Distal vs proximal qpAdm: what temporal order protects against, and what it cannot

Two qpAdm models of the same Iron Age genome can both pass every numeric bar and rest on very different assumptions. Model one explains the target from deep-time ancestries: hunter-gatherers, first farmers, steppe pastoralists, all comfortably older than the target. Model two explains it from near-contemporary neighbours, populations a few centuries removed, some possibly younger. The literature calls these distal and proximal models, and the largest simulation study of qpAdm protocols so far, Flegontova et al. 2025, measured how much the difference matters.

The answer depends on what kind of history is being modelled, which is the part this article used to leave out. The study simulated two kinds: histories shaped like admixture graphs, and continuous stepping-stone landscapes. Temporal order was the biggest lever in the first and was not enough in the second.

Definitions first. A protocol is distal when a target "postdates (or is as old as) all its proxy sources", which is what temporal stratification means for the left set, and proximal when at least one proxy source is younger than the target, or more loosely when no stratification is enforced at all.

Two kinds of history#

An admixture graph is the familiar picture: discrete populations that split, drift apart and occasionally mix in pulses, so that every target is a mixture of branches you could in principle sample. A stepping-stone landscape is the other extreme: a grid of local groups, each exchanging migrants with its neighbours generation after generation, so that ancestry flows continuously across space and no small set of discrete source populations exists at all.

Real regions and periods can look like either, or like both at once. The authors are explicit that the common assumption that human history is approximated by complex admixture graphs "remains unproven". Every rule in this article holds in one of these worlds and needs qualifying in the other.

On admixture-graph histories, order is the biggest lever#

Flegontova and colleagues simulated qpAdm screens over random admixture-graph histories where the truth was known, and scored how often the models a screen accepted were false (their Fig. 3):

Three findings sharpen the point. More data made proximal screens worse: their false-discovery rate grew significantly with the amount of data, because more data rejects models more efficiently, including true simple models whose assumptions are quietly violated, and a rejected simple model sends the screen on to more complex ones that often fit. Much of the damage is that false rejection of simple models: of the models their screens consistently rejected, 40% belonged to targets with no admixture in their history at all. And corroboration worked for distal models: validating distal results with PCA and unsupervised ADMIXTURE brought their false-discovery rate to about 0% in these simulations.

Two cautions come with these numbers, and the authors give both. The false-discovery rate varied enormously from one simulated topology to the next, with many topologies at 0% or 100%, so they write that it is hard to predict whether real histories are prone to false admixture signals. And any false-discovery rate depends on the pre-study odds, the ratio of plausible to implausible models among those tested, which is a property of the model set rather than of the method. The table ranks protocols inside one simulated world. It is not an error rate for real studies.

On stepping-stone landscapes, order is not enough#

The same study then ran screens on simulated two-dimensional landscapes, with groups sampled at several time points, and the picture changed. Across the qpAdm protocols compared there, the false-discovery rate showed no statistically significant differences (their Fig. 7 and supplementary figures 24 to 26): the protocols traded false positives against false negatives, and the differences mostly cancelled out. What did move the false-discovery rate was the density of sampling across the landscape, which the authors call probably the single most important predictor of qpAdm performance there, the complexity of the models, the pre-study odds, and the conditions put on the estimated admixture fractions.

Temporal order does not rescue a landscape screen. In one systematic design where every source was older than the target, a distal protocol throughout, the false-discovery rate was 0% for two-way models, 40% for three-way models and 98.4% for six-way models (supplementary table 13). The protocol did not change between those rows. The odds did: plausible models to implausible ones fell from about 1 to 7, to 1 to 27, to 1 to 8,300 as the models grew. Rotation explained walks through why the odds, more than the protocol, set the error rate.

Why order helps on one and not the other#

On an admixture graph, complexity accumulates with time: every admixture event adds to the histories of everything downstream of it. Sources older than the target therefore tend to carry simpler histories than the target itself, which keeps them clear of the gene flows that break qpAdm's assumptions, such as a right population that received ancestry from a source after the two separated. Proximal sources can carry more admixture than the target, and that is where the violations concentrate.

