The latest declines raise a broader question: what predicts gains and losses in PISA scores across successive waves? Here I examine changes from 2006 to 2025, alongside persistent differences between countries. The recent downturn is the starting point for that question; explaining it specifically would require a separate analysis.

Changing immigrant composition is one possible contributor to falling national averages. Where immigrant students score lower on average, a rising immigrant share can lower the overall score even if performance within each group is unchanged. The effect depends on the country and its immigrant population. I therefore also examine non-immigrant students separately.

In “Can Genes Help Countries Get Rich?”, I examined country-level educational-attainment polygenic scores and long-run economic growth. This follow-up asks whether prosperity during childhood helps explain PISA performance. I also examine whether country-level educational attainment (EA4) polygenic scores help explain persistent differences in PISA performance, and how much of that association is accounted for by prosperity.

Read more

That leaves a different question: where does my genome sit when we compare it directly with ancient individuals from different places and periods? Here I explore that question visually, using two ancient-DNA PCA comparisons. I ask which affinities recur—and how much the picture changes when modern rather than ancient people define the axes.

This also extends the visual approach in my earlier post, Painting Geographic Maps with DNA, and the accompanying PCA Atlas. Those geographic maps described variation among people living today; this comparison turns to people who lived in the past.

Two ways to draw the genetic map

Principal component analysis compresses genetic variation into a few axes. Nearby points have similar scores on the dimensions shown. These are genetic plots, not geographic maps: moving left or right does not translate directly into travelling west or east.

In the first analysis, 269 modern individuals define the axes. Ancient genomes and my own genome are then placed onto that fixed reference frame. In the second, 44 high-call-rate ancient individuals define a separate frame, and the remaining genomes are placed onto it. My genome is held out of both fits, so it does not determine either set of axes.

Both analyses use the same 76,159 SNPs. This matters because it holds the marker set constant while changing the people who define the axes. The two panels are independently centred and scaled; their PC1 and PC2 coordinates are not interchangeable. This modern-reference comparison is also a separate analysis from the existing Europe-wide atlas PCA.

Below, I examine which ancient populations fall closest to my genome—and why the answer changes with the reference panel.

Read more

The interesting part, for me, was how much of that northern contribution would show up. Would it be a small addition to an otherwise local Italian profile, or would it account for a substantial part of the result? And could the comparison tell me anything about whether that affinity was closer to the Celtic or Germanic side of my expectations? I wanted to see how far the data would take me.

Then I got a result I had not anticipated. My genome fitted an ancient Roman/Italic source on its own, without needing an added northern source to pass the test. Expressed as an ancestry bar, that is a rather arresting 100% Roman. Hence the question in the title.

I had gone looking for a mixture, so this was an unexpected place to start. Where was the Celtic or Germanic contribution I had expected? Would it emerge in other comparisons, or would those also place me close to the ancient Italians? Before deciding what to make of my result, I wanted to know how often the same thing happened to other northern Italians. A result shared by many of them would be a different story from one peculiar to me.

So this post follows my genome through those comparisons, with other northern Italians alongside it. I start with where I sit among living populations, then turn to the ancient samples and the ancestry models. The question that interests me is how much of the history I had in mind can actually be recovered from my DNA, and what happens to that surprising Roman result when I look more closely.

How I made the comparison

The baseline Roman/Italic source contains four people from Castel di Decima, Palestrina, Ardea and Boville Ernica. They represent Roman, Latin and Volscian or Hernician contexts, spanning roughly 900–200 BCE. Some lived before the Republic, so I use the broader name Roman/Italic. Etruscans are tested separately.

For this article’s PCA, all reference genotypes come from the Allen Ancient DNA Resource (AADR), which includes modern as well as ancient people. I fitted the axes to a selected subset of 693 modern references in 90 groups, using 61,521 filtered autosomal SNPs, then projected my genome and the 19 study individuals onto those fixed axes. None of us helps define the coordinate system. [2]

In the first qpAdm comparison, I tested the four-person Roman/Italic pool and a separate ten-person Etruscan sample from Tarquinia, both alone and with each of nine alternative sources. Etruscans provide a neighbouring central Italian comparison. Repeating the twenty-model comparison with the six-person Roman/Italic pool gives 40 models per person and 800 across all 20 genomes. Each mixture contains one Italian source and one additional source; I do not fit Gaulish and Germanic contributions simultaneously. All memberships stay fixed between targets.

I use p ≥ 0.05 as the model-fit threshold. For an added source, I also require valid proportions, distinguishable source populations and an improvement over the corresponding Italian-only model after multiple-testing correction. The appendix gives the sample lists and settings.

A second comparison adds Imperial Italian communities and eastern Mediterranean sources. Here I keep the twelve comparison populations fixed for every model, rather than rotating unused candidate sources into that set. This gives the four-person, six-person and Imperial source models a common basis for comparison. Figures 1–2 use the first design; Figures 3–4 use this fixed-panel test.

For model selection, I start with each Italian source alone and test the permitted additions one at a time. I retain an addition only when the resulting model passes (p ≥ 0.05), its proportions are valid, its sources are distinguishable, and the nested improvement test has Holm-corrected p < 0.05. I stop a branch when no tested addition qualifies; if its model fails, that branch supplies no adequate model. When several alternatives qualify, I report them separately. A larger overall model p-value is not enough to retain a source. Figure 3 applies this rule; Figure 4 separately shows the highest-p-value candidates. [1]

Who the comparison populations were

The Gaulish reference consists of 40 people from Bucy-le-Long in northern France, a La Tène community of Iron Age Gaul. The Gauls were Celtic peoples; this group tests the Celtic side of my expectation using an actual ancient community, rather than modern French people. It represents northern Gaul, not every Celtic population. [9, 10]

Alken Enge (14), Asnæs (10), Lille Vadsby (9) and Simonsborg (19) are Roman-period groups from Denmark. I use them as Scandinavian proxies for Germanic-related ancestry, keeping the sites separate to see whether the result depends on one community. Wielbark (17), from two sites in northern Poland, adds a Baltic comparison. Its archaeological culture is commonly associated with the Goths, although that does not establish the ethnic identity of every person buried there. [10, 12]

The Longobards, or Lombards, were a Germanic people whose migration from Pannonia into Italy in 568 CE makes them especially relevant here. The Hungarian references come from their associated cemetery at Szólád. I retain its main northern/central-European group (13) and two additional AADR groups (7 and 4) separately. “Core”, “1” and “2” identify those genetic groupings, not different tribes. The community contained people of differing ancestry, so a percentage assigned to a Szólád group is not automatically a percentage of northern ancestry. [11]

Every source is a group of people, not a single ancient genome. Together, these nine alternatives contain 133 distinct individuals. They were chosen for their historical relevance, archaeological context and available qualifying genomes, not for the highest fit to my DNA.

My genome is closest to northern Italian samples, particularly Bergamo and Torino.

Migration brings ancestry into a population; later gene flow and drift change how that population relates to its neighbours. PCA and FST summarise the resulting differences. They do not separate Celtic, Roman and Germanic contributions. To ask about those historical sources, I need an explicit model.

What the single Roman bar means

With one source, qpAdm fixes its coefficient at 100%. The test is whether that source can account for my genetic rFigure 1. Every model of my genomeelationships with the comparison populations. My four-person Roman/Italic model passes at p = 0.2257. The Etruscan-only model fails at p = 0.0080. [1]

Adding the main Szólád Longobard group to the four-person Roman/Italic baseline gives a Longobard-source estimate of 2.4% ± 13.4 percentage points and p = 0.1463. The Gaulish alternative produces a negative coefficient, −5.7%, so it is not a valid ancestry mixture. Figure 1 shows every model, including the rejected and invalid fits.

Figure 1. Every model of my genome

Including the Roman outliers

Figure 1 compares the full model sets. The six-person pool adds two individuals whom the metadata identify as eastern-Mediterranean-shifted outliers, one from Palestrina and one from Ardea. They belong to Roman/Latin archaeological contexts but differ genetically from the four-person baseline. Including them tests a broader version of the ancient reference, while keeping Etruscans outside it.

With this pool, my one-source p-value is 0.1122, which still passes. Adding the Gaulish group gives 18.2% ± 13.6 percentage points and p = 0.1293. The main Szólád group gives 14.8% ± 10.4 percentage points and p = 0.1081. Both mixtures pass, but neither added source survives the corrected improvement test. The one-source model remains sufficient under this comparison.

Including eastern-Mediterranean-shifted individuals raises the estimated contribution from several northern alternatives. A northern source can partly compensate for that change in the Italian reference. Yet the uncertainty is large enough that the added contribution is not required. The difference between a passing mixture and evidence for an additional source matters here: I get the former, but not the latter.

The two additional Szólád groups give negative personal coefficients and, in several fits, extremely large errors. Those unstable estimates are not usable ancestry percentages. Keeping the groups separate makes this visible instead of hiding the cemetery’s diversity inside one Longobard average.

The Etruscan-only model fails in both analyses, as do all eighteen Etruscan-plus-additional-source fits for my genome. This result concerns the particular Tarquinia sample and comparison populations used here. Ancient-DNA studies also find substantial genetic similarity between Etruscans and neighbouring central Italians, so the archaeological labels should not be read as sharply separated genetic categories. [3, 4]

Below, I compare my result with 19 other northern Italians and test what happens when Imperial-era genomes replace the Iron Age references. Upgrade to read the full results, including the Celtic and Longobard-associated models and how the choice of ancient samples changes the answer.

Read more

National stereotypes certainly assume that such differences exist. Americans are said to be outgoing and brash, the British reserved, Germans orderly, Italians expressive and Finns quiet. These portraits might contain observation, caricature or both. In a famous 49-culture study, national-character stereotypes were internally consistent but generally failed to match the average personality scores reported by the people who lived in those countries.

The measurement problem runs deeper than inaccurate stereotypes. When people answer a statement such as ‘I am outgoing’, they rarely compare themselves with humanity as a whole. They compare themselves with friends, colleagues and other people around them. An Italian and a Finn with the same observable behaviour may therefore choose different answers because each is judging against a different local norm. Psychologists call this the frame-of-reference, or reference-group, effect. It can compress, erase or even reverse apparent differences between countries in self-report data.

Until recently, the genetic side of the comparison was also too weak for a serious test. That changed with the new Nature personality GWAS. It is large enough to identify 1,260 lead variants across the Big Five, making it possible to ask whether population differences in those variants predict population differences in reported personality. I tested that question globally, in Europe, and in a broader whole-genome frequency panel.

Explore the results yourself: Open the interactive Extraversion map to switch between the aggregate-frequency and individual-level genotype maps, zoom around the world, and compare genetic and self-reported scores.

Only one trait survived every test. Which countries came out on top, and did the stereotypes get it right? Upgrade to see the full global and European results.

