28 — Population structure and inbreeding
Before this: Ch 26 · Ch 27 · Time: ~45 min
Statistics needed: S3 Sampling and estimation · S5 Variance and regression · S7 High-dimensional data
Hardy–Weinberg assumed one population mating at random. Neither half of that is ever true. People mate with relatives more often than chance dictates, and the "population" you sampled is nearly always a bag of partly separated groups. Both violations produce the same statistical signature, both are summarised by the same coefficient, and failing to model either is the single most common way to get a wrong answer out of human genetic data.
What you'll be able to do
- Compute the inbreeding coefficient F from any loop by path counting, and say what the number is measured relative to
- Derive frequencies under inbreeding, predict both the heterozygote deficit and the shift in mean , and quantify how much a cousin marriage raises disease risk — in relative terms and in absolute terms, which differ enormously
- Read a run-of-homozygosity profile and infer how many generations back the shared ancestor sat
- Derive the Wahlund heterozygote deficit, and define FIS, FST, FIT and prove they chain
- Distinguish what FST ≈ 0.12 among human continental groups does license from what it does not, and explain why Lewontin's variance partition and Edwards' classification argument are both correct
- Classify what , , an FST outlier scan and local-ancestry inference each actually measure, and diagnose the artefact each is prone to — smooth gradients and sampling-induced clusters, an arbitrary K, and heavy tails manufactured by demography
- Explain, as an ordinary omitted-variable problem, why unmodelled structure inflates false positives more as sample size grows
The core idea
Two at a locus can look alike for two different reasons. They can be the same by coincidence — the same arose twice, or the variant is simply common. Or they can be the same because they are physical copies of one ancestral molecule, descending through a loop in the . Only the second kind carries information about ancestry, and only the second kind drags along the entire surrounding .
The inbreeding coefficient F is the probability that an individual's two alleles at a randomly chosen locus are copies of one ancestral allele. That single probability is enough to derive frequencies, disease risk, the expected length of homozygous chromosome tracts, and the amount of trait depression. And when you pool distinct groups into one sample, you get a number with exactly the same algebra.
Inbreeding and population structure are the same phenomenon at different scales. Both are non-random mating that makes an individual's two alleles more likely to be identical by descent than a random pair drawn from the population. One F describes both, and at a single locus nothing in the data can tell them apart.
1. Identity by descent, and why "identical" is not enough
Identity by state (IBS) — two have the same sequence. That is all it says. Two copies of the reference allele at a common SNP are IBS, and so are three billion other people's.
(IBD) — two alleles are copies of a single allele in a specific shared ancestor, with an unbroken chain of replication between.
IBD implies IBS (barring a on the path). IBS emphatically does not imply IBD. The distinction is what makes F a coherent quantity, and it carries a caveat that trips people up:
IBD is only defined relative to a chosen base population. Trace far enough back and every pair of human alleles is IBD. When we say F = 1/16 for the child of first cousins, we mean relative to the 's founders — treated as unrelated by convention, with F = 0. Change the base population and F changes. It is a modelling choice, not a property of the DNA.
An individual whose two alleles at a locus are IBD is autozygous there. Autozygosity is homozygosity with a provenance.
2. Computing F from a pedigree
Any loop in a — any ancestor reachable from both parents — creates an opportunity for autozygosity. Consider one common ancestor A of the two parents, sitting n₁ generations above parent P₁ and n₂ above parent P₂:
First cousins P₁ and P₂. Married-in spouses are omitted — they contribute no loop. G₁ and G₂ are two separate common ancestors, so this pedigree contains two paths.
