and difference (1843) had given philosophers a logic of causal reasoning, and in 1899 the epidemiologist Wade Hampton Frost and others were already reasoning informally about confounding, but statistics as a discipline had no formal machinery.

The germ of one appeared in 1923, when the geneticist and statistician Sewall Wright published a method of path coefficients—diagrams of arrows between variables whose strengths could be estimated from correlations. Wright used it to disentangle the contributions of heredity and environment to guinea-pig coat color and, later, to interpret corn and livestock breeding data. His work was largely ignored by statisticians for decades, but it was the first serious attempt to represent causal structure in a formal, estimable model2.

Randomization and the Potential Outcomes Idea

The decisive break came from experimental design. In the 1920s at Rothamsted Experimental Station, Ronald Fisher developed the randomized experiment: assign treatments to plots by chance, and the treatment groups become comparable in expectation, so differences in outcomes can be attributed to the treatment. Fisher's The Design of Experiments (1935) codified randomization, blocking, and the analysis of variance. Randomization solved confounding by construction—but it was a design principle, not a theory of what a causal effect is.

That theory arrived from two directions. In 1923, in a Polish-language paper on agricultural experiments, Jerzy Neyman introduced what we now call potential outcomes: each unit has a pair of outcomes, one under treatment and one under control, and the causal effect is their difference32. Neyman used this to justify the analysis of randomized experiments and to derive confidence intervals. In the 1970s, Donald Rubin rediscovered and generalized the idea—now often called the Rubin causal model—extending it to observational studies by treating the treatment assignment itself as something to be modeled. The framework's core object is the counterfactual: the outcome a unit would have had under a treatment it did not receive. Since only one potential outcome per unit is ever observed, causal inference is fundamentally a missing-data problem.

Observational Studies and the Problem of Confounding

Randomization is often impossible—you cannot randomly assign people to smoke4. In 1950, Austin Bradford Hill and Richard Doll published the first strong epidemiological evidence linking cigarette smoking to lung cancer. The finding was attacked on the grounds that smokers might differ from non-smokers in other ways. Hill's response, elaborated in his 1965 essay on "The Environment and Disease: Association or Causation?", was a set of causal criteria—strength, consistency, temporality, biological gradient, plausibility, and others—for judging when an association is likely causal. These criteria were heuristics, not a formal theory, but they shaped epidemiology for a generation1.

The formal theory of observational studies came from William Cochran and, above all, Donald Rubin, whose 1974 paper "Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies" laid out the potential-outcomes framework for observational data. Rubin's key assumption was strong ignorability: treatment assignment is independent of potential outcomes once we condition on observed covariates. Under that assumption, adjusting for the right variables identifies the causal effect. The framework made precise what "controlling for confounders" had always meant informally.

Matching, Stratification, and the Propensity Score

If adjustment requires conditioning on many covariates, the practical problem is that exact matching on many variables becomes impossible as their number grows. Paul Rosenbaum and Donald Rubin solved this in 1983 with the propensity score: the probability of receiving treatment given observed covariates4. They proved that, under ignorability, conditioning on this single scalar balances the covariates between treated and control groups—so matching or stratifying on the propensity score is enough. The result turned a high-dimensional adjustment problem into a one-dimensional one and spawned a large matching literature (Rosenbaum and Rubin's 1985 paper on bias reduction, and later work by Guido Imbens, Donald Rubin, and others on optimal matching and its properties).

Natural Experiments: IV, Regression Discontinuity, and Difference-in-Differences

A parallel tradition asked what can be learned when treatment is not assigned by a researcher but by circumstance. Instrumental variables originated in economics: Philip Wright's 1928 work on supply and demand for butter and flax used an external variable to identify a demand curve, and the method was formalized in the 1940s by Trygve Haavelmo and the Cowles Commission. An instrument is a variable that affects treatment but affects the outcome only through treatment; it identifies a causal effect for the subpopulation whose treatment it moves. In the 1990s, Joshua Angrist, Guido Imbens, and others connected IV to the potential-outcomes framework, clarifying that it estimates a local average treatment effect—the effect for "compliers" whose treatment responds to the instrument5.