Read more

Online ancestry reports love a dramatic cast list: Viking, Celt, Roman, Western Hunter-Gatherer, Anatolian Farmer. The percentages look wonderfully precise. The trouble is that an analyst can keep swapping ancient populations until a seductive combination appears. If the failed attempts vanish and only the prettiest bar survives, the result tells us as much about the search as it does about the genome.

I chose modern Italians as my test case, but the same procedure can be used for any population. I begin with sources that make historical sense, test deep, prehistoric and historical ancestry separately, and add components one at a time. When a new source fails to improve the simpler model, I stop.

I wanted to ask those questions in order. First: which deep ancestry components werThanks so much for supporting my work!e actually needed? Second: which Neolithic, Bell Beaker, Corded Ware or Yamnaya-era populations could model the prehistoric layer? Only then did I ask whether a northern historical source improved an Iron Age Italian baseline. The source set changed at each step, so a Roman population never shared an ancestry bar with WHG or Anatolian farmers.

Figure 1. Digging through the layers of ancestry

qpAdm, in plain English: a statistical test that asks whether a target genome can be represented by one or more ancient populations chosen in advance. It compares patterns of shared genetic variation against a separate reference panel and reports both model fit and source weights.

A model with p < .05 is rejected. A model above that threshold is merely not rejected: it is not automatically true, unique or historically best. Its percentages are weights assigned to the chosen genetic sources.

Layer 1: Deep ancestry

I began with the deepest ancestry layer: Western and Eastern hunter-gatherers, Anatolian Neolithic farmers, Iranian Neolithic ancestry and a western Yamnaya-related proxy. These are broad genetic building blocks, not historical peoples, and no Etruscan, Roman, Celtic or Germanic source appears in this part of the analysis.

Figures 2 and 3 compare like with like: five northern individuals—one each from Ferrara, Cremona, Domodossola, Taio and Treviso—and five Central Italian individuals from Bolsena, Ascoli Piceno, Montepulciano, Perugia and Chiusi. The regional pools appear only in the historical section.

All ten individuals retained a two- or three-source deep model. Anatolian Neolithic and western Yamnaya-related ancestry appeared throughout. Iranian Neolithic ancestry was also retained for the Emilia-Romagna and Veneto representatives and all five Central Italian individuals; it was not needed for the representatives from Lombardy, Piedmont or Trentino–Alto Adige. WHG and EHG were available to the search but did not earn a separate place in the retained models.

Figure 2. Deep ancestry models for northern and Central Italy

Layer 2: Prehistoric ancestry

I then moved one rung closer in time and changed the source set completely. The northern targets were tested with Neolithic Italian, Bell Beaker, Corded Ware and regional Yamnaya sources. For the Central Italian targets I declared a separate, period-matched menu: Central Italian Neolithic and Bronze Age groups from Lazio and Sicily, Adriatic-Balkan groups from Cetina and Mokrin, and Aegean Late Bronze Age groups from Aegina, Koukounaries and Chania. Historical populations were excluded.

This layer was much less forgiving. Two northern individuals retained a passing, feasible model: the Lombardy representative was estimated as 53% Central Italian Neolithic plus 47% Corded Ware [model p = .309], while the Trentino–Alto Adige representative was estimated as 32% Central Italian Neolithic plus 68% Bell Beaker [model p = .775]. With the separate Central-Mediterranean menu, Bolsena also passed: 44.4% Mokrin Early Bronze Age plus 55.6% Aegina Late Bronze Age [model p = .0714]. The other seven individuals had no passing, feasible prehistoric model.

Figure 3. Prehistoric ancestry models for northern and Central Italy

Layer 3: Historical ancestry

Only at the final layer did I introduce Etruscans and Roman-Republic Italians, then ask whether Celtic- and Germanic-period sources improved those Iron Age Italian baselines. The figures that follow answer that narrower historical question. Deep and prehistoric sources do not appear in these bars.

Historical models: the northern regional pools

I first pooled modern AADR individuals by region: Emilia-Romagna (n = 7), Lombardy (n = 3), Piedmont (n = 4), Trentino–Alto Adige (n = 2) and Veneto (n = 3). Pooling gives a regional signal, although the last two pools are small and should not be mistaken for population surveys.

Exactly what each pool contains: Emilia-Romagna: 7 individuals; Lombardy: 3; Piedmont: 4; Trentino–Alto Adige: 2; Veneto: 3—19 people in total. Every individual row later in the article represents one person (n = 1), not another pool.

Figure 4. Historical additions retained for five northern Italian regional pools

Emilia-Romagna, Lombardy and Piedmont retained no northern addition in any of the four historical families. That result does not deny Celtic or Germanic ancestry in those regions. It says that, with these sources and this statistical panel, none of the additions improved an already viable Iron Age Italian baseline strongly enough to clear the stopping rule.

Trentino–Alto Adige was different. One model assigned 72.9% to the Roman-Republic baseline and 27.1% to a Hungary La Tène proxy [model p = .658; Holm p = .0227]. A separate, later comparison assigned 81.5% to late Lazio Etruscans and 18.5% to an early Roman-period group from Alken Enge in Denmark [model p = .158; Holm p = .0306]. These are alternative temporal models, not components to be stacked together.

Veneto also retained a Roman-period northern proxy: 75.0% late Lazio Etruscan and 25.0% Alken Enge [model p = .0515; Holm p = .000125]. Its early-medieval family produced 61.4% Roman-Republic and 38.6% Saxon [model p = .780; Holm p = .0227], but that bar deserves an asterisk because the early-medieval calibration behaved abnormally.

What happens when every pool member is tested?

The only fair way to interpret a pooled result is to test everyone inside it. I therefore repeated all four historical families for each of the 19 pool members—912 additional qpAdm models under the same SNP panel, source rotations and corrected stopping rule. No person was represented by a regional stand-in.

The negative pools were internally consistent. None of the seven Emilia-Romagna individuals and none of the three Lombardy individuals retained a northern addition. In Piedmont, three of four individuals—including Cuneo—retained nothing, but the Domodossola individual retained 24.8% Alken-Enge-related ancestry [model p = .480; Holm p = .0157]. A 42.0% Saxon-like early-medieval model also passed narrowly [model p = .0530; Holm p = .0434], but belongs to the technically unreliable early-medieval family.

Trentino now makes sense. The Roman-period pool estimated 18.5% Alken-Enge-related ancestry. Nogarè (Nei27) retained 23.9 ± 8.1% [model p = .0804; Holm p = .0422], whereas Taio (Nei18) estimated 14.6 ± 8.1% but did not improve the Italian baseline after correction [Holm p = .683]. The pooled Roman-period result is therefore consistent with the stronger signal in Nogarè.

The Trentino La Tène result is a different case. Nogarè estimated 29.1 ± 11.1% and Taio 25.3 ± 13.0%, but neither cleared Holm correction alone [Holm p = .0796 and .148]. The two-person pool estimated a similar 27.1 ± 8.9%; its smaller uncertainty allowed the addition to pass [Holm p = .0227]. Pooling did not invent a different coefficient—it made a shared but noisy signal precise enough to retain.

Veneto supplied the cleanest replication. All three members retained a Roman-period northern proxy: Bassano del Grappa 24.6%, Treviso 30.5% and Sappada 27.5%. Their pooled estimate was 25.0%. The pool’s 38.6% Saxon-like early-medieval result was not retained by any member separately and remains provisional because of the early-medieval calibration problem.

Figure 5. Northern Italian regional pools and all 19 people inside them

Historical models: Central Italy

The Central Italian comparison used five modern individuals from Bolsena, Ascoli Piceno, Montepulciano, Perugia and Chiusi. Four retained no northern historical addition at all. Their results were compatible with the sampled Iron Age Italian baselines without requiring Hallstatt, La Tène, Roman-period northern or early-medieval sources.

Bolsena was the sole exception, and only barely in ancestry terms: 98.9% late Lazio Etruscan plus 1.1% Hungary Longobard-like [model p = .335; Holm p = 1.44 × 10⁻⁹]. Alternative early-medieval models put the northern coefficient between 0.6% and 0.9%. A statistically sharp distinction can therefore correspond to a biologically tiny coefficient.

I would not build a migration story around that one percent. The synthetic early-medieval tests printed zero jackknife standard errors in every replicate, an obvious warning that this family was overconfident. The honest summary is that the five Central Italian genomes were almost entirely compatible with Iron Age Italian baselines, with one tiny and technically fragile exception.

Who were the Etruscans? They lived in city-states centred on Etruria—roughly modern Tuscany, northern Lazio and western Umbria—from about the eighth to third centuries BCE. They spoke a non-Indo-European language, strongly influenced early Rome and were gradually incorporated into the Roman Republic.

What the comparison actually shows

The cleanest contrast is not ‘Roman Central Italians versus Germanic Northern Italians.’ It is uniformity versus heterogeneity. The Central Italian individuals were consistently explained by Iron Age Italian baselines. Across the 19 northerners, robust retained additions were concentrated in Domodossola, Nogarè and all three Veneto localities, while fourteen individuals retained none.

Figure 6. Historical tests for five Central Italian individuals

*Early-medieval estimates are shown for completeness but treated cautiously because the synthetic calibration returned zero printed standard errors.

The retained northern weights were substantial where they appeared: about 18–30% in the most reliable La Tène and Roman-period results. Closely related ancient sources can sometimes substitute for one another, so each source label is best read together with its period and comparison family.

The sensitivity experiment was reassuring for the Roman-period family: simulated 5–30% northern mixtures were usually detected after correction, and no false northern component was retained at 0%. Hallstatt and La Tène models were less stable because some simulated mixtures failed the overall model test. The early-medieval family remains provisional because of the zero-standard-error problem.

Why the stopping rule matters

If I had simply displayed every passing two-way model, nearly every target could have acquired a colourful Celtic, Danish, Saxon or Longobard slice. A high p-value would then look like discovery. The nested test asks the more useful question: did that slice improve the simpler model enough to justify the extra story?

That rule removed most additions. It also preserved Roman-period northern signals in Domodossola, Nogarè and all three Veneto individuals, plus the corresponding Trentino and Veneto pools. A method that sometimes says yes and often says no is more informative than one designed to produce an ancestry cocktail for everyone.

So how Roman is Italy today? The cautious answer is that these modern genomes remain strongly compatible with sampled Iron Age Italian ancestry, especially in Central Italy. Northern Italy shows additional northern historical affinity in some places and people, but not as a universal layer.

Who is hiding in your DNA?

An Etruscan? A steppe herder? A suspiciously persistent Longobard?

Most ancestry websites produce a colourful cocktail of Romans, Vikings and hunter-gatherers, then quietly neglect to mention how many alternative recipes they tried first. I take the opposite approach: I test your genome against real ancient populations, discard models that fail, and stop adding ancestors when the evidence stops improving.

To celebrate this project, I am offering five new yearly subscribers a complimentary Personal Ancient-Ancestry Modelling Report.