Derivation. Take the individual I received from P₁ and trace it backwards. At each it came from one of that parent's two alleles, each with probability ½, independently. There are n₁ + 1 meioses from A down to I on this side, so the chance the allele traces back to one specified allele of A is (½)n₁+1. Likewise (½)n₂+1 on the other side. So:
- P(both of I's alleles descend from A's first allele) = (½)n₁+n₂+2
- P(both descend from A's second allele) = the same
- P(one from each) = 2 × (½)n₁+n₂+2, and these are IBD only if A was itself inbred — probability FA
Adding the first two gives (½)n₁+n₂+1, and the third contributes (½)n₁+n₂+1FA. Summing over every distinct path:
An equivalent bookkeeping: count the individuals in the loop excluding I — that count is n₁ + n₂ + 1.
| Relationship of the parents | Paths | n₁, n₂ | F of offspring |
|---|---|---|---|
| Full sibs | 2 | 1, 1 | 2·(½)³ = 1/4 |
| Parent × offspring | 1 | 0, 1 | (½)² = 1/4 |
| Half sibs | 1 | 1, 1 | (½)³ = 1/8 |
| Uncle × niece | 2 | 1, 2 | 2·(½)⁴ = 1/8 |
| Double first cousins | 4 | 2, 2 | 4·(½)⁵ = 1/8 |
| First cousins | 2 | 2, 2 | 2·(½)⁵ = 1/16 |
| First cousins once removed | 2 | 2, 3 | 2·(½)⁶ = 1/32 |
| Second cousins | 2 | 3, 3 | 2·(½)⁷ = 1/64 |
In the parent × offspring row the common ancestor is one of the parents, so n₁ = 0 and the loop contains just two individuals besides I. And note the coincidences: half-sib, uncle–niece and double-first-cousin unions all give F = 1/8 by completely different routes. F summarises the loop structure and discards everything else.
3. What F does to genotype frequencies
Condition on the two events. With probability F the are IBD, so the individual is homozygous for whatever that single ancestral allele was — A₁ with probability p. With probability 1 − F they are independent draws, giving Hardy–Weinberg proportions:
| Frequency | |
|---|---|
| A₁A₁ | Fp + (1−F)p² = p² + Fpq |
| A₁A₂ | (1−F)·2pq |
| A₂A₂ | Fq + (1−F)q² = q² + Fpq |
They sum to 1 (the two +Fpq terms exactly cancel the −2Fpq in the heterozygote). Two consequences do most of the work in this chapter:
Heterozygosity is scaled by (1 − F). Hobs = (1−F)·2pq, which inverts to the estimator every downstream method uses:
Allele frequencies are unchanged. Count alleles: p² + Fpq + ½(1−F)2pq = p. Inbreeding is not one of the four forces of Ch 27; it redistributes genotypes without moving frequencies. It matters evolutionarily only indirectly — by exposing alleles to selection, which then does move frequencies.
4. Inbreeding depression, and recessive disease
Give the three values +a, d, −a (midpoint at zero, d the dominance deviation). The population mean under inbreeding is:
M(F) = a(p²+Fpq) + d·2pq(1−F) − a(q²+Fpq) = a(p−q) + 2pqd − 2pqdF
The Fpq terms from the two homozygotes cancel, leaving
Three things fall out. The depression is exactly linear in F — which is why regressing a trait on F or FROH is the standard estimator. It requires directional dominance: if d = 0 everywhere (pure additivity) there is no depression, and if the dᵢ point in random directions they cancel. And it says nothing about which mechanism supplies that directional dominance. Two candidates:
| Partial dominance | ||
|---|---|---|
| Claim | Deleterious are recurrently generated by and are mostly partially ; inbreeding unmasks them | Heterozygotes are genuinely fitter at the locus itself; any homozygosity costs |
| Predicts | Yes — slow inbreeding lets selection strip the recessive load, and depression declines | No — the load is regenerated every generation regardless |
| Apparent overdominant should resolve into two linked recessives in repulsion | Should resolve to a single locus |
Partial dominance is much better supported. Purging is observed in slowly inbred lines; apparent overdominant QTL in maize and Drosophila have repeatedly resolved into pseudo-overdominance (recessive deleterious alleles in repulsion on the two ); and the mutational load model predicts the observed magnitudes without extra assumptions. True overdominance exists — HBB and malaria is the textbook case — but it is rare and cannot carry the bulk of the effect. Associative overdominance in low-recombination regions blurs the line.