A landscape has no such arrow. A group sampled three hundred generations ago was exchanging migrants with its neighbours exactly as later groups do, so an older source is not automatically a cleaner one, and every population is connected to every other through the groups in between. There the questions that matter are geometric: how densely the landscape around the target is sampled, whether the candidate sources sit around the target or off to one side, and how far the right set sits from the left. The paper found that a small average distance between the left and right sets lowers p-values and pushes screens toward more complex models, because shared drift between nearby left and right groups is exactly what qpAdm's test reads as misfit.

What a fitting model means on a landscape#

On a landscape there is generally no unique true decomposition of a target into a few discrete sources. Models lie on what the authors call an "optimality continuum", and qpAdm's estimates follow geometry. Admixture fractions land inside (0,1) and track real gene flow mainly when the sources sit symmetrically around the target, while a small but important share of lopsided models with distant sources also receive plausible fractions, which is how false long-distance migrations appear. p-values, by contrast, relate only weakly to how close or how symmetric the sources are.

So a feasible model such as "60% A + 40% B" can be a useful low-dimensional summary of where a target sits between sampled groups without populations resembling A and B ever having mixed. The paper spells out three readings of the same result, a two-way model that fits while every alternative of its size is rejected: a true model of a relatively simple graph-like history; a suboptimal model on a poorly sampled landscape; or an optimal but overly simple model on a densely sampled one. The output alone does not say which of the three you are looking at. The history around it does.

What a modern genome modelled from ancient sources is, and is not#

A modern genotype file as target with ancient sources is distal by construction: every source predates the target. That is worth having, because it protects against the class of violations the admixture-graph simulations exposed. It is not a guarantee. If the history of a region behaves more like a landscape than a graph, the protocol class stops being what matters, and sampling density, the candidate pool and the acceptance criteria take over.

That is why the honest reading of a consumer qpAdm model is compositional. "42% Anatolian-farmer-related" says your genome is modelled as sharing that much ancestry with the sampled Anatolian farmers, relative to the reference populations used: similar ancestry composition, not descent from those particular people and not a dated migration. A proximal "42% medieval Population X" sounds even more like history, and needs even more of that caution.

Proximal models still have a proper job. When the question itself is recent ("does this early-medieval genome need a Slavic-related component on top of the local Iron Age base?"), a proximal model is a focused hypothesis, worth testing one at a time with era-appropriate outgroups and with the simpler nested models given first refusal. A handful of hypotheses chosen for reasons outside qpAdm keeps the pre-study odds high; an automated sweep across every recent population keeps them low, whatever its protocol.

How our reports use this#

Every model we publish is temporally ordered and measured against one house bar in every era, and the report says per era whether it is met: p above 0.05, every source's |Z| above 3 and every standard error below 0.10, with every weight inside (0,1). That bar is our own convention, not a criterion the audit literature validated. One property is worth stating: with every weight positive, |Z| above 3 places each weight more than three standard errors from both 0 and 1, which implies the paper's stricter condition on admixture fractions, that each estimate plus or minus two standard errors lies inside (0,1).

Candidate sources are fixed before any model runs: a pool curated for the customer's declared origins and era. Our analyst tests models drawn from that pool and re-tests the survivors on alternative right sets and against their simpler nested models, and a person reviews the candidates, chooses the model to publish and approves it. The pool is our main lever on pre-study odds. The search inside it is large, which is why a published model is best read as one adequate description of a genome relative to the sampled groups, not as proof that those groups mixed.

The worked example shows the sequence on a real file, and the Ancestry Workbench lets you probe proximal hypotheses yourself on your own merged dataset, with the distal model as the anchor it should be.

Terms used here are defined in the glossary.

References#

• Flegontova, O. et al. (2025). Performance of qpAdm-based screens for genetic admixture on graph-shaped histories and stepping stone landscapes. Genetics, 230(1), iyaf047.
• Harney, É., Patterson, N., Reich, D. & Wakeley, J. (2021). Assessing the performance of qpAdm. Genetics, 217(4), iyaa045.
• Williams, M. P. et al. (2024). Testing times: disentangling admixture histories in recent and complex demographies using ancient DNA. Genetics, 228(1), iyae110.

Editorial note: revised on 24 September 2026. An earlier version presented false-discovery rates measured on simulated admixture-graph histories as if they applied to qpAdm protocols in general, and left out the same study's results on stepping-stone landscapes. This version separates the two and describes our publish bar as our own convention.