Your report will explore three distinct chapters of your genetic history: your deepest hunter-gatherer and early-farmer foundations; your Neolithic, Bronze Age and steppe-related ancestry; and your affinities to Iron Age, Roman and early-medieval populations.

You will receive the models that worked, the ones that failed, and a plain-English explanation of the results. The same stopping rule used in this article will be applied before the report is written.

Become a yearly subscriber and apply: email pifferdavide@gmail.com with the subject ‘Ancient DNA report’ and the address used for your subscription. I will reply with private upload instructions. Please do not attach your genome to the email.

Raw DNA files will be submitted privately, used only to prepare the report and deleted afterward. The analysis concerns population history—not health, paternity or medical risk.

Methods note

To limit model fishing, I kept the 173,185-SNP AADR v66.HO-aligned panel, right-population set, source definitions and 200 bootstrap replicates fixed within each declared model family. Deep, prehistoric and historical layers remained separate, and source groups generally required at least four unrelated, quality-controlled individuals. Figures 2 and 3 use ten single-person targets: five northern and five Central Italian individuals. Each faced 31 deep combinations. The northern prehistoric menu contained 1,023 combinations per person; the revised Central-Mediterranean menu contained 47 predeclared combinations per person, or 235 models in total. Its groups contained four Central Italian Neolithic individuals, four Lazio Bronze Age, eight Sicily Early Bronze Age, eight Cetina Middle Bronze Age, twelve Mokrin Early Bronze Age, four Koukounaries Late Bronze Age, four Aegina Late Bronze Age and seventeen Chania Late Bronze Age individuals. The historical section then switches explicitly to five northern regional pools, each screened across 98 historical combinations, before testing all 19 constituent individuals separately. That individual extension was historical only: 912 models and 1,444 nested comparisons.

Expanded historical models were retained only when the overall fit was p ≥ .05, every coefficient lay between zero and one, the sources were distinguishable, and the improvement over the matching one-source Italian model survived both Benjamini–Hochberg and Holm correction. In Figures 4 and 6, a full red bar shows the retained one-source Iron Age Italian baseline; a blue segment appears only when a northern addition clears the full rule. Figure 5 uses a dash for a northern addition that was not retained.

References

Antonio, M.L., Gao, Z., Moots, H.M. et al. (2019). Ancient Rome: A genetic crossroads of Europe and the Mediterranean. Science, 366(6466), 708–714. https://doi.org/10.1126/science.aay6826

Harney, É., Patterson, N., Reich, D. & Wakeley, J. (2021). Assessing the performance of qpAdm: a statistical tool for studying population admixture. Genetics, 217(4), iyaa045. https://doi.org/10.1093/genetics/iyaa045

Mallick, S., Micco, A., Mah, M., Ringbauer, H., Lazaridis, I., Olalde, I., Patterson, N. & Reich, D. (2024). The Allen Ancient DNA Resource (AADR) a curated compendium of ancient human genomes. Scientific Data, 11, 182. https://doi.org/10.1038/s41597-024-03031-7

I have come a long way since, as a teenager, I first encountered the maps of human genetic variation created by Luigi Luca Cavalli-Sforza and his colleagues. Their work opened the way for using DNA to reconstruct our migration histories and address two of the oldest questions: Who are we, and where do we come from? It showed me that population history could be made visible through patterns preserved in genes.

At the time, no ancient human genome had yet been sequenced, so the past had to be inferred almost entirely from genetic variation among living populations. Today, thanks to the massive collective effort of researchers around the world, we have DNA from ancient Homo sapiens across every inhabited continent, beginning with individuals who lived during the Upper Paleolithic. Researchers have also recovered DNA from Neanderthals and Denisovans, reaching even further back in time.

We no longer have to reconstruct every migration and population change backward from the present. We can observe parts of that history unfolding through time.

Buried in the Genome grew out of that scientific transformation, and out of my own long engagement with its questions. Today I am publishing it.

The book follows the genomic record from deep prehistory to the rise of complex societies—and asks what it can tell us about the populations and people who built them.

History Has Entered the Genomic Age

For most of history, bones and artifacts could tell us where people lived and what they made, but not precisely who they were or how they were related. Ancient DNA changed the scale of the question. A skeleton can now reveal ancestry, biological sex, family relationships, pathogens, pigmentation variants, and fragments of a population’s demographic history.

The result is not simply more information about the past. It is a different kind of history. Genomes preserve events that left no chronicle: migrations whose languages disappeared, local populations absorbed by newcomers, adaptations assembled gradually from variants with different origins, and demographic reversals hidden beneath familiar national labels.

I wrote Buried in the Genome because these discoveries are too important to remain scattered across specialist papers, technical debates, and isolated datasets. They belong in a single historical narrative.

What the Book Follows

The book begins with the tools. Genetic distance, principal-component maps, ancient genomes, and polygenic scores are powerful precisely because they answer narrow questions. Understanding what their numbers measure is the price of using them well.

Populations in motion. The migrations, replacements, mixtures, and local continuities that assembled Europe and reshaped Eurasia.

Traits through time. Pigmentation, stature, and cognition each have their own chronology. Human traits did not emerge as finished packages, and the populations carrying them continued to change long after the great prehistoric migrations.

Genes and civilization. Agriculture, climate, institutions, and selection created new environments for human reproduction. The final chapters bring those questions into Rome and Italy, where ancient DNA allows regional differences to be followed across unusually fine historical scales.

Across twelve chapters, the argument moves from the basic problem of reading a genome to a larger question: how did biological and cultural evolution interact while human societies became larger, denser, and more complex?

Written to Be Read—and Examined

Buried in the Genome is written for readers interested in genetics, archaeology, history, psychology, and the origins of civilization. You do not need specialist training to follow the main narrative.

But I did not want to hide the machinery. The chapters are accompanied by technical appendices explaining the datasets, models, ancestry corrections, robustness tests, and evidentiary limits behind the central results. Readers who want the argument can read straight through. Readers who want to inspect how it was built can go further.

This is the book I wanted to read: broad enough to connect ancient genomes with the history of civilization, and technical enough to show where the claims come from.

Why This Book, Now?

The ancient-DNA revolution is moving quickly. Each new dataset can redraw a migration, expose hidden family structure, or change the timeline of a trait. At the same time, public discussion still lags behind the evidence. Old categories are often repeated as if populations stood still, while new genomic results are reduced to headlines that detach them from their methods.

A better account has to put movement, ancestry, traits, and institutions into the same frame. That is what I have tried to do here. The genome is not a complete history of humanity, but it is now one of history’s most revealing archives.

Read Buried in the Genome

The book is available now as a direct digital edition containing both EPUB and PDF. It is also available on Kindle and in paperback, with links for international Amazon stores.

GET THE BOOK → book.davidepiffer.com

Direct EPUB + PDF: $12.99 · Kindle: $12.99 · Paperback: $29.99

If my work has interested you, buying the book—and sharing this post with someone who would want to read it—is the most direct way to support it.

The past is not silent. It survives in us.

Buried in the Genome is my attempt to listen to it.

It also created crowded settlements, unfamiliar diets, new disease environments and much larger populations. In The 10,000 Year Explosion, Gregory Cochran and Henry Harpending argued that this upheaval accelerated human evolution. Culture had not replaced natural selection; it had built a new arena for it. Hawks and colleagues later reported genomic patterns consistent with unusually rapid adaptation during the Late Pleistocene and Holocene.

The cognitive version of that argument is straightforward. If farming societies rewarded learning, planning and sustained effort, variants associated today with educational attainment might have increased. If they rewarded a greater willingness to wait for future rewards, variants associated with delay discounting might have decreased.

This is the East Asian test. The genome collection used here is centered on East Asia, where rice and wheat can be matched to local archaeological adoption dates. Europe, Africa and the Americas each require their own agricultural clock, built from their own crops, archaeological sites and ancient genomes. They are the natural next tests.

This is the first direct test of the farming hypothesis with ancient DNA. Previous studies tracked genetic change over calendar time or compared broad archaeological groups. Here, the date of local crop adoption enters the model itself, making it possible to ask whether the genetic trajectory changes when farming arrives.

Ancient DNA had already revealed half the story

Piffer and Kirkegaard found rising educational-attainment and IQ polygenic scores across European genomes spanning roughly 12,000 years. Akbari and colleagues found evidence of directional selection across West Eurasia and higher mean scores among early European farmers than Western hunter-gatherers.

These studies make a Holocene rise plausible, but farming itself remains in the background: calendar time and broad archaeological labels stand in for the event. The innovation here is to synchronize every genome to its own local farming transition. Each person is placed on an agricultural clock—centuries before or after the first local evidence of domesticated crops—so the analysis can ask whether the genetic trajectory actually bends when farming arrives.

Read more

About 3,000 years ago, migrants with Northeast Asian ancestry brought wet-rice agriculture during the Yayoi transition. Centuries later, as political power consolidated during the Kofun period, another substantial influx introduced ancestry related to other East Asian populations. The groups mixed. Modern Japanese origins are therefore not a story of a timeless, isolated population—or of one population wholly replacing another.

For decades, the leading model was simpler: indigenous Jōmon plus continental farmers. Ancient DNA complicated it. Genomes spanning the Jōmon, Yayoi and Kofun periods support a tripartite origin for present-day Japanese: Jōmon ancestry, a Northeast Asian component associated with the spread of farming, and a later East Asian component associated with state formation (Cooke et al., 2021). A study involving more than 250,000 modern participants later showed that this Jōmon legacy still helps structure genomic variation across the archipelago (Yamamoto et al., 2024).

This turns a present-day Japanese genome into a historical mosaic. One stretch of chromosome can be more Jōmon-like; the next can be more mainland-like, because recombination has repeatedly cut and reshuffled ancestral chromosomes. Every Japanese chromosome is, in that sense, an archaeological site.

And that creates the real mystery. Did these ancestries merely mix, or did natural selection edit the mosaic after admixture—preserving Jōmon-derived segments in some regions while thinning them in others? The question has become especially timely after Watanabe and colleagues analysed 42 Jomon genomes and reported signals consistent with cold adaptation in pathways involving thermogenesis, lipid metabolism and cold sensation (Watanabe et al., 2026). Here I ask whether Jōmon ancestry is unusually enriched or depleted around variants used in three polygenic scores: educational attainment, height and skin colour.

A genome is a mosaic, not a smoothie

A whole-genome ancestry percentage compresses a great deal of history into one number. Local ancestry does the opposite: it moves along each chromosome and asks which reference source is more compatible with each segment. Because recombination breaks inherited chromosomes at every generation, admixed genomes become mosaics of ancestry tracts (Gravel, 2012). Here, a two-state hidden Markov model assigned a probability of Jōmon-like or mainland-like ancestry to phased chromosome segments in 104 Japanese-in-Tokyo participants from the 1000 Genomes Project (1000 Genomes Project Consortium, 2015).

