Phylogeography studies how genetic lineages are distributed across places and what those distributions can tell us about population history. Its gold standard is a historical explanation that survives competing explanations, appropriate sampling and uncertainty checks. A coloured map of DNA groups is a starting observation; it does not, by itself, reveal where a species originated or which route its ancestors travelled.
To read a phylogeographic study well, follow five questions: what was sampled, what part of the genome was measured, what genealogy was inferred, how geography and time entered the model, and which alternative histories remain possible? This guide develops those questions with original teaching cases. All numerical cases are invented and deliberately simplified; none reports a real species distribution or recommends moving organisms.
THE GOLD STANDARD · PHYLOGEOGRAPHY
Test histories written in genes and geography
Choose a route through lineage histories, sampling, competing explanations and the limits of geographic reconstruction.
ROUTE 1 · CHAPTERS 1–3
Read histories carefully
Separate lineage evidence from geographic stories.
ROUTE 3 · CHAPTERS 7–10
Compare explanations
Test time, contact and alternative histories.
ROUTE 4 · CHAPTERS 11–14
Repair weak inference
Check models, locations and reproducibility.
ROUTE 5 · CHAPTERS 15–17
Transfer and continue
Try a changed-context case and learning route.
Full chapter index · Try the transfer task · Gold Standard subject hub
Chapter index
Read histories carefully · Chapters 1–3
Build the evidence · Chapters 4–6
Compare explanations · Chapters 7–10
Repair weak inference · Chapters 11–14
Transfer and continue · Chapters 15–17
Imagine a freshwater animal occurring in two neighbouring drainage basins. Individuals in the western basin share several genetic variants, while eastern individuals carry others. One possible explanation is long isolation after a drainage divide formed. Another is colonisation from different source populations. A third involves older separation followed by contact, with some genetic exchange surviving today.
These explanations make different claims about events, not merely different descriptions of the same map. A useful question is therefore specific: did separation precede the present drainage divide, or did the divide plausibly initiate it? Answering requires information about genealogical history and geography at the relevant times. Present basin boundaries alone cannot date biological separation.
Write down the proposed event, organisms, places, time interval and observable consequences before choosing software. For example: “If these populations remained isolated after a recent barrier formed, we expect the genetic evidence to be compatible with that timing under an explicit demographic model.” This is more testable than “the barrier explains the genetics.”
The broader Biogeography root concerns distributions of life, environmental conditions and historical processes. Phylogeography contributes a particular form of evidence: relationships among sampled genetic lineages. Its strength comes from combining that evidence with geography without pretending either one is a complete historical record.
CHAPTER 2 OF 17 · Read histories carefully
2. Separate three objects that look similar on a map
A sampling map records where specimens were collected. A genetic map summarises where measured variants or inferred groups occur in that sample. A historical reconstruction proposes where ancestors lived or how lineages changed location. These are three different products with different uncertainties.
A specimen collected at a coastal market may have originated inland. A museum label naming a district may refer to a broad collecting area. A present population may descend from organisms transported by people. Before interpreting coordinates as biological geography, establish what each location means. Collection, residence, birthplace and ancestral origin are not interchangeable.
Likewise, a genetic cluster is a result of an analytical representation. Its boundary may shift when sampling fills a geographic gap, when more loci are used, or when the assumed number of groups changes. A sharp colour change on a map can conceal a gradual transition in allele frequencies.
For a practical audit, place three headings beside any figure: observed, estimated and assumed. Under observed, list specimens, dates and sequence measurements. Under estimated, list trees, clusters and ancestral locations. Under assumed, list mutation models, geographic states and demographic histories. If an arrow crosses these categories without explanation, the claim needs repair.
CHAPTER 3 OF 17 · Read histories carefully
3. Follow inheritance before following a migration arrow
Genes are transmitted through reproduction, and their ancestry can be traced backward through copies in earlier generations. The history of one genomic region is not automatically the history of every individual or every population. Recombination allows different regions to have different ancestral histories. A population also contains multiple lineages rather than one representative ancestor.
