Contents
29 — Linkage disequilibrium and haplotypes Part 5 — Population genetics

29 — Linkage disequilibrium and haplotypes

Before this: Ch 14 · Ch 26 · Ch 27 · Ch 28 · Time: ~35 min

Statistics needed: S5 Variance and regression · S2 Distributions

Genome-wide association studies do not measure the variants that cause disease. They measure a few hundred thousand cheap markers and rely on those markers being correlated with whatever the causal variants are. That correlation is . Its magnitude sets how many people you need, its extent sets how precisely you can localise a hit, and its variation between populations is why a trained in one degrades in another.

What you'll be able to do

The core idea

Lay out a population's as a matrix: one row per chromosome copy, one column per variable site, entries 0 and 1. is nothing more exotic than correlation between columns of that matrix.

Two things create the correlation and one thing destroys it. Every new is born on a single existing chromosome, so at the instant it arises it is perfectly associated with every variant already on that chromosome — correlation 1 with its entire neighbourhood. Demography and selection add more correlation by non-random sampling of whole chromosomes. Recombination, one crossover at a time, grinds it back down. The observed correlation structure of a genome is the current balance of those processes, and it is a record of population history.

is a property of chromosomes. Linkage disequilibrium is a property of a population at a moment in time. Two loci 1 kb apart are linked in every organism that has ever lived, but can sit at zero LD. Two loci on different chromosomes are never linked, yet can show strong LD in an admixed population. Linkage governs how fast LD decays; it does not create it and does not guarantee it.


1. Haplotypes, and the phase problem

A is the set of carried on one physical copy — alleles that travelled together through the last . You have two haplotypes at every autosomal region, one from each parent.

A is what a standard assay reports: the unordered pair. That loses information. An individual heterozygous at two sites is written A/a and B/b, which is consistent with two different chromosome configurations:

   phase 1                  phase 2
   ── A ──── B ──           ── A ──── b ──      (paternal)
   ── a ──── b ──           ── a ──── B ──      (maternal)
   haplotypes AB, ab        haplotypes Ab, aB

Genotype data cannot distinguish them; only double heterozygotes are ambiguous, but they are exactly the informative ones. Recovering requires trios (which parent contributed what), reads long enough to span both sites (Ch 46), or statistical phasing — inferring the most probable haplotypes given that the population carries a limited number of common ones. That last method is circular in a productive way: it uses LD to infer haplotypes and haplotypes to quantify LD.

Fortunately the main statistic can be computed without phasing at all. We will get to why in §4.

2. Linkage is not linkage disequilibrium

This confusion is universal and worth killing before any algebra.

Genetic
What it describesTwo loci on the same , close enough that crossovers between them are rareA statistical association between the at two loci
Property ofThe physical genome (plus the hotspot landscape)A specific population at a specific time
Quantified byRecombination fraction c, genetic distance in cMD, D', r²
TimescaleOne Thousands of generations of population history
Changes whenEssentially neverEvery generation — and instantly on
Detected in (Ch 14)Unrelated individuals

Linked but at equilibrium. Two common SNPs 5 kb apart, both millions of years old, on lineages that have recombined and back-mutated enough times that the association has washed out. c is tiny; r² is ~0.

In LD but unlinked. Merge two populations that differ in at loci on chromosomes 1 and 19. The merged population shows LD between them, with no linkage anywhere in sight (§7). It decays at c = 0.5 — halving every generation — but it is real, and it is the source of confounding in association studies (Ch 51).

Linkage is the brake on decay, not the cause of association.

3. Deriving D

Two biallelic loci. Locus 1 has A/a at frequencies p_A, p_a = 1 - p_A; locus 2 has B/b at p_B, p_b. Four exist, with frequencies p_{AB}, p_{Ab}, p_{aB}, p_{ab} summing to 1.

If the loci were associating at random, the frequency of the AB haplotype would be the product of the marginals. Define the coefficient of as the departure from that product:

D = p_{AB} - p_A p_B

One quantity suffices for all four cells. Since p_{Ab} = p_A - p_{AB}:

p_{Ab} = p_A - (p_A p_B + D) = p_A(1 - p_B) - D = p_A p_b - D

and by the same substitution p_{aB} = p_a p_B - D and p_{ab} = p_a p_b + D. The whole table is one parameter deep:

Bbtotal
Ap_Ap_B + Dp_Ap_b - Dp_A
ap_ap_B - Dp_ap_b + Dp_a
totalp_Bp_b1

Expanding gives the equivalent cross-product form, which is how you compute it from counts:

D = p_{AB}\,p_{ab} - p_{Ab}\,p_{aB}

D = 0 is equilibrium: the four haplotype frequencies are exactly the products of the allele frequencies.

Why D alone is a bad summary

Every cell must be non-negative, which bounds D immediately:

\max(-p_Ap_B,\; -p_ap_b) \;\le\; D \;\le\; \min(p_Ap_b,\; p_ap_B)

The bound depends on the allele frequencies. If p_A = p_B = 0.5, D can reach 0.25. If p_B = 0.01, D cannot exceed 0.005 no matter how perfect the association. So D = 0.004 means "almost nothing" in the first case and "as strong as physically possible" in the second. A statistic whose scale moves with the marginals is useless for comparing loci.

