The Average Treatment Effect on the Treated

Why the Treated Group Defines the Question

An observational study often begins with a practical question: among people who received a treatment, how much did it change their outcomes, on average? This question does not ask for the average effect among everyone eligible for treatment. It asks about the people who were actually treated. That distinction determines the target population, and therefore the causal quantity being estimated.

For example, a health system might ask whether a new care program improved outcomes for patients who enrolled. The patients who enrolled may differ from those who did not—in measured characteristics such as baseline health, and possibly in unmeasured ways as well. The observed difference between the two groups does not automatically answer the causal question. First define the target effect; then assess whether the data and assumptions support its estimation.

Definition: Let $T=1$ indicate that a unit receives treatment and $T=0$ indicate that it does not. Let $Y(1)$ and $Y(0)$ denote that unit’s potential outcomes under treatment and no treatment, respectively. The average treatment effect on the treated (ATT) is $$ \operatorname{ATT}=E\{Y(1)-Y(0)\mid T=1\}. $$ It is the mean difference between treated units’ potential outcomes under treatment and under no treatment. Its target population is the population of units with $T=1$, not all eligible units and not the untreated group.

Only one potential outcome is observed for each unit: $Y=Y(1)$ when $T=1$, and $Y=Y(0)$ when $T=0$, assuming the treatment and outcome are consistently defined. Thus, for each treated unit, the outcome under treatment is observed, but the same unit’s outcome under no treatment is not. The ATT averages these individual causal contrasts across treated units, even though those individual contrasts cannot generally be observed directly.

What the ATT Does—and Does Not—Describe

The ATT is anchored to treatment receipt. Its interpretation is: the average change in outcome that the treated units experienced because they received treatment, compared with what those same units would have experienced without it. The final comparison is counterfactual; it is not simply a comparison between treated and untreated people as they happened to be observed.

This target can be important when a decision concerns people already receiving a service, or when treatment is selected by patients or clinicians. For instance, a program that attracts patients with more severe symptoms may have a different ATT from the effect one would expect for the full population. The ATT does not claim to describe people who were not treated. Nor does its definition guarantee that the effect can be estimated credibly from an observational comparison.

The average treatment effect (ATE) instead averages the contrast $Y(1)-Y(0)$ over the full target population. The two estimands differ in whose potential outcomes are averaged: the ATT conditions on $T=1$, while the ATE does not. No numerical equality between them follows just from their definitions. Detailed comparisons of how their values relate under treatment-effect variation belong to a separate analysis; here the key point is that choosing the ATT is a substantive choice of target population.

Worked Example: Defining the ATT in a Finite Treated Group

Consider a hypothetical program with four treated participants. Suppose, for illustration, that both potential outcomes are known for each participant, and that the outcome is a score in points.

Participant	$Y(1)$, treated	$Y(0)$, untreated	Individual effect $Y(1)-Y(0)$
A	15	10	5
B	12	10	2
C	9	10	-1
D	14	10	4

The ATT for these four treated participants is the mean of their four individual effects:

\operatorname{ATT} =\frac{5+2+(-1)+4}{4} =\frac{10}{4} =2.5\text{ points}.

In this hypothetical complete-potential-outcomes example, treatment increased the score by an average of 2.5 points for the treated participants. In a real observational data set, the $Y(0)$ values for treated participants would be missing, so this direct calculation would not be available. The example clarifies the estimand; it does not describe a generally available estimation procedure.

Identifying the ATT From Observational Data

To estimate the ATT from observed data, one needs a way to learn what treated units’ outcomes would have been without treatment. Let $X$ denote pretreatment covariates that may be related to treatment assignment and outcomes. Conditional exchangeability for the untreated potential outcome states that, at each relevant covariate value, treated and untreated units have the same mean $Y(0)$. Consistency connects observed outcomes to potential outcomes, and overlap requires untreated observations at covariate values represented among treated units.

Identification conditions for the ATT:

Consistency and a well-defined treatment: the observed outcome is $Y=Y(T)$; treatment versions are sufficiently well defined for the potential outcomes to be meaningful.
No interference: one unit’s outcome does not depend on other units’ treatment assignments.
Conditional exchangeability for $Y(0)$: $Y(0)$ is independent of $T$ given $X$. In words, after conditioning on $X$, treated and untreated units are comparable with respect to their potential outcomes under no treatment.
Overlap for the treated covariate distribution: for covariate values occurring among treated units, the probability of being untreated is positive, $P(T=0\mid X=x)>0$.

Under these conditions, the ATT is identified by standardizing the conditional outcome difference over the covariate distribution of treated units.