This matters when two populations share a variant. They may share it because genetic exchange occurred after separation. They may instead have inherited it from an ancestral population and retained it. Similarity does not uniquely identify recent contact; difference does not uniquely identify ancient isolation.
The Population Genetics root develops mutation, drift, migration and population structure. Here those processes become competing historical explanations. Drift can remove variants without a geographic movement. Migration can introduce a lineage without replacing the whole population. Mutation supplies changes, but the time represented by those changes depends on the rate model and genealogy.
A good reading habit is to replace “these animals came from there” with a more precise provisional statement: “this sampled genomic region has a relationship compatible with ancestry shared with those samples.” The stronger movement claim requires additional evidence. Precision at this stage prevents an attractive map from becoming a story that the data never actually tested.
Sampling should make rival histories distinguishable. If the question concerns separation by a mountain range, sampling only two distant towns leaves distance, habitat and the mountain confounded. Intermediate sites, alternative passes and neighbouring regions may be more informative than collecting many more specimens from the original towns.
There is no universal number of samples that makes a phylogeographic analysis reliable. The relevant quantity depends on geographic variation, relatedness, marker information, the expected frequency of variants and the question. Sampling ten siblings can represent much less independent information about a population than sampling ten unrelated individuals across the relevant habitat.
An original planning case makes the tradeoff concrete. A team can sequence 60 specimens. Plan A samples 30 at each end of a valley. Plan B samples ten at each of six sites, including the proposed contact zone. Plan A estimates variation at its endpoints more intensively; Plan B observes spatial transitions more directly. Neither is universally superior. If the main question concerns a contact zone, Plan B addresses information that Plan A cannot supply at all.
Record unsuccessful visits and inaccessible regions, not only successful collections. Otherwise absence of samples can be mistaken for absence of organisms. Sampling effort is part of the observation process and belongs beside the biological data. A later analyst needs to know whether a blank region was searched, excluded or never visited.
Worked design check: missing a variant is possible
Suppose a variant truly occurs in 10% of independently sampled individuals in a defined population. Assume random sampling, perfect measurement and a constant frequency. The chance that one sample lacks it is 0.9. The chance that all ten independent samples lack it is 0.9 raised to the power ten, approximately 0.349, or 34.9%. Thus a sample of ten can miss a variant that is genuinely present at that frequency.
With 30 independent samples, the corresponding probability is 0.9 raised to the power thirty, approximately 0.0424, or 4.24%. The probability of detecting it at least once is therefore about 95.76%. These are consequences of the stated model, not a recommendation that every locality needs 30 specimens. If the variant is rarer, samples are related, measurement is imperfect or accessibility biases collection, the calculation changes.
This case teaches two distinct lessons. First, an observed zero is not automatically biological absence. Second, increasing the sample size helps only with the uncertainty represented by the model. Thirty specimens from an unrepresentative family or a different habitat do not become a random sample merely because thirty was a convenient target.
When reporting an absence, retain the denominator and the sampling context: “not observed among these ten specimens from these sites” communicates the result more accurately than “absent from the region.” If absence is central to a proposed historical barrier, assess how plausible non-detection remains under the alternative history. The design should test the contrast that carries the conclusion.
CHAPTER 5 OF 17 · Build the evidence
5. Understand what a marker can and cannot represent
A marker is a measured genomic region or set of variants. Its inheritance, mutation behaviour and relationship to other markers determine what historical information it can supply. An organellar genome and many independent nuclear regions need not tell identical stories because their transmission and genealogical histories differ.
Do not count thousands of tightly linked positions as thousands of independent historical witnesses. More positions can improve reconstruction of one genealogy while adding little independence among genealogies. Conversely, short independent regions can contain little signal individually. Both sequence information and genomic independence matter.
Before analysis, specify whether the aim is identifying variants, reconstructing a genealogy, estimating population parameters or comparing historical models. A marker set chosen for identification may not contain the variation needed to estimate recent demographic change. A panel selected because variants were common in one region may underrepresent variation elsewhere.