The disease arithmetic. For a recessive disorder with q, risk goes from q² to q² + Fpq, so the relative risk is
The F/q term is the whole story: multiplies risk enormously for rare recessives and barely at all for common ones. For cystic fibrosis in northern Europeans (q ≈ 0.02), first-cousin offspring are ~4× at risk; for a disorder with q = 0.001, ~63×.
Yet the absolute excess is modest, because each rare recessive is rare. Summed over loci the excess is F·Σqᵢ, i.e. F times the individual's total recessive load. Empirically — and the two outcome classes must be quoted separately, because they are not the same number — first-cousin offspring carry an excess of significant congenital anomaly of about 1.7–2.8 percentage points above a population background of 2–3%, and, separately, roughly 1.1 points of excess infant mortality and 3.5 points of excess pre-reproductive mortality (Bittles & Black 2010). The two categories overlap, since many of those deaths are anomaly-related, and residual socioeconomic confounding inflates both. So: a few percentage points on each axis, against backgrounds of a few percent — a near-doubling from a small base, not a transformation of it. Both framings are true; quoting only one is how this topic gets misreported in either direction.
Where this gets used on a whole population rather than a family. Ch 35A §7 runs M(F) backwards: if is recessive load unmasked by homozygosity, then immigration that restores heterozygosity should mask it again, which is the mechanism of . Three consequences of this section carry over intact. Inbreeding there is arithmetic rather than behaviour — in a closed population F rises by 1/(2Ne) per generation whatever the mating system. Partial dominance being the mechanism means the load is unmasked, not created, so a lineage's history sets how much masked load it still has to express. And the purging that this section reports in slowly inbred lines is worth less than it looks in a small wild population: across 119 captive pedigreed populations its average effect is under 1%, because the small Ne causing the inbreeding is exactly what blinds selection (Ch 27 §6).
5. Runs of homozygosity: measuring F from the genome
An IBD segment is not a point. Both copies descend from one ancestral , so the entire surrounding region is autozygous until recombination cuts it — producing a long, unbroken run of homozygosity (ROH).
chr4 ═══════════════════════════════════════════════════════════
·······█████████████████████·············█████████···········
|<---- 21 Mb ---->| |<- 8 Mb ->|
autozygous tract autozygous tract
Statistics: the Poisson process, and the exponential distribution of the gaps between its events, are covered in S2 §2 and §6.
Model crossovers as a Poisson process at rate 1 per Morgan per . If the loop from the common ancestor down to the individual and back up contains m meioses, breakpoints accumulate at rate m per Morgan, so the surviving segment length is Exponential with mean 1/m Morgans:
where g is the number of generations back to the common ancestor (m = 2g, one chain of g meioses down each side). The m = 2g shorthand assumes the two descent chains are of equal length; when they are not, count meioses directly. Tract length is therefore a clock:
| Parents' relationship | g | Mean tract | Interpretation |
|---|---|---|---|
| Full sibs | 2 (m = 4) | ~25 cM | extreme, unmistakable |
| Parent–offspring | asymmetric (m = 3) | ~33 cM | extreme, unmistakable |
| First cousins | 3 | ~17 cM | recent |
| Second cousins | 4 | ~13 cM | recent consanguinity |
| Distant shared ancestor | 20 | ~2.5 cM | endogamy / small community |
| Ancient | 100+ | <0.5 cM | background LD, everyone has these |
Inverting: g ≈ 50 / (mean tract length in cM). This is why ROH callers bin by length. Long tracts date recent loops; a genome dense with short tracts indicates a small long-term , not a consanguineous marriage.
