Loading field map
Causal inference is the branch of statistics concerned with answering questions about what would happen under different interventions, treatments, or exposures—not merely what is associated with what. Its central problem is that the data we can observe are almost never generated by the controlled experiments we would ideally run. A physician wants to know whether a drug reduces mortality, an economist whether a job-training program raises wages, an epidemiologist whether air pollution causes asthma. Each faces the same fundamental difficulty: the people who received the treatment are not directly comparable to those who did not, because the decision to treat was itself influenced by factors that also affect the outcome.
The field’s defining intellectual move is to make this difficulty precise. Causal inference treats the question "What is the effect of X on Y?" as a well-defined counterfactual question: "What would have happened to the same individuals if, contrary to fact, they had received a different treatment?" Since we can never observe both the actual and the counterfactual outcome for the same unit, the task becomes one of using assumptions, study design, and statistical methods to estimate the unobservable from the observable.
The dominant formal language for causal questions is the potential outcomes framework, sometimes called the Rubin causal model after its most prominent developer, Donald Rubin, though its roots trace to earlier work by Jerzy Neyman in the 1920s on randomized experiments. For each unit in a study, we imagine two (or more) potential outcomes: the outcome that would occur under treatment, and the outcome that would occur under control. The causal effect for a unit is the difference between these two potential outcomes. The fundamental problem of causal inference is that we observe only one of them for any given unit.
This framework makes the logic of randomized experiments transparent. In a properly randomized experiment, treatment assignment is independent of all potential outcomes, so the average outcome among treated units is an unbiased estimate of the average potential outcome under treatment for everyone, and similarly for controls. The difference in observed averages therefore estimates the average causal effect. Randomization works not because it balances every confounder perfectly in any finite sample, but because it makes the assignment mechanism ignorable: the reason a unit received treatment carries no information about its potential outcomes.
Most of causal inference concerns what can be done when randomization is impossible or unethical. The key concept is the confounder: a variable that influences both treatment assignment and outcome. If we can measure all confounders, then within groups defined by those confounders, treatment assignment is as good as random—a condition formalized as unconfoundedness or ignorability. Under this assumption, the average causal effect can be estimated by comparing treated and untreated units who share the same values of the confounders. The practical difficulty is that confounders are often numerous, and the data may be sparse: with many confounders, there may be few or no treated and untreated units with exactly the same values.
Several estimation strategies have been developed to handle high-dimensional confounding, each with a distinct logic and set of trade-offs.
Outcome regression is the most direct approach: model the outcome as a function of treatment and confounders, then predict the outcome under treatment and under control for every unit and average the difference. Its strength is simplicity and efficiency when the model is correct. Its weakness is that the model can be misspecified, and the results can be sensitive to the functional form chosen—especially when treated and untreated units differ substantially in their confounder distributions.
Propensity score methods take the opposite route. The propensity score is the probability of receiving treatment given the confounders. A theorem by Rosenbaum and Rubin shows that if treatment is unconfounded given the confounders, it is also unconfounded given the propensity score alone. This reduces the dimensionality of the adjustment problem: instead of matching on many variables, one can match, stratify, or weight on a single score. The most common implementation is inverse probability weighting, in which each unit is weighted by the inverse of its probability of receiving the treatment it actually received, creating a pseudo-population in which treatment is independent of confounders. Propensity score methods are intuitive and widely used, but they require the propensity score model itself to be correct, and they can be unstable when some units have very low probabilities of receiving the treatment they did receive.
Doubly robust methods combine outcome regression and propensity score weighting in a way that yields consistent estimates if either model is correct. The most common form is the augmented inverse probability weighting estimator, which adds a regression-based correction term to the weighted estimator. These methods are attractive because they offer two chances to be right, but they inherit the risk of both models being wrong, and their finite-sample performance can be poor when weights are extreme.
Matching methods pair each treated unit with one or more control units that are similar in terms of confounders, then compare outcomes within the matched pairs. Matching is conceptually transparent and does not require specifying a model for the outcome. Its limitations include the difficulty of matching well when confounders are high-dimensional, and the fact that unmatched units are discarded, which can reduce precision and change the estimand to the effect on the treated rather than the whole population.
A recurring theme across these approaches is the bias-variance trade-off and the danger of model dependence. A method that relies on a single parametric model can give badly wrong answers if that model is wrong. This has motivated a shift toward nonparametric and machine learning methods that estimate the regression or propensity score flexibly, and toward estimators with double robustness properties that remain valid under weaker conditions.
A second major tradition, associated with Judea Pearl and the structural equation modeling literature, approaches causality through directed acyclic graphs (DAGs). In this framework, each variable is a node, and a directed edge from one variable to another represents a direct causal influence. The graph encodes a set of conditional independence relationships: a variable is independent of its non-descendants given its parents. This provides a language for reasoning about which variables must be measured to identify a causal effect, and which must not be.
The central graphical concept is the back-door criterion. A set of variables satisfies the back-door criterion if it blocks every path between treatment and outcome that contains an arrow into treatment. Conditioning on such a set identifies the causal effect. The graphical approach also reveals a subtlety that the potential outcomes framework handles less transparently: the danger of conditioning on colliders—variables that are caused by both treatment and outcome. Conditioning on a collider can induce a spurious association between treatment and outcome, even when no confounding exists. This insight has practical importance: adjusting for variables that are affected by treatment, or that are common effects of treatment and outcome, can bias estimates rather than remove bias.