Researchers sometimes use the phrase “local admixture,” but local ancestry is the more precise term: admixture is the historical mixing event, while local ancestry is the inferred source of one particular chromosome segment. Figure 1 follows the full logic. Ancient genomes define the reference endpoints; recombination creates alternating ancestry tracts; the hidden Markov model uses marker patterns and continuity along the chromosome to estimate a Jōmon-like probability; and the score loci are compared with the same locus pattern shifted to other positions on the chromosome.

Figure 1. How local ancestry works.

This schematic follows one chromosome from reference panels to the selection test. Blue and orange segments represent probabilistic Jōmon-like and mainland-like assignments, not cultural identities. The hidden Markov model (HMM) combines allele patterns at nearby markers with the expectation that ancestry usually continues along a tract. The lower-right panel uses the primary equal-weight EA4 result as a numerical example: 13.1% Jōmon ancestry at the score loci versus a 12.6% chromosome-matched null, a difference of about +0.5 percentage points. A positive difference is enrichment; a negative difference is depletion.

The words “Jōmon-like” and “mainland-like” matter. The model compares modern haplotypes with allele-frequency endpoints estimated from ancient reference panels. The primary Jōmon endpoint used 45 independently represented ancient individuals. The mainland endpoint pooled 65 ancient people from Shandong dated about 2,016–2,374 years before present—used as geographically and chronologically plausible Yayoi-related proxies—with eight Gaya-period Koreans dated about 1,525–1,550 years before present, used as Kofun-related proxies. Shandong-only and Gaya-only models tested how much the conclusion depended on this subjective proxy choice.

Three scores, two different questions

I examined three polygenic scores. The first comes from the fourth major genome-wide association study of educational attainment, commonly called EA4. It uses variants associated with years of schooling in a modern sample of about three million people, mostly of European genetic ancestry (Okbay et al., 2022). It is not an intelligence score, and the original study showed that family-level controls substantially reduce its association with education. The second score uses height-associated variants from a multi-ancestry GWAS of 5.4 million people (Yengo et al., 2022). The third uses skin-colour variants discovered in 48,433 East Asians by Kim and colleagues (2024); higher values are oriented toward lighter measured skin colour.

For each score, two questions must be kept separate:

1. Question 1: Do alleles on Jōmon-like and mainland-like tracts receive different PGS weights? This describes ancestry-associated differences in the score. It does not establish selection after admixture.

2. Question 2: Is Jōmon ancestry unusually common or rare around the score loci? This is the local-ancestry test relevant to post-admixture enrichment or depletion. Even here, demographic and reference-panel explanations must be excluded before calling selection.

Read more

After Troy

“Tell me, oh Muse, of that ingenious hero who travelled far and wide after he had sacked the famous town of Troy.”

—Homer, The Odyssey, Book I, translated by Samuel Butler [1]

The Odyssey begins after the victory. Troy has fallen; the fleet has scattered; kings return to homes that no longer behave as they did before the war. Christopher Nolan’s 2026 film makes the historical subtext explicit. His Odysseus says: “Our age of bronze is collapsing.” [2] The film looks beyond the victory at Troy toward the loss of palaces, trade and writing—and the Greek Dark Age that followed. Beneath the gods and monsters lies an old historical question: what happens to people when a wealthy, interconnected world breaks apart?

Around 1200 BCE, the Mycenaean palatial system of mainland Greece disintegrated. Palaces were destroyed or abandoned, and Linear B—the writing system used by palace scribes to keep administrative records of goods, land, workers and taxes—disappeared with them. Political and economic life was reorganized, and Greece entered a period in which writing effectively vanished for several centuries. Yet “collapse” is a label for uneven regional processes, not a single night of destruction. Archaeologists still debate the mix of warfare, migration, internal political failure, trade disruption, climate stress and other causes [3].

The idea I want to test is simple. Mycenaean palaces needed skilled people. Scribes had to learn Linear B, officials had to keep records and accounts, and administrators had to coordinate production, storage and trade. If the population became less genetically predisposed toward traits related to learning and problem-solving, the pool of people able to perform these jobs may have shrunk. That would not by itself explain the collapse, but it could have made an already fragile system less able to cope with the pressures it faced.

If this played a role, we should be able to see some trace of it in ancient DNA. In particular, a genetic index associated with educational attainment should begin to decline before the palatial system disappeared. I tested this using EA4, an index that combines thousands of genetic variants associated with educational attainment in modern populations, and tracked how it changed in ancient Greek genomes before and after the collapse.

Read more

What surprised me later was how rarely I encountered that visual language. Modern population-genetics papers are full of PCA scatterplots, ancestry bars, and formal mixture models. Those are powerful tools and usually more explicit about uncertainty. Yet the red-green-blue maps that first captured my imagination seem to have all but disappeared from scientific articles and genetics blogs.

So I decided to reproduce the idea for Italy. The country is almost designed for such an experiment: a long north-south peninsula, two large islands, mountain barriers, and a position at the junction of continental Europe and the Mediterranean. Its history contains repeated movements of farmers, pastoralists, merchants, soldiers, enslaved people, colonists and refugees. If genetic geography can be painted anywhere, Italy should leave a vivid canvas.

Alongside this article, I am launching pca.davidepiffer.com, a growing interactive atlas of genetic geography. Italy is the first detailed entry, and a Europe-wide map provides the broader continental view. Readers can move between analyses, zoom into regions, inspect the geographic anchors beneath each map, and switch between the RGB composite and its constituent principal components.

The atlas is designed to grow across countries, regions and continents. Because a world PCA, a continental PCA and a country-scale PCA reveal different levels of structure, the site keeps them as separate coordinate systems. It identifies the reference panel, scaling and interpolation used for every map. The colors are therefore tools for exploration, not ancestry percentages, and each analysis includes downloadable figures, locality data and validation notes.

How the RGB maps were created

I used a principal-component analysis built from 918 unrelated comparison individuals: 690 Europeans and 228 people from Middle Eastern or Caucasus reference populations. The model used 3,113 shared autosomal SNPs. We held out all 348 Italian targets while fitting the axes, then projected them into that fixed coordinate system, so the Italian samples did not rotate the map around themselves.

Of the 348 projected Italian targets, 193 could be assigned to defensible geographic anchors: 129 using regional-capital coordinates under the published Raveane rule, 51 using individual sampling localities, and 13 Sicilians using four disclosed subregional representative anchors. The other 155 targets had only broad population or regional labels and were excluded from the spatial interpolation, although they remained in the PCA. I also mapped 28 Sardinian fitting references at one population-level anchor. The final map therefore represents 221 individuals—193 projected targets and 28 Sardinian references—at 66 geographic anchors. I averaged people assigned to the same locality before interpolation, preventing a heavily sampled city from exerting more spatial pull simply because it contributed more individuals. PC1 became red, PC2 green, and PC3 blue. Each channel was clipped at the 2.5th and 97.5th percentiles of the locality means and rescaled from zero to one.

The first map shows only the observed locality means. The continuous map uses inverse-distance weighting on a five-kilometre grid. The interpolation was limited to the same landmass as an observed point and to a maximum of 150 kilometres from the data. That keeps the sea from becoming an imaginary genetic bridge and leaves unsupported land grey. Full technical details appear in the appendix.

Results: a genetic landscape, not a genetic checkerboard

The observed localities already reveal the main result. Nearby places often have related hues, but not identical ones, and the peninsula changes progressively rather than breaking into neat blocks (see Figure 1). The broad geographic signal is measurable: across all pairs of localities, geographic distance and RGB distance have a Spearman correlation of 0.403. The median color distance is 0.255 for pairs within 100 kilometres and 0.349 for pairs separated by more than 300 kilometres.

That is strong enough to make geography visible, but weak enough to preserve surprise. The Alps, the Apennines, coastlines, islands, historic roads and uneven sampling all matter. So do drift and the fact that two people from one town need not have identical family histories. The point map is the honest foundation: it shows what was observed before a smooth surface encourages the eye to invent continuity.

Figure 1. From genotypes to colors: observed locality means.

The peninsula as a cline

When the locality means are interpolated, the peninsula resolves into a broad transition from greener and teal shades in much of the north to mauve, magenta and purple shades in the south (Figure 2). Central Italy is not a hard dividing line. It is a zone of overlap in which the two dominant gradients meet, with local deviations layered on top.

This north-south pattern agrees with earlier genome-wide studies of Italy and with the wider European observation that genetic variation often mirrors geography (Novembre et al., 2008; Raveane et al., 2019). But a cline is not a migrating population frozen in place. It is the accumulated covariance produced by many episodes of movement, mating, and isolation. The smooth background should therefore be read as a hypothesis about spatial continuity between observations, not as measured DNA at every pixel.

Figure 2. A smoothed RGB surface across modern Italy.

What each color channel contributes

The RGB map becomes easier to interpret when its channels are separated. These single-component maps do not reveal three pure ancestries. They show three perpendicular statistical directions chosen by this particular reference panel. Changing the panel or marker set can rotate the axes, especially beyond the strongest geographic direction.

PC1: the strongest peninsular gradient

PC1 increases strongly toward the south and east of the peninsula (Figure 3). At the locality level, its Spearman correlation with latitude is -0.815 (-0.822 when Sardinia and Sicily are excluded). In the wider reference panel, the positive end of this axis points toward the eastern and southern Mediterranean side of West Eurasian variation, while the negative end points toward northern and northeastern Europe.

That orientation makes the Italian pattern historically suggestive. Southern Italy and Sicily have repeatedly received gene flow through Aegean, Balkan, Anatolian and wider Mediterranean connections, while northern Italy has had stronger continental links across the Alps and the Po basin. Yet PC1 cannot tell us whether a given shift came from Neolithic farmers, Bronze Age movements, Greek colonization, Roman mobility or later migrations. Several processes can move populations in the same direction in PCA space.

Figure 3. PC1: a pronounced north-south and west-east gradient.

PC2: Sardinia and the western side of the map

PC2 supplies much of the green channel in the composite (Figure 4). It tends to be higher in the north and west and lower toward the south and east. Sardinia is the conspicuous extreme: its reference centroid has the highest PC2 value in this comparison panel, ahead of Basques, Balearic Islanders and Corsicans.

This is where ancestry history and drift become inseparable. Ancient DNA shows that Sardinia retained unusually high affinity to early western Mediterranean farmers and experienced long periods of relative isolation, followed by later Mediterranean gene flow (Chiang et al., 2018; Marcus et al., 2020). Isolation also magnified random allele-frequency change. In PCA, such drift can pull an island population away from a mainland cline even when the underlying ancestry is not exotic. The bright Sardinian color is therefore not a single migration signal; it is a composite of ancestry retention, later contacts, and drift.

Figure 4. PC2: a western-insular and north-south contrast.

PC3: local texture and a warning against overstory

PC3 adds the blue channel and is much less tightly aligned with latitude (Spearman rho = -0.248; Figure 5). It separates Sardinia strongly from the peninsula and contributes smaller patches of contrast within mainland Italy. In the wider panel, this axis distinguishes western and insular populations at one end from parts of eastern and northeastern Europe at the other.