Regression discontinuity exploits a threshold: if a scholarship is awarded above a test score cutoff, students just above and just below the cutoff are comparable, and the jump in outcomes at the cutoff estimates the effect6. The design was introduced by Donald Thistlethwaite and Donald Campbell in 1960 in a study of scholarship awards, and was revived and formalized in the late 1990s and 2000s by Guido Imbens, Thomas Lemieux, Justin McCrary, and others.

Difference-in-differences compares the change in outcomes in a treated group to the change in a control group, differencing out fixed differences between them7. It became a workhorse of applied economics after John Snow's 1855 analysis of a London cholera outbreak is often cited as an early instance, and its modern form was developed in labor economics from the 1980s onward (notably by Orley Ashenfelter and David Card). The key assumption—parallel trends absent treatment—is untestable in general, which has driven a large recent literature on its diagnostics and on staggered adoption designs.

Causal Graphs and the Structural Turn

A third tradition grew from Sewall Wright's path diagrams, revived by computer scientists and philosophers. In the 1980s and 1990s, Judea Pearl developed causal graphical models—directed acyclic graphs (DAGs) in which an arrow means "directly causes." Pearl's Causality (2000) gave the framework its mature form, including the do-calculus, a set of rules for determining whether a causal effect is identifiable from observational data given the graph. The graph makes confounding visible: a backdoor path is a non-causal path between treatment and outcome, and adjusting for a set of variables that blocks all backdoor paths identifies the effect. Peter Spirtes, Clark Glymour, and Richard Scheines's Causation, Prediction, and Search (1993) addressed the inverse problem—learning graph structure from data. The graphical and potential-outcomes traditions were shown to be closely related, and both are now standard.

Sensitivity Analysis

Because ignorability is an assumption, not a fact, a central question is how robust a conclusion is to its violation. Paul Rosenbaum's 1987 work on sensitivity analysis introduced a formal way to ask: how strong would an unmeasured confounder have to be to overturn the finding? The approach, developed with Rosenbaum's Observational Studies (1995), quantifies the departure from randomization needed to explain away an effect, and remains the standard tool for assessing unmeasured confounding.

Doubly Robust and Targeted Estimation

In the 2000s, attention turned to estimators that combine outcome modeling and treatment modeling. Heuristics for combining them had circulated, but James Robins, Andrea Rotnitzky, and Lue Ping Zhao (1994) gave the rigorous theory of doubly robust estimators: consistent if either the outcome model or the treatment model is correctly specified, not necessarily both. This work grew out of Robins's broader development of g-methods—g-computation, inverse probability weighting, and g-estimation—for time-varying treatments, where standard regression fails because treatment and confounders feed back on each other over time. In the 2010s, Mark van der Laan and Sherri Rose developed targeted maximum likelihood estimation (TMLE), a doubly robust, semiparametric-efficient approach built on machine-learning fits of the nuisance models, and Aad van der Vaart, Peter Bickel, and others supplied the underlying semiparametric efficiency theory.

Heterogeneous Effects and Causal Machine Learning

Classical methods estimated an average effect. But a drug may help some patients and harm others, and a policy may work in some regions and not others. The modern literature asks how effects vary—conditional average treatment effects—and how to estimate them with flexible models. Charles Manski's work on partial identification (1990) established what can be learned when assumptions are weak. Susan Athey and Guido Imbens's causal forests (2016) and the broader causal machine learning program adapted tree ensembles, neural networks, and other tools to estimate heterogeneous effects and to select covariates for adjustment, while preserving valid inference. Victor Chernozhukov, Denis Chetverikov, and colleagues' double/debiased machine learning (2018) showed how to use high-dimensional or nonparametric nuisance estimates without contaminating the causal estimate—a direct descendant of the doubly robust tradition.

