The Average Treatment Effect in Causal Designs

What the Average Treatment Effect Means

A treatment can help some people, have little effect for others, and harm still others. The average treatment effect (ATE) summarizes those individual differences by averaging them over a specified population. It is therefore not simply the difference between the outcomes observed in the treated and untreated groups. Those groups may differ before treatment, and their observed outcomes do not reveal every individual's response to both treatment options.

Let $A$ denote treatment assignment, with $A=1$ for treatment and $A=0$ for control. Let $Y(1)$ and $Y(0)$ denote the outcomes a unit would have under treatment and control, respectively. These are called potential outcomes. For any one unit, the individual treatment effect is $Y(1)-Y(0)$, but only one of the two potential outcomes is observed.

Definition: For a specified target population $S$, the average treatment effect is the mean individual treatment effect among members of that population: $$ \mathrm{ATE}_S = E\left[Y(1)-Y(0)\mid S=1\right]. $$ Here $S=1$ indicates membership in the target population. If the target is a finite population of $N_S$ units, its ATE is $N_S^{-1}\sum_{i:S_i=1}\{Y_i(1)-Y_i(0)\}$. The ATE is a population quantity, or estimand; an estimate calculated from data is not the estimand itself.

The population must be named to interpret the ATE. For example, an ATE among people eligible for a clinical service is not automatically the ATE among everyone who might someday be eligible. A change in the target population changes which individual effects are averaged, and can change the resulting value even when the treatment and outcome definitions stay fixed.

The outcome scale also matters. If $Y$ is a number of hospital days, the ATE is measured in days. If $Y$ is a binary indicator, the ATE is a difference in probabilities, often expressed in percentage points. In either case, the ATE describes an average contrast, not the treatment effect for every individual.

When Observed Data Identify the ATE

The potential-outcome definition specifies the causal quantity, but it does not by itself say how to calculate that quantity from observed data. Identification requires assumptions that connect potential outcomes to treatment assignment and observed outcomes. Write $X$ for measured pretreatment covariates, and state the assumptions for the target population itself.

Assumptions for covariate-adjusted identification within $S$:

Consistency and a well-defined treatment: each unit's observed outcome equals its potential outcome under the treatment received, $Y=Y(A)$. The treatment conditions being compared must be sufficiently well defined for these potential outcomes to have meaning.

No interference: one unit's outcome under a treatment condition does not depend on other units' assignments. Together with consistency, this rules out ambiguity about the potential outcome represented by $Y(a)$.

Conditional exchangeability within the target: $Y(1),Y(0)\perp A\mid X,S=1$. Among target members with the same measured $X$, treatment assignment carries no additional information about either potential outcome.

Positivity within the target: for covariate values occurring in the target population, $0<P(A=1\mid X=x,S=1)<1$. Both treatment conditions must be possible at those values.

The conditioning on $S=1$ is essential. An exchangeability statement in a broader source population does not, on its own, establish exchangeability among members of a selected target population. Likewise, positivity must apply to the covariate patterns represented in the target, not merely to some broader population.

Under these assumptions, the ATE can be expressed using observed conditional outcome means. The equality follows by first averaging the individual effects over the target's covariate distribution, then using exchangeability to relate each potential-outcome mean to its treatment group's observed outcome mean, and finally using consistency.

\begin{aligned} \mathrm{ATE}_S &= E_{X\mid S=1}\left[E\{Y(1)-Y(0)\mid X,S=1\}\right]\\ &= E_{X\mid S=1}\left[E\{Y(1)\mid X,S=1\}-E\{Y(0)\mid X,S=1\}\right]\\ &= E_{X\mid S=1}\left[E(Y\mid A=1,X,S=1)-E(Y\mid A=0,X,S=1)\right]. \end{aligned}

The final expression says to compare treatment and control outcome means within covariate values, then average those contrasts over the covariate distribution in the target population. It is not enough to average over whichever covariate distribution happens to occur in one observed treatment group. If treatment is randomized within the target, randomization can justify exchangeability there; with unequal assignment probabilities, the design still needs to assign both conditions with positive probability in the relevant target strata.

