Type qpAdm into any forum search and a genre appears: the method is broken, the method is fine, the papers using it are all wrong, the critics misunderstand it. Both camps cite the same three audits. This post is the referee's version: what Harney 2021, Williams 2024 and Flegontova 2025 actually established, numbers rather than vibes, what survived, and what a careful user does differently now. Spoiler for the impatient: the estimator survived; several popular workflows did not; and one finding is about what a fitting model means rather than how to find it.
What the audits established#
The machinery is sound. Harney and colleagues' baseline validation still stands unrefuted: on clean simulated data, weights are unbiased (99.3% of estimates within three standard errors of truth), p-values are uniform when the model is true, and the estimates are robust to pseudohaploid data, uniform damage, random missingness and single-individual samples. Nobody in the criticism literature disputes this layer. The identity underneath is arithmetic, and the arithmetic works.
Model selection by p-value does not work. The same audit's most consequential number: among plausible candidates, the true model carries the best p-value in only 48% of cases. Every workflow that sweeps models and keeps the top p, which described a lot of forum practice and some published practice, was measured and found to be roughly a coin flip.
On simulated admixture graphs, unstratified screening fails. Flegontova and colleagues put false-discovery rates on whole protocols over random admixture-graph histories: 72.5 to 100% for proximal rotating screens, degrading further as data grows, largely because simple true models get rejected and the screen climbs to complex ones that fit (40% of the models their screens consistently rejected belonged to targets with no admixture at all). Temporally stratified (distal) protocols had medians of 16.4 to 31.2%, and corroboration by PCA and unsupervised ADMIXTURE brought them to about 0% in those simulations. Both halves of the study, side by side.
On simulated landscapes, the protocol stops mattering and the odds take over. The same study ran screens on stepping-stone landscapes, groups exchanging migrants with their neighbours continuously, and found no statistically significant differences in false-discovery rate between the protocols it compared. What mattered there was sampling density, model complexity, the conditions on admixture fractions and, above all, the pre-study odds: the ratio of plausible to implausible models among those tested. Screens built like typical published protocols had odds of about 0.18 for two-way models and 0.015 for four-way ones, and their false-discovery rates tracked the odds almost perfectly. The authors' remedy is to test a few hypotheses made plausible by evidence outside qpAdm, rather than to search widely and keep what passes (the arithmetic).
On a landscape, a fit is a position, not an event. There is generally no unique true decomposition of a target into a few discrete sources on a landscape. Admixture fractions land inside (0,1) mainly for sources placed symmetrically around the target, a small share of lopsided models with distant sources pass as well, and a feasible "60% A + 40% B" can summarise where a target sits without A-like and B-like groups ever having mixed. The authors conclude that protocols built to find one or a few best-fitting models are hardly suitable as evidence for long-distance migrations in exploratory studies.
Resolution has a floor. Williams and colleagues measured it: sources separated by FST under roughly 0.002 cannot be distinguished by any right set, and at Iron-Age-scale differentiation the true model is only ~22% of the plausible set. Fine-grained historical claims, this medieval population rather than its neighbour, sit at or below the floor more often than users want.
Single-winner claims overreach. Maier and colleagues found alternative topologies fitting as well as or better than 19 of 22 published admixture graphs: the same lesson at the graph scale. The family of fitting models is the honest result, not one winner.
What that means, and does not mean#
Read carefully, the criticism literature is a manual, not an obituary. None of it found the estimator biased; all of it found usage patterns with measurable error rates, plus one finding about meaning. The analogy that fits: nobody concluded thermometers are broken on discovering that waving one around the room does not measure the patient. The audits priced the workflows, and the priced-out ones (p-value ranking, unstratified rotation at scale, wide exploratory screens, fine splits below the FST floor) were exactly the convenient ones. The landscape finding is different in kind. It does not say qpAdm computes anything wrongly; it says a fitting model can be a good description without being a historical event.
What genuinely changed in careful practice: composite feasibility criteria replaced p-thresholds alone, cutting false positives more than fourfold in Williams' simulations of recent history (from 27.5% to 6.0%), though the false-discovery rate stayed near 60% there, which is the pre-study odds problem in miniature; temporal stratification hardened from advice into a rule for graph-like questions; right sets are built for the specific contrast rather than by piling on; corroboration by orthogonal methods, PCA and unsupervised ADMIXTURE, became part of the protocol; and pre-study odds became the first question to ask of any search, before its results.
Where that leaves a reader in 2026#
Trust a qpAdm result in proportion to its protocol and to the question it answers, and both are checkable from the report. Is it stratified? Where did the candidate sources come from, and how many were tried? Is the right set stated in full? Were simpler models given first refusal? Is acceptance composite? Are the alternatives that also passed reported? And is the claim a composition ("this much ancestry related to these sampled groups") or a migration story, which needs evidence from outside qpAdm? A result carrying all of that, the model record discipline, is as checkable as ancestry modelling gets. A result carrying none of it gives you nothing to check.
That conditional is also the honest description of how we run the method: candidate sources fixed for the customer's declared origins before any model runs, every published model temporally ordered and measured against one house bar, a person reviewing the candidates and approving what is published, the model record on screen, and the result presented as ancestry composition rather than as a migration. Not because the critics were wrong: because they were right, the fixes are known, and applying them is work someone has to actually do.
Terms used here are defined in the glossary.
References#
- 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.
- 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.
- Maier, R. et al. (2023). On the limits of fitting complex models of population history to f-statistics. eLife, 12, e85492.
Editorial note: revised on 24 September 2026. An earlier version reported the admixture-graph false-discovery rates as if they applied to qpAdm protocols in general and left out the same study's stepping-stone landscape results, including what they imply for reading a fitting model. This version adds them.