The genomic estimator is FROH = (total length in ROH) / (autosomal genome length), conventionally counting tracts above ~1.5 Mb. It is preferable to pedigree F for a reason that is exactly the expectation-versus-realisation distinction of S3 §2: pedigree F is an expectation; FROH is the realisation. Mendelian sampling and recombination are stochastic, so realised autozygosity scatters around the pedigree value. For a first-cousin offspring: 6.25% of a ~3,500 cM map is ~219 cM, in tracts averaging ~17 cM, so roughly 13 segments. Treating that as compound Poisson gives Var ≈ 13 × 2 × 17² ≈ 7,500 cM², SD ≈ 87 cM ≈ 0.025 in F units. So a first-cousin child's realised F commonly lands anywhere between about 0.04 and 0.09. The pedigree tells you the mean; only the genome tells you the draw.
6. The Wahlund effect
Now the other scale. Suppose k subpopulations, each internally in perfect Hardy–Weinberg, with frequencies pᵢ and weights wᵢ. You pool them and treat the result as one population.
Let p̄ = Σwᵢpᵢ and Var(p) = Σwᵢpᵢ² − p̄². Then the pooled heterozygote frequency is
Σwᵢ·2pᵢqᵢ = 2𝔼[p(1−p)] = 2(𝔼[p] − 𝔼[p²]) = 2(p̄ − p̄² − Var(p))
Each homozygote class is in excess by exactly Var(p). Pooling differentiated groups always produces a heterozygote deficit, and the deficit is exactly twice the among-group variance in . It is not an approximation and it does not require the groups to be exotic — any Var(p) > 0 does it.
Rewrite the deficit as a fraction of expectation and you have recovered F:
which is algebraically identical to the inbreeding result of §3. At one locus, a heterozygote deficit from structure and one from inbreeding are indistinguishable in principle. They separate only across the genome: inbreeding leaves contiguous ROH tracts and no low-dimensional structure in the individual-by-individual covariance; structure leaves no long tracts but a low-rank covariance and correlated deficits at ancestry-informative loci. (A third cause is mundane and common: genotyping error and null alleles also eat heterozygotes. Check that before you invoke biology.)
7. Wright's F-statistics
Three heterozygosities, three levels of nesting:
| Symbol | Definition |
|---|---|
| HI | heterozygosity observed in Individuals, averaged over subpopulations |
| HS | heterozygosity expected within Subpopulations from their own frequencies, averaged |
| HT | heterozygosity expected in the Total pooled population, 2p̄q̄ |
Each F is one minus a ratio of two of them:
FIS is non-random mating within subpopulations — classical inbreeding, and it can be negative if mating is disassortative. FST is differentiation among subpopulations. FIT is total departure from panmixia. The identity is then immediate:
It telescopes because 1 − F is a probability of non-identity, and non-identity at the total level requires non-identity at every nested level. The multiplication is the multiplication of independent probabilities, not a coincidence of notation.
Substituting HS = 2(p̄q̄ − Var(p)) into the definition recovers FST = Var(p)/p̄q̄ from §6. So FST is a variance ratio: the fraction of total allele-frequency variance attributable to among-group differences. It is an intraclass correlation. It is an R² — variance explained, in exactly the sense of S5 §4.
Three practical notes for anyone computing it. Estimate FST across loci as a ratio of averages (mean numerator over mean denominator), never as an average of per-locus ratios — the latter is badly biased when some loci have tiny denominators. Use an estimator that corrects for finite and unequal sample sizes (Weir & Cockerham's θ, or Hudson's estimator). And under Wright's island model the –migration equilibrium derived in Ch 27 gives FST ≈ 1/(1 + 4Nm) — meaning roughly one migrant per generation holds FST near 0.2 regardless of population size, because larger populations drift more slowly by exactly the factor that larger Nm offsets.
8. How much structure is there in humans — and what that number means
Human FST among continental groups is roughly 0.10–0.15, depending on the marker set, the populations included, and the estimator. This number is used carelessly in both directions, so it is worth being precise about what it says.
