The average treatment effect (ATE) is the average causal difference between the outcomes that would occur under two specified interventions in a defined population. It is a central quantity in causal inference, used to evaluate policies, programs, experimental interventions, and other changes. Unlike a difference between observed group averages, the ATE compares alternative outcomes for the same target population. Establishing that an observed comparison identifies this causal quantity requires assumptions about treatment assignment, outcome measurement, and the intervention itself.
Definition and potential outcomes
In the potential outcomes framework, let indicate a binary treatment. For each unit, define:
- : the outcome that would occur under treatment;
- : the outcome that would occur under the comparison condition.
“Treatment” is a general term: it may mean participation in a training program, exposure to a policy, or assignment to one version of a product. The comparison condition need not mean doing nothing; it may be another active intervention.
The individual treatment effect is
For a population represented by a probability distribution, the ATE is the expected value
assuming the expectations exist. For a fixed population of units, the finite-population ATE is
These definitions distinguish the estimand, the quantity of interest, from an estimator, a rule for calculating an estimate from observed data.
The ATE has the same units as the outcome. If an outcome is measured in annual earnings, the ATE is an average earnings difference. For a binary outcome, it is a difference in probabilities, commonly expressed in percentage points—not a risk ratio or an odds ratio.
The missing counterfactual
Ordinarily, a unit is observed under only one treatment condition. Under consistency, its observed outcome satisfies
Thus, one of its two potential outcomes is unobserved. This is the fundamental problem of causal inference: the individual causal effect cannot generally be calculated directly.
Nevertheless, an average effect can be identified without observing both potential outcomes for every unit. The essential task is to justify comparisons that recover the two population means.
Causal effects versus observed differences
The observed difference
is an association. It is not automatically the ATE, because the treated and untreated groups may differ before treatment.
Under consistency, this difference can be decomposed as
The first term is the average effect among treated units. The second describes how their untreated potential outcomes differ from those of untreated units. For example, participants in a voluntary training program may have different prior skills or motivation from nonparticipants.
Such differences can produce confounding. Increasing the sample size reduces sampling uncertainty but does not, by itself, eliminate this source of bias.
Identification assumptions
Identification concerns whether the ATE is uniquely determined by the observed-data distribution together with stated assumptions. It is distinct from estimation accuracy in a finite sample.
A common identification strategy uses pretreatment covariates and the following conditions.
Consistency and well-defined interventions
If a unit actually receives treatment , its observed outcome equals .
The intervention must be specified sufficiently clearly. If “receiving training” combines substantially different programs, durations, or delivery methods, the meaning of a single may require further specification.
The usual unit-level formulation also assumes no interference: one unit’s outcome does not depend on other units’ treatment assignments. When spillovers occur, potential outcomes must instead account for the relevant assignment pattern, and the causal contrast must be defined accordingly.
Conditional exchangeability
A standard sufficient condition is
where the notation expresses conditional independence.
This means that, within covariate strata, treatment assignment carries no additional information about the potential outcomes. It is often called no unmeasured confounding. Conditional mean exchangeability is sufficient for identifying the ATE, so full conditional independence is stronger than strictly necessary.
In observational studies, exchangeability is a substantive assumption and cannot generally be verified from observed data alone. Covariates must be chosen using knowledge of the causal structure; adjusting for every available variable is not necessarily valid.
Positivity
For almost every covariate value in the target population,
Both treatment conditions must be possible within the relevant covariate strata. If some subgroup always receives treatment, the data provide no untreated comparison for that subgroup without additional extrapolation assumptions.
The adjustment formula
Under these conditions, defining
gives
This is the adjustment formula, also called a form of the g-formula. It averages differences in conditional expectations over the covariate distribution of the target population.
Estimation in randomized experiments
In a randomized controlled trial, randomization makes treatment assignment independent of potential outcomes according to the experimental design. Under complete randomization with fixed treatment-group sizes, the difference in sample means,
is unbiased for the finite-sample ATE, provided outcomes are observed and the treatment contrast is appropriately defined.
