AI Glossary / Causal inference and experimentation / Intermediate
Average treatment effect
Definition
The average treatment effect, or ATE, is the average difference between the outcome a defined population would experience under one intervention and the outcome that same population would experience under a specified alternative. It is a causal target, not simply a difference between two observed groups.
What it means
Treatment means an intervention and need not be medical. It could be a training program, a website design or an operating policy. The ATE asks how outcomes would differ, on average, if everyone in the target population followed one strategy rather than the other.
For each person or unit, potential outcomes describe what would happen under each alternative. Usually only one is observed. The missing alternative is why estimating a causal effect requires a research design and assumptions.
An average does not describe every individual. Positive and negative effects can cancel, and an estimate for one population may not apply to another with different characteristics.
How it works
Define the population, intervention, comparison, outcome and follow-up period. Using Y(1) and Y(0) for the two potential outcomes, the ATE on a difference scale is E[Y(1)−Y(0)].
In a well-conducted randomized study, appropriate group comparisons can estimate a corresponding average effect, with attention to adherence, missing data and the exact strategy being evaluated. Observational estimates need assumptions such as adequate control of confounding, overlap and consistency between the defined intervention and observed treatment. Estimation methods do not create those conditions automatically. Report statistical uncertainty and assess whether the sample and design support the intended target population.
Three practical examples
- A company estimates the average effect of offering a training course on a defined skill score among eligible staff.
- A website experiment estimates the average effect of a new page design on completion probability among eligible visits.
- An operations study evaluates the average effect of a scheduling policy on delivery time for a specified set of orders.
Think of it this way
Imagine two parallel versions of the same population, one following each policy, and compare their average outcomes. The ATE describes that contrast. In reality, both versions cannot usually be observed for every unit, so the analogy defines the question rather than supplying the answer. A credible design is needed to learn about the unobserved alternative.
Worked example
For an invented four-unit population, suppose the outcomes under treatment would be 8,6,5 and 7, while outcomes under the comparison would be 5,5,6 and 4. The individual effects are therefore 3,1,−1 and 3. Their average is 6/4=1.5 outcome units.
This ATE is also the difference between the treatment-potential-outcome mean, 6.5, and the comparison-potential-outcome mean, 5. The equality follows from averaging the same population under both alternatives.
The table is hypothetical because an actual study ordinarily observes only one potential outcome for each unit. It also shows heterogeneity: one unit would fare worse even though the population average is positive. An estimated ATE should therefore not be described as the guaranteed benefit for every recipient.
Illustrative case study
Fictional case: Alder Learning wants to know whether offering a short course improves employees' assessment scores. Volunteers score higher than non-volunteers, but they also had more prior experience and motivation.
The evaluation team defines the target as eligible employees and the intervention as an offer of the course, then designs a randomized offer study with a stated follow-up period. It distinguishes the effect of being offered training from the effect of actually completing it. The report includes uncertainty and the population to which the result applies. This case illustrates a clearer causal question; it does not assert that the course improved scores.
Common misconceptions and mistakes
A treated-versus-untreated mean difference is not automatically an ATE estimate with a valid causal interpretation. Selection, confounding and missing outcomes can distort it. The effect of an offer differs from the effect of receipt when people do not follow their assignment.
Do not confuse the ATE with the average treatment effect among those treated. The target populations differ. An average effect is not an individual effect, and an effect on one outcome scale may look different on another. A precise estimate can still be biased.
Business use
The ATE is useful when deciding whether to implement a policy across a defined population. It gives a clear target for experiments and observational evaluations, provided the policy alternatives and outcome are concrete.
Before analysis, document eligibility, treatment versions, follow-up and how outcomes will be measured. Consider spillovers, because one person's treatment may affect another person's outcome. Compare the effect with costs and practical consequences using the same target population. If deployment will be targeted to particular groups, investigate whether a population-wide average answers that decision or whether a different estimand is needed.
People also ask
Is treatment always a medicine?
No. It can be any clearly defined intervention.
Can the ATE be zero when some people benefit?
Yes. Benefits and harms can offset in the average.
Does randomization make the result apply everywhere?
No. It supports comparison within the study design; generalizing to another population needs additional justification.
Check your understanding
A program's ATE is estimated as an increase of two score points. Does that mean every participant gains two points?
Answer: No. It is an average causal contrast for a defined population. Individual effects may differ, and the estimate also has uncertainty and relies on the study's assumptions.
Related terms
- Causal estimand
- Average treatment effect on the treated
- Confounding
- Heterogeneous treatment effect
- Potential outcome
- randomised experiment
- positivity
Learn next
Learn potential outcomes and estimands next. Write a one-sentence causal question naming population, intervention, comparison, outcome and time horizon. Then separate the definition of the effect from the assumptions and data required to estimate it.
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, author page consulted: https://miguelhernan.org/whatifbook. Direct book downloads failed or exceeded retrieval limits; indexed primary-text passages were available. Accessible primary methodological research by Wang, Wu and Chen, Sections 2.1–2.2, consulted for potential-outcome contrasts and identification assumptions: https://nejsds.nestat.org/journal/NEJSDS/article/89/read. The original four-unit average and distinction between an estimand and observed-group contrast were independently checked.
Previous: Average reward Next: Average treatment effect on the treated