Learn qpAdm: the complete guide, from zero to defensible models

qpAdm has a strange learning curve: the software is free, the papers are public, and yet the road from curiosity to a defensible model is undocumented enough that most people learn it from forum folklore. This page is the road, drawn properly — every piece we have published on the method, ordered so each one stands on the last, with what it teaches and when to read it. Bookmark it, walk it at your own pace, and skip nothing in stage two.

Stage one: what qpAdm is (read these first)#

• What a qpAdm ancestry test actually is — the buyer's-level orientation: what the method claims, what it returns, what no method returns.
• Understanding qpAdm — the method itself, for readers who want the mathematics: the admixture identity, the least-squares solution, the jackknife.
• f4-statistics explained — the arithmetic everything rests on. If you read only one technical piece, read this one; qpAdm is f4-statistics arranged into a testable model, and every later confusion traces back here.
• How qpAdm changed ancient DNA — the history: one 2015 supplement to the field's workhorse, and the decade of stress-testing that wrote its operating manual.

Stage two: the discipline (where models are won and lost)#

• How to choose sources and outgroups — the left set and the right set, the drift rule, temporal logic. The single most consequential page in the sequence.
• The standard right sets — O9 and its extensions, the arithmetic floor, the measured ceiling, and the published case where one added outgroup cut standard errors threefold.
• qpWave explained — counting ancestry streams before naming them; the rank test that referees every model.
• Distal vs proximal models — the protocol choice with measured false-discovery rates attached: 16–31% stratified, 72–100% not.
• Rotation explained — what screening is for, and the bill it runs when it picks winners.
• qpAdm best practices — the checklist that compresses this whole stage into something you can hold while working.

Stage three: hands on#

• Data preparation: from raw file to AADR merge — formats, builds, strand hygiene, convertf and Poseidon; the unglamorous half of every real analysis.
• The AADR's sample labels — what .SG, .DG and .HO mean and how to pick samples without poisoning your statistics.
• Running qpAdm in R with ADMIXTOOLS 2 — the working tutorial: install, extract, run, read.
• f2 extraction and the parameters that change results — allsnps, fudge, blgsize and friends, with the measured stakes of each.
• Classic ADMIXTOOLS vs ADMIXTOOLS 2 — why your numbers differ from a 2016 paper's, and how to match them when you must.
• The free AdmixTools 2 Lab — everything above, runnable in a browser on the public panel while you learn.

Stage four: reading results like an analyst#

• How to read p-values, standard errors and z-scores — the three numbers, in the order that matters.
• Why models get rejected — the good failures and what each one teaches.
• Troubleshooting the weird outputs — negative weights, SE 9.99, everything-passes and everything-fails, decoded.
• The model record explained — every number in a full record, and a worked report walkthrough beside it.
• How many SNPs qpAdm needs — the information budget that sets your error bars before anyone models anything.

Stage five: applying it to real ancestries#

• Recipes, with sources and right sets from the published literature: European · South Asian · Middle Eastern — and the worked tutorial on a real file.
• The neighbouring instruments, for the questions qpAdm hands off: qpGraph for whole-history topologies, f4-ratios for one-number ancestry fractions, admixture dating for when the mixing happened, and qpAdm vs the ADMIXTURE software for where description ends and testing begins.
• The honest audit to finish on: Is qpAdm still reliable? — what the criticism literature established, and how practice absorbed it.

The shortcuts, stated honestly#

If you want the results without the pipeline, that fork exists at every stage: the paid analysis is stages two through four done by hand on your genotypes merged into AADR v66 — roughly four hundred composed runs behind a typical order, published against one bar (p above 0.05, every |Z| above 3, every SE below 0.10) with the full record downloadable. The Model Lab then hands you stage three on your own merged dataset. And if you are still deciding whether qpAdm is your instrument at all, qpAdm vs Global25 and is qpAdm worth it are the two forks in the road, mapped.

Terms used here are defined in the glossary.