Module 02 of 15

Confounding Variables

How a third factor can create or hide an apparent relationship.

AMathematical Foundation

A confounder Z is a variable that influences both the supposed cause X and the outcome Y. If Z is ignored, the observed association between X and Y mixes the effect of X with the effect of Z. Adjusting for Z (for example, comparing X and Y within groups that share the same value of Z, or including Z in a regression) removes the bias from that measured confounder. The bias of the unadjusted estimate depends on how strongly Z affects X and how strongly Z affects Y.

\hat\beta_{\text{unadjusted}} \approx \beta_{\text{true}} + \underbrace{\gamma\cdot\delta}_{\text{omitted-variable bias}},\qquad \delta=\frac{\operatorname{Cov}(X,Z)}{\operatorname{Var}(X)}

In words: In a linear model fitted by least squares, the slope of Y on X alone is approximately the true effect of X plus gamma, the effect of Z on Y, times delta, the slope you get by regressing Z on X. The approximation becomes exact as the sample grows (and holds exactly in any sample with no other error).

$\beta_{\text{true}}$: causal effect of X on Y
$\gamma$: effect of the confounder Z on the outcome Y, holding X fixed
$\delta$: slope from regressing Z on X: how much Z differs, on average, per unit of X (descriptive, not causal)

Assumptions

The formula assumes Y = α + βX + γZ + error with the error unrelated to X and Z; δ is a descriptive slope (how Z moves with X), not a causal effect of Z on X. The qualitative lesson holds more generally.
Adjustment only removes bias from confounders that are measured, and measured accurately.
Adjusting for a variable caused by X (a mediator) or by both X and Y (a collider) can create bias rather than remove it.

BIntuitive Explanation

Imagine that students who study more score higher. But students who did well last year both study more (they are motivated) and score higher (they already know the material). Compare students with similar prior attainment, and much of the study-hours advantage may shrink. The third variable was doing part of the work you credited to study time.

CConnection to Debate

Confounding is the most common reason a statistic quoted in a round does not support the policy conclusion drawn from it. When the Proposition says that students in virtual schools score lower, the Opposition can ask whether families choosing virtual schooling differ (health, mobility, prior disengagement). When a speaker says homegrown players perform better, ask whether the best clubs both run strong academies and win more. The skill is to name the confounder, explain its path to both X and Y, and say which way it would bias the estimate.

DReal-World Application

Virtual school outcomes
Competing interpretation
Students who enrol in full-time online charter schools are not a random group: switching schools, health needs and earlier struggles are not evenly distributed. CREDO's 2015 study therefore compared each online charter student with a “virtual twin” built from traditional-school students with the same demographics and prior test scores, and still found much weaker growth (about −0.25 standard deviations a year in maths). Matching removes confounding by measured characteristics such as prior achievement, but not by unmeasured ones, such as the reasons a family chose to switch.
Sources
- Online Charter School Study 2015Center for Research on Education Outcomes (CREDO), Stanford University, 2015-10
Consultants and organisational performance
Competing interpretation
Organisations that hire consultants are often in crisis or undergoing change, so their later performance mixes the consultants' contribution with the circumstances that prompted the hiring. Bloom et al. avoided this confounding by randomly assigning free, intensive management consulting among Indian textile plants (28 plants in 17 firms): treated plants raised productivity by about 11% relative to controls. Random assignment answers the causal question for that setting; it does not show that typical paid consulting engagements have similar effects.
Sources
- Does Management Matter? Evidence from IndiaNicholas Bloom, Benn Eifert, Aprajit Mahajan, David McKenzie, John Roberts (NBER Working Paper 16658; Quarterly Journal of Economics 128(1), 2013), 2011-01

These examples describe real policies and research. The numbers in the interactive below and in the practice questions are hypothetical teaching data unless a question says otherwise.

EInteractive Visualization

Set the true effect and the strength of a hidden variable, then reveal the variable and compare unadjusted and adjusted slopes.

Illustrative example — hypothetical data.Simulated data. X = weekly study hours, Y = final score, Z = prior attainment (hidden at first).

Dashed red: unadjusted trend (ignores Z)

True causal effect of one extra study hour0 points

Strength of the confounder0.80

How strongly prior attainment raises both hours and scores.

Unadjusted slope (points per hour)

4.30

Adjusted for Z (within-group slope)

0.02

True effect you set

0.0

Bias from ignoring Z

4.28

Ignoring prior attainment, each extra hour appears to add 4.3 points. Within each attainment group the slope is 0.0, close to the true effect of 0. Students who already did well both study more and score higher, so part of the raw association belongs to Z, not to study hours.

Adjustment only removes confounders you have measured. An unmeasured Z would still bias the estimate.

FDebate Practice

GLimitations

Adjustment is not magic. You can only adjust for variables you have measured; unmeasured confounding remains. Adjusting for the wrong variables (mediators, colliders) can introduce bias. And the choice of which variables to adjust for depends on a causal model, which is itself an assumption that should be stated and defended.