The exchangeability condition is deliberately stated for $Y(0)$. The treated units’ $Y(1)$ outcomes are observed under consistency, so identifying their mean under treatment does not require predicting those outcomes from untreated units. The missing component of the ATT is the mean $Y(0)$ among treated units. Conditional exchangeability allows that component to be learned from untreated units with the same $X$, while overlap ensures those untreated comparisons exist.

\operatorname{ATT} =E\{Y(1)\mid T=1\}-E\{Y(0)\mid T=1\} =E\left[ E(Y\mid T=1,X)-E(Y\mid T=0,X) \mid T=1 \right].

The first conditional outcome mean in the last expression is evaluated among treated units; by consistency it represents their mean treated outcome at a given $X$. The second is evaluated among untreated units; conditional exchangeability makes it a valid mean for treated units’ unobserved $Y(0)$ at that same $X$. Averaging the within-$X$ differences over the covariate distribution of treated units targets the ATT.

Deriving the Standardization Formula

The derivation makes explicit why the treated covariate distribution appears in the formula. Start with the definition and condition the missing untreated potential outcome on $X$. Then use exchangeability to replace its conditional mean among treated units with the corresponding mean among untreated units. Finally, apply consistency to express both conditional potential-outcome means using observed outcomes.

Start from the estimand.
$\operatorname{ATT}=E\{Y(1)\mid T=1\}-E\{Y(0)\mid T=1\}$.

Condition on pretreatment covariates.
By iterated expectation, the second mean is $E[E\{Y(0)\mid T=1,X\}\mid T=1]$.

Use conditional exchangeability.
Replace $E\{Y(0)\mid T=1,X\}$ with $E\{Y(0)\mid T=0,X\}$. This replacement concerns the untreated potential outcome only.

Use consistency and overlap.
Among treated units, $Y=Y(1)$; among untreated units, $Y=Y(0)$. Overlap ensures the untreated conditional mean is available at covariate values represented among treated units. Substitution gives the standardization formula.

The formula is a population identification result, not a guarantee that any particular finite-sample estimator will be accurate. An estimator must estimate the conditional outcome means adequately, represent the treated covariate distribution, and account appropriately for sampling uncertainty. The identifying assumptions themselves cannot be proved from the observed outcomes alone.

Worked Example: Standardizing Within Two Covariate Strata

Suppose a hypothetical observational study compares a care program with no program. The outcome is symptom-free days over a fixed follow-up period. Let baseline severity $X$ have two categories, high and low. The study finds the following conditional mean outcomes:

Baseline severity	Share of treated group	Mean among treated	Mean among untreated	Conditional difference
High	0.60	18 days	14 days	4 days
Low	0.40	12 days	11 days	1 day

Assume, for this example, that treatment and outcome are consistently defined, there is no interference, conditional exchangeability for $Y(0)$ holds within the two severity strata, and both strata include untreated units. The treated-group weights are $0.60$ and $0.40$, so standardization gives:

\widehat{\operatorname{ATT}} =0.60(18-14)+0.40(12-11) =0.60(4)+0.40(1) =2.4+0.4 =2.8\text{ days}.

Under the stated assumptions, the estimated ATT is an average increase of 2.8 symptom-free days for the treated population. The weights come from the treated group, not the untreated group and not a pooled distribution. This matters: treated participants are more often in the high-severity stratum, so the target average should give that stratum more weight.

For comparison, the overall observed treated mean is $0.60(18)+0.40(12)=15.6$ days. If the untreated group is 20% high severity and 80% low severity, its overall observed mean is $0.20(14)+0.80(11)=11.6$ days. The unadjusted difference is therefore $15.6-11.6=4.0$ days, not 2.8 days. The difference arises because the groups have different severity distributions as well as different conditional outcomes. The raw difference is not automatically the ATT.

Checking Whether the Target Is Supported

The overlap condition is specific to the ATT. It does not require untreated observations for every covariate value in the entire population; it requires them where treated units occur. If a group of treated units has no comparable untreated units, the observed data do not reveal that group’s mean untreated potential outcome without additional assumptions or extrapolation. A model may still return a number, but that number is not identified from the observed comparisons alone.

Worked Example: Recognizing a Failure of Overlap

Consider a hypothetical study with 40 treated high-risk patients and 60 untreated low-risk patients. There are no untreated high-risk patients in the data. The treated patients’ observed mean outcome is 20 hospital-free days, and the untreated patients’ mean is 15 days. The outcome units are hospital-free days.

A tempting calculation is $20-15=5$ days. But the 15-day mean describes untreated low-risk patients, not the missing untreated outcomes for treated high-risk patients. Since there are no untreated high-risk observations, the required conditional mean $E(Y\mid T=0,X=\text{high risk})$ is unavailable.

