Methodology deep-dives, population spotlights, and engineering notes from the Ancestrify team, qpAdm modelling, ancient DNA panels, and how genetic ancestry analysis really works.
41 stories
A first-person answer to "is Ancestrify legit": the registered trader behind it, what each of the three analyses delivers and costs, where the data lives, how refunds work, what it never claims, and how to verify every statement yourself.
A structured learning path through everything qpAdm — what to read in what order, from the f4-statistics underneath to running models in R, choosing outgroups, reading results and knowing the method's measured limits.
The definitive working checklist for qpAdm in 2026 — temporal stratification, right-set construction, lowest-rank-first search, composite feasibility and reporting standards — each rule carrying its measured justification from the 2021–2025 audit literature.
qpAdm has been audited harder than any tool in ancient DNA: measured false-discovery rates, resolution floors, protocol failures. What the 2021–2025 criticism literature actually established, what survived it, and how practice changed.
qpAdm says how much; linkage-disequilibrium decay says when. How DATES and ALDER read generation counts out of chromosome fragment lengths, what the dates mean, and how a date corroborates or breaks a qpAdm model.
Before qpAdm there was the f4-ratio — one number, two f4-statistics, an ancestry proportion. How the classic estimator works, the famous results built on it, and the precise trade against qpAdm.
qpAdm's bigger sibling models whole population histories as trees with admixture edges. How qpGraph works, what find_graphs automates, and the 2023 finding that reshaped how graph results should be read.
Same model, different p-value: the real differences between original qpAdm and ADMIXTOOLS 2 — allsnps semantics, fudge_twice, f2 precomputation — and the settings that reproduce classic behaviour when you need to.
The suffixes on AADR population labels encode how each genome was produced — capture, shotgun, diploid, array — and mixing them carelessly biases f-statistics. What each label means and the selection rules that keep models honest.
The settings tutorials skip and reviewers ask about: extract_f2 versus genotype input, allsnps, maxmiss, fudge and fudge_twice, blgsize, boot, afprod and constrained — what each does, with the measured stakes.
The undocumented half of every qpAdm analysis: file formats, genome builds, strand hygiene, convertf and Poseidon, and the merge arithmetic that decides your standard errors before any model runs.
f2, f3, f4 and D-statistics are the shared-drift arithmetic beneath qpAdm, qpWave and admixture graphs. What each statistic measures, how a four-population test works, and how to read Z-scores like the papers do.
From genotype files to a tested model: f2 extraction, qpadm() and its output tables, the arguments that silently change results (allsnps, fudge_twice, constrained), and the protocol discipline the code will not enforce for you.
Negative weights, SE 9.99, every model rejected, every model passing, infeasible popdrop rows, allsnps confusion — the standard failure gallery of qpAdm runs and what each symptom actually indicates.
qpAdm's sibling asks a prior question — how many independent streams of ancestry does a set of populations need? The rank test, cladality checks, and why every good qpAdm search starts with a qpWave answer.
The canonical nine outgroups from Lazaridis 2016, the o9a and o9aamcn extensions, the arithmetic floor and the ~30-population ceiling, and the published example where one added outgroup cut standard errors threefold.
The rotating protocol tests every candidate as source and outgroup in turn — elegant, recommended by the method's auditors, and carrying a measured 72–100% false-discovery rate when run without temporal discipline. What rotation is actually for.
Whether your sources predate your target is not a style preference — measured false-discovery rates run 16–31% for temporally stratified protocols and 72–100% for proximal rotating screens. What each model type is for.
The number that governs a qpAdm model is never your file's marker count — it is the intersection with the ancient panel, per statistic. The published coverage numbers, the allsnps table, and what a consumer chip can honestly support.
ADMIXTURE describes structure with K inferred components; qpAdm tests an explicit historical model and can reject it. What each estimates, why papers use both, and which answers which of your questions.
The fit distance measures how far a mixture still sits from your coordinate — not whether the model is right. Calibrated bands for reading one, why lower stops being better, and the checks that actually validate a model.
Every Global25 row exists in two forms, and mixing them is the hobby's most common silent error. What scaling does mathematically, which form each tool expects, and the tell-tale signs of a mixed comparison.
Adding sources always lowers a G25 fit distance and routinely worsens the model — a worked demonstration where the fit improved from 0.0110 to 0.0068 while the story collapsed. The rules that keep a panel honest.
Global25 is a projection, not a test — so 'accuracy' has layers: the coordinate's own fidelity, the references it is compared against, and the models built on top. An honest audit of all three.
Ger Huijbregts's nMonte turned Global25 rows into ancestry percentages and founded a whole tool tradition. How the Monte-Carlo search works, what nMonte3 changed, what the fit distance is, and the discipline the script never enforces.
What a Global25 PCA projection shows, what the axes are and are not, why plot distance lies, the difference between projecting onto a fixed view and fitting your own — and the honest uses of both.
What a Global25 distance ranking actually measures, why era choice changes everything, how close is 'close', and the four misreadings that turn a good tool into a wrong conclusion.
Every admixture calculator — GEDmatch's classics, Global25 fits, testing-company estimates — is an optimiser that cannot say no. How the three families work, what the percentages mean, and the questions a calculator can and cannot answer.
Accuracy has three different meanings for an ancestry calculator, and the tools do well on exactly one of them. Where percentages are trustworthy, where they are noise, and how to tell which regime you are in.
The same genome scores 44% North European in one calculator and 51% in another. Component anchoring, reference panels, the calculator effect and K explain the spread — and none of the numbers is 'the real one'.
The program behind the bar plots in population-genetics papers: how ADMIXTURE estimates K components jointly from all samples, supervised versus unsupervised runs, choosing K, and why hobbyist calculators are its frozen shadows.
A step-by-step qpAdm tutorial: check your raw file, merge it into AADR v66, choose sources and outgroups, run it in the browser or in R, and read the result.
What an admixture calculator optimises, why it always answers, what a qpAdm model adds, when each is the right tool, and when qpAdm is not worth paying for.
How to choose qpAdm left and right populations: the classic outgroups, the drift rule, the rank test, temporal logic, proxy labels, rotation and a worked example.
What a qpAdm p-value below 0.05 does and does not mean, the five usual causes of a rejection, negative weights, low Z-scores, what to try next and when to stop.
An illustrative qpAdm report for one era, read from the model line through the weights, p-value, right set, nested models and rank test to the explanation.
What the full qpAdm model record in an Ancestrify report means — chi-square, degrees of freedom, f4 rank, SNP counts, jackknife blocks, 95% confidence intervals, the nested-model table and the rank test — and how to read the plain-text download.
A plain reading guide to the three numbers in every qpAdm result — what the p-value tests, what a standard error bounds, what a Z-score rules out — with worked examples and the mistakes that make a passing model wrong.
Global25 fits your coordinate to a mixture and always returns percentages. qpAdm tests a model against allele-frequency statistics and can reject it. A practical comparison of when each one is the right instrument.
What the 25 numbers mean, why scaled and unscaled forms must never be mixed, how a closest-population ranking is computed, what a fit distance does and does not tell you, and the failure modes that make confident results wrong.
The method, for readers who want the maths: what qpAdm computes, what the p-value, standard error and z-score each mean, why outgroup choice decides whether a model is worth anything, and how to read a rejection — with a worked example.