What is Average Treatment Effect?
Average Treatment Effect is the mean causal contrast in a declared target population between each unit's Potential Outcome under treatment and its Potential Outcome under a comparator, commonly written `E[Y(1)-Y(0)]`.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Name the population and effect scale
A risk difference, risk ratio, odds ratio, mean difference, and survival contrast answer different questions even with the same treatment and outcome. Also specify whether the target is the study population, all eligible users, treated users, or a deployment population. Aggregating heterogeneous effects with the wrong population weights silently changes the estimand.
Identification connects ATE to observed data
Naimi and Whitcomb define ATE with Potential Outcomes and describe the common identification conditions: well-defined and consistent treatment, no relevant interference, Exchangeability, and Positivity. Randomization can establish assignment independence by design; observational estimates require a justified adjustment set or another identification strategy.
Estimator choice comes after the estimand
Difference in means, standardization, matching, inverse-probability weighting, and Doubly Robust methods can target ATE under different designs and assumptions. Report finite-sample uncertainty, overlap, attrition, missing outcomes, treatment adherence, and subgroup heterogeneity. A small average can hide large positive and negative subgroup effects, while a precise association can still be causally biased.
Key Characteristics
- Averages `Y(1)-Y(0)` over an explicitly declared target population
- Is an estimand rather than a specific regression or weighting method
- Depends on treatment, comparator, outcome, horizon, and effect scale
- Differs from ATT, ATC, CATE, and individual treatment effects
- Requires identification assumptions before observed data can estimate it
- Can conceal meaningful treatment-effect heterogeneity
Common Use Cases
- Quantifying average lift from a randomized product experiment
- Estimating a policy effect for all eligible users
- Comparing ATE with ATT when adoption is selective
- Standardizing stratum-specific effects to a deployment population
- Defining the target parameter for a Doubly Robust analysis
Example
Loading code...Frequently Asked Questions
What is the difference between ATE and an observed mean difference?
ATE compares outcomes under two interventions for the same target population. An observed mean difference compares people who actually received different treatments. They coincide under a valid randomized design or sufficient identification assumptions, not by definition.
How do ATE, ATT, ATC, and CATE differ?
ATE averages over the target population, ATT over treated units, ATC over untreated units, and CATE within covariate-defined groups. With heterogeneous effects and selective treatment, these quantities can differ materially and require different weights.
Can ATE be estimated from observational data?
Yes, if the intervention is well defined and assumptions such as conditional Exchangeability, Positivity, Consistency, and valid measurement are defensible. Regression, weighting, matching, or Doubly Robust estimators implement those assumptions; they do not prove them.
Does an ATE describe every individual's treatment effect?
No. ATE is a population average. It may hide zero, harmful, and beneficial effects across subgroups, and the individual pair of Potential Outcomes is normally unobserved. Use prespecified CATE analyses and decision-focused validation when heterogeneity matters.
Should ATE always be reported as a difference?
No. Continuous outcomes often use mean differences, while binary or time-to-event outcomes may use risk differences, risk ratios, survival probabilities, or other scales. The scale and horizon must be prespecified because they change interpretation.