MerQuR+Statistics Glossary

Statistics Glossary → Average treatment effect (ATE, ATT, ATC)

Average treatment effect (ATE, ATT, ATC)concept

Three versions of the average causal effect of an intervention, depending on the population it is defined for: the whole sample (ATE), the treated (ATT) and the untreated (ATC).

In causal inference each unit has two potential outcomes: the value it would have with treatment and the value it would have without; the unit-level effect is their difference, but only one of the two is ever observed. The average treatment effect (ATE) is the mean of this difference over the whole population, 'everyone treated versus no one treated'. The average treatment effect on the treated (ATT) defines this difference only for those who actually received the treatment, and the average treatment effect on the controls (ATC) for those who did not.

When the effect varies across individuals and those who take the treatment differ systematically, the three values diverge. Which one to target depends on the research question: 'did participants benefit?' calls for the ATT, 'what if we rolled the programme out to everyone?' calls for the ATE. In propensity-score methods the target estimand determines how matching and weighting are done (for the ATT, for instance, control units are made to resemble the treated group). All these estimates rest on the assumptions of no unmeasured confounding and of common support between the groups.

An everyday example

A company runs a voluntary time-management course, and the employees who sign up tend to be those keen to boost their productivity. The course may be found to raise participants' productivity by 10% on average (ATT). Less motivated staff might benefit less if made to attend, though; if the course were compulsory for all, the average effect (ATE) might be 6%, and for those who currently stay away (ATC) 4%.

In MerQur

In Propensity Score Analysis, the 'Estimand (ATT / ATE / ATC)' option targets the average effect in the treated group, the whole sample or the control group.

Analyses where this concept appears

Propensity Score Analysis

Related concepts

Propensity score matching (nearest neighbour, caliper)Inverse probability weighting (IPW)