AI Glossary / Causal inference and experimentation / Intermediate

Average treatment effect on the treated

Definition

The average treatment effect on the treated, or ATT, is the average causal effect of an intervention among the units that actually received it. It compares their observed treatment outcomes with the outcomes those same treated units would have had under the specified alternative.

What it means

The ATT answers a different population question from the average treatment effect across everyone eligible. It asks what the intervention did for its actual recipients. If people who receive treatment differ from those who do not, those two average effects can differ.
The missing comparison is not simply the observed outcome among untreated people. It is the untreated potential outcome for the treated population. Untreated records may help estimate it, but only with a suitable design and assumptions.
Treatment can mean a business program, product feature or policy. Define receipt carefully, particularly when there are multiple versions, partial participation or changes over time.

How it works

Let A indicate actual treatment and let Y(1),Y(0) denote potential outcomes. On a difference scale, ATT=E[Y(1)−Y(0)|A=1]. The treated group's average Y(1) is observed under consistency, but its average Y(0) is not.
Methods such as matching, weighting or outcome modeling can estimate that missing counterfactual under appropriate assumptions. They require credible adjustment for relevant pre-treatment confounders and sufficient comparable untreated observations for the treated units. The overlap requirement is tailored to this target population. Measured balance does not prove that unmeasured confounding is absent, and excluding treated units with poor matches changes the population being estimated.

Three practical examples

• An employer evaluates the effect of a voluntary mentoring program for employees who actually joined.
• A service team estimates a support intervention's effect among customers who received that intervention.
• A policy analyst studies the effect of a grant among organizations awarded funding, with a carefully justified comparison design.

Think of it this way

Think of asking how the people who took a particular route would have traveled if they had taken the alternative. Other travelers can offer comparison information, but their journeys are not automatically the missing journeys of the original group. The analogy highlights the target population; statistical identification still depends on the route-selection process and available evidence.

Worked example

Suppose an invented population contains four units. The first two receive treatment and have individual potential-outcome effects of 4 and 2. The other two would have effects of −1 and −1. The ATT is (4+2)/2=3, whereas the ATE across all four units is (4+2−1−1)/4=1.
Now suppose the treated group's observed mean outcome is 12 and the untreated group's observed mean is 7. Their raw difference is 5, not necessarily the ATT of 3. If the treated units would have averaged 9 without treatment, their causal effect is 12−9=3.
The unobserved value 9 is supplied only for this hypothetical illustration. In real data it must be identified and estimated using a defensible design. The example shows both why the target population matters and why a raw treated-versus-untreated comparison can be misleading.

Illustrative case study

Fictional case: Fern Services offers optional mentoring. Participants later receive higher assessment scores than nonparticipants, so management initially attributes the entire difference to mentoring.
An analyst defines the question as the effect among participants and examines pre-program experience, earlier scores and eligibility. Matching reveals that some participants have no comparable nonparticipants. The team reports this limitation and distinguishes an estimate for the matched subset from the ATT for all participants. It also discusses possible unmeasured motivation differences. This case illustrates estimand discipline rather than a finding that mentoring was effective or ineffective.

Common misconceptions and mistakes

ATT is not simply the observed treated-group average, and it is not automatically the raw difference between treated and untreated groups. A comparison must estimate what would have happened to the treated units without treatment.
Do not call an estimate the full ATT after discarding treated units unless the remaining target is stated. Matching and propensity scores do not remove every source of bias. Avoid adjusting indiscriminately for variables measured after treatment, because mediators or colliders can alter the question or introduce bias. A result for current recipients may not generalize to future recipients.

Business use

The ATT is useful when reviewing a program already delivered to a particular group: what difference did it make for those recipients? That can inform accountability and program assessment.
It may not answer whether extending the program to everyone would be worthwhile. Expansion could reach people with different expected effects or require a different implementation. Document who received treatment, which comparison observations support the analysis and whether any recipients were excluded. Pair the estimate with uncertainty, sensitivity analysis and a clear account of possible unmeasured confounding. Keep the evaluation population aligned with the business decision being made.

People also ask

Can ATT differ from ATE?

Yes, especially when effects vary and treatment receipt is selective.

Does matching automatically identify ATT?

No. Its causal interpretation depends on assumptions and adequate comparison support.

Is ATT only for observational studies?

No. It is a causal target, although actual receipt and random assignment must be distinguished carefully.

Check your understanding

An analysis estimates an effect only for program participants who could be matched to untreated records. Is that necessarily the ATT for every participant?

Answer: No. If unmatched participants were excluded, the estimate concerns a restricted treated population unless additional assumptions justify extending it to everyone treated.

Related terms

• Average treatment effect
• Propensity scores
• Confounding
• Causal estimand
• potential outcome
• propensity score
• matching
• overlap
• treatment-effect heterogeneity

Learn next

Learn the difference between defining an estimand and choosing an estimator. Compare ATE and ATT using a small potential-outcome table, then examine whether the available untreated data support the counterfactual outcomes of the actual treated population.

More in Causal inference and experimentation

• Automatic evaluation - Automatic evaluation uses software to assess a model, its outputs, or an AI-enabled process according to specified tests or scoring rules.
• Autorater evaluation - Autorater evaluation uses an automated rating system, often a language model, to judge outputs against defined criteria or compare competing responses.
• Average precision at k - Average precision at k, commonly written AP@k, evaluates how early relevant items appear within the first k positions of a ranked result list.
• Balanced accuracy score - Balanced accuracy score measures classification performance by averaging the recall of each class with equal class weight.
• BLEU (Bilingual Evaluation Understudy) - BLEU, short for Bilingual Evaluation Understudy, is an automatic metric originally designed to compare machine translations with human reference translations.
• Box-Benkhen experiment - A Box - Behnken experiment uses a response-surface design to study how several controllable factors affect an outcome and to fit a quadratic model.

Sources

Hernan and Robins, Causal Inference: What If, indexed Fine Point 4.1 and Technical Point 4.1 consulted for the treated-population target and partial exchangeability: https://content.sph.harvard.edu/wwwhsph/sites/1268/2024/04/hernanrobins_WhatIf_26apr24.pdf. Full download was unavailable. Wang, Wu and Chen, Sections 2.1–2.3, also consulted directly: https://nejsds.nestat.org/journal/NEJSDS/article/89/read. Its stronger sufficient assumptions are not confused with the narrower requirements for ATT. Hypothetical arithmetic was independently checked.

Previous: Average treatment effect Next: Averaged one-dependence estimators