When ATE, ATT, and ATC Coincide

Why Compare These Three Effects?

The average treatment effect (ATE), average treatment effect on the treated (ATT), and average treatment effect on the controls (ATC) describe effects for different target groups. They need not have the same value: if treatment effects vary across units, the average effect among treated units can differ from the average effect among controls, and both can differ from the overall average. A key special case is a constant individual treatment effect, which guarantees that all three estimands agree.

The earlier material on ATE and ATT under treatment-effect heterogeneity emphasizes that group composition can matter. Here the focus is the exact condition that makes the three averages coincide, why that condition is sufficient but not necessary, and how to check the arithmetic without confusing estimands with estimates from observed data.

These are population estimands, not automatically quantities identified by a particular observed-data analysis. The definitions use both potential outcomes for each unit, although only one is observed for that unit. Identification from observed data requires additional design or assumptions; equality among the estimands alone does not supply them.

The Relationship Among ATE, ATT, and ATC

Write \(p=P(D_i=1)\) for the treated proportion in the target population, with \(0<p<1\). Every member of the population is either treated or a control. Partitioning the population average according to these two groups gives the following identity.

This is the population version of a weighted-average identity: the overall mean effect is the treated-group mean effect weighted by the treated proportion, plus the control-group mean effect weighted by the control proportion. The weights are nonnegative and sum to one. Consequently, ATE lies between ATT and ATC, inclusive. It equals both if and only if ATT and ATC are equal, when both groups have positive population proportions.

Why Constant Effects Guarantee Equality

If \(\tau_i=\tau\) for every unit, then averaging \(\tau_i\) over any nonempty subset of the target population produces the same value. In particular, the whole population, the treated group, and the control group each have average effect \(\tau\). This establishes the result directly.

The weighted-average identity gives a second justification. Under constant effects, \(ATT=ATC=\tau\), so \(ATE=p\tau+(1-p)\tau=\tau\). Both arguments rely on the same substantive condition: no unit in the target population has a different treatment effect from any other unit. The potential outcomes themselves need not be constant across units; it is their difference that must be constant.

Heterogeneity Can Separate the Estimands

When effects vary, treated and control groups may have different average effects. If, for example, the treated group contains more units who benefit greatly, ATT may exceed ATC. The overall ATE remains the weighted average of the two subgroup means, so its location depends on both their values and the group proportions.

Heterogeneity by itself does not force the estimands to differ. The relevant comparison is whether the average effects in the treated and control groups differ. Each group can contain units with different effects while the two group averages still match.

Checking the Claim and Avoiding Common Errors

When assessing whether these estimands coincide, separate two questions. First, is the individual treatment effect constant? That is a strong unit-level claim: every \(\tau_i\) must be the same. Second, are the treated-group and control-group mean effects equal? That is an average-level claim. The first implies the second, but the second does not imply the first.

In many applications, both potential outcomes are not observed for any one person, so the individual effect cannot be directly computed from the observed record. Constant effects may be a substantive modeling assumption, but equality of ATE, ATT, and ATC is not established merely by seeing similar observed outcome averages in treatment groups. Such a comparison may reflect baseline differences, confounding, or other features of the data rather than equality of causal effects.

A useful algebraic diagnostic is to check the weighted identity whenever all three estimands are reported. If \(p=0.3\), \(ATT=6\), and \(ATC=2\), for example, the implied ATE is \(0.3(6)+0.7(2)=3.2\), not 4. This check can reveal an arithmetic error, inconsistent target populations, or a mismatch in the stated group proportions. It cannot, by itself, verify that the estimands have been identified correctly.

A common mistake is to reason, “Treatment effects are heterogeneous, so ATE, ATT, and ATC must differ.” The correction is that heterogeneity alone says effects vary across units; it does not say the treated and control groups have different average effects. Conversely, assuming a constant effect can make the three estimands equal, but it does not remove confounding or guarantee that a statistical procedure estimates their common value. Estimand equality is an algebraic and population-composition statement, whereas causal identification concerns whether the target effect can be learned from the available design and data.

The main alternative to assuming constant effects is to allow heterogeneity and define the target group explicitly. Then report the estimand appropriate to that group rather than treating the ATE, ATT, and ATC as interchangeable. These are different target-population averages in general; choosing among them is a scientific decision, not a way to make estimates agree. Equality may be a justified simplification in a particular application, but it should not be inferred from notation or from an unadjusted comparison.