Calculating and Interpreting the ATE

Worked Example: Averaging Individual Effects in a Finite Target

Suppose a hypothetical target population consists of four patients, and the treatment effect is measured in symptom-score points, with a positive value indicating a higher score under treatment. Their potential outcomes are shown below. In practice, both potential outcomes for the same patient cannot be observed; this table is hypothetical and is used to define the finite-population estimand.

Patient	$Y(1)$	$Y(0)$	$Y(1)-Y(0)$
1	8	6	2
2	7	6	1
3	4	5	-1
4	9	5	4

The target-population ATE is the mean of the four individual differences:

\mathrm{ATE}_S =\frac{2+1+(-1)+4}{4} =\frac{6}{4} =1.5\text{ symptom-score points}.

For this hypothetical finite population, treatment raises the mean symptom score by 1.5 points relative to control. That average does not imply that every patient has a higher score under treatment: patient 3 has an effect of $-1$ point. This calculation defines the estimand directly; it does not show that both potential outcomes can be recovered from ordinary observed data.

Worked Example: Identifying the Target ATE by Covariate Adjustment

Consider a hypothetical observational study in a target population of eligible adults. Let $X=0$ and $X=1$ denote two baseline risk groups. Suppose the target population is 60% in group $X=0$ and 40% in group $X=1$. The observed treatment probabilities are 0.8 and 0.2 in those groups, respectively. The observed mean outcomes are:

Risk group	Target proportion	$P(A=1\mid X,S=1)$	Mean $Y$ if $A=1$	Mean $Y$ if $A=0$
$X=0$	0.60	0.80	8	6
$X=1$	0.40	0.20	5	2

Assume consistency, no interference, exchangeability conditional on $X$ within this target, and positivity. Positivity holds in both groups because 0.8 and 0.2 are each strictly between 0 and 1. Under conditional exchangeability and consistency, the observed within-group differences identify the corresponding conditional average effects: $8-6=2$ outcome units for $X=0$, and $5-2=3$ units for $X=1$. Averaging over the target's stated group proportions gives:

\widehat{\mathrm{ATE}}_S =0.60(8-6)+0.40(5-2) =0.60(2)+0.40(3) =2.4\text{ outcome units}.

The calculation is an estimate from observed data; its causal interpretation depends on the assumptions just stated. To see why the target weights matter, the treated group has covariate masses $0.60(0.80)=0.48$ and $0.40(0.20)=0.08$, which sum to 0.56. Its risk-group proportions are therefore $0.48/0.56$ and $0.08/0.56$, and its mean outcome is approximately $7.571$. The control group has masses $0.60(0.20)=0.12$ and $0.40(0.80)=0.32$, summing to 0.44; its mean outcome is approximately $3.091$. The crude difference is thus about $7.571-3.091=4.481$ units, not 2.4. The discrepancy arises because treatment and control groups have different mixtures of risk groups. The crude contrast does not average the conditional effects over the specified target distribution.

Why Target-Specific Assumptions Matter

Selection into a target population can alter the relationship between treatment and potential outcomes. The following hypothetical example illustrates why it is unsafe to transfer an exchangeability claim from a source population to a selected target without justification.

Worked Example: Source Randomization Does Not Guarantee Target Exchangeability

In a hypothetical source population, let $A$ and $Y(0)$ each be binary, mutually independent variables with probability 0.5 of being 1. Let $Y(1)=Y(0)$, so treatment has no effect for any unit. Treatment is randomized in the source: $A$ is independent of both potential outcomes. Now define a selected target by including a unit with probability 0.9 when $A=Y(0)$, and with probability 0.1 when $A\ne Y(0)$. Selection is therefore related to both treatment and the control potential outcome.

Among selected units, the probability of the cell $A=1,Y(0)=1$ is proportional to $0.25(0.9)=0.225$, while the probability of $A=0,Y(0)=1$ is proportional to $0.25(0.1)=0.025$. Thus, among selected units with $Y(0)=1$, the probability of treatment is $0.225/(0.225+0.025)=0.9$. For $Y(0)=0$, the corresponding treatment probability is 0.1. The treatment probability is positive in both outcome strata, but treatment and $Y(0)$ are associated within the selected target.

