Bagley ran the ancient-DNA test. Here is what the resolution actually is.

Disclosure: True News is a Bagley Enterprises platform, and this story is about the company's owner, Len Bagley, and his independent research. It was reviewed under the same standards as every story we publish. Bagley's findings are presented as his own and attributed to him.

In the first two parts of this series, True News documented what the DNA companies told Len Bagley. Living DNA called ancient-DNA tools "community-assembled resources" that "are not peer-reviewed instruments." AncestryDNA said such tools "often show high 'noise' or coincidental matches." Neither company said it had run the comparison he asked for.

So Bagley ran a version of it himself, using the tools the ancient-DNA field itself uses. This is what he found.

How the test was built

Bagley designed and ran the testing using open, published software: ADMIXTOOLS 2 and PLINK, run against the Allen Ancient DNA Resource (AADR v66.p1, 1240K format). The base dataset contained 23,089 individuals before his was added.

His raw AncestryDNA file held about 677,876 genotyped positions. After matching it to the ancient dataset and excluding 36,264 strand-ambiguous sites, about 380,000 remained. Those 380,000 markers set the limit on everything that follows.

The pipeline had to earn its credibility before it produced any result about him. Bagley set two controls. A modern French individual modeled as a descendant of the 12,700-year-old Anzick-1 sample had to be rejected, and it was, at p ≈ 3×10⁻¹⁸⁴ to 3×10⁻²⁵². A known Yoruba individual modeled as Yoruba had to fit, and under the final configuration it did, at p = 0.606.

It did not pass on the first try, and the failures are in the record. An early merge script produced impossible statistics, with f3 scores between +172 and +206. A self-consistency test caught them, and Bagley rebuilt the merge with standard tools. An intermediate configuration then rejected a known Yoruba individual as Yoruba (p = 0.0009). His paper says a pipeline that rejects African ancestry in a known African "cannot be trusted to detect it in anyone," and that the discovery invalidated every earlier configuration's apparent "low African" signal. He fixed the blind spot by adding a deeply diverged African outgroup. The final outgroup set was Mbuti, Ju|'hoan North, Dinka, Han, Papuan and Chimp.

Only then was any result about Bagley interpreted.

The result

The model asked how much of Bagley's ancestry fits three sources: French for European, Anzick-1 for Indigenous American and Yoruba for West African.

• Bagley (model p = 0.988): French 23.5% ± 7.8%; Anzick-1 (Indigenous) 77.7% ± 35.2%; Yoruba (West African) −1.2% ± 32.1%.

• Second subject, a collaborator (model p = 0.303): French 14.5% ± 12.2%; Anzick-1 (Indigenous) 109.0% ± 67.1%; Yoruba (West African) −23.4% ± 58.2%.

The model fits both subjects. What it cannot do is pin down the weights. Bagley's West African estimate carries an uncertainty of ±32.1 points, a 95 percent interval of roughly −64 to +62 percent. The second subject's model returned an Indigenous weight of 109 percent and a West African weight of −23 percent. Those are mathematical artifacts, not biological quantities, and they are what an underpowered model looks like.

The direct f4 tests, which ask whether a person shares more DNA with a candidate source than a European anchor does, agree. Yoruba scored +0.064. Anzick-1 scored −0.805, Kennewick +0.439 and Sumidouro −1.86. None reached the conventional threshold of |Z| = 3.

Bagley's paper states what this does and does not show. The data does not support a confident "African" determination, and it does not support a confident "Indigenous" one. "That symmetry is itself the finding."

The wall

Three limits explain why:

• Array density. The consumer array captures a fixed set of markers, and about 380,000 usable ones is not enough to separate deeply related ancestry components in a single person. No reprocessing can add information the array never captured. Whole-genome sequencing is the one change that would tighten the error bars.

• Missing references. The available ancient American genomes, such as Anzick-1 and Kennewick, are deep, western samples. Ancient DNA from the Eastern Woodlands and the Southeast, where the documentary record points, is scarce and largely unavailable to researchers.

• One genome. A single person's genome carries limited information for separating correlated deep ancestry.

Bagley also reports a choice he declined to make. Adding extra related populations to the model would have shrunk the error bars on paper. He says that precision would have been an artifact of a mis-specified model, not real information, so he published the wide interval.

What it means for the companies' claims

The result does not settle where Bagley's family comes from. It shows that the instrument sold as settling it cannot.

Here is the contrast. A controlled pipeline with explicit controls, run on array data from AncestryDNA's own kit, returned an uncertainty of tens of percentage points. The commercial reports built on comparable data return 85 or 87 percent, sometimes to the decimal, and name five peoples. Neither carries an interval.

Either the companies have a method that resolves what a controlled f-statistic analysis could not, in which case Bagley says they should publish it with its validation, or their stated precision is not supported by their data. They have so far declined to publish how their panels were built.

Put together, Bagley says, commercial products return high African percentages while the match geography stays American. When the same question is put to an open qpAdm and f4 design with Anzick-1, Kennewick and Yoruba, the result is non-significant and underpowered. The percentage should not be treated as a settled continental verdict until a transparent comparison against Southeastern and Eastern Woodlands ancient samples is run and published.

What he is demanding

• Publish the interval. Report an uncertainty range with every percentage, on the consumer report itself.

• Publish the method. Release how each reference group was built, who was excluded, and the validation behind the named peoples.

• Allow "undetermined." Let the software say so when a person's DNA fits no group well.

• Sequence the question. Run the comparison on whole-genome data, against ancient DNA from the Eastern Woodlands and the American Southeast, with the method published.

• Test it against recent African arrivals. Compare match rates for Black Americans with documented roots before 1900 against those of African-born testers who arrived in the United States after 1990, on the same products. Bagley says the second group finds continental relatives and his family does not, and he wants the companies to show the comparison.

True News is asking AncestryDNA and Living DNA to comment on this analysis. Any responses will be added here.

Bagley's research is available through his Academia profile, linked with this story.

References

• Mallick, S. et al. "The Allen Ancient DNA Resource (AADR): A curated compendium of ancient human genomes." Scientific Data 11 (2024).

• Maier, R. et al. "ADMIXTOOLS 2: A Fast and Flexible R Package for Population Genetic Analyses." (2023).

• Chang, C.C. et al. "Second-generation PLINK: rising to the challenge of larger and richer datasets." GigaScience 4 (2015).

• Patterson, N. et al. "Ancient Admixture in Human History." Genetics 192 (2012).

Originally published by True News: https://truenews.bagleyenterprises.net/stories/29