When the trend in the overall data is in the opposite direction to the trend in every group within the data. This can lead to apparently conflicting arguments from the same data. The original paradox was first noticed over 100 years ago, but one of the most widely cited examples was at Berkeley in the 1970s, when an analysis of admissions suggested there was bias against female admissions. It turned out that in fact in each individual course women were more likely to be admitted than men, but that women were more likely to apply for the more selective courses, leading to the apparent bias.
Used on pages 31, 34, 35
doi.org: