COURSE 16L2100% FREE
Verified 2026-08-14

Confounding examples

**Confounders** are factors that drive both “treatment” and “outcome,” faking causation. UCLA’s coffee story: coffee appeared linked to pancreatic cancer until smoking—the shared cause—was accounted for.

What it is

Confounders are factors that drive both “treatment” and “outcome,” faking causation. UCLA’s coffee story: coffee appeared linked to pancreatic cancer until smoking—the shared cause—was accounted for.

MEDIUM PRIORITYDIAGRAM
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

Three mini DAG cards of real-world-ish confounders (education/income/health toy; ad spend/season/sales). Each with fork shape highlighted.

Educational Focus: Practice recognizing confounders in stories.

Why it matters

AI features correlated with success may just mark already-successful teams. Acting on the correlation wastes money or causes harm. Hernán & Robins organize much of *What If* around making confounding and identification explicit.

How it works (plain)

Ask what else could cause both sides. Draw a simple graph (fork: \(X\leftarrow Z\rightarrow Y\)). Prefer experiments when you can. If not, adjust carefully—not by dumping every variable into a regression (collider bias and mediator bias are real; UCLA “four elemental confounds”).

Everyday example

Shoe size and reading ability both rise with child age—age confounds them.

Try it

Invent a confounder for “teams using AI code assistants ship more.” Say whether you’d experiment, adjust, or admit “unknown.”

Myths

⚠️ Myth: Controlling for everything possible is always good.
✓ Reality: Adjusting for mediators or colliders can bias estimates (UCLA pipe/collider lessons; Cinelli–Forney–Pearl “good and bad controls” pointer on UCLA page).
⚠️ Myth: “We matched on 200 features” proves causality.
✓ Reality: Matching needs a credible adjustment set / design story (*What If*; STATS 361 propensity topics).

Sources