Correlation Is Not Causation

“Correlation is not causation” is a cliche precisely because it is so easy to forget. In health policy it is forgotten constantly, because the correlations are real, the stakes are high, and the causal story is appealing. This lesson is about the specific trap behind the cliche: confounding.

What confounding is

A confounder is a third factor that influences both who is in a program and how they turn out, creating a correlation that is not causal. The classic value-based example:

Practices that volunteer for a value-based program are often already better organized, better capitalized, and caring for healthier patients. They would have had lower costs and better scores anyway. When the program’s patients look better than everyone else’s, some or all of that gap was there before the program started.

The program did not necessarily cause the difference. The same trait that made a practice join also made it look good. That is confounding, and it makes an ineffective program look effective.

Why smart people fall for it

• The correlation is real. Program participants genuinely do have lower costs. The data is not wrong; the causal interpretation is.
• The story is plausible. “The program improved care, so costs fell” sounds right, which makes it easy to stop asking questions.
• The alternative is invisible. The pre-existing difference between joiners and non-joiners does not announce itself in the results table.

Worth remembering: whenever a group chose to participate, assume they differed from non-participants until proven otherwise. The burden is on the evaluation to show the groups were comparable, not on you to imagine how they differed.

The direction of the bias matters

Confounding does not always flatter. A program that attracts the sickest, most complex patients (because those are the ones who need it) can look worse than it is, since its population would have had high costs regardless. So confounding can hide a real benefit as easily as it can manufacture a fake one. Either way, the raw comparison is untrustworthy until the groups are made comparable, which is exactly what the next module is about.

Key takeaways

• A confounder is a third factor that drives both participation and outcome, creating a non-causal correlation.
• Voluntary participation is the classic source: joiners differ from non-joiners before the program begins.
• Confounding can make a program look better or worse than it is, so raw group comparisons cannot be trusted on their own.