Lewontin (1972) partitioned variation at 17 protein loci and found ~85% within populations, ~8% among populations within continental groupings, ~6% among those groupings. Larger modern datasets push the within-population share higher: Rosenberg et al. (2002), a three-level AMOVA on 377 highly polymorphic microsatellites in 1,056 individuals from 52 populations, attributes 93–95% of variance to within-population differences, roughly another 2% to differences among populations within major geographic groups, and only 3–5% to differences among those groups.
Reconcile that 3–5% against the 0.10–0.15 above before reading on, because they are different statistics rather than a contradiction. Two things differ. The 3–5% is the top level of a three-level partition, so the intermediate among-populations-within-group level absorbs variance that a two-level calculation charges to "among groups". And FST is deflated by marker heterozygosity — its ceiling is 1 − HS — so highly polymorphic microsatellites give systematically lower values than biallelic SNPs on the same samples, ~0.05 against ~0.10 in human data. (Hedrick's G′ST and Jost's D exist to standardise this away.) Marker type and the number of hierarchical levels each move the number by a factor of two or more, so quote the statistic with both. The consistent finding across all of them is that most human genetic variation is within populations, not between them, and that nearly all common are found on every continent — the differences are in frequency, not in presence or absence. By comparison with most other large mammals, humans are strikingly homogeneous, which is what you expect from a recent common origin plus continuous gene flow.
Edwards (2003) pointed out a genuine logical gap in the strongest version of the inference. That per-locus variance is mostly within-group does not imply that individuals cannot be classified by ancestry. Allele-frequency differences are small at each locus but correlated across loci, so a classifier accumulating thousands of weakly informative features achieves near-perfect accuracy. This is not surprising to anyone who has built a classifier: weak individually-informative features that share a common signal combine to near-certainty. Both statements are correct because they answer different questions — "how is variance partitioned?" and "can ancestry be inferred?" — and neither answers a third question that people repeatedly smuggle in.
So, precisely: FST ≈ 0.12 means the among-group component of allele-frequency variance is about 12% of the total. It does not mean two individuals differ in 12% of their sequence; the actual figure is ~0.1% and most of that is shared variation. It does not imply discrete groups — human variation is largely clinal (§9), and the clusters that appear in analyses often reflect discrete sampling of a continuous gradient. It says nothing whatsoever about the genetic architecture of any complex trait or about between-group differences in one; those require entirely separate evidence and are addressed — mostly negatively — in Ch 53 and Ch 58. And the "populations" in any such calculation are operational sampling constructs, not natural kinds; is continuous and only loosely correlated with self-identified categories.
9. Isolation by distance
Structure usually is not a set of boxes. When dispersal is local, individuals near each other are more related than individuals far apart, and genetic differentiation grows smoothly with geographic distance — isolation by distance. Differentiation rises roughly linearly with distance in one dimension and roughly with the logarithm of distance in two.
The empirical demonstration is striking: the first two principal components of European data reproduce the map of Europe closely enough to place an individual's origin within a few hundred kilometres. Nothing discrete is happening; the gradient is the signal. And under pure isolation by distance, of the resulting data produces smooth sinusoidal gradients as its leading components — a mathematical property of the model, not evidence of any migration event. Reading structure off PC plots without that caution has generated a lot of confident false history.
10. Measuring structure
Statistics: principal components, eigenvalues and the genetic relationship matrix they are computed from are covered in S7 §5.
on the matrix. Build G, n individuals × m SNPs, entries in {0,1,2} counting one . Under HWE the entry at SNP j has mean 2pⱼ and variance 2pⱼ(1−pⱼ), so normalise:
g_ij - 2p_j
X_ij = ─────────────────────
sqrt( 2 p_j (1 - p_j) )
Centring removes the mean; the denominator equalises each SNP's contribution in units of its own variance (a Wright–Fisher step changes p with variance proportional to p(1−p), so this makes the drift signal homoscedastic across markers). Then XXᵀ/m is an estimate of the genetic relationship matrix: off-diagonal entries estimate twice the kinship between two individuals relative to the sample mean, diagonals ≈ 1 + F.