Quality control should include specimen identity, missing data patterns, alignment decisions and possible contamination. If low-quality samples mostly come from one region, technical differences can imitate geographic structure. Plot quality measures by locality before interpreting biological differences. The first repair may be laboratory or data-processing work rather than a more elaborate evolutionary model.
The practical question is not “how many markers?” alone. Ask which histories they can represent, how independently they represent them, and whether their selection favoured one part of the geographic comparison.
CHAPTER 6 OF 17 · Build the evidence
6. Read a genealogy without turning it into a travel diary
A genealogy represents ancestral relationships among sampled genetic copies. A network can be useful when relationships are uncertain or when a single branching depiction hides alternatives. Neither diagram directly records the journeys of organisms. Branch lengths also need a legend: they may represent substitutions, time, or a different quantity.
Consider four artificial sequences, each ten positions long. A differs from B at one position; A differs from C at three positions; B differs from C at two. Those pairwise differences describe the sequences. They do not alone determine the location of their common ancestor, direction of dispersal or date of population separation.
A small calculation with a large limitation
If two sequences differ at 4 of 200 comparable sites, the observed proportion of differences is 4 ÷ 200 = 0.02, or 2%. This is an observed sequence distance, not automatically a corrected number of substitutions. The same site can change more than once, and different sites may evolve differently. Missing or ambiguously aligned positions also require explicit handling.
The denominator matters. Four differences among 200 valid comparisons is not the same measurement as four differences among 150. Keep the aligned positions, exclusion rule and missing-data treatment with the result. A polished tree without those details is harder to audit than a simple calculation with a transparent denominator.
CHAPTER 7 OF 17 · Compare explanations
7. Distinguish gene coalescence from population separation
Looking backward in time, two sampled genetic lineages coalesce when their ancestry meets in a common ancestral copy. That event need not coincide with the time their populations separated. Under a simple isolation model, their common ancestral copy can predate the population split, sometimes substantially.
This distinction is essential when comparing genetic dates with geological events. If a gene genealogy is older than a new channel or ridge, that does not automatically refute a population split associated with the channel or ridge. The ancestral population may already have contained divergent lineages before the geographic separation.
Dunn Lab’s explanation of gene trees and species trees describes how ancestral variation can persist across splits and yield genealogies different from the population or species branching history. It also distinguishes biological discordance from errors in estimating a tree. Collecting more sequence for one region can reduce estimation error without removing real differences among genomic histories.
For a reading exercise, label two dates separately: the estimated common ancestry of the sampled gene copies and the estimated population separation. If the study reports only the first, do not silently substitute it for the second. If migration is allowed after separation, ask how that changes the relation between the two. The meaning of the date comes from the model, not from the calendar unit alone.
CHAPTER 8 OF 17 · Compare explanations
8. Use molecular clocks as conditional models
A molecular clock relates genetic change to elapsed time through a rate model and calibration information. It is not a universal stopwatch. Rates can differ among genomic regions, lineages and estimation contexts; calibration uncertainty must be considered alongside uncertainty in genetic change.
Original arithmetic example
Assume two lineages diverged simultaneously, each accumulated changes at a constant rate r, and a corrected genetic distance d represents change along both branches. Ignoring ancestral variation and other complications, d = 2rt. If d = 0.02 substitutions per site and r = 0.00000001 substitutions per site per year, t = 0.02 ÷ 0.00000002 = 1,000,000 years.
The factor of two appears because both descendant branches accumulate change. If the supplied rate already describes pairwise divergence, multiplying it by two again would be wrong. Always read the rate definition and units before substituting numbers.
Now retain d but use twice the single-lineage rate. The calculated time becomes 500,000 years. Halve the rate and it becomes 2,000,000 years. This sensitivity exercise is not a confidence interval; it merely demonstrates how strongly the answer depends on an assumed input.
Real analysis must address the assumptions omitted here, including ancestral genealogical depth, uncertainty in calibration, variation in rates and the sequence model. A precise decimal output does not compensate for uncertain calibration. Present the conditional calculation separately from any historical conclusion.
