Skip to content
Tutorials Center
My learning

Topics

Lesson 13 of 37 ยท Potential Outcomes

Regression for Causal Effects

Use ordinary least squares to estimate causal effects under ignorability: choosing controls, interactions for heterogeneous effects, g-computation and what the treatment coefficient really means.

Linear regression is the workhorse of applied causal inference. Used carefully, the coefficient on the treatment variable is a causal effect. Used carelessly, it is just a partial correlation. This lesson shows what makes the difference.

When the coefficient is causal

In the model

TEXT
Y = a + tau * T + b1 * X1 + ... + bk * Xk + error

tau is the causal effect of T if

  1. the controls X block every backdoor path (conditional ignorability, see the assumptions),
  2. the functional form is roughly right (outcome is linear in X, effect is constant), and
  3. no control is a collider or a consequence of treatment (see mediators and bad controls).

The regression "controls for" X by comparing treated and untreated units with the same values of X.

Python OLS with controls

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.

Leaving out health (a confounder) biases the effect; including both confounders recovers about 1.5. Omitting age, which has a weaker link, changes things less. Which variables matter is a causal question.

Heterogeneous effects with interactions

If the effect depends on a covariate, add an interaction T * X. Centering X makes the main coefficient the effect at the average X, i.e. an estimate of the ATE.

Python Interaction and the average effect

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.

With the interaction, the model tells you how the effect changes with x (slope about 0.8) and gives the ATE at the mean of x.

G-computation (standardization with a model)

A more flexible recipe, g-computation, uses the regression as a prediction machine:

  1. Fit a model of Y on T and X.
  2. Predict every unit's outcome as if treated (T=1) and as if untreated (T=0).
  3. Average the difference over the population (ATE) or over the treated (ATT).
Python G-computation with a fitted model

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 same fitted model gives both ATE and ATT just by averaging over different groups: another view of why estimand and estimator are separate.

Pitfalls of regression adjustment

  • Wrong functional form: if Y is curved in X, a straight line leaves residual confounding. Add polynomials or splines, or use matching.
  • Extrapolation: with poor overlap, regression silently predicts in regions without data.
  • Bad controls: adding mediators or colliders ruins the estimate.
  • Coefficient on a control is not causal: health's coefficient is just a nuisance parameter (the "Table 2 fallacy").
  • Standard errors: use robust (heteroskedasticity-consistent) errors in practice.

Regression with a binary treatment and many covariates weights strata by how much treatment variation they contain, so with heterogeneous effects it does not equal the ATE. G-computation and weighting avoid that.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. The coefficient on T in a regression is a causal effect if:

    Causal meaning comes from the assumptions, not from fit or significance.
  2. 2. In g-computation the ATT is obtained by:

    Predict Y(1) - Y(0) for each unit, then average over the treated.
  3. 3. Centering X before adding the T*X interaction makes the T coefficient:

    With a centered X, X = 0 corresponds to the mean.
  4. 4. Which is a bad control?

    Conditioning on a mediator removes part of the effect you want.

Summary

  • A regression coefficient is causal only under ignorability, correct form and sensible controls.
  • Interactions describe effect heterogeneity; centering gives the average effect.
  • G-computation predicts both worlds and averages, producing ATE or ATT.
  • Watch for misspecification, poor overlap and bad controls.

Next: Randomization Inference and the Bootstrap