That reframes what the PCs are. The principal components of a normalised genotype matrix are the eigenvectors of the empirical relatedness matrix. They are not arbitrary axes of maximal variance in some abstract space — they are the main axes along which people in your sample are related. Under a model of k populations split by drift, the top k − 1 eigenvectors span the space of population membership; a Tracy–Widom test on the eigenvalues (the null being the largest eigenvalue of a random Wishart matrix) tells you how many exceed noise. Two practical traps: LD-prune first and mask known long-range LD regions — the 17q21 MAPT , the HLA region, the LCT region — or a top PC will faithfully represent an inversion polymorphism rather than ancestry; and remember that discrete-looking clusters can come from discrete sampling.
Model-based clustering (STRUCTURE, ). An explicit generative model with K ancestral populations: population k has allele frequencies fkj, individual i has ancestry proportions qᵢ on the simplex, and each of the two alleles at each locus is generated by first drawing a source population with probability qik, then drawing an allele from Bernoulli(fkj). It assumes HWE within ancestral populations and equilibrium between markers given ancestry. STRUCTURE samples the posterior by MCMC; ADMIXTURE maximises the same likelihood by block relaxation and is orders of magnitude faster. Seen from the linear-algebra side it is G ≈ QF — the same low-rank factorisation PCA performs, with non-negativity and sum-to-one replacing orthogonality.
The universal error is treating the selected K as the number of real populations. There is no true K; the model is a description, not a discovery. Cross-validation error picks the K that predicts held-out genotypes best, which is a different thing. Worse, sampling imbalance and distinct demographic histories can produce identical bar plots: a bottleneck in one group and recent admixture in another are not distinguishable from the plot alone.
FST outlier scans. Compute per-locus FST and take the extreme tail as candidates for — the canonical hits (LCT, SLC24A5, EDAR) do come out this way. But the null distribution depends entirely on demography: bottlenecks, hierarchical structure and isolation by distance all generate heavy tails with no selection at all. Methods that fit an explicit null help (a trimmed χ² fit to the FST distribution, or hierarchical island models). And because FST is a ratio whose denominator is within-population diversity, deflates the denominator and manufactures "islands of divergence" where nothing adaptive happened. Corroborate with an absolute divergence measure before believing any of it (Ch 33).
11. Admixture and local ancestry
When previously separated groups interbreed, each descendant becomes a mosaic of segments from different source populations. Global ancestry is the genome-wide proportion from each source; local ancestry is the assignment at each position.
Local ancestry inference is a hidden Markov model along the chromosome, which should feel familiar: hidden states are the ancestry (or ordered pair of ancestries) at a position, transitions are governed by recombination rate times generations since , and emissions come from reference-panel frequencies or . RFMix, HAPMIX and LAMP-LD differ in the emission model, not the skeleton.
The tract-length argument of §5 runs again: after g generations of admixture, ancestry tracts are roughly exponential with mean 100/g cM, so the tract-length distribution dates the admixture. Admixture also generates long-range LD between unlinked loci whose frequencies differ between the sources, decaying with the same clock — the basis of admixture dating methods.
The payoff is admixture mapping: scan for regions where local ancestry in cases deviates from the genome-wide average. When disease risk differs between the source populations, this is powerful and involves only a few thousand effectively independent ancestry blocks rather than millions of SNPs. APOL1 and kidney disease is the standard success story.
12. Why unmodelled structure breaks association studies
This is the setup for Ch 51, and it is an ordinary omitted-variable problem wearing biological clothes.
Statistics: omitted-variable bias, and what "adjusting for a covariate" does and does not remove, are worked through in S5 §6.
