Simpson's paradox is what happens when a trend that holds in every subgroup of the data reverses once the subgroups are combined. Nothing has been miscalculated. Both the split figures and the combined figure are correct, and they point in opposite directions.
The error is not in the arithmetic but in deciding which of the two answers to quote.
Two surgeons, both treating the same condition:
Dr B is better with easy cases and better with hard cases. Dr A has the better overall rate. Both statements are true, and the reason is that Dr B took ten times as many hard cases.
The reversal is driven by a lurking variable that differs between groups — here, case difficulty. Where no such variable exists, the paradox cannot occur.
Simpson's paradox matters because it makes the question “what do the data say?” genuinely ambiguous. Someone quoting the aggregate and someone quoting the breakdown can both be accurate and reach opposite conclusions, and no amount of recomputation will settle it.
What settles it is causal reasoning about why the groups differ in size. In the surgeons' case the split figures are the fair comparison, because Dr B's worse aggregate is produced by taking the harder work. If instead the grouping variable were something the surgeons caused — if one of them created the difficulty rather than inheriting it — the aggregate might be the honest number.
That is the uncomfortable part: the data alone cannot tell you which figure to use. You need to know how the groups came to be the sizes they are, and that is a question about the world rather than about the spreadsheet.
In practice the paradox is most common wherever a population is unevenly distributed across conditions — hospital outcomes, university admissions, hiring rates, treatment effects by age. Any time an aggregate figure contradicts every breakdown you have seen, this is the first thing to check.