More Confounders

A sharp meta-science essay. The puzzle: many studies show sleeping pills (sedatives/benzodiazepines) raise all-cause mortality 3-5x and cancer rates, even after controlling for confounders (Kripke 2012, Kao 2012, Welch 2017) — seemingly disproving the obvious confounding-by-underlying-illness explanation. But Patorno et al 2017 (4.2M users, 35.6M controls) found minimal/zero mortality difference, because they adjusted for ~300 confounders via Glynn's automated propensity-score algorithm, vs the usual 10-20. The lesson: if Patorno is right, everyone using 'control for the 10-20 confounders you can think of' is doing it wrong — you can't assume you know which confounders matter, and a whole genre of 'we adjusted for confounders' observational studies is suspect. Ties to how genetics mysteries vanished at n=100,000+.

Quality 75 · Excellent. Excellent floor. A genuinely important, clarifying methodological insight that changes how a careful reader reads observational research ('we controlled for confounders' is probably inadequate unless you control for hundreds), well-explained via a concrete, high-stakes case. Held at the Excellent floor as a focused single-insight essay, appropriately hedged ('I don't want to throw out decades of studies').

Claude’s paradigm shift 55 · Moderate. Moderate–Notable (55). The '300 automated confounders flips the answer; standard confounder-control is inadequate' framing was a fresh, important methodological point with broad implications, built on the existing confounding literature.

Real-world impact 2 · Minor. Minor (2). An influential meta-science point in how EBM-minded readers interpret observational studies; no material-world change.