For a completely randomized experiment, its randomization variance is
where , , and are finite-population variances of the two potential outcomes and individual effects, calculated with denominator . The last quantity is generally not identifiable because both potential outcomes are not observed for any unit. Omitting that nonnegative term yields the usual conservative variance-estimation approach.
The analysis must reflect the assignment design. Cluster randomization, stratification, and unequal assignment probabilities can require different estimators or standard errors.
Random assignment also does not guarantee that all participants comply with their assigned treatment. Comparing groups by assignment estimates an intention-to-treat effect—the effect of assignment—not necessarily the ATE of treatment actually received.
Estimation in observational studies
When treatment is not randomized, several methods implement the identification assumptions.
Outcome regression and standardization
Estimate and , then average the fitted contrasts:
The outcome models may use linear regression, other parametric models, or flexible statistical methods. Reliability depends on model adequacy and support in the data. A single treatment coefficient from an arbitrary regression is not automatically the population ATE, especially with nonlinear models or heterogeneous effects.
Inverse probability weighting
The propensity score is
Under the identification conditions,
Inverse probability weighting estimates this expression using estimated treatment probabilities. It reweights observed groups so that their pretreatment covariate distributions represent a common target population.
Propensity scores near zero or one can generate large weights and unstable estimates. Trimming observations or changing the weighting scheme may improve precision, but can also change the target population or estimand.
Doubly robust estimation
An augmented inverse probability weighted estimator combines outcome models and a propensity-score model:
Under appropriate regularity conditions, it is consistent if either the propensity-score model or both outcome models are correctly specified. This property is known as double robustness.
Double robustness does not protect against unmeasured confounding, poorly defined interventions, or failed positivity. When flexible machine-learning methods estimate the component models, sample splitting or cross-fitting can help control overfitting-related bias, but valid inference still requires additional conditions.
Related causal quantities
Different averages answer different questions.
The average treatment effect on the treated (ATT) is
The average treatment effect on the untreated (ATU) is
If , with both groups present, then
The conditional average treatment effect is
It describes effect variation across measured characteristics, and the population ATE is its average over the target distribution of .
Other research designs can identify different causal quantities. Under standard assumptions, instrumental-variable methods may identify a local average treatment effect for units whose treatment changes with the instrument. A regression discontinuity design usually identifies an effect at an assignment threshold. Common difference-in-differences designs identify effects among treated units under parallel-trends and related assumptions. These quantities are not generally interchangeable with the population ATE.
Interpretation, uncertainty, and limitations
The target population matters. The ATE depends on whose effects are averaged. An experiment can have strong internal validity without representing a broader population. Random treatment assignment and random sampling address different problems; generalization may require additional assumptions about participation and effect differences.
Averages can conceal heterogeneity. A zero ATE may coexist with positive effects for some units and negative effects for others. Even when both marginal potential-outcome distributions are identified, the distribution of individual effects usually is not, because the joint relationship between and remains unknown.
Timing and scale are part of the estimand. Short-term and long-term effects are different quantities. An ATE on a logarithmic outcome is not generally the logarithm of the ATE on the original scale.
Statistical precision is not causal validity. A narrow confidence interval can accompany an estimate biased by confounding, attrition, or model misspecification. Sensitivity analysis can describe how conclusions depend on departures from assumptions, but cannot establish those assumptions as true.
Post-treatment adjustment can change the question. Controlling for variables affected by treatment may remove part of the total effect or introduce bias. Estimating direct or mediated effects requires separate causal definitions and assumptions.
Historical development
The potential-outcomes approach has roots in Jerzy Neyman’s work on randomized experiments in the 1920s and was developed extensively by Donald Rubin for experimental and observational studies in the 1970s. Ronald Fisher made foundational contributions to randomization and experimental design.
Subsequent work in econometrics, statistics, and related fields clarified how treatment-effect estimands differ across populations and assignment mechanisms. These developments made explicit a distinction central to modern analysis: defining the causal effect of interest, establishing conditions that identify it, and selecting an estimator are separate tasks.