Consequently, source-population randomization does not establish exchangeability within this selected target. The example does not mean that every restriction to a target population creates bias; it shows that the target-specific assumption must be established or justified. If target selection is determined before treatment and the randomized assignment is preserved within the selected group, exchangeability can hold there. Otherwise, additional assumptions or a design that supports transport to the target are needed.

In real applications, the potential-outcome independence in an observational study cannot usually be verified from observed data alone. Researchers can inspect whether measured covariates overlap across treatment groups, examine balance after adjustment, check how much the estimate depends on modeling choices, and use subject-matter knowledge to assess whether important pretreatment causes of treatment and outcome are missing. These checks may reveal problems, but they cannot prove the absence of unmeasured confounding.

Interpretation, Alternatives, and Limitations

The ATE is a population-average causal contrast, while a crude difference in observed means is a descriptive contrast between the units who happened to receive each treatment. The two can coincide under suitable designs, such as treatment randomization in the target population, but need not coincide in observational data. Covariate adjustment can identify the ATE when its assumptions hold; adjustment does not guarantee identification merely because many covariates were included.

The ATE also differs from causal effects defined for other populations, such as effects averaged among treated units or in a population with particular treatment overlap. These are distinct estimands, not alternative calculations that must yield the same answer. The appropriate choice depends on the scientific question and the population to which a decision is intended to apply.

Several limitations follow directly from the assumptions. A treatment that has multiple versions may not correspond to a single well-defined $Y(1)$. Interference violates the assumption that one unit's treatment does not affect another unit's outcome. Lack of positivity means some target covariate groups provide no observed comparison between treatment conditions; the covariate-adjusted formula then requires unsupported extrapolation for those groups. Unmeasured confounding can make conditional exchangeability false even when measured covariates appear balanced.

Check the target: state who is included before interpreting the average.
Check the scale: report the ATE in the outcome's units and explain which direction is favorable.
Check support: confirm that both treatment conditions occur across covariate patterns in the target.
Check the design: distinguish what randomization or adjustment supports from what remains an assumption.

Common mistake: “The source study randomized treatment, so the adjusted contrast identifies the ATE in any subgroup or selected target.” This conclusion does not follow automatically. Selection may change treatment–potential-outcome relationships; the relevant exchangeability and positivity conditions must hold within the target, or there must be a justified way to transport the source information. Randomization is powerful, but its protection applies to the population and design for which it was carried out.

Key takeaway: Define the ATE as the average of $Y(1)-Y(0)$ in a named target population. Observed data identify it only when the design and assumptions link both potential-outcome means to observed outcomes in that same target.

Check Your Understanding

Answer each question, then compare with the solutions below.

A target population has individual treatment effects of 3, 0, $-2$, and 7 minutes. What is its ATE, and what does the value mean?
In a target population, two covariate groups have proportions 0.25 and 0.75, with conditional treatment-control outcome differences of 4 and 1 points. Calculate the ATE using the target distribution.
Why is conditional exchangeability in a source population not sufficient by itself to identify an ATE in a selected target?
A covariate pattern occurs in the target, but every unit with that pattern receives treatment. Which identification condition fails, and why does that matter?
Explain why a crude difference between the observed treatment and control means may differ from the target ATE.

Solutions. (1) The mean is $\{3+0+(-2)+7\}/4=8/4=2$ minutes. In this target population, the average outcome under treatment is 2 minutes higher than under control; individual effects need not all equal 2. (2) The target-weighted contrast is $0.25(4)+0.75(1)=1+0.75=1.75$ points. (3) Selection can change the association between treatment and potential outcomes, so exchangeability must hold within the target or be connected to it by justified transport assumptions. (4) Positivity fails: there is no observed control outcome for that covariate pattern, so the conditional treatment-control contrast cannot be learned there from the data without extrapolation. (5) Treatment and control groups can have different covariate compositions; their crude means then combine outcome differences with population-composition differences, whereas the target ATE averages conditional effects over the target's covariate distribution.

Patient	\(Y(1)\)	\(Y(0)\)	\(Y(1)-Y(0)\)
1	8	6	2
2	7	6	1
3	4	5	-1
4	9	5	4