Because the third component is weaker and more panel-dependent, it is tempting to turn every local patch into a story. That would be a mistake. Some texture may reflect real regional drift, mountain-valley isolation or historically localized gene flow; some may reflect sparse local sampling or the averaging of people at a coordinate. PC3 is most useful here as a reminder that Italy is not explained by a single north-south axis.

Figure 5. PC3: island separation and finer regional texture.

What migrations could have painted this landscape?

The safest historical reading begins with layers rather than labels. Present-day Italians, like other Europeans, descend in varying proportions from hunter-gatherers, early farmers ultimately connected to Anatolia, and later groups carrying steppe-related ancestry. Italy received those transformations unevenly. Maritime Neolithic routes, Alpine and Adriatic corridors, and different Bronze Age contacts left gradients that later populations inherited.

The southward PC1 shift is compatible with stronger affinities to southeastern Europe and the eastern Mediterranean documented in modern and ancient-DNA studies. That broad affinity may contain several chronologically distinct processes: Neolithic settlement, post-Neolithic movements related to the Aegean and Caucasus, Greek and other colonial networks, and the intense mobility of the Roman world. Ancient genomes from Rome show that the city became a genetic crossroads with substantial Mediterranean diversity; later central Italian transects show further transformation after antiquity (Antonio et al., 2019; Posth et al., 2021). None of these episodes owns a color channel, but together they provide plausible historical mechanisms for the direction of the cline.

Northern Italy’s cooler green-teal region is likewise not simply ‘more northern European.’ The Po Valley is a corridor as well as a boundary zone, and the Alps are crossed as well as isolating. Continental affinities can reflect Bronze Age and later movements from central Europe, while local Alpine and Apennine communities may acquire distinctive signals through endogamy and drift. A smooth continental gradient and sharp microregional differentiation can coexist.

Sardinia demonstrates why drift matters. A population can become genetically distinctive without a dramatic replacement: long-term small effective population size and reduced gene flow allow chance changes in allele frequencies to accumulate. Sicily illustrates the complementary process. Its position in the central Mediterranean made it repeatedly connected, so its colors are more naturally read as layered gene flow than as isolation alone. The two islands are both distinct, but for different demographic reasons.

What the colors say

Cavalli-Sforza’s great visual insight was that human genetic variation has a geography. When the major dimensions of genetic variation are turned into colors, Italy does not break neatly into regional blocks. Instead, gradients run across the peninsula, interrupted by islands and more isolated populations.

That pattern makes historical sense. Italy has always been both a corridor and a patchwork of local worlds. People moved along coasts, valleys and plains, while mountains and distance slowed movement and isolation allowed some communities to drift apart. Ports and cities repeatedly brought in people from elsewhere. The result is not a mosaic with hard borders but a landscape in which genetic similarity changes gradually across space, with a few conspicuous exceptions.

When I first saw Cavalli-Sforza’s maps as a teenager, I thought of them almost as maps of the past. Recreating them now, I see them somewhat differently. They are maps of relationships among people living today, but those relationships are themselves the product of history. Geography shaped who mixed with whom, migrations altered the gradients, and isolation preserved or exaggerated local differences. The RGB map compresses all of that into three colors.

It cannot tell us when those differences arose or which migrations produced them. For that we now have ancient DNA and explicit demographic models. But as a way of seeing the genetic geography of Italy at a glance, Cavalli-Sforza’s idea remains remarkably effective.

Technical appendix

PCA model and projection

The RGB figures use the fully held-out Europe plus Middle East/Caucasus sensitivity PCA because it contains PC1 through PC20 for the same 348 Italian targets. The fit comprised 690 unrelated European and 228 unrelated Middle Eastern/Caucasus references at 3,113 shared SNPs. Italian targets were projected after fitting. This differs from the later 20-replicate consensus export, which used 42,242 SNPs but retained only PC1 and PC2; a three-channel RGB map requires PC3.

Geographic units and special anchors

The mapping unit is the mean PCA coordinate at each distinct audited latitude-longitude pair. Of the 348 projected Italian targets, 193 were mappable at 65 target localities: 129 were assigned regional-capital coordinates under the published Raveane rule, 51 had individual sampling localities, and 13 Sicilians were assigned to four disclosed subregional representative anchors. The remaining 155 targets had only broad population or regional labels; they remained in the PCA but were excluded from spatial interpolation rather than being assigned invented point coordinates. The map also includes 28 Sardinian fitting references at one additional anchor, producing a total of 221 individuals at 66 geographic anchors. The Sicilian targets have source-defined western, central, eastern and southern labels but no individual coordinates; they are placed at representative anchors in Palermo, Caltanissetta, Catania and Agrigento. The Sardinians are not members of the 348-target export and all share the AADR/HGDP population coordinate 40° N, 9° E. Consequently, Sardinia is shown as one spatially uniform anchor and cannot reveal within-island structure.

RGB encoding

Stored PC signs were used verbatim: PC1 to red, PC2 to green and PC3 to blue. No manual sign flip was chosen to make the map resemble geography. For each component, locality means below the 2.5th percentile were clipped to zero, values above the 97.5th percentile were clipped to one, and intermediate values were linearly rescaled. This robust scaling improves contrast but means that RGB distances are visual distances within this map, not distances in the original PCA units.

Interpolation and validation

Inverse-distance weighting was applied separately to PC1, PC2 and PC3. Candidate powers 1.5 and 2.0 were compared by leave-one-locality-out prediction; 1.5 had the lower mean standardized RMSE across the three components. The surface uses a 5 km grid, is masked to Italian land, never crosses between disconnected land polygons, and is suppressed more than 150 km from the nearest observation. The realized maximum distance was 137.4 km. The relationship between geographic and RGB distance was moderate rather than deterministic (Spearman rho = 0.403).

Limits

The reference panel determines the PCA axes, and a different scope or marker set can rotate them. Geographic coordinates vary in precision; several are regional or population anchors rather than individual sampling locations. Interpolation creates visually plausible values where no genome was measured. Sample density is uneven. PCA itself is descriptive: historical claims require ancient DNA, explicit admixture models, dates and sensitivity analyses. For those reasons the interpretations above are hypotheses consistent with the maps and prior research, not causal assignments of a color to a migration.

References and further reading

Cavalli-Sforza, L.L., Menozzi, P. & Piazza, A. (1994). The History and Geography of Human Genes. Princeton University Press.

Novembre, J., Johnson, T., Bryc, K. et al. (2008). Genes mirror geography within Europe. Nature 456, 98-101.

Di Gaetano, C., Voglino, F., Guarrera, S. et al. (2012). An overview of the genetic structure within the Italian population from genome-wide data. PLOS ONE 7, e43759.

Fiorito, G., Di Gaetano, C., Guarrera, S. et al. (2016). The Italian genome reflects the history of Europe and the Mediterranean basin. European Journal of Human Genetics 24, 1056-1062.

Raveane, A., Aneli, S., Montinaro, F. et al. (2019). Population structure of modern-day Italians reveals patterns of ancient and archaic ancestries in Southern Europe. Science Advances 5, eaaw3492.

Chiang, C.W.K., Marcus, J.H., Sidore, C. et al. (2018). Genomic history of the Sardinian population. Nature Genetics 50, 1426-1434.

Marcus, J.H., Posth, C., Ringbauer, H. et al. (2020). Genetic history from the Middle Neolithic to present on the Mediterranean island of Sardinia. Nature Communications 11, 939.

Antonio, M.L., Gao, Z., Moots, H.M. et al. (2019). Ancient Rome: A genetic crossroads of Europe and the Mediterranean. Science 366, 708-714.

Posth, C., Zaro, V., Spyrou, M.A. et al. (2021). The origin and legacy of the Etruscans through a 2000-year archeogenomic time transect. Science Advances 7, eabi7673.

Page

The DNA points to a history between those two extremes. Most of the ancestry of modern Turks still comes from Anatolia and neighbouring West Asia, showing deep local continuity. Yet the Turkic-speaking arrivals left a genuine Central-Asian-related strand as well. Anatolia’s people were not swept away, but neither did the region simply change its language while its population stood still: migration became part of the transformation.

So the questions are straightforward, even if the history is not. How much of the ancestry of present-day Turks was already present in Anatolia before the medieval migrations? What arrived with Turkic-speaking newcomers from Central Asia, and when did those populations begin to mix? Was the process the same across Turkey, or did different regions inherit different versions of it? And can the genetic evidence distinguish a single conquest-era event from a longer history of movement, marriage and absorption? Those are the questions I will take up below.

Anatolia before the Turks

The Byzantine individuals in this analysis come from several Anatolian sites. They were not one genetically uniform population, just as the inhabitants of modern Turkey are not. This matters because any model that squeezes all Byzantine Anatolians into one box begins by simplifying the past.

Ancient DNA nevertheless shows substantial continuity in Anatolia from pre-Roman through Roman and Byzantine times. This was never an isolated land: Anatolia connected the Balkans, Caucasus, Levant and Mesopotamia, and migrants appear in its cemeteries. But the local population did not repeatedly vanish and restart. Turkic migration added a new layer to an already mixed population (Lazaridis et al., 2022; Koptekin et al., 2023).

I use “Central-Asian-related” deliberately. The ancient groups in the model stand in for parts of the migrants’ ancestry; they are not genomes of a single, pure “Turkic people.”

The evidence is uneven: the mixture date is much clearer than the ancestry percentage. Table 1 summarizes what the evidence supports.

Table 1. What the analyses do—and do not—tell us about the transformation of Anatolia.

An ancient glimpse of the transition

Read more

Italy makes this intuition almost unavoidable. Travel from the Alps toward Sicily and nobody needs a genetics paper to notice that average appearance changes. Yet sit in a café in Cuneo, Macerata or Cosenza and the overlap is enormous. A fair-haired Sicilian is not a paradox. A dark-haired Lombard is not an exception requiring explanation. The regional averages are real, but they are not rules for individuals.

The intuitive question is not really about regional averages or statistical correlations. It is personal: if I look at one individual—their complexion, hair and eyes—how much can I tell about where their ancestors came from?

At a very broad level, appearance plainly carries some information. If two strangers were drawn at random, one from northern Europe and one from the eastern Mediterranean, pigmentation might help us guess which was which more often than chance. But that is the easy version of the problem. Italy lets us divide the harder question in two. First, does pigmentation help distinguish a northern Italian from a southern Italian? Second, once we compare people only within northern, central or southern Italy, does the person with lighter pigmentation also tend to have more northern-shifted ancestry?

Those are not equivalent tests. A feature can differ reliably between population averages while being a poor guide to the ancestry of any particular person. Men are taller than women on average, yet height alone cannot identify every individual’s sex. In the same way, northern and southern populations can differ in average pigmentation while their individual distributions overlap heavily.