Let Z index subpopulation, g be dosage and y the . Decompose:
Cov(g, y) = 𝔼[Cov(g,y | Z)] + Cov(𝔼[g|Z], 𝔼[y|Z])
Suppose the SNP has genuinely zero effect within every subpopulation, so the first term vanishes. With two subpopulations in proportions w and 1−w, frequencies p₁, p₂ and phenotype means μ₁, μ₂:
Non-zero whenever both the and the mean phenotype differ between groups. Ancestry is a common cause of genotype and phenotype: a confounder in the strict sense. The textbook illustration is the "chopsticks gene" — in a mixed East Asian and European sample, any SNP differing in frequency between the groups will associate with chopstick use.
The consequence that matters most:
This is bias, not variance. It does not shrink with n. The estimate stays put while its standard error falls, so the non-centrality parameter grows linearly in n and the p-value marches toward zero. Bigger studies make a stratification artifact more significant, not less.
Which is why genome-wide inflation of the test statistic (λGC) is a poor diagnostic in large samples: it grows with n under confounding and grows with n under genuine polygenicity, so λ > 1 proves nothing on its own. Separating the two requires something that distinguishes confounding from real signal by its LD structure — the intercept.
Statistics: λGC and the it summarises are S7 §4; the mixed model that carries the GRM as a random-effect covariance is §8 of the same chapter.
The standard defences, in ascending order of strength: include the top principal components as covariates; fit a linear mixed model with the genetic relationship matrix as a random-effect covariance, which handles fine relatedness that PCs miss; or use a within-family design, where the Mendelian coin-flip at randomises genotype conditional on parents and confounding is removed by construction rather than adjusted for. The first two are approximations, and residual stratification demonstrably survives them: several published geographic gradients in polygenic trait scores shrank dramatically when re-analysed in sibling designs. Structure is not a nuisance you subtract once; it is the risk in the entire enterprise.
Common misconceptions
| What people think | What's actually true |
|---|---|
| Inbreeding causes harmful | It creates no mutations. It makes existing homozygous. The load was already there, hidden in heterozygotes |
| F is a property of a person's DNA | It is a probability defined relative to a chosen base population, and F is only an expectation — the realised value varies by a standard deviation of ~0.02 even for first cousins |
| Inbreeding changes allele frequencies | It changes frequencies only. Allele frequencies are invariant; inbreeding acts on evolution only by exposing recessives to selection |
| A heterozygote deficit means inbreeding | It equally means pooled structure (Wahlund), assortative mating, or — most often in practice — null alleles and genotyping error |
| FST = 0.12 means people from different continents differ in 12% of their DNA | It is a ratio of variance components. Two humans differ at ~0.1% of sites, and the overwhelming majority of alleles are present everywhere; only frequencies differ |
| Lewontin and Edwards contradict each other | They answer different questions. Variance is mostly within-population and multi-locus ancestry classification is accurate, because small per-locus differences are correlated across loci |
| The best K in is the number of real populations | There is no true K. Cross-validation picks the best predictive description. Different sampling designs and different histories can give indistinguishable bar plots |
| Clusters in a plot mean discrete populations | Sample a continuous discretely and you get clusters. Under pure isolation by distance the leading PCs are smooth gradients by mathematical necessity |
| Adding 10 PCs removes stratification | PCs capture broad structure. Fine-scale structure, recent relatedness and assortative mating survive, and the residual bias grows in significance with sample size |
Worked example
(a) A first-cousin couple and a rare . q = 0.002 (p = 0.998), so 2pq = 0.00399 ≈ 1 in 250.
- F from the : two paths through the shared grandparents, n₁ = n₂ = 2, founders assumed unrelated. F = 2 × (½)⁵ = 1/16 = 0.0625.
- Baseline risk under random mating: q² = (0.002)² = 4.0 × 10⁻⁶ = 1 in 250,000.
