Skip to content
Tutorials Center
My learning

Topics

Lesson 12 of 37 ยท Potential Outcomes

Ignorability, SUTVA and Positivity

Learn the key identification assumptions of causal inference: conditional ignorability (exchangeability), positivity (overlap), consistency and SUTVA, with Python checks and violations.

Data alone never prove a causal effect. To turn observed differences into causal effects you need assumptions that link the observed world to the potential outcomes. Four of them appear in almost every observational analysis. Knowing them lets you judge, and honestly state, what a result relies on.

1. Conditional ignorability (no unmeasured confounding)

Also called exchangeability or unconfoundedness. Given measured covariates X, treatment assignment is as good as random:

TEXT
(Y(1), Y(0))  is independent of  T   given X

In words: among people with the same X, who got treated does not depend on how they would have responded. Every common cause of treatment and outcome is in X.

This is untestable: it concerns unobserved potential outcomes. You defend it with subject knowledge, graphs and sensitivity analysis.

Python Ignorability holds only after conditioning

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.

Overall, the treated have a lower baseline Y(0) (more high-risk people), violating exchangeability. Inside each x, treated and untreated have the same Y(0), so ignorability holds given x.

2. Positivity (overlap)

Every kind of person must have a chance of both treatments:

TEXT
0 < P(T=1 | X=x) < 1   for all x in the population

If some covariate values always get treated (or never), there is nobody to compare and the effect for them is unidentifiable without extrapolating by a model. Check it by inspecting the treatment rate within groups, and later with overlap diagnostics.

Python Detecting a positivity violation

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.

Treatment rates of 0 and 1 in two age bands mean the effect for those bands cannot be estimated from comparisons: there is no overlap. Honest options: restrict the question to the overlapping region, or rely on a model assumption you state explicitly.

3. Consistency

The potential outcome under the treatment you received equals the outcome you observe: Y = Y(T). It also requires the treatment to be well defined. "Exercise" is vague (running? yoga? how long?); "30 minutes of jogging three times a week" is a well-defined treatment. Vague treatments make Y(1) ambiguous.

4. SUTVA: no interference

The stable unit treatment value assumption says (a) one unit's treatment does not change another's outcome, and (b) there are no hidden versions of the treatment. Violations are common in social settings:

  • Vaccination: your neighbors' vaccination protects you.
  • Marketplace experiments: discounting some buyers changes prices for others.
  • Social networks: a feature used by a friend changes your behavior.
Python Interference breaks the simple comparison

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 direct effect is 4, yet the simple comparison gives about 4.3. Treated people live in households with slightly more treated members, so their spillover benefit is mixed into the difference, while the untreated also benefit from treated housemates. Under interference the comparison no longer isolates "treated versus untreated" (here it is biased in whichever direction the spillover pushes). Designs that randomize groups (clusters) can estimate total effects.

Summary of assumptions

flowchart TD
  A["Observed data"] --> B{"Assumptions"}
  B --> C["Ignorability given X"]
  B --> D["Positivity"]
  B --> E["Consistency"]
  B --> F["SUTVA"]
  C --> G["Identified causal effect"]
  D --> G
  E --> G
  F --> G

Remove any one and the effect is no longer identified from the data alone.

When you report a causal result, write one sentence per assumption saying why you find it plausible, and what you did to test what can be tested (overlap, balance, placebo outcomes).

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Conditional ignorability means:

    Among people with the same X, treatment assignment is as good as random.
  2. 2. Which assumption can you partly check by tabulating treatment rates across covariate groups?

    Treatment rates of 0 or 1 in a group show a lack of overlap.
  3. 3. A vaccine protects unvaccinated neighbors. Which assumption fails?

    One person's treatment affects another's outcome.
  4. 4. Why is 'exercise' a poor treatment definition for causal analysis?

    Different versions of exercise may have different effects.

Summary

  • Identification needs assumptions: ignorability given X, positivity, consistency and SUTVA.
  • Ignorability is untestable; defend it with subject knowledge and sensitivity analysis.
  • Positivity can be checked by looking at treatment rates within groups.
  • Interference and vague treatments make "the effect" ill-defined.

Next: Regression for Causal Effects