The ATT for the treated group requires the mean outcome those 40 high-risk patients would have had without treatment. The observed data provide no untreated patients in that covariate stratum, so overlap fails for the treated covariate distribution. The ATT is not identified by the observed comparison in this example; 5 days is not a justified ATT estimate. More flexible modeling cannot create the absent comparison without relying on extrapolation assumptions.

Diagnostics, Alternatives, and Limitations

Before interpreting an ATT estimate, inspect whether the data support the comparison it relies on. Examine the distributions of important pretreatment covariates by treatment group, including ranges and sparse regions rather than only overall averages. In a stratified analysis, check that treated strata have untreated observations. In a model-based analysis, note where predictions for treated units depend on covariate combinations with little or no untreated data. Strong imbalance or sparse support is a warning that estimates may depend heavily on modeling choices.

Also assess whether the covariates used for adjustment are measured before treatment and plausibly capture important common causes of treatment and outcome. Conditional exchangeability is an assumption about potential outcomes, not a property verified by a balance table. Unmeasured confounding may remain even when recorded covariates appear balanced. Measurement error, missing covariates, treatment misclassification, and poorly defined treatment versions can also undermine the causal interpretation.

Approach or quantity	Target or interpretation	Key caution
ATT	Mean causal effect among treated units	Requires credible information about treated units’ untreated counterfactual outcomes.
ATE	Mean causal effect in the full target population	Targets a different population; the ATT does not automatically estimate it.
Unadjusted treated–untreated difference	Difference in observed group means	Can combine causal effects with differences in group composition.
Randomized comparison	Can support causal comparisons through randomized assignment	The target and the analysis still need to be specified; observational treatment selection is not randomized.

Regression adjustment and other covariate-adjustment approaches can be used to estimate conditional outcome means and standardize them over treated units. Their credibility depends on the same causal identification conditions and on suitable estimation of those means. The ATT definition does not prescribe one estimator. The choice among estimators is separate from the choice of estimand.

A further limitation is that ATT results apply to the treated population represented by the study and analysis. If treated units are selectively included in the data, or if the treated group changes across settings or time, the target population may not carry over. Restricting analysis to units with adequate overlap can also change which treated units are represented; any resulting estimate must be interpreted for its actual target, not silently generalized to all treated units.

Common Mistake

Wrong reasoning: “The ATT is the mean outcome among treated people minus the mean outcome among untreated people, because that is the observed treatment effect.” The subtraction is an observed group difference, not a mean of treated units’ two potential outcomes. If treatment selection is related to baseline characteristics that also predict outcomes, the groups’ outcome means can differ even without treatment.

Correction: The ATT compares treated units’ observed outcomes with the counterfactual mean those same treated units would have had untreated. Under conditional exchangeability and overlap, estimate the missing counterfactual by comparing within covariate levels, then average those conditional differences using the treated group’s covariate distribution. If that support or exchangeability is not credible, the observational contrast does not by itself establish the ATT.

Key takeaway: The ATT is defined by both an outcome contrast and a target population: it averages $Y(1)-Y(0)$ among units who received treatment. In observational data, identifying it requires a credible untreated counterfactual for those treated units, supported by conditional exchangeability and overlap.

Check Your Understanding

Answer each question, then compare your reasoning with the solution provided.

In one sentence, define the ATT and name its target population.
Solution: The ATT is $E\{Y(1)-Y(0)\mid T=1\}$, the average treatment effect among units who received treatment.
A study has treated-group covariate proportions of 0.25 in stratum A and 0.75 in stratum B. The estimated conditional effects are 6 and 2 outcome units. Calculate the standardized ATT.
Solution: Weight by the treated proportions: $0.25(6)+0.75(2)=1.5+1.5=3.0$ outcome units.
Why can the difference between the overall treated and untreated means fail to equal the ATT?
Solution: The observed groups can have different covariate distributions, and the untreated group’s overall mean need not represent treated units’ missing untreated outcomes. The raw difference can therefore reflect both treatment effects and group-composition differences.
For which covariate values does overlap need to hold to identify the ATT?
Solution: Untreated observations must be possible at covariate values represented among treated units: $P(T=0\mid X=x)>0$ over the treated covariate distribution.
Suppose no untreated units share a covariate stratum with treated units. Can the observed data alone identify the ATT for those treated units? Explain.
Solution: No. Their mean untreated potential outcome is not represented by untreated observations in that stratum. An estimate would require additional assumptions or extrapolation beyond the observed comparison.