Search for "qpAdm test" and you will find two very different things sold under one name: a formal population-genetics method that can reject its own answer, and a growing number of products that print percentages with a scientific label attached. This guide is for someone deciding whether to buy one, and it is written from the inside — we sell a qpAdm analysis ourselves — so it tries to be exact about what the method can and cannot do rather than persuasive.
What qpAdm is, in one paragraph#
qpAdm is a method from the ADMIXTOOLS package, developed in David Reich's laboratory and used in most published ancient-DNA admixture studies since 2015. It takes a target — here, your genome — and asks whether it can be explained as a mixture of a chosen set of ancient source populations, measured against a set of distant outgroups. It returns a weight for each source, a standard error on each weight, and a single p-value for the whole model. Crucially, the p-value can say no: the proposed mixture is not compatible with the data. A method that can fail is what separates a test from an estimate. The maths is set out in Understanding qpAdm; the reading guide for the three numbers is How to read qpAdm results.
What happens to your file#
A qpAdm ancestry test starts with a raw-data export from a consumer testing company — 23andMe,
AncestryDNA, MyHeritage, FamilyTreeDNA and similar, as .txt, .csv, .zip or .gz up to
50 MB. Whole-genome sequencing customers can upload a .vcf or .vcf.gz up to 1 GB instead, via
the €10 Whole-genome upload add-on; the file is converted to the reference panel's markers at
upload, lifted from GRCh38 to GRCh37 where needed, and the VCF itself is not stored. The details are
in Upload a whole-genome VCF.
Your genotypes are then merged with the reference panel — the Allen Ancient DNA Resource (AADR) v66, roughly 23,265 samples across roughly 6,015 population labels — so that every ancient sample and your own file are read at the same positions. The number of positions that survive that merge is your coverage, and it is the single most important property of your file. Coverage bounds how small a standard error can get, and no amount of analyst effort can shrink an error the file has already fixed. Before paying, you can see your own file's coverage verdict for free in the Raw DNA File Check.
What the report contains#
An honest qpAdm report is mostly numbers you can check, arranged so that a reader who understands the method could reproduce the result. Ours publishes, for each era:
- the model: which source populations the target was modelled as a mixture of;
- the weights, the share of ancestry assigned to each source;
- a standard error and a z-score per source, never rounded away;
- the model's p-value;
- the complete right set — every outgroup the model was run against, with a "show all" control rather than a truncated sample.
Since August 2026 the report's Reading view goes further, and publishes the complete record of the run behind each era: the chi-square and degrees of freedom the p-value is computed from, the f4 rank, the minimum SNP count per f4 statistic and the merged SNP count, the jackknife block count, each source's panel label, sample count and 95% confidence interval, each outgroup's sample count, the nested-model table (every simpler model, its refitted weights, p-value and feasibility) and the rank test, plus the tool's own warnings and the run id. Beside it sits Reading your model, a written explanation from the analyst who built the model of why your genome resolved into these sources in these proportions, and what each label does and does not mean. The whole record downloads as a plain-text file, free, at every tier; every number in it is read in The model record explained.
That last point about the right set is the tell. A qpAdm weight without its outgroup list cannot be evaluated by anyone, because outgroup choice is where models are won or lost. If a product prints percentages and calls them qpAdm but shows no right set, no standard errors and no p-value, you are looking at a percentage generator with a borrowed name.
Ours models two eras — Hunter-Gatherer & Neolithic Farmer, and Classical Antiquity — from a curated set of 47 source populations across the two, each with its own page in the ancestry population directory, and covers every sovereign state: 199 countries and 1,115 regions. The report also carries maps and rendered videos; the numbers are the product, the rest is presentation.
The publish bar, stated once#
Every model we publish, at every tier, must clear the same numeric bar: p > 0.05, and for every source in every era, |Z| > 3 and a standard error below 0.10. A model that misses on any one of those is not published, whatever it looks like. This is why some files cannot be given a deep model at all: if coverage fixes the standard errors above the bar, the honest answer is to say so before purchase rather than to print the numbers anyway.