CHAPTER 9 OF 17 · Compare explanations
9. Compare isolation, expansion and secondary contact
Isolation means restricted exchange over a specified period. Expansion concerns changes in population size or occupied range, which are related but distinct. Secondary contact means populations that were previously separated came into contact again. A single present-day pattern may be compatible with more than one of these histories.
A teaching case involves a western and eastern lineage meeting near a river. A simple explanation is that the river separates them. Another is that they expanded from different areas and met near it. Another involves differential survival after earlier contact. The location of the meeting point does not determine which process produced it.
Write predictions before viewing every result. Would isolation predict differences across many independent loci? Would secondary contact predict a spatial transition in ancestry? What observations would distinguish recent movement from retained ancestral variation? The appropriate tests depend on the organism and model; a checklist cannot substitute for a designed comparison.
Do not make the most elaborate scenario the default winner. A model with more parameters can accommodate more patterns while remaining weakly identified. Ask whether the available data actually distinguish the added events. If several histories remain plausible, report that ambiguity and identify the most informative next observation.
The existing How Science Works guide to Biogeography connects dispersal, barriers and historical geography. Use it for the larger setting, while retaining the distinction here between dates of gene ancestry and dates of population events.
CHAPTER 10 OF 17 · Compare explanations
10. Interpret geographic states and reconstructed origins
Some phylogeographic models represent locations as discrete states, such as islands or regions. Others represent position continuously. Each choice imposes a representation of geography. A broad region can hide internal movement, while very fine categories can outstrip the information available to estimate transitions.
An inferred ancestral location is conditional on sampled sequences, their locations, the genealogy and the geographic model. An unsampled source is especially important: a reconstruction may distribute support among the available states even when the true historical origin is poorly represented. “Highest support among included states” is not synonymous with “demonstrated origin.”
In New Routes to Phylogeography, De Maio and colleagues compared approaches and demonstrated serious sensitivity of discrete-trait inference to sampling and model choice in their investigated settings. Their work motivates checking whether a model’s treatment of population processes fits the question rather than choosing a method solely because it produces an attractive map.
Before accepting an origin claim, ask whether plausible source regions were represented, whether location definitions were defensible and whether the inference changes under alternative sampling. A report should show uncertainty across locations or regions where the method supports it. Replacing an uncertain area with a single map pin communicates more precision than the analysis has earned.
CHAPTER 11 OF 17 · Repair weak inference
11. Diagnose sampling imbalance numerically
Suppose 90 specimens were collected in region A and ten in region B. Sixty-three A specimens and four B specimens carry a particular variant. The pooled sample frequency is 67 ÷ 100 = 67%. Within-region frequencies are 63 ÷ 90 = 70% and 4 ÷ 10 = 40%.
If a teaching question gives the two regions equal weight, their equally weighted mean is (70% + 40%) ÷ 2 = 55%. The difference between 67% and 55% does not establish which estimate describes the biological population. It establishes that weighting changes the estimand: the quantity being summarised. Equal regional weighting is not automatically correct either.
This arithmetic is a simple demonstration of representation, not a substitute for a phylogeographic likelihood. It explains why the sampling scheme must be visible. The larger sample can dominate a pooled summary without representing a larger biological population or a more likely ancestral source.
Kalkauskas and colleagues’ continuous phylogeography study used simulations to investigate how sampling and model choice affect reconstructed locations. Their results show why apparently strong location support can still be misleading under biased sampling. That finding does not imply every imbalanced dataset is unusable; it makes sensitivity analysis and observation-process reasoning necessary.
