Understanding Confounding

Slide 1 of 8

What is Confounding?

The Central Challenge of Observational Studies

Confounding is perhaps the most important concept in causal inference from observational data. It occurs when the observed association between treatment and outcome differs from the true causal effect because of the influence of other variables.

Confounding Defined

Confounding occurs when there exists a common cause of both the treatment and the outcome. This common cause creates a non-causal association that mixes with—and potentially overwhelms—the causal effect we wish to estimate.

The Confounding Triangle

The classic graphical representation of confounding shows a variable L that affects both treatment A and outcome Y:

        L (Confounder)
       ↙ ↘
      A   →   Y

In this diagram:

L → A: The confounder affects who receives treatment
L → Y: The confounder affects the outcome
A → Y: The causal effect we want to estimate

The Problem: When we compare outcomes between treated and untreated individuals, we're not just seeing the effect of treatment—we're also seeing the effect of L, which differs between groups.

Why Confounding Matters

Confounding can:

Hide true causal effects (make them appear null)
Exaggerate small effects (make them appear larger)
Reverse the direction of effects (make beneficial treatments appear harmful)

From What If

"Confounding is present when the treated and the untreated are not exchangeable. In other words, the treated would have had a different outcome than the untreated, even if neither had received treatment."

Source: What If (Hernán & Robins), Chapter 7

Use ← → arrow keys to navigate