"Are admixture calculators accurate?" is really three questions wearing one word, and the honest answer differs by question. A calculator can be repeatable, it can be faithful to real genetic structure, and it can be true as a statement about your ancestors. The tools score well, mixed, and poorly on those three — in that order — and knowing which regime a given number lives in is the entire skill of reading one.
Background on what the tools are doing: what an admixture calculator is.
Repeatability: high#
The same file through the same calculator returns the same numbers, and closely related inputs land close together. Chip differences nudge results — two testing companies' files for one person can disagree by a few points per component, because different marker sets survive — but the tools are deterministic machines, not horoscopes. If someone's percentages moved, the input moved.
Fidelity to structure: real, with conditions#
The big strokes are genuinely there. The components that dominate West Eurasian calculators track the three deep ancestry layers ancient DNA later proved — hunter-gatherer, early farmer and steppe-related — and a calculator's farmer-versus-steppe balance for a European genome broadly agrees with what formal methods recover. Distances between profiles mirror geography, the famous genes mirror geography result. On resemblance, the tools are honest instruments.
The conditions are the two structural biases. The calculator effect means fidelity is highest for populations well represented in the training references and degrades silently off the home field. And anchoring means the absolute numbers are calculator-specific: the same British genome runs ~44% or ~51% "north European" in two respected calculators. Within one tool, comparisons are sound; the numbers themselves are coordinates in that tool's space, not measurements of a shared quantity.
Truth about ancestors: mostly out of reach#
Here is where the word "accurate" quietly breaks. A percentage from an optimiser is a best fit, not a tested claim, and three specific failures follow:
- No error bars. A real 12% and a meaningless 12% print identically. Standard errors are not hiding in the interface; the method does not produce them.
- Trace components are noise until proven otherwise. Below a few percent, entries mostly reflect the optimiser distributing residual noise across available profiles. The famous sub-1% "exotic" component is the least trustworthy number on any results screen.
- No rejection. Feed a calculator references that exclude your real ancestry and it fits you confidently to what remains. The output looks identical to a good fit. Only the fit distance in coordinate tools even gestures at the problem, and it is a gap measure, not a test.
So "am I really 18% East Med?" has no answer within the tool that printed it. The component is an anchored construct; the number has no uncertainty attached; and nothing checked whether a model without it explains you just as well. Those are exactly the three things a formal method adds: qpAdm returns a weight with a standard error, a z-score that says whether the source is distinguishable from zero, and a p-value that can reject the model outright. The difference in kind — not degree — is the subject of is qpAdm worth it.
The practical calibration#
| Claim type | Example | Trust a calculator? |
|---|---|---|
| Resemblance, big strokes | "My genome leans farmer over steppe" | Yes — this is what they do |
| Relative comparison | "More Baltic than my cousin, same tool" | Yes, within one calculator |
| Absolute percentage | "I am 18% East Med" | Only as that calculator's coordinate |
| Trace component | "My 0.8% Siberian is real" | No — noise until tested |
| Fine splits | "North Sea vs Fennoscandian ratio" | No — correlated profiles trade freely |
| Descent | "I descend from population X" | Never — wrong tool for the claim |
Two free habits raise the ceiling considerably. Cross-check resemblance with a plain distance ranking — no mixing, no components, just nearest reference populations era by era — and look at the space itself in a PCA view so you can see whether "close" is crowded or lonely. Where a claim needs to survive argument, take it to a formal model. Everything short of that is exploration — which is what the calculators are accurate at.
Terms used here are defined in the glossary.
References#
- Lawson, D. J., van Dorp, L. & Falush, D. (2018). A tutorial on how not to over-interpret STRUCTURE and ADMIXTURE bar plots. Nature Communications, 9, 3258.
- Novembre, J. et al. (2008). Genes mirror geography within Europe. Nature, 456, 98–101.
- Alexander, D. H., Novembre, J. & Lange, K. (2009). Fast model-based estimation of ancestry in unrelated individuals. Genome Research, 19(9), 1655–1664.