The structural approach also gives a precise meaning to direct versus total effects, and to more complex quantities such as the natural direct effect and the controlled direct effect, which decompose a total effect into the part mediated by intermediate variables and the part that operates through other pathways. Mediation analysis, which asks how a treatment works, has been a major growth area, though it requires stronger assumptions than estimating total effects.
The two frameworks—potential outcomes and graphical models—are not rivals in the sense of making incompatible claims. They are complementary languages. The potential outcomes framework is better suited to defining estimands precisely and to guiding estimation; the graphical framework is better suited to clarifying the assumptions needed for identification and to exposing the consequences of conditioning on the wrong variables. Modern practice increasingly uses both: a graph to decide what to adjust for, and a potential-outcomes-based estimator to do the adjusting.
When unmeasured confounding is present, no amount of adjustment can recover the causal effect from observational data alone. The instrumental variable (IV) approach provides an alternative identification strategy. An instrument is a variable Z that satisfies three conditions: it affects the treatment; it has no direct effect on the outcome except through the treatment; and it is independent of unmeasured confounders. Under these assumptions, the IV estimator—the ratio of the association between Z and Y to the association between Z and treatment—identifies the causal effect of treatment on outcome.
The canonical example is from economics: whether an individual serves in the military affects later earnings, but military service is chosen, so it is confounded with unobserved ability. The Vietnam War draft lottery assigned draft numbers essentially at random, and draft number affected the probability of military service. Since draft number plausibly affects earnings only through military service, it serves as an instrument. The IV estimate then captures the effect of military service on earnings for the subset of people whose service was actually changed by the draft—the compliers.
This last point is crucial. IV methods do not generally estimate the average treatment effect for the whole population. They estimate the local average treatment effect (LATE) for the subgroup of units whose treatment status is changed by the instrument. This is a limitation when the instrument affects only a small or atypical subgroup, but it is also a strength: the LATE is often the policy-relevant quantity, because it describes the effect on the people who would be induced to change behavior by a policy that shifts the instrument.
The IV framework has been enormously influential in economics, where natural experiments—policy changes, weather shocks, institutional rules—are used as instruments. The credibility of an IV analysis rests entirely on the validity of the exclusion restriction (the instrument affects the outcome only through treatment) and the independence assumption, both of which are often arguable. The field has therefore developed a strong culture of robustness checks and sensitivity analysis around IV estimates.
Several developments have reshaped the field in recent decades. One is the growing recognition that average effects can hide important heterogeneity. The same treatment may help some people and harm others, and the average may be near zero. This has motivated methods for estimating conditional average treatment effects—the effect as a function of observed covariates—using machine learning techniques such as causal forests and targeted maximum likelihood estimation. These methods aim to identify subgroups with particularly large or small effects, though they require careful validation to avoid overfitting.
A second development is the increased emphasis on sensitivity analysis. Since unmeasured confounding can never be ruled out in observational studies, researchers increasingly report how large an unmeasured confounder would have to be to overturn their conclusions. The E-value and related measures quantify the minimum strength of association that an unmeasured confounder would need to have with both treatment and outcome to explain away the estimated effect. These tools do not prove that confounding is absent, but they make the assumptions behind a causal claim more transparent.
A third development is the revival of design-based thinking. Rather than relying solely on statistical adjustment, many researchers now prefer to exploit natural experiments, regression discontinuity designs, and other quasi-experimental settings where the assignment mechanism is known or nearly known. In a regression discontinuity design, for example, treatment is assigned by a cutoff on a continuous variable—such as a test score or an age threshold—and units just above and below the cutoff are compared. Because assignment near the cutoff is nearly random, this design can yield credible causal estimates without strong modeling assumptions. The emphasis on design reflects a broader epistemic stance: the most reliable causal knowledge comes not from clever statistical corrections applied after the fact, but from settings where the data-generating process itself provides a natural experiment.
It is important to be clear about what causal inference can and cannot deliver. All observational methods rely on assumptions that are untestable from the data alone. Unconfoundedness cannot be verified; the exclusion restriction for an instrument cannot be proven; the functional forms in a regression can be wrong in ways that no diagnostic can fully detect. The field’s contribution is not to eliminate these problems but to make them explicit, to provide methods that are valid under clearly stated assumptions, and to develop tools for assessing how sensitive conclusions are to violations of those assumptions.
A further limitation is that causal effects are always defined relative to a specific intervention and population. The effect of a drug in a clinical trial may not generalize to a different population, a different dosage, or a different setting. The effect of a policy in one country may not hold in another. Causal inference provides estimates of effects under the conditions studied; extrapolation beyond those conditions is a judgment that no statistical method can fully justify.
Finally, causal inference is not a substitute for scientific understanding. A statistically identified causal effect is a relationship that holds under the assumptions of the analysis; it does not by itself explain the mechanism, nor does it guarantee that the relationship will persist under changed circumstances. The best causal analyses combine statistical rigor with substantive knowledge about the process under study. The field’s enduring contribution is to have made the logic of causal questions precise enough that these judgments can be made transparently, and its methods continue to evolve as new data sources and new computational tools expand the range of questions that can be addressed.