Ecological Bias in Federated COVID-19 Analysis
Ecological bias distorts effect sizes in federated COVID-19 analysis: evidence from 42 million patient records
Multi-site studies and federated networks pool site-level summaries to draw conclusions about individual patients. In 38 million CDC records grouped into seven racial and ethnic pseudo-sites, ecological regression detects an age–mortality relationship (β = +0.55, p = 0.02), but patients over 80 have 9.9 times the odds of dying at the individual level; in 4 million confirmed cases across Mexico’s 32 states the ecological coefficient is again a different quantity from the individual-level odds ratio of 11.3. A power simulation shows that grouping, not sample size, determines ecological power: grouping by race and ethnicity gives 100% power, and random grouping of the same records gives 4.8%. In the 4CE consortium’s neurological COVID-19 data (22,000 patients, 21 hospitals, six countries), meta-analytic methods diverge 350-fold on the pooled effect.