CHAPTER 12 OF 17 · Repair weak inference
12. Build a repair table before believing the story
Different problems need different repairs. Adding sequence length does not fill a missing region. Adding specimens does not correct a mislabeled locality. A longer computational run does not make an inappropriate demographic model appropriate.
| Warning sign | What it may mean | A useful next check |
|---|---|---|
| Two clusters separated by a sampling gap | A gradual transition may be invisible | Sample or analyse intermediate localities where possible |
| One marker supports a dramatic origin story | A single genealogy may not represent population history | Compare suitable independent genomic evidence |
| Geographic groups differ in missing data | Processing may be confounded with location | Audit quality and filtering by locality |
| An inferred origin moves after subsampling | Sampling composition influences the result | Report sensitivity and revisit the observation process |
| Gene date is equated with barrier date | Different historical quantities are conflated | Separate coalescence, population separation and geological timing |
| All alternative histories fit similarly | The available data may not identify the event | Report unresolved alternatives and design an informative follow-up |
Keep repairs proportional to the problem. If coordinates were rounded to a district centre, use appropriate spatial uncertainty rather than inventing precise points. If the distinction between two histories is unsupported, narrow the conclusion. “The data show geographic structure” can be a valuable result even when “the population originated here” cannot be defended.
CHAPTER 13 OF 17 · Repair weak inference
13. Test models with data they did not merely describe
A fitted model can reproduce a pattern because it was adjusted to that pattern. Stronger evaluation asks whether it also captures relevant features not used in the same way during fitting. Depending on the method, this may involve simulations under fitted parameters, held-out observations, alternative summaries or independent sources of historical evidence.
For an original classroom example, suppose three proposed histories predict counts of a designated variant in a held-out sample of 20: history A predicts 12, B predicts 8 and C predicts 15. The observed count is 9. Absolute deviations are 3, 1 and 6. B is closest by this chosen descriptive score.
That ranking is not a probability that B is true. Predictions have distributions, counts vary through sampling, and one held-out summary may ignore other important features. The example illustrates the discipline of stating a comparison rule rather than announcing that a story “looks right.” A real comparison needs the appropriate uncertainty model and a justification for the summaries examined.
If every candidate fits poorly, choosing the least poor one does not validate its history. Missing populations, unmodelled contact, incorrect marker assumptions or location errors may require a new model set. A useful conclusion distinguishes evidence against a particular explanation from evidence uniquely establishing another.
Do not use the same event twice
A geological event can provide independent context, or it can be used as a calibration assumption. Those roles must be distinguished. If an analysis assumes that a population split occurred when a particular barrier formed, a resulting date near that barrier’s age does not independently demonstrate that the barrier caused the split. The agreement was partly built into the analysis.
A useful repair is to label each historical datum by role: calibration, model definition or external comparison. Then ask whether an alternative calibration or a model without the assumed event gives a compatible result. Independence is about the information used, not whether the evidence appears in a different figure or paragraph.
CHAPTER 14 OF 17 · Repair weak inference
14. Keep the investigation reproducible and ethically usable
A reproducible record should connect specimen identifiers, location meanings, dates, sequence accessions or approved data access, filtering rules, alignment versions, model settings and analysis outputs. Keep the link from the final figure back to the data transformation that produced it. A colleague should be able to discover why a sample disappeared between the original table and the final map.
Save software versions and random seeds where applicable, but do not confuse repeatability with correctness. A consistently reproduced error remains an error. Use independent checks on units, locality coding and specimen identities as well as computational reruns. Document alternative analyses that materially changed the conclusion.
Exact coordinates can expose vulnerable populations or sensitive collecting sites. Share sufficient spatial information for scientific interpretation while respecting access conditions and legitimate protection needs. When studying people, genetic patterns must not be converted into unsupported claims about cultural identity, belonging or individual worth. Different questions require different evidence and governance.
Historical relatedness alone also does not authorise conservation translocations. Present-day ecological conditions, health risks, adaptation, population viability and relevant permissions need separate consideration. For contemporary environmental connections and their tests, continue with Landscape Genetics. Historical reconstruction and present connectivity can inform each other without being treated as interchangeable measurements.
CHAPTER 15 OF 17 · Transfer and continue
15. Transfer check: a third island changes the question
You first study two islands, North and South. Ten North specimens all carry variant X; ten South specimens all carry variant Y. A colleague says the species began on North and migrated south because North appears closer to the mainland.
