Skip to content
Tutorials Center
My learning

Topics

Lesson 11 of 37 ยท Potential Outcomes

ATE, ATT, ATU and Conditional Effects

Define and compute the average treatment effect, the effect on the treated and untreated and conditional effects, see why they differ, and learn which one to report for a decision.

With potential outcomes defined, we can name the effects people actually want. They are all averages of the individual effect tau_i = Y_i(1) - Y_i(0) over different groups. Choosing the right group is part of asking the right question.

The main estimands

Name Formula Question it answers
ATE E[Y(1) - Y(0)] What if we treated everyone instead of no one?
ATT E[Y(1) - Y(0) given T=1] How much did the treated gain?
ATU (ATC) E[Y(1) - Y(0) given T=0] What would the untreated gain if treated?
CATE E[Y(1) - Y(0) given X=x] What is the effect for people like x?
LATE effect for "compliers" Effect for those moved by an instrument (IV)

The ATE is a weighted mix of ATT and ATU:

TEXT
ATE = P(T=1) * ATT + P(T=0) * ATU
Python ATE, ATT and ATU from a science table

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

People who benefit most choose treatment (high x), so the ATT is larger than the ATE and the ATU is smaller. A program evaluation that reports the ATT may overpromise for a nationwide rollout, which needs the ATE (or the ATU for the people not yet covered).

Which estimand for which decision?

  • Evaluating a program that already ran: ATT ("did it help those who used it?").
  • Deciding whether to roll out to everyone: ATE.
  • Expanding to people not currently served: ATU.
  • Personalizing who gets treated: CATE, covered in the track overview.
  • Offers people may decline: ITT, and LATE for compliers.

Pick the group first

ATE Everyone
ATT Those treated
ATU Those not treated
CATE People like x

Same Y(1) minus Y(0), different averaging group.

1 / 2

Conditional effects (CATE)

When effects vary with a covariate X, the conditional average treatment effect describes the pattern. Averaging CATEs over the whole population gives the ATE.

Python CATE by group

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The overall effect hides a strong pattern: large for the young, almost none for older people. Reporting only an average can conceal who benefits.

Estimating each effect

Which estimand a method delivers depends on how it is set up:

  • Randomized experiment: the difference in means estimates the ATE (and ATT, since groups are alike).
  • Regression with one treatment coefficient (see regression for causal effects): a weighted average of stratum effects, not necessarily the ATE.
  • Matching: often targets the ATT.
  • Inverse probability weighting: can target ATE or ATT through the weights (IPW).
  • IV: the LATE.

If effects are heterogeneous and treatment is not random, a method may recover an average over a group you did not choose. Always connect the method back to the estimand.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. The ATT is the average effect:

    It conditions on T = 1.
  2. 2. If the people who benefit most are the ones who choose treatment, then:

    The treated group is a selected subset with larger-than-average effects.
  3. 3. Which estimand is the right target for deciding whether to roll a program out to everyone?

    The ATE compares treating everyone with treating no one.
  4. 4. ATE can be written as:

    The population is the mixture of treated and untreated units.

Summary

  • ATE, ATT and ATU average the same individual effect over different groups.
  • With heterogeneous effects and selection into treatment they differ; ATE is the mixture of ATT and ATU.
  • CATE describes how the effect varies with covariates.
  • Pick the estimand that matches the decision, then pick a method that targets it.

Next: Ignorability, SUTVA and Positivity