Confounding Variables

The hospital paradox

Countries with more hospitals have higher death rates. More hospitals → more deaths.

Should we close hospitals?

Obviously not. Sicker populations build more hospitals because they need more care. Poor population health drives both the death rate and the number of hospitals. The variable you weren't looking at — underlying health of the population — explains both outcomes.

This is a confounding variable: a third factor that causes (or is associated with) both the thing you're measuring and the outcome you care about. Ignore it, and the relationship you observe between X and Y is not the relationship that actually exists.

Coffee and cancer

For decades, studies showed that coffee drinkers had higher rates of certain cancers. The implication was troubling: should we stop drinking coffee?

Then researchers noticed something: coffee drinkers of that era also smoked significantly more than non-drinkers. Smoking is a powerful carcinogen. When researchers controlled for smoking — compared coffee drinkers to non-coffee-drinkers who smoked at the same rate — the coffee-cancer association largely disappeared.

Smoking was the confounder. It caused people to both drink more coffee (social correlation at the time) and develop cancer. The coffee-cancer relationship was real in the data and false as a causal claim.

What a confounder looks like

Draw it as a simple graph. A confounder Z has arrows pointing to both X (your predictor) and Y (your outcome):

Z → X
Z → Y

This means Z influences who gets the treatment and what happens to the outcome. When you compare treated and untreated groups, you're not just comparing treatment vs. no treatment — you're also comparing different levels of Z.

Example: Does exercise cause better health outcomes, or do healthier people exercise more? Both. And wealth, time, neighborhood safety, and baseline fitness all push on both exercise and health. Every one of those is a potential confounder.

The deeper problem: you can only control for confounders you know about and have measured. The ones you don't know about — unmeasured confounders — are the most dangerous, because there's no statistical adjustment that can save you.

Why randomized trials solve this

A randomized controlled trial (RCT) is the cleanest solution. Randomly assign people to treatment or control. If the assignment is truly random, every confounder — measured and unmeasured — gets distributed equally between the groups on average.

You don't need to know what the confounders are. Random assignment makes them irrelevant. Any difference in outcomes can then be attributed to the treatment.

This is why the RCT is called the gold standard for causal inference. It's not because it's sophisticated — it's because randomization is the only mechanism that guarantees confounders are balanced without having to identify them.

The catch: RCTs are expensive, sometimes unethical, and often impossible. You can't randomly assign people to smoke for 20 years, or randomly assign countries to adopt different economic policies. Observational data — where you watch what people do naturally — is what most research actually works with. And in observational data, confounding is almost always present.

How to spot confounding in practice

When you see a claimed relationship between X and Y, ask:

1. Is there a plausible third variable that causes both? Think about what determines who gets X and what else that variable predicts about Y.
2. Does the effect change when you control for other variables? If controlling for Z makes the X-Y relationship stronger or weaker, Z was doing something.
3. Does the direction make physical sense? More hospitals causing more deaths should trigger immediate skepticism — the causal story is backwards.

In data science, confounding shows up constantly in A/B tests when randomization is imperfect, in dashboards that compare cohorts formed by behavior (not by random assignment), and in any analysis comparing groups of people who chose different things.

Confounder DAG

PooledBy season

r = 0.86

Pooled r

r = -0.08

Within-season r

More ice cream, more drownings — strong correlation. Sounds causal.

Key takeaways

• A confounder is a third variable that influences both your predictor and your outcome — it creates a fake correlation that looks real and dissolves only when you account for it.
• The most dangerous confounders are the ones you didn't think to measure.
• Randomized trials work because random assignment balances all confounders — measured and unmeasured — without you having to identify them.
• Observational data almost always has confounding. Treat any X → Y claim as provisional until someone asks "what else affects both?"
• Common confounders hide in plain sight: season, age, wealth, baseline health, neighborhood, time of day.