qpAdm vs Global25: what each method can and cannot tell you

Two methods dominate amateur ancient-ancestry analysis, and they are constantly compared as if they were competing answers to one question. They are not. Global25 and qpAdm answer different questions, fail in different ways, and disagree for reasons that are usually informative rather than alarming.

This is the comparison written out properly: what each one computes, what a number from each one actually licenses you to say, and how to tell which one you need.

The one-sentence version#

Global25 measures position. qpAdm tests a hypothesis.

Global25 places your genome as a point in a 25-dimensional reference space and finds the weighted mixture of source populations whose combined point sits closest to yours. It will always return percentages.

qpAdm starts from a model you propose — this target, these sources, these outgroups — and asks whether the allele-frequency data are compatible with it. It can answer no.

That single difference drives everything below.

What Global25 actually does#

A Global25 coordinate is 25 numbers. Each one is your position along an axis of genetic variation derived from a large reference set. Comparing two coordinates means measuring the distance between two points.

A closest-population ranking is a straight Euclidean distance: square the difference in each of the 25 dimensions, add, take the square root, sort. An admixture model is a search — usually a Monte-Carlo fit in the nMonte tradition — for the combination of sources whose weighted average lands nearest your point. The leftover gap is reported as a fit distance.

Three properties follow, and all three are routinely misread:

1. It always succeeds. Give it a panel of sources with no historical relationship to you and it will still distribute 100% of the weight and report a fit distance. Nothing in the output says "these sources are wrong."
2. The panel does more work than the maths. Which populations you offer determines the answer far more than the fitting algorithm does. Two honest analysts with different panels get different breakdowns from the same coordinate, and both are "correct" arithmetic.
3. Nearby sources trade off freely. If two source populations sit close together in the reference space, the fit can move weight between them almost without penalty. Their individual percentages are much less stable than the total they share.

None of this makes Global25 unreliable. It makes it a descriptive instrument: fast, reproducible, excellent for exploration, and honest about position. It simply has no mechanism for saying no.

What qpAdm actually does#

qpAdm works from f-statistics — summaries of shared genetic drift between sets of populations — rather than from coordinates. You supply:

• a target (the genome being modelled),
• a left set: the candidate source populations,
• a right set: outgroups, used as reference points rather than as candidate ancestors.

The method then asks whether the pattern of shared drift between your target, your sources and those outgroups is consistent with the proposed mixture. The output is not just percentages. For each source you get a weight, a standard error and a z-score; for the model as a whole you get a p-value.

The p-value is the part with no Global25 equivalent. A low p-value means the data are not compatible with the model you proposed. That is a real answer — arguably the most valuable one the method produces — and it is why qpAdm results are what published ancient-DNA research is built on.

Two things about qpAdm that surprise people coming from coordinate tools:

• Rejection is the normal case. Most models you can think of will fail. A rejected model is the method working, not the tool malfunctioning.
• The right set matters as much as the left. Outgroups are what give the test its power to discriminate. A right set that is too small, or too closely related to your candidate sources, will accept almost anything you propose — the commonest way to produce a confident and meaningless result.

Side by side#

Why they disagree — and why that is usually fine#

When a Global25 fit and a qpAdm model disagree, the reason is almost always one of these:

Different questions. The coordinate fit found the nearest combination available in your panel. qpAdm asked whether a specific ancestry story is statistically supportable. "Nearest available" and "statistically supportable" are not the same property, and neither implies the other.

The panel contained a proxy. Coordinate methods happily use a population that sits near the real source without being it. The fit closes; the history is wrong. qpAdm tends to expose this because the proxy's drift pattern relative to the outgroups differs from the true source's.

Resolution. Twenty-five dimensions compress a great deal. Two populations that are genuinely distinct in allele-frequency terms can sit almost on top of each other in coordinate space, so the coordinate fit cannot tell them apart and qpAdm can.

One of them was run badly. A coordinate fit against an anachronistic panel, or a qpAdm model on a thin right set, will produce confident nonsense. Rule this out before reaching for an interesting explanation.

Which one do you want?#

Use Global25 when you want to explore quickly, rank your closest populations, compare yourself against many groups, see how your position shifts across eras, or produce something visual. It is instant, it is reproducible, and for orientation it is genuinely the better instrument. You can do all of that free in our G25 distance calculator and admixture calculator.

Use qpAdm when you have a specific claim you want tested — that a population descends from these particular sources in these proportions — and you need an answer that can come back negative. It is also what you want if you intend to cite the result anywhere that formal standards apply.

Use both when you are doing this seriously. The productive workflow is to explore with coordinates and test with qpAdm: use Global25 to generate candidate sources cheaply, then put the resulting hypothesis in front of a method that is able to refuse it.

Two honest caveats#

Coverage sets the ceiling for both. Neither method can recover information a raw DNA file never contained. Standard errors in qpAdm depend far more on how many markers survive the merge than on how long anyone searched for a model, and a sparse file simply cannot reach the tightest statistical bars. Our free raw DNA file check reports what a file actually holds.

Neither identifies an ancestor. Both work with population averages and statistical patterns. A close distance or a well-fitting weight is evidence about populations, never about individuals, and no output from either method licenses the phrase "your ancestor".

Where we stand#

We sell both, which is the reason this comparison exists rather than a case for one of them. Our Global25 analysis is a coordinate product: distances, admixture and PCA across six eras. Our qpAdm analysis is formal modelling against the Allen Ancient DNA Resource, published with its p-value, per-source standard errors and z-scores, and the complete right set every model was run against — because a weight without those three numbers cannot be evaluated by anyone — and, since August 2026, the complete model record behind each era (chi-square, degrees of freedom, confidence intervals, the nested-model table, the rank test) with a written analyst explanation, downloadable as plain text. Global25 has no such record because it computes no such test.

If you want to try the formal machinery yourself first, the AdmixTools 2 Lab runs real f-statistics, qpWave and qpAdm on a reference panel, free, in the browser. Expect rejections. That is the point.

Unfamiliar terms are defined in the glossary.