The Present

Causal inference today is a meeting point of several lineages: the potential-outcomes framework (Rubin, Neyman) supplies the definition of a causal effect; graphical models (Pearl) supply the logic of identification; design-based methods (randomization, IV, regression discontinuity, difference-in-differences) supply credible strategies when randomization is unavailable; and semiparametric and machine-learning theory supplies estimators that are flexible yet valid1. The field's central open questions are the ones its history kept returning to: how to justify ignorability in a given application, how to detect and bound the effects of unmeasured confounding, how to handle time-varying treatments and interference between units, and how to make heterogeneous-effect estimates trustworthy enough to guide individual decisions. The 2021 Nobel Prize in Economic Sciences, awarded to David Card, Joshua Angrist, and Guido Imbens for work on natural experiments and methods for causal inference, marked how far the field had traveled from the days when causation was considered outside the statistician's remit.

References

  1. Causal inference (introduction, Approaches in epidemiology, General), Wikipedia
  2. Causal model (History), Wikipedia
  3. Rubin causal model, Wikipedia
  4. Propensity score matching, Wikipedia
  5. Local average treatment effect, Wikipedia
  6. Regression discontinuity design, Wikipedia
  7. Difference in differences, Wikipedia

The frameworks of Causal Inference

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

  1. 1923-Present

    Rubin causal model

    What is a causal effect for a single unit?

    Fly to its tower in Nousfarer

  2. 1925-Present

    Randomized Experimental Design

    How does random assignment identify a treatment effect?

    Fly to its tower in Nousfarer

  3. 1928-Present

    Instrumental Variables

    How can an external variable identify a causal effect?

    Fly to its tower in Nousfarer

  4. 1960-Present

    Regression Discontinuity Design

    When does a cutoff create a usable comparison?

    Fly to its tower in Nousfarer

  5. 1965-1990

    Bradford Hill criteria

    When is an observed association causal?

    Fly to its tower in Nousfarer

  6. 1980-Present

    Natural experiment

    What can be learned when chance assigns treatment?

    Fly to its tower in Nousfarer

  7. 1983-Present

    Propensity Score and Matching Methods

    How can many covariates be balanced at once?

    Fly to its tower in Nousfarer

  8. 1986-Present

    Causal graph

    Which variables must be adjusted for to identify an effect?

    Fly to its tower in Nousfarer

  9. 1986-Present

    G-Methods for Time-Varying Treatments

    How can effects be estimated with time-varying treatments?

    Fly to its tower in Nousfarer

  10. 1987-Present

    Sensitivity Analysis for Unmeasured Confounding

    How strong must a hidden confounder be to overturn a finding?

    Fly to its tower in Nousfarer

  11. 1990-Present

    Partial Identification

    What range of effects is consistent with the data?

    Fly to its tower in Nousfarer

  12. 1994-Present

    Doubly Robust and Semiparametric Estimation

    How can estimators remain valid when models are wrong?

    Fly to its tower in Nousfarer

  13. 2010-Present

    Causal Machine Learning

    How can flexible algorithms estimate causal effects?

    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.

Rubin causal modelRandomized Experimental DesignInstrumental VariablesRegression Discontinuity DesignBradford Hill criteriaNatural experimentPropensity Score and Matching MethodsCausal graphG-Methods for Time-Varying TreatmentsSensitivity Analysis for Unmeasured C…Partial IdentificationDoubly Robust and Semiparametric Esti…Causal Machine Learning

Beyond the field

Where Causal Inference's ideas came from in other fields, and what they led to there: the strongest 12 of 65 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.

Causal Inference is a planet in Nousfarer

Its city has a tower for each of these 13 frameworks. Land, walk in, and climb one concept at a time.

Land on Causal Inference

Related fields

Fields its ideas are linked to

More in Statistics

See where Causal Inference sits on the Map of Human Knowledge