Skip to content
Tutorials Center
My learning

Topics

Lesson 10 of 37 ยท Potential Outcomes

The Potential Outcomes Framework

Learn the potential outcomes framework: Y(1), Y(0), individual treatment effects, the fundamental problem of causal inference and the role of selection bias, with simulations.

The potential outcomes framework (also called the Rubin causal model or Neyman-Rubin model) gives causal inference a precise language. Instead of vague talk about "effect", it defines the effect as a comparison between two possible worlds for the same unit.

Two potential outcomes

For each unit i (a person, a store, a patient) and a binary treatment:

  • Y_i(1) is the outcome if unit i receives the treatment.
  • Y_i(0) is the outcome if unit i does not.

The individual treatment effect is

TEXT
tau_i = Y_i(1) - Y_i(0)

The treatment indicator is T_i (1 or 0). The outcome we actually see is

TEXT
Y_i = T_i * Y_i(1) + (1 - T_i) * Y_i(0)

This is the consistency link between observed and potential outcomes: we observe Y(1) for the treated and Y(0) for the untreated.

The fundamental problem

For any unit we see one of the two potential outcomes. The other is missing. So tau_i is never observed: causal inference is a missing data problem.

unit T Y(1) Y(0) Y observed tau
1 1 7 ? 7 ?
2 0 ? 4 4 ?
3 1 9 ? 9 ?
4 0 ? 6 6 ?

Two worlds, one seen

Y(1) Outcome if treated
Y(0) Outcome if not
Effect Y(1) minus Y(0)

We only ever observe one of the two.

1 / 2

A simulation where we can see both

A simulation can store both columns, which real data never can. This is a science table (the full table) next to the observed table.

Python Science table versus observed data

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 Y0 of treated rows and Y1 of untreated rows are exactly what you cannot observe in practice.

The population effect and the naive comparison

Since individual effects are out of reach, we aim at averages. The average treatment effect is

TEXT
ATE = E[Y(1) - Y(0)]

A naive comparison of observed groups gives E[Y | T=1] - E[Y | T=0], which can be decomposed:

TEXT
naive = ATT + selection bias
selection bias = E[Y(0) | T=1] - E[Y(0) | T=0]

The selection bias says: would the treated have had the same untreated outcomes as the untreated? If people who choose treatment are different at baseline, the term is not zero.

Python Naive difference = effect + selection bias

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 naive estimate equals the ATT plus the selection bias term (exactly, up to rounding). Randomization sets the bias to zero because treated and untreated have the same Y(0) distribution.

Why this notation matters

  • It separates the estimand (an average of Y(1) - Y(0)) from the estimator (what you compute).
  • It makes assumptions explicit (see ignorability, SUTVA and positivity).
  • It handles effect heterogeneity: tau differs between units.
  • It extends to many treatments, continuous doses and time.

Potential outcomes and the graphs of causal DAGs are two languages for the same ideas. Economists and statisticians often prefer the first, computer scientists and epidemiologists often the second. Knowing both is an advantage.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. The fundamental problem of causal inference is that:

    The other potential outcome is a counterfactual and never observed.
  2. 2. The observed outcome is Y = T*Y(1) + (1-T)*Y(0). This statement is known as:

    It links the observed outcome to the potential outcome of the treatment received.
  3. 3. The naive difference in means equals:

    naive = ATT + E[Y(0)|T=1] - E[Y(0)|T=0].
  4. 4. Randomization makes the selection bias term:

    Treated and untreated then have the same distribution of Y(0).

Summary

  • Each unit has two potential outcomes, Y(1) and Y(0); the effect is their difference.
  • We observe only one, so causal inference is a missing data problem.
  • A naive group comparison equals the ATT plus selection bias.
  • Randomization removes the selection bias; other designs need assumptions.

Next: ATE, ATT and ATU