What does it mean to 'control' for variables in a scientific study? A statistician explains

Mark Louie Ramos, Assistant Research Professor of Health Policy and Administration at Penn State, writing for The Conversation, 25 September 2026. A short public-facing explainer, prompted by news coverage of GLP-1 drug studies that says a result "controlled for" some list of variables and leaves the reader to decide whether that makes it credible.

T2 — not a textbook or a peer-reviewed paper, but an editorially reviewed piece by a named statistician explaining an established methodological distinction rather than reporting a new result or selling anything. No primary data or citation trail of its own (the GLP-1 studies and the 2021 diabetes trial it references are named only in passing, not linked or dated precisely), which is why it sits at T2 rather than T1 and why nothing here should be cited as evidence about GLP-1 drugs specifically — only about the statistical distinction it explains.

The core distinction

The word "control" means two different things depending on how the data was produced, and the article's whole argument is that conflating them is why the phrase misleads:

• Experimental control. Researchers assign a treatment to some participants and withhold it (often via placebo) from others — the untreated arm is "the control group." Here control means the researcher directly shaped how the data was generated, via random assignment.
• Statistical control. Researchers analyze data they did not generate — health records, survey panels — and use a statistical model to adjust for a variable such as age or sex. Here control means accounting for something the researcher had no influence over.

See confounding for the concept both senses are protecting against.

Confounding, randomization, stratification

The worked example: if smokers are more likely to take a GLP-1 drug, and smoking independently raises bone-injury risk, then an analysis that never adjusts for smoking will make GLP-1 use look riskier than it is. That is confounding — a third variable driving both the exposure and the outcome.

In an experiment, random assignment is what breaks that link, at least in expectation. Stratification — randomizing separately within subgroups (the article's example is patients grouped by baseline blood sugar before randomizing a 2021 GLP-1 diabetes trial) — prevents one arm from being accidentally loaded with people who share a trait, and sharpens the effect estimate. That is the same mechanism larsen-marx-mathematical-statistics teaches as randomized block designs (ch. 13): block on a nuisance variable, then randomize within each block. The article gives that classroom procedure a concrete clinical reason to exist.

Why observational studies happen at all, and what can go wrong

Trial participants are rarely representative of everyone who will eventually take a treatment, which is the article's stated reason researchers turn to observational ("real-world evidence") data despite its weaker footing. Two named failure modes, symmetric and both about how many variables a model controls for:

• Too few controls. Missing a real confounder produces a spurious association — the example given is healthier people self-selecting into taking a supplement, making the supplement look effective when it is the pre-existing health difference doing the work. The article notes such findings often fail to survive a randomized trial. That is the mechanism side of the replication problem herzog-understanding-statistics documents from the literature side (the Test for Excess Success, the Open Science Collaboration's 36% replication rate) — one source shows an association can be manufactured by an absent control, the other shows how often published literatures look exactly like they were.
• Too many controls. Adding unrelated variables makes an estimate less precise — a precision cost, not just a bias risk — and controlling for a collider (a variable that is itself caused by both the exposure and the outcome) can distort the effect being estimated rather than clarify it. Nothing else in this corpus names the collider failure mode; it is new to the spoke.

Where this sits in the corpus

This is the spoke's first source that is explicitly about study design and causal inference rather than about the probability or inference machinery itself — closer to the "research methods" material synthesis has been watching since herzog-understanding-statistics (growth-edge question 5, the mathematics-vs-research-methods fork). It carries no derivation and no theorem; what it adds is vocabulary and a worked failure case (confounding, colliders, stratification-as-blocking) that the corpus's two statistics texts assume the reader already has. Routed here over ../machine-learning-wiki (the runner-up: GLP-1/health-records analysis is exactly the kind of observational-data pipeline that spoke's practitioners run) because the piece teaches the general statistical distinction from first principles, with the drug-trial case as illustration rather than as the subject — the same rule that kept herzog-understanding-statistics and mari-measurement-across-sciences here.

Related

confounding · mathematical-statistics · larsen-marx-mathematical-statistics · herzog-understanding-statistics · illowsky-introductory-statistics · mark-louie-ramos · pennsylvania-state-university · the-conversation · mathematics