- Risk for this couple's child: q² + Fpq = 4.0×10⁻⁶ + (0.0625)(0.998)(0.002) = 4.0×10⁻⁶ + 1.2475×10⁻⁴ = 1.2875 × 10⁻⁴ = 1 in 7,767.
- Relative risk: 1.2875×10⁻⁴ / 4.0×10⁻⁶ = 32.2×. Cross-check with the formula: 1 + F(1−q)/q = 1 + 0.0625 × 0.998/0.002 = 1 + 31.19 = 32.19. ✓
- Absolute excess for this one disorder: 1.2875×10⁻⁴ − 4.0×10⁻⁶ ≈ 0.012%. A 32-fold relative increase on a tiny base. Summed over all recessive loci, the honest counselling figures for first-cousin offspring are ~1.7–2.8 percentage points of excess congenital anomaly and, separately, ~3.5 points of excess pre-reproductive mortality (§4) — and the reason the two framings must always be given together.
- Expected genomic signature: 6.25% of a ~3,500 cM autosomal map = 219 cM autozygous, in tracts of mean 100/(2×3) = 16.7 cM ⇒ roughly 13 ROH tracts averaging ~14 Mb. Any ROH caller will see this immediately; long tracts of that size are not produced by background relatedness.
- Realised F is a draw, not a constant: Var ≈ 13 × 2 × 16.7² ≈ 7,250 cM², SD ≈ 85 cM ≈ 0.024 in F units, so FROH for this child will plausibly fall anywhere in ~0.04–0.09.
(b) The same signature, from structure instead. Two subpopulations of equal size, each in perfect HWE internally, at a SNP with p₁ = 0.1 and p₂ = 0.5. Sample both, pool, and analyse as one population.
| Quantity | Calculation | Value |
|---|---|---|
| p̄ | ½(0.1) + ½(0.5) | 0.30 |
| HT = 2p̄q̄ | 2(0.3)(0.7) | 0.420 |
| HS observed | ½·2(0.1)(0.9) + ½·2(0.5)(0.5) = ½(0.18) + ½(0.50) | 0.340 |
| Var(p) | ½(0.01) + ½(0.25) − 0.09 | 0.040 |
| Deficit 2Var(p) | 0.420 − 0.340 | 0.080 ✓ |
| Homozygote A₁A₁ | ½(0.01) + ½(0.25) = 0.130 vs p̄² = 0.090 | excess 0.040 = Var(p) ✓ |
| FST = Var(p)/p̄q̄ | 0.040 / 0.210 | 0.190 |
| Apparent F = 1 − Hobs/Hexp | 1 − 0.340/0.420 | 0.190 |
The two routes to 0.190 are numerically identical, and no test applied to this locus alone can separate them. You separate them by looking elsewhere in the genome: the inbred child in (a) has 13 multi-megabase homozygous tracts and no low-rank covariance with anyone; the pooled sample in (b) has no long tracts at all but a covariance matrix with an obvious leading eigenvector. The single-locus statistic is ambiguous; the genome-wide pattern is not.
Connections
- Back to: Ch 26 — every result here is a departure from HWE, derived by relaxing exactly one assumption · Ch 27 — generates the Var(p) that becomes FST, and migration erases it · Ch 09 — the meioses being counted · Ch 14 — where cM comes from, and why 1 Morgan means one crossover per
- Forward to: Ch 29 — structure and are major generators of LD, including between unlinked loci · Ch 31 — the same GRM, used as a random-effect covariance · Ch 51 — §12 is the entire justification for the covariate matrix · Ch 53 — why transfer poorly across ancestries · Ch 54 — homozygosity mapping in consanguineous families turns §5 into a gene-finding method · Ch 33 — FST outliers as selection evidence · Ch 35A — §4's M(F) run backwards as , §5's ROH used to diagnose it, and §§7–9's FST and isolation by distance turned into the outlier scan — where isolation by distance is one of the three neutral processes that manufacture the outliers · Ch 58