It describes a situataion where the observed trend reverses depending on whether we pool the data or divide into strata (Wikipedia). A simple illustrative example: Pasta, BMI, and Simpson’s Paradox
Also related to the differences across pooled, random effects, and fixed effects Regression analysis.
- Hernán2011simpsons: The Simpson’s paradox unraveled
In Europe, the excess death rate (vs. summer) increases as the winter average temperature goes up. However, the colder the weather is in each country, the more deaths occur: https://twitter.com/AinslieJstone/status/1599786478300590080