Later you sample a third island, Central. Five specimens carry X and five carry Y. You also learn that the original South collection came from one family group, whereas the North specimens came from several sites. No time-calibrated genealogy has yet been estimated.
Answer four questions. What did the initial data establish? Did they establish migration direction? What does Central change? What should be repaired before drawing a historical arrow?
Answer and reasoning
The initial samples established a difference in observed variant composition between those collections. They did not establish species origin or direction of movement. Present distance from a mainland is contextual geography, not an ancestral-location measurement.
Central shows that the two variants coexist in at least the sampled collection there. This creates new questions about contact, retained variation, colonisation and sampling. It does not by itself prove that Central was the origin or meeting point of two expansions. All three historical interpretations require evidence beyond variant presence.
The family sampling at South weakens its representation of the wider population. The immediate repair is to document relatedness and obtain a more appropriate geographic and biological sample where feasible. Then choose genomic evidence and historical models suited to distinguishing the proposed alternatives. A larger number printed beside the original family would not fix the design.
The transferable lesson is to reopen the inference when new evidence changes the observation process. Preserve the valid initial observation, withdraw the unsupported direction claim, and specify which new evidence could discriminate among the remaining histories.
Is phylogeography the same as population genetics?
They overlap. Population genetics supplies mechanisms and models for genetic variation; phylogeography puts lineage relationships into geographic and historical questions. The boundary is practical rather than absolute. A study can contribute to both while making clear which parameters and events it estimates.
Does a distinct genetic lineage mean a distinct species?
Not automatically. A lineage at one genomic region can be distinct without resolving species boundaries. Species interpretation may require additional genomic, ecological, reproductive and taxonomic evidence. Avoid turning a colour category on one figure into a formal classification without the relevant analysis.
Does the region with greatest diversity have to be the origin?
No. Diversity can reflect sampling, population size, admixture, persistence, loss elsewhere and other processes. High diversity may be informative in a specified historical comparison, but it is not a universal origin detector. Ask what competing histories could generate the same observation.
Can present-day DNA tell us everything about extinct populations?
No. Extinction and incomplete sampling remove information. Surviving lineages provide a partial record shaped by the processes that allowed them to persist. Historical or ancient material may add evidence when available and appropriately analysed, but no method guarantees recovery of every lost population or route.
Is a strongly supported result necessarily historically correct?
Support is conditional on the analysis and its assumptions. A narrow interval can coexist with systematic error from missing regions, poor calibration or an inappropriate model. Report support alongside design limitations and sensitivity checks, not as a replacement for them.
CHAPTER 17 OF 17 · Transfer and continue
17. A practical learning and research route
Begin with a small, clearly labelled sequence-and-locality dataset, preferably synthetic for early exercises. Calculate observed differences, map sampling effort and distinguish records from inferences. Before estimating historical events, explain which aspects of the observations each proposed history must account for.
Next, compare alternative sampling designs on paper. Remove one locality, oversample another and ask which interpretations become impossible to distinguish. This develops a useful intuition before software produces complex outputs. Keep the numerical examples in this article separate from empirical claims: they are exercises in reasoning, not parameter recommendations.
For methodological reading, use the original De Maio and colleagues study to examine model choice and the original Kalkauskas and colleagues study to examine spatial sampling sensitivity. Their conclusions concern investigated models and conditions; use the methods and limitations sections to understand the scope. For computational entry points, the BEAST documentation provides maintained tutorials, including phylogeographic analysis and diagnostics. Match instructions to the actual software and model you use.
Within the subject collection, read Biogeography for distributions and environment, Population Genetics for evolutionary mechanisms, and Landscape Genetics for spatial connections and resistance hypotheses. The specialist Biogeography guide supplies further routes through islands, barriers and range change.
Finish a study exercise with three sentences: what the observations establish, what the model suggests under stated assumptions, and what remains unresolved. Then change one important condition and see whether the explanation still holds. That practice connects this subject to The Gold Standard of Education: understanding that can survive a changed problem. Continue through The Gold Standard master or How X Works with the same separation between evidence, assumptions and inference.