Aggregation as Microscope: Why Both Selection and Structure Hide at the Single Locus
2026-08-07 · AI-generated insight
These two notes describe the same statistical move applied to two different questions, and the parallel is worth making explicit: signals that are invisible at any single genetic locus become unmistakable when you aggregate correlated small effects.
Winegard and Carl's rebuttal of Lewontin turns on this. At any one locus, between-population variation is dwarfed by within-population variation — just as nose size alone can't sex a face. But correlated differences across thousands of loci make population structure legible, the way a whole face makes sex obvious. The information was always there; it just doesn't live in any single variable.
The ancient DNA study faces the mirror-image problem in the time dimension. A single allele's frequency shift over millennia is ambiguous — it could be drift, migration, or population structure. The method works by testing for consistent trends over time, and the headline findings involve combinations of alleles: standard-deviation-scale shifts in polygenic predictors of body fat, schizophrenia, and cognitive performance. Again, aggregation across many weak, correlated signals is what turns noise into detection.
But there's a deeper layer: in the selection case, aggregation isn't just our epistemic trick — it's apparently how evolution itself operated. The study confirms that classic hard sweeps (one big mutation driven to fixation) were rare. Recent human evolution instead worked through polygenic selection: nudging hundreds of alleles simultaneously, each by a little. Evolution distributed its signal across many loci for the same reason Lewontin could hide population structure there — individual loci are cheap, noisy, and low-stakes, while the aggregate carries the phenotype.
The synthesis: the genetic architecture of complex traits and the statistics needed to study them are the same shape. Lewontin's error was treating loci as independent witnesses rather than a correlated chorus. The ancient DNA method succeeds precisely by refusing that error in the temporal domain. Any future debate about recent human evolution — including uncomfortable ones about cognition, which the paper explicitly flags with its caveat about past versus present phenotypes — will be fought on this terrain: not "is there a gene for X?" but "do the small correlated shifts point the same direction?"