This study cannot literally test how accurately a stranger’s face reveals ancestry, because I have neither photographs nor measured pigmentation. It tests the closest genetic version available here: whether genome-wide ancestry coordinates predict genetic propensities for lighter skin and blonde hair. If the association is strong only when populations are pooled, appearance contains information mainly about broad group differences. If it remains equally strong within each population, it could also distinguish finer ancestry differences among individuals.

The distinction matters well beyond pigmentation. Human beings are natural ecological reasoners. We notice that two populations differ in ancestry and in a trait, and then silently assume that the same ancestry–trait slope operates inside each population. Sometimes it does. Sometimes it is much smaller. Sometimes it changes direction or moves onto another dimension of ancestry altogether.

I began with Italy, where the intuition is especially strong, and then widened the comparison. The aim was simple: test whether the appearance–ancestry relationship visible across regions survives when people are compared only with others from the same population.

That distinction turns out to matter. A pattern that looks obvious on a continental map may tell us much less about two neighbours. The rest of this post shows how I turned genomes into a genetic map, how I measured pigmentation without photographs, and what happened when I separated population averages from individual differences.

Read more

The history of modern Greeks is too often compressed into rival slogans: timeless continuity on one side, sweeping replacement on the other. Neither does justice to a country that preserved deep links with its ancient past while continuing to receive newcomers. One much-debated chapter began in the early Middle Ages, when Slavic-speaking groups moved south into the Balkans and settled in parts of Greece. Did they leave the same mark everywhere? How much of the older population endured, and where is later migration most visible today?

The answer is clear, but not tidy. Greece was neither replaced nor left untouched: different regions absorbed later migration to different degrees. The early-medieval legacy is clearest in the mainland samples and in Thessaloniki, smaller in the southern mainland, and weak or absent in Athens and the islands examined here. Modern Greeks also remain close to ancient people from Greece, although later migrations left their mark. There is no honest single “Slavic percentage” or one date for the whole country: the interesting answer lies in the differences from one region to another.

Four questions, in plain English

I break that larger story into four practical questions about Greece. Where is the early-medieval signal strongest? Can the published Balkans mixture recipe produce believable ancestry shares? Does any proposed date survive reasonable changes to the ancient comparison groups? And how does the distance between ancient and modern Greeks vary across time and place?

Read more

It is a ridiculous premise, which is exactly why it is memorable. It is also surprisingly easy to translate into a real hypothesis: if dogs substitute for children, countries with more dogs should have lower fertility. More importantly, increases in dogs should come before declines in fertility. And if the arrow runs the other way (if falling fertility makes dogs more attractive) then fertility should move first.

Brad Hargreaves made a very similar real-world argument in his July 2025 Thesis Driven essay “On Urban Dogs,” and in the tweet that circulated alongside it. He pointed to U.S. dog-owning households rising from 48.2 million in 2016 to almost 60 million in 2024, pet spending more than doubling, and fertility falling below replacement. His stronger claim was directional: young urban couples acquire a dog instead of having a child, and the dog’s later demands on money, space, and time can even help prevent a second child. That is a testable version of the Boss Baby hypothesis: not merely “dogs and fertility changed at the same time,” but “dogs came first and pushed fertility down” (Hargreaves, 2025).

The joke also has a real scholarly shadow. A recent Taiwanese working paper linking pet-registry records to tax records reports that adopting a dog was followed by a 33% increase in the probability of subsequent childbearing, while births also changed later dog adoption. This evidence is more compatible with dogs and babies as complements than substitutes (Chen et al., 2025). Other studies point toward substitution, but mostly for fertility intentions rather than completed births: stronger pet attachment was associated with lower fertility intentions among higher-SES respondents in a Chinese study (Guo et al., 2021), while interviews with childfree companion-animal owners and an ethnography of Israeli couples describe pets as “prechildren” or satisfying substitutes for some people (Laurent-Simpson, 2017; Shir-Vertesh, 2012). Those studies are important, but they concern intentions, narratives, or selected couples, not population-level fertility effects.

That leaves three competing mechanisms. Substitution is the Boss Baby story: dogs absorb money, time, care, or emotional investment that might otherwise support children. Complementarity is the opposite possibility: acquiring a dog may be part of a transition toward parenthood, a “practice child,” or a family lifestyle that later includes children. Reverse causality runs from family formation—or from declining fertility, childlessness, or weaker social networks—to dogs: households with children may be more likely to acquire or retain pets, as life-course research has long suggested (Albert and Bulcroft, 1988).

So: does the Boss Baby have a point?

Read more

Like many of you, I have a weakness for Neanderthal stories. They make evolution feel wonderfully close: a vanished cousin, a few fragments of DNA still carried in our genomes, and suddenly a possible explanation for our muscles, our pregnancies, our immune systems, or even the shape of our noses.

But that fascination also makes us easy targets for hype. And I am increasingly tired of journalists exploiting it, turning intriguing but often modest genetic associations into sweeping stories about what Neanderthals supposedly “gave” us.

That was my reaction to an MDR WISSEN / RR article from 9 August 2026, Muskeln, Kiefer, Schmerzen: Was der Neandertaler in unserem Körper hinterlassen hat (“Muscles, jaw, pain: what Neanderthals left in our bodies”). It begins with a new finding about a Neanderthal-derived growth-hormone receptor, then opens out into a much larger, irresistible catalogue of things Neanderthals may have left us.

Here is the story the article invites us to picture:

• A Neanderthal growth-hormone-receptor variant is linked to more muscle mass and is found in about 20% of South Asians, but only about 0.5% of Europeans (this was covered in my previous post)

• Nearly one in three women in Europe carry a Neanderthal progesterone-receptor version and have fewer early-pregnancy bleeds and miscarriages.

• Three major risk factors for Dupuytren’s contracture—the hand condition nicknamed “Viking disease”—come from Neanderthals.

• A chromosome-3 Neanderthal segment is a major early-COVID severity risk locus, appears in roughly one in six Europeans, and lowers HIV-infection risk by 27%.

• Other Neanderthal fragments improve defence against bacteria and fungi but raise allergy risk; make blood clot faster but raise stroke and pulmonary-embolism risk; and affect nasal height, fat metabolism and the roundness of the internal braincase.

That is a very good list of questions. It is also a very compressed way of telling several different kinds of genetic story. MDR directly links the new growth-hormone-receptor research; the rest appears as background in the article, without individual study citations. So I used the MDR piece as the doorway, then went back to the studies behind the claims.

I have already written about the first, GHR/muscle claim in Did the ‘Neanderthal Muscle Gene’ Really Turn Modern Humans into Muscle Heads? I will not repeat that analysis here. The short version is that the molecular and body-composition association is real, but it does not turn a carrier into a “muscle head,” explain athletic strength, or explain Neanderthal physique, or even differences between modern populations. The nine remaining claims are what I want to walk through here.

Before doing so, one piece of scale helps me keep my footing. In a large study of 96 traits in 291,273 unrelated White British UK Biobank participants, introgressed variants collectively explained an average of 0.12% of trait variation. That is not nothing. But it is a useful antidote to the idea that a surviving Neanderthal fragment is a miniature destiny (Wei et al. (2023)

The question I keep asking

When I read “a Neanderthal gene,” I pause. Usually the phrase means a stretch of introgressed DNA—or a modern marker that travels with it—not a single gene with one simple effect. The causal variant may still be uncertain. The population in which a finding was made may be very different from the population in which it is being repeated. And “white,” “South Asian,” and “Northern European” are not interchangeable genetic categories.

For each claim, I came back to three questions:

• Is this really an introgressed Neanderthal segment?

• What, exactly, did the study measure, and in whom?

• Does the evidence earn the big words: prevents, protects, makes, causes or adaptive?

Read more

That left an obvious regional question. What happens when the same three-source design is applied to the Dhaka Bengali panel, eleven Pakistan panels, including two Punjabi cohorts, and Ror from Haryana? These groups were outside the previous post’s primary India comparison or were raised by readers afterward.

This follow-up runs two linked tests. First, I transfer the India post’s three-source model—an Onge-based proxy for deep South Asian ancestry, the Indus-Periphery-West genome and 35 Bronze Age Steppe individuals—to 13 regional panels on one common SNP set, using the same comparison populations as before. For four panels with substantial East Asian-related ancestry, I also test a prespecified four-source extension. Second, I repeat the regional experiment after replacing Indus-Periphery-West with the Sarazm farmer-related source proposed by Kerdoncuff and colleagues.

The transferred three-source model rejects for all 13 regional panels, although an East-Asian-aware version fits the Bengali panel from Dhaka. The Sarazm alternative produces accepted models for Bengali, Kalash, two Punjabi panels and Ror, while most other Pakistani panels still reject. Below I report every pass and failure, and show how much the inferred Steppe-related proportions depend on the western source used.

Read more

A new genome-wide association study by Thorpe and colleagues identified 11 loci associated with delay discounting in 134,935 US participants who clustered with European genetic reference panels. Of nine genome-wide-significant signals supplied with the GWAS materials, six could be aligned and scored in the international genetic datasets used here. I asked a deliberately simple question: do countries with a higher delay-discounting polygenic score also tend to look less patient on independent behavioral, survey, and cultural measures?

At the descriptive level, every patience measure with enough country matches correlates with the genetic score in the predicted negative direction.

What the score means

A higher score means more weight on the six genetic variants associated with steeper delay discounting—the tendency to choose a sooner reward over a later one. I will call this the impatience direction for shorthand.

The score is deliberately small and transparent: it is a weighted sum of six genome-wide-significant variants from the discovery GWAS, rather than a large predictor tuned to maximize accuracy in a new sample. The discovery study was conducted in a European-ancestry sample, so the same score may not transport equally well to every population.

I estimated the score in pooled genotypes and in a 51-panel frequency resource. Where several represented populations contributed to one country, I used documented population shares where available rather than treating every panel as equally large. The result is 49 country-level estimates.

A few countries required special population-share reconstructions, including Israel and Russia.

Read the score, then, as a transparent geographic signal with substantial uncertainty.

The geography of the score

Figure 1 maps the standardized DD PGS across all 49 country-level estimates, and Figure 2 ranks them. In the steeper-discounting direction, Peru is the clear outlier (3.26 SD above the 49-country mean), followed by Bolivia, South Sudan, Kenya and Nigeria. At the other end are Taiwan, Vietnam, Japan, Tunisia and China. In the score’s own direction, the first group is genetically more immediate-reward oriented and the second group less so.

There is a broad continental pattern, but not a tidy continental league table. The highest values cluster in the Andean and sub-Saharan estimates; the lowest cluster in East Asia. Europe mostly sits nearer the middle, while the Americas are highly heterogeneous: Peru and Bolivia are at the high end, whereas the United States is close to the sample mean. North Africa and West Asia also span both sides of the distribution. The map is therefore more informative than a single continental average.

Figure 1. Delay-discounting PGS in the audited 49-country score sample

Figure 2 turns the map into a complete ranking: every one of the 49 country estimates is labeled and ordered from the highest to the lowest standardized DD PGS.