What the four tiers buy — and what they do not#
Our qpAdm analysis is sold at four tiers: €29.99, €39.99, €49.99 and €59.99. It is important to understand what changes between them, because the natural assumption is wrong.
The tiers do not buy a different standard. The publish bar above is identical at every tier. They do not buy more content. The report, its maps and its videos are the same at every tier.
What the tiers buy is search effort: how far the analyst keeps going past the first model that clears the bar. At the base tier the search is focused; at the top it ends only when nothing beats the model on the table. A longer search means more candidate models composed and run, more alternative outgroup sets to re-test them against, and more nested models checked so that no source is carried that does not earn its place. Deeper tiers therefore buy certainty, not complexity — the hardest-won models are often the simplest ones.
If your question is "will the top tier give me a lower standard error?", the answer is no. Standard error is set by coverage. If your question is "will it give me a more defensible model?", the answer is yes, because more of the alternatives will have been tried and rejected.
Every model is composed by hand#
Automated model rotation — trying combinations of sources and outgroups until one passes — is the obvious way to scale a qpAdm product, and it is a trap. Run enough models and some will clear any threshold by chance; the best-scoring model is frequently not the right one. Harney and colleagues (2021, Genetics) documented how easily qpAdm accepts wrong models when sources and outgroups are chosen carelessly. We built automated rotation, measured its false-discovery rate, and removed it. Every published model is composed, run and audited by one person, and that is what the price pays for.
What a result never means#
Three sentences that belong in every buyer's head before the report opens:
- A source population is a statistical proxy, not a family. Being modelled as 45% "Western Steppe Herders" means your genome is compatible with drawing that share of ancestry from a population like that one — it does not say those particular buried people were related to you.
- A passing model is "not refuted", never "confirmed". Several contradictory models can pass; that is why the search matters and why the right set is published.
- Your result describes your genome, not your identity. It says nothing about nationality, language or belonging.
After the report: running it yourself#
Once a report is published you can unlock the Model Lab for €10 — one unlock, one time — which lets you run your own qpAdm models on your own merged sample, choosing sources and outgroups from the panel, up to 100 runs per rolling 24 hours, and download the exact EIGENSTRAT bundle the report was computed from. Expect rejections; they are the method working. The full walkthrough is in Run your own qpAdm models and on the Model Lab page. If we later find a better model for a published report, it is offered as a second version for €15 — the original stays exactly as published, and you switch between them in the report. Optional add-ons at checkout are €10 each: Fast compute, the paternal (Y-DNA) haplogroup and the maternal (mtDNA) haplogroup.
How it compares with a coordinate test#
If you already hold Global25 coordinates, a Global25 analysis at €29.99 answers a different question with a different instrument: it fits your coordinate row to a mixture and always returns percentages, with a fit distance and no p-value. It cannot reject a model; qpAdm can. The comparison is set out in qpAdm vs Global25, and the terms used throughout this guide are defined in the glossary. If you want to try the method before buying anything, the free AdmixTools 2 Lab runs real qpAdm over the public reference panel in your browser — not on your own sample, but enough to learn how a rejection feels.
A checklist before you buy#
- Does the product show a p-value, a standard error and z-score per source, and the complete right set? If not, it is not a qpAdm result you can check.
- Can you download the full record — chi-square, degrees of freedom, confidence intervals, the nested-model table — as a file you can hand to another analyst?
- Does it tell you what your file's coverage supports before you pay?
- Does it say what the tiers change — and is the answer "effort", not "a different bar"?
- Does it claim any population is your ancestor? It should not.
- Does it name and version its reference panel?
If those five are answered, you are buying a formal model. Ours is at /qpadm. Ready to order? The buying page lists the tiers and what each includes.
References#
- Haak, W. et al. (2015). Massive migration from the steppe was a source for Indo-European languages in Europe. Nature, 522, 207–211.
- 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.
- Maier, R. et al. (2023). On the limits of fitting complex models of population history to f-statistics. eLife, 12, e85492.
- Mallick, S. et al. (2024). The Allen Ancient DNA Resource (AADR): a curated compendium of ancient human genomes. Scientific Data, 11, 182.
