When you’re studying causal inference, experimental design, or hypothesis testing, one phrase shows up everywhere: “Randomization eliminates confounding.” On exams, you’re expected to explain what that means and how to justify it without hand-waving—especially when the confounders could be unobserved.
This page gives a clean what + how answer you can use in problem sets, midterms, and research writeups.
Let \(T\in\{0,1\}\) be treatment assignment, \(Y(1)\) and \(Y(0)\) be potential outcomes, and \(X\) be baseline covariates (observed confounders). If assignment is randomized, then:
\[ T \perp (Y(1), Y(0), X). \]That independence implies the treated and control groups have the same distribution of covariates and potential outcomes (in expectation), so differences in outcomes can be attributed to treatment rather than pre-existing differences.
Why confounding disappears under random assignment
Confounding happens when treatment assignment is related to factors that also affect the outcome. In notation, a typical confounding structure is:
\[ X \rightarrow T \quad \text{and} \quad X \rightarrow Y, \]so \(T\) and \(Y\) are associated even when the treatment has no causal effect. Randomization breaks the arrow \(X \rightarrow T\) by design—assignment is driven by a random mechanism, not by \(X\).
How to show randomization eliminates confounding on an exam
- State what is randomized: random assignment \(T\), not random sampling of participants.
- Write the independence statement: \(T \perp (Y(1),Y(0),X)\) (or \(T \perp X\) as a simpler version).
- Convert independence into equal expectations: \[ \mathbb{E}[Y(0)\mid T=1]=\mathbb{E}[Y(0)\mid T=0], \quad \mathbb{E}[Y(1)\mid T=1]=\mathbb{E}[Y(1)\mid T=0]. \]
- Link to unbiased causal estimation: the difference in sample means estimates the ATE: \[ \mathbb{E}[Y\mid T=1]-\mathbb{E}[Y\mid T=0] = \mathbb{E}[Y(1)-Y(0)]. \]
- Add the practical caveat: balance holds “on average”; in finite samples you can still see imbalance by chance.
Numerical example: selection vs randomization
Suppose baseline ability \(X\) affects exam score \(Y\). If motivated students self-select into tutoring (\(T=1\)), then \(T\) is correlated with \(X\), and a naive comparison overstates the effect of tutoring.
- Selection: \( \mathbb{E}[X\mid T=1] > \mathbb{E}[X\mid T=0] \Rightarrow \) confounding.
- Random assignment: \( \mathbb{E}[X\mid T=1] = \mathbb{E}[X\mid T=0] \Rightarrow \) no confounding (in expectation).
That’s the core idea: randomization makes the “treated vs control” comparison behave like a fair coin flip rather than a choice driven by \(X\).
Common mistakes
- Mixing up random sampling vs random assignment: random sampling helps external validity; random assignment identifies causal effects.
- Claiming perfect balance: randomization balances in expectation; finite samples can look imbalanced.
- Forgetting attrition/noncompliance: if people drop out or don’t follow assignment, independence can break.
- Saying “we controlled for everything”: the key benefit is that it addresses unobserved confounders too.
Why this matters in graduate coursework
This concept is a backbone of experimental design, causal inference, and regression interpretation. It shows up in: A/B testing logic, difference-in-means estimators, IV intuition (as “as-if random”), and why observational studies require adjustment. If you can write the independence statement and translate it into unbiased estimation, you’re usually in excellent shape for exams.
This idea connects directly to:
- Statistics Tutoring
- Statistical Inference
- Biostatistics & Experimental Design
- Statistics Blog
- Question Hub
Speak Directly to a Tutor — Send Your Message Below
No call centers. No delays. Your message goes straight to the tutor.
- Call/Text:
- Email:
- WhatsApp: