Throwing Cold Water on the Cold Winters Hypothesis: A Global Test
One of the oldest evolutionary explanations for global cognitive differences is the cold winters hypothesis. In its modern form, the theory is most strongly associated with Richard Lynn, who argued that populations living in colder, more seasonal environments faced recurrent survival challenges that disproportionately rewarded planning, long-term foresight, and problem-solving ability (Lynn, 2006; Rushton, 2012). Over many generations, this would lead to selection favoring higher average cognitive ability in colder climates.
The basic logic is straightforward. In warm, relatively stable environments, food can often be obtained year-round with limited storage, planning, or technological complexity. In contrast, cold or highly seasonal environments impose sharp survival bottlenecks: winters must be anticipated, food stored, shelter constructed, and resources managed months in advance. Failure to do so carries immediate fitness costs. Under such conditions, traits related to long-term planning, impulse control, and learning efficiency should matter more for survival and reproduction.
Versions of this idea long predate modern intelligence testing. Nineteenth- and early twentieth-century geographers and anthropologists repeatedly noted correlations between climate, subsistence complexity, and social organization. What Lynn contributed was a systematic attempt, controversial but influential, to link these ecological pressures to population differences in measured cognitive outcomes.
In recent years, however, this debate has shifted. Instead of relying on phenotypic measures like IQ tests, we can now ask a sharper question: do genetic variants associated with educational attainment and cognition show systematic geographic patterning consistent with cold-climate selection?
This post is an attempt to answer that question using ancient and modern DNA.
Why use latitude as a proxy for climate?
In principle, the cold winters theory should be tested using historical measures of winter severity and seasonality at each population’s location and time. In practice, that is difficult to do consistently for ancient samples spanning tens of thousands of years.
Latitude offers a pragmatic alternative. It is precisely measured for every sample, strongly correlated with long-run temperature and seasonality, and, crucially for ancient DNA, temporally stable. It also avoids a key weighting problem: what matters for selection is not the temperature in the exact year an individual died, but the average environment experienced over many generations, often across nearby locations as populations moved.
For that reason, latitude has long been used as a first-pass proxy in evolutionary and ecological studies of human variation. Here it allows a narrow test of the cold winters prediction: if colder, more seasonal environments favored cognitive traits, polygenic scores associated with education and cognition should increase with absolute latitude, all else equal.
This analysis uses polygenic scores (PGS) for educational attainment (EA), constructed from two large GWAS (EA3 and EA4; Lee et al., 2018 and Okbay et al., 2022, respectively) , averaged and standardized within the combined ancient-modern sample. For comparison, I also analyze PGS for height (Yengo et al., 2022).
Height is a useful positive control because latitude–body size gradients are a classic ecogeographic pattern (often discussed under Bergmann’s rule): larger bodies lose less heat per unit mass, since surface area rises more slowly than volume. If latitude is capturing anything climate-like here, height is where it should show up.
Educational attainment is a noisy phenotype. It reflects cognitive ability, but also noncognitive traits such as motivation, self-regulation, persistence, and planning that help translate ability into years of schooling.
This motivates a common objection to ancient-DNA time series of education polygenic scores: even if EA PGS rose over time, the change might mostly reflect noncognitive traits rather than cognition itself.
Recent work makes it possible to test this directly. Using GWAS-by-subtraction, the genetic signal for education can be decomposed into a Cog component tied to cognitive performance and a NonCog component capturing the remainder (Demange et al., 2021). Here I examine latitude patterns not only for EA PGS, but also for these Cog and NonCog components, allowing a cleaner test of what kind of traits, if any, line up with climate-related gradients.
I start with the raw latitude patterns, then progressively tighten the test by adjusting for ancestry (the first 10 genome-wide PCs) and technical quality (genomic coverage). I run the same logic on ancient-only, modern-only, and pooled data to see which results are robust rather than driven by who is sampled where.
Below the paywall, I present the full set of results and figures, including nonlinear latitude tests, and which traits, if any, retain a genuine latitudinal signal.