Figure 2. Standardized DD PGS for every country in the audited 49-country genetic sample.

Patience was measured in several different ways

The main behavioral benchmark is the Global Preferences Survey, which used standardized elicitation in representative samples from 76 countries. Its patience measure combines a staircase of hypothetical sooner-versus-later monetary choices with self-assessed willingness to wait (Falk et al., 2018; Sunde et al., 2022).

I also used the international compilation assembled by Rieger, Wang, and Hens (2021). It includes direct INTRA monetary choices, inferred long-run discount and present-bias parameters, a pace-of-life field index, World Values Survey long-term orientation, and GLOBE future-orientation practices and values. Their published UP-time index is itself a cross-study principal component. I report it as a benchmark but do not put it inside my own factor, which would double-count its component measures.

The distinction between direct and indirect measures is important. Choosing a later monetary payment is conceptually close to delay discounting. WVS and GLOBE measure broader cultural orientation. Agreement across them is interesting, but it does not mean they are interchangeable observations of one perfectly defined psychological trait.

The next question is whether the genetic score lines up with all nine measures, whether those measures can be summarized as one broader patience signal, and whether the pattern changes when we account for ancestry-related structure. Finally, I test whether this score is correlated to the educational attainment PGS.

Read more

No participant in the study was shown to have been turned into a “muscle head.” The result closest to that claim was 150 g more arm-and-leg lean mass per haplotype copy (95% CI 64–236 g) and 121 g more fat-free trunk mass (50–192 g)—about 271 g combined. Across five biobanks, adult carriers also averaged 0.30 cm taller and 285 g heavier per copy. Those are persuasive associations and relatively large effects for one common-variant signal. They are not a bodybuilding phenotype. The study measured no muscular appearance, grip strength, torque, power or athletic performance. [2]

The news story also changes a cell-count result into a cell-size claim. It says cells with the Neanderthal receptor “grew about 40% larger.” The paper reports approximately 39% more cells by day five, following a 6% increase in exponential growth rate—not cells that were 40% larger. Nor did the experiment show a statistically significant shift in apparent pituitary-growth-hormone sensitivity: the dose-response EC50 values were 0.07 and 0.06 nM (P=0.13). It showed increased signaling output and proliferation in a particular engineered cell system. That is biologically interesting, but it is a different result. [1,2]

Science’s visual framing turns a population-level association into an individual before-and-after-style contrast. Maričić is larger and carries one copy of the haplotype; Kanis is slighter and does not. But a staged comparison of two co-authors cannot separate genotype from height, age, training, diet, body fat, ancestry or the rest of their genomes. The evidence is the adjusted carrier–noncarrier comparison across cohorts, not which scientist looks more muscular in a photograph. The possibility suggested by the caption cannot be tested from the photograph itself. [1,2]

To its credit, the Science report quotes three researchers warning readers not to explain Neanderthal morphology with one variant. Those cautions are right; the headline, opening rhetoric, cell-size error and flexing photograph nevertheless push in the opposite direction. Correcting that framing does not require pretending the variant is trivial. The molecular effect is real, and the lean-mass and height associations are relatively large by single-SNP standards. What they do not provide is an explanation of strength, individual physique, population differences or the overall Neanderthal body plan. [1,2]

The mistake is not believing that the variant matters. It is confusing a relatively important SNP with an explanation of a highly polygenic phenotype.

But how important is “relatively important” for lean mass? The direct benchmark is Pei et al.’s 2020 genome-wide association study (GWAS) of 450,243 UK Biobank participants. [3] It measured appendicular lean mass—the estimated fat-free mass of the arms and legs—and selected 1,059 primary and conditional associations at a stringent genome-wide threshold. rs6184 was not among them. Yet its standardized point estimate was larger than 80.6% of that selected set, equivalent to rank 206 by magnitude. Its P value, 6.35×10⁻⁴, was far above the study’s threshold.

How can rs6184 look fairly large by single-SNP standards, fail genome-wide significance in the lean-mass GWAS and explain only a few thousandths of one percent of lean-mass variation? And why does the broad population score point in the opposite direction from the haplotype’s frequency? Resolving those apparent contradictions requires separating effect magnitude from statistical certainty, one allele from the polygenic background, and lean mass from actual strength. [3]

Read more

The difficult questions are how much they contributed, when they arrived, and why the answer differs so sharply among Indian groups.

Ancient DNA now gives a fairly clear broad answer. Ancestry related to Middle and Late Bronze Age Steppe pastoralists entered northern South Asia after the mature Indus period, probably during the second millennium BCE. It became widespread, but it is not dominant in any of the published Modern Indian Cline groups analyzed here. In those data, it ranges from only a few percent to roughly one quarter.

There is no single Indian “Aryan percentage.” There is a gradient shaped by geography, language, endogamy and social history.

What Counts as “Aryan” Ancestry Here?

In this article, “Aryan ancestry” refers operationally to ancestry related to Central Steppe populations of the Middle-to-Late Bronze Age (MLBA): the eastern Steppe genetic cluster associated with populations such as Sintashta and related groups. Archaeology and historical linguistics make populations from this broader horizon plausible carriers of early Indo-Iranian languages, while genetics establishes the underlying movement and mixture (Narasimhan et al., 2019).

The Steppe source was itself an admixed population assembled on the Eurasian Steppe. “Steppe-related” therefore denotes statistical genetic affinity to sampled ancient reference populations.

India Before the Steppe

Before this Steppe ancestry arrived, the subcontinent already contained deeply divergent populations.

One ancestry layer is conventionally called AASI, or Ancient Ancestral South Indian. No unadmixed AASI genome has yet been sampled. It is a statistically reconstructed lineage, and present-day Andamanese hunter-gatherers are used as a distant proxy in some models. This does not mean that mainland Indians descend from a migration out of the Andaman Islands.

A second layer is Iranian-related ancestry. Here too the name is easy to misunderstand. It does not mean ancestry from modern Iranians, and it is not identical to the ancestry of the sampled early farmers of western Iran. The direct Harappan genome from Rakhigarhi carried substantial ancestry related to an ancient lineage connected to Iran, but that lineage had split before the sampled Iranian hunter-gatherers, herders and farmers differentiated (Shinde et al., 2019).

These two deep layers had already mixed before the Steppe-related movements. Eleven Bronze Age people from Gonur and Shahr-i-Sokhta formed what Narasimhan and colleagues called the Indus Periphery Cline. They were not excavated inside the Indus Valley Civilization (IVC), but archaeology and genetic modeling connect them to the IVC sphere. The single usable genome from an actual Harappan burial at Rakhigarhi fell on the same broad gradient and showed no detectable Steppe ancestry. One individual cannot represent every Harappan city or community, but it supplies a crucial pre-Steppe anchor.

Earlier modern-genome research summarized the main ancestry gradient as mixture between two statistically reconstructed endpoints: Ancestral North Indians (ANI) and Ancestral South Indians (ASI) (Reich et al., 2009). These were model-based ancestral populations, not ancient individuals sampled directly or present-day groups. Ancient DNA later clarified their composition. The ASI endpoint is an AASI-rich mixture: it combines Indus-related ancestry—which already contained AASI-related and Iranian-related ancestry—with ancestry from groups carrying still more AASI-related ancestry. The ANI endpoint is also composite, combining Indus-related and Steppe-related ancestry (Narasimhan et al., 2019). Put simply, AASI is a deep ancestry lineage and one ingredient in ASI, whereas ASI is a later, AASI-rich mixture.

When the Steppe Signal Appears

The ancient sequence is more informative than a modern ancestry map alone.

In the supplementary data reanalyzed here, 11 deduplicated Indus Periphery individuals have date midpoints spanning approximately 3175–2056 BCE. The Central Steppe MLBA source group spans approximately 2036–1100 BCE. The later South Asian rows comprise 85 individuals from archaeological sites in the Swat Valley—a mountain valley in present-day Khyber Pakhtunkhwa, northwestern Pakistan—assigned to the study’s main Late Bronze and Iron Age cluster (SPGT), plus two genetic outliers (SPGT_o), with date midpoints spanning approximately 1263–808 BCE.

Figure 1 places those groups and sites in both geographic and chronological context. Panel A shows where the named locations lie; Panel B shows when the analyzed individuals lived.

Figure 1. Ancient geography and chronology behind the South Asian Steppe signal.

The sampled skeletons from Swat lived after the relevant gene flow began. In a separate analysis of 86 SPGT individuals, ancestry-covariance dating placed their Steppe-related mixture earlier, with an approximate interval of 1900–1500 BCE. Steppe ancestry also appears in outlying individuals in Central Asia before it becomes visible in the South Asian transect. Together, the chronology, geography and source matching support movement into South Asia during the second millennium BCE (Narasimhan et al., 2019).

Modern-genome linkage-disequilibrium estimates place major ANI–ASI-related mixture events roughly 1,900–4,200 years ago (Moorjani et al., 2013). Those dates do not independently date the first Steppe arrival: they average mixture between already complex ancestral populations, sometimes across more than one episode.

This is consistent with a route for Indo-Iranian languages, but it is not a genetic recording of speech. Nor does the evidence support treating the main population associated with the Bactria–Margiana Archaeological Complex (BMAC)—a Bronze Age urban cultural sphere centered in southern Central Asia—as the principal genetic source of present-day South Asians. BMAC lay along the relevant cultural corridor, but genetic and cultural transmission need not have identical sources (Narasimhan et al., 2019).

There Is No Single Indian Percentage

For the main modern comparison I reanalyzed the hierarchical qpAdm estimates published by Narasimhan and colleagues for 140 named groups on the Modern Indian Cline. My primary Indian subset contains 132 groups whose sampling location is an Indian state or union territory. Groups sampled in Pakistan, Nepal, Bangladesh, Myanmar, Iran or the United States were excluded from that primary subset, although they remain in the full South Asian figure and data table. All ranges, medians, maps and regressions in this section use those published hierarchical estimates; the separate Allen Ancient DNA Resource (AADR) reconstruction below is not appended to them.

The unit is a named sampled group—not an individual and not a population-weighted share of India.

Across all 140 cline groups, Central Steppe MLBA-related ancestry ranges from 2.7% to 27.4%, with an unweighted median of 10.9%. In the 132-group primary Indian subset, the range is 2.7% to 25.0% and the median is 10.3%.

Table 1 brings the main descriptive comparisons together: the full cline, the primary India subset, the two largest language families, and the workbook’s Brahmin/Bhumihar split. Each row summarizes named groups with equal weight.

Table 1. Unweighted Steppe-related ancestry summaries across named sampled groups. “Other label” means groups not marked Brahmin or Bhumihar in the source workbook; it is not a complete caste category.

Those medians are useful descriptions of the sampled groups. They are not estimates that “the average Indian is 10% Aryan.” The underlying samples do not mirror India’s population sizes, and the modeled cline deliberately excludes several groups with substantial East Asian-related, Austroasiatic-related or other atypical ancestry.

The first visual comparison is geographic. A conventional heatmap would imply estimates for unsampled territory, but the 132 groups occupy only 104 source-table coordinates, with dense sampling in some areas and large gaps in others. Figure 2 therefore uses a point heat map: color represents the median among groups assigned to the same coordinate, point size records how many groups share it, and no ancestry surface is interpolated.

Figure 2. Point heat map of Steppe-related ancestry at sampled Indian locations.

Even the lower end needs care. Eight southern Dravidian-speaking tribal groups could be modeled without Steppe ancestry in additional qpAdm tests reported by Narasimhan and colleagues. Their hierarchical point estimates in the main clinal model are slightly above zero because the joint model shrinks noisy group estimates toward the fitted gradient. A small positive point estimate is therefore not proof that every community received direct Steppe migrants.

The main result is not a national number but heterogeneity. The Steppe contribution is substantial in some northern and traditionally upper-status groups, modest across many other populations, and minimal or statistically unnecessary in some southern tribal groups. Across all 140 published cline point estimates, most ancestry nevertheless derives from older South Asian and Indus-related layers.

The map establishes the gradient. It does not tell us what drives it. Does the Indo-European–Dravidian difference survive once latitude and the Brahmin/Bhumihar label are considered together? Does the paternal signal tell the same story as autosomal ancestry? And can the published qpAdm pattern be recovered from AADR v66?

Below, I test those questions, compare autosomal and paternal signals, examine source-model sensitivity, and run a fresh qpAdm reconstruction. One of the apparent contrasts changes sharply after adjustment.

Read more

There is also a narrower possibility that economists rarely put into cross-country growth models: populations may differ slightly, on average, in inherited traits related to learning and human-capital accumulation. That does not mean genes determine the destiny of countries. Development depends on policy, health, migration, institutions, incentives, conflict, and chance. The question is whether education-linked genetic differences contribute any predictive information once some of that history is taken into account.

I tested that question in 38 countries with genetic estimates and comparable economic data from 1970 to 2019.

Why growth is a harder test than wealth

The genetic evidence begins with a genome-wide association study, usually shortened to GWAS. A GWAS scans genetic variants in a large sample and asks whether each variant is statistically associated with a trait. Here the trait is educational attainment—roughly, years of schooling completed. Each individual association is tiny, uncertain, and not usefully interpreted on its own.

A polygenic score, or PGS, combines many of those small estimated associations into one index. For each sampled population, I multiplied the education-GWAS weight at each variant by its estimated frequency and added the results. A PGS is a statistical predictor, not an education gene, a measure of human value, or a fixed forecast of anyone’s life. Its meaning is especially indirect when it is averaged across population samples and assigned to a country.

I used two published scores. EA3 comes from the 2018 study led by James Lee, which analyzed about 1.1 million people. EA4 is the 2022 successor led by Aysu Okbay, based on roughly 3 million people and including within-family analyses. I standardized EA3 and EA4 across the final country sample, averaged them equally, and standardized the average again. Equal weighting is simple and transparent. It also avoids pretending that the two studies are independent replications, because some participants and much of the underlying genetic signal overlap.

The outcome is not how rich a country is today. It is average annual growth in real GDP per person between 1970 and 2019: the change in log GDP per person, divided by 49 years and expressed as a percentage. In ordinary language, this measures how quickly each country’s average income rose over the period.

The first look at the data

The unadjusted pattern is positive. Across the 38 countries, the Pearson correlation between the composite score and annual real-GDP-per-person growth is 0.49. In a simple regression, a one-standard-deviation higher score is associated with 0.71 percentage points faster annual growth. The 95% uncertainty interval runs from 0.38 to 1.04 percentage points. Figure 1 shows the individual countries behind that line.

Figure 1. Raw education-PGS composite and 1970–2019 annual growth in real GDP per person across 38 countries.

Accounting for where countries started

The raw relationship is not the most informative test. Countries did not start 1970 at the same level of development. Poorer countries can often grow quickly by adopting machinery, knowledge, and organizational practices already used elsewhere. Economists call this convergence. Rich countries usually have less room for that kind of easy catch-up.

I therefore control for GDP per person in 1970. The comparison becomes: among countries beginning at roughly similar income levels, did the country with the higher education-linked score subsequently grow faster?

In this conditional-convergence model, the estimated association is 1.16 percentage points per score standard deviation, with a 95% interval from 0.76 to 1.57.

A simpler ancestry correction adds five broad ancestry-geography blocks. Its estimate is 0.66 percentage points, with a 95% interval from 0.25 to 1.07. Figure 2 introduces this adjusted relationship before any more elaborate genetic model: both axes remove starting-income and five-block differences. It is useful as an intuitive comparison, but broad categories cannot capture the fine-grained relatedness among populations.

Figure 2. Relationship after removing 1970 income and five broad ancestry-geography blocks from both axes.

Using the rest of the genome as a control

The hardest objection is genetic autocorrelation. Countries are not independent genetic units. Nearby and historically connected populations tend to resemble one another across the genome. They may also share institutions, migration histories, languages, and exposure to regional shocks. An education-PGS association could therefore be a disguised ancestry or history association.

The main robustness test uses 2,959 LD-pruned markers spread across the genome, with the education-score loci excluded. From their allele frequencies I constructed a country-by-country relatedness matrix. The model then allows the unexplained growth outcomes of genetically similar countries to be correlated. Instead of deciding that all members of a continent are equivalent, it uses the full pairwise pattern of genome-wide similarity.

In this genome-wide relatedness model, a one-standard-deviation higher within-block composite is associated with 0.62 percentage points faster annual growth in real GDP per person. The conditional 95% interval runs from 0.09 to 1.14. In 4,999 datasets simulated under a null model with no score effect, a result this strong appeared about 0.9% of the time.

That simulation is a demanding check, but it is not magic. It treats the country scores, allele frequencies, and relatedness matrix as known rather than propagating every layer of sampling error. It also cannot remove an omitted historical factor simply because that factor is correlated with ancestry in a complicated way. Figure 3 compares the main relatedness result with the five-block estimate and ancillary specifications. Its horizontal axis is the estimated association with annual growth, not a country-level genetic-similarity score.

Figure 3. Estimated annual-growth association across model specifications.

How a national genetic score is constructed

Turning population panels into national averages is an important limitation, not a clerical detail. A panel carrying a country label is rarely a random sample of the whole country. Whenever several identifiable groups represent a country and defensible demographic information exists, I weight the represented groups by documented population shares. When the available groups cover only part of the population, the primary rule normalizes their shares within that covered portion; where a broad residual proxy exists, it represents the documented remainder. The identical weights are used for the PGS and the genome-wide marker frequencies. The full component weights, coverage shares, sources, and sensitivity rules are preserved in the accompanying methods CSV.

The United States and China illustrate the approach. For the United States, the primary construction uses 1970 population shares and three imperfect genetic proxies: CEU for non-Hispanic White residents, ASW for non-Hispanic Black residents, and MXL for Hispanic or Latino residents. Those categories cover 98.82% of the 1970 population; within the represented share, their weights are 84.47%, 11.01%, and 4.52%, respectively.

For China, Han populations receive the overwhelming majority of the weight: the census Han share is 91.11%. Several Han panels are first combined into the Han component, while sampled minority panels receive their documented shares and the remaining unsampled minority share follows the stated residual rule. Across the five matched U.S. demographic scenarios, the main relatedness-model coefficient ranges from 0.59 to 0.62 percentage points.

These are still approximations. Census race and ethnicity are social self-identifications, not genetic ancestries. CEU, ASW, MXL, and the Chinese panels are narrow samples, not miniature national censuses. The same caveat applies elsewhere, even after demographic reweighting. A carefully documented national proxy is better than silently giving every available panel equal weight, but it does not become a probability sample.

What the secondary checks say

The five-block model is a readable secondary check; the genome-wide matrix is the main one. I also report principal-component specifications only as ancillary diagnostics. Principal components compress major axes of genetic variation into a few variables, but here they can remove variation in the education score itself as well as unrelated structure. With two PCs the estimate is 0.64; with four it is 0.31 percentage points. These results show sensitivity to how ancestry is parameterized, but I do not treat the PC models as more authoritative than the full relatedness model.

No single country determines the sign. Re-estimating the main model after omitting each country in turn leaves the coefficient between 0.41 and 0.68 percentage points. Every leave-one-out estimate remains positive. Yet the uncertainty interval crosses zero in 19 of the 38 runs; the estimate falls most when United States or Brazil is omitted. That check can detect a spectacularly influential country; it cannot detect a bias shared across many countries.

Catch-up, innovation, and institutions

A correlation is more informative when its pattern matches a plausible mechanism. Education-linked traits might matter most in poorer countries if they help populations absorb existing technologies. That is the technological catch-up hypothesis. Alternatively, they might matter most near the frontier, where growth depends more on research, entrepreneurship, and creating new technology.

I tested this by interacting the score with 1970 income. The interaction is negative at −0.28 percentage points, with a 95% interval from −0.76 to 0.20. Its sign leans toward a catch-up interpretation, in which the association is stronger among initially poorer countries. The interval includes zero, so the data do not clearly distinguish the two mechanisms. Figure 4 shows predicted growth across score values for countries at lower, median, and higher starting-income levels.

A second interaction asks whether stronger initial institutions let education-linked traits produce larger economic returns. Its estimate is −0.26 percentage points, with a 95% interval from −0.67 to 0.15. The interval includes zero, so this sample provides no clear evidence that the association is stronger under better institutions. With only this many countries, both interaction tests are imprecise and should be read as clues about mechanism rather than decisive findings.

Figure 4. Predicted annual real-GDP-per-person growth across education-PGS values at lower, median, and higher 1970 income.

Conclusion

The narrow result is that countries with higher EA PGS tended to experience faster annual growth in real GDP per person from 1970 to 2019. The association survives adjustment for starting income and a model in which genome-wide relatedness shapes the covariance of unexplained growth.

Sources and further reading

Lee, J.J., Wedow, R., Okbay, A. et al. (2018). Gene discovery and polygenic prediction from a genome-wide association study of educational attainment in 1.1 million individuals. Nature Genetics, 50, 1112-1121. https://doi.org/10.1038/s41588-018-0147-3

Okbay, A., Wu, Y., Wang, N. et al. (2022). Polygenic prediction of educational attainment within and between families from genome-wide association analyses in 3 million individuals. Nature Genetics, 54, 437-449. https://doi.org/10.1038/s41588-022-01016-z

The 1000 Genomes Project Consortium. (2015). A global reference for human genetic variation. Nature, 526, 68-74. https://doi.org/10.1038/nature15393

Feenstra, R.C., Inklaar, R. & Timmer, M.P. (2015). The Next Generation of the Penn World Table. American Economic Review, 105(10), 3150-3182. https://doi.org/10.1257/aer.20130954