4. Two normalisations: D' and r²

D' — how far along the physical maximum

Divide by the bound in the direction D actually points:

D' = \frac{D}{D_{\max}}, \qquad D_{\max} = \begin{cases} \min(p_Ap_b,\; p_ap_B) & D > 0\\[4pt] |\max(-p_Ap_B,\; -p_ap_b)| & D < 0\end{cases}

D' \in [0, 1]. D' = 1 means one of the four is absent — the pair has not been separated by a recombination or a recurrent since the younger arose. That is a statement about history, which makes D' the right tool for asking "has this interval recombined?" It is also why D' is used to define .

Its weakness is the mirror image of that strength: D' is 1 whenever a haplotype is missing, including when it is missing purely because an allele is rare. Small samples of rare alleles give D' = 1 constantly. D' is upward-biased and noisy exactly where genomes are most variable.

r² — the squared correlation

Here is the framing that makes the rest of the chapter obvious. Draw one at random from the population and define two indicator variables:

X = \mathbb{1}[\text{carries } A], \qquad Y = \mathbb{1}[\text{carries } B]

Both are Bernoulli; their moments are given in S2 §1:

\mathbb{E}[X] = p_A,\quad \operatorname{Var}(X) = p_Ap_a, \qquad \mathbb{E}[Y] = p_B,\quad \operatorname{Var}(Y) = p_Bp_b

And \mathbb{E}[XY] = \Pr(X{=}1, Y{=}1) = p_{AB}, so

\operatorname{Cov}(X, Y) = p_{AB} - p_Ap_B = D

D is literally the covariance between the two columns of the haplotype matrix. Normalise it the standard way for a covariance:

r = \frac{\operatorname{Cov}(X,Y)}{\sqrt{\operatorname{Var}(X)\operatorname{Var}(Y)}} = \frac{D}{\sqrt{p_Ap_a\,p_Bp_b}}, \qquad \boxed{\;r^2 = \frac{D^2}{p_Ap_a\,p_Bp_b}\;}

This is the ordinary Pearson correlation of S5 §3, and everything there transfers without modification — including that r^2 is the fraction of variance in one column explained by the other, and that n r^2 is a \chi^2_1 statistic for testing D = 0.

Two practical consequences fall out immediately:

Statistics: the additivity of variances and covariances over independent contributions — what makes the factors cancel in the first bullet — is covered in S5 §2.

D'r^2
AsksHas recombination separated these alleles?How well does one predict the other?
=1 whenOne haplotype is absentOne haplotype absent and allele frequencies equal
Rare allelesInflated, often 1 by chanceCorrectly small
Use forBlock definition, historical recombination, power, ,

5. Why r² is the number that sets your sample size

Let the causal variant have dosage G_c and the genotyped tag have dosage G_t, with correlation r between them. The truth is additive: y = \mu + \beta G_c + e. You do not observe G_c; you regress on G_t.

Statistics: fitting a regression and reading R^2 as variance explained is covered in S5 §4; the non-centrality parameter \lambda and how it converts into power is S4 §4.

R^2_t = \frac{\operatorname{Cov}(y, G_t)^2}{\operatorname{Var}(G_t)\operatorname{Var}(y)} = \frac{\beta^2 \operatorname{Cov}(G_c, G_t)^2}{\operatorname{Var}(G_t)\operatorname{Var}(y)}

Multiply and divide by \operatorname{Var}(G_c) and regroup:

R^2_t = \underbrace{\frac{\beta^2\operatorname{Var}(G_c)}{\operatorname{Var}(y)}}_{R^2_c} \cdot \underbrace{\frac{\operatorname{Cov}(G_c,G_t)^2}{\operatorname{Var}(G_c)\operatorname{Var}(G_t)}}_{r^2} \;=\; r^2 R^2_c

The non-centrality parameter of the association test is \lambda \approx N R^2. Equal power requires equal \lambda, so

N_{\text{tag}} = \frac{N_{\text{causal}}}{r^2}

Testing a tag instead of the causal variant costs a factor of 1/r^2 in sample size. The conventional array-design threshold of r^2 \ge 0.8 is precisely a decision to accept a 25% inflation. At r^2 = 0.2 you need five times the cohort; at r^2 = 0.05, twenty times.

Assumptions worth naming: one causal variant, additive effect, tag not itself causal, and r measured in the study population. The last one is where dies (§10).

6. How LD decays

Large randomly-mating population, no selection, no , no ; recombination fraction c. Consider a gamete produced in generation t+1. With probability 1-c no crossover separates the loci, so it carries a sampled intact from the parental pool: \Pr(AB) = p_{AB}(t). With probability c it is a recombinant, taking its A from one parental haplotype and its B allele from the other — independent draws, so \Pr(AB) = p_A p_B. Recombination does not change allele frequencies, so p_A and p_B are constant. Therefore:

p_{AB}(t+1) = (1-c)\,p_{AB}(t) + c\,p_Ap_B

Subtract p_Ap_B from both sides:

D(t+1) = (1-c)p_{AB}(t) - (1-c)p_Ap_B = (1-c)\,D(t)
\boxed{\;D_t = D_0(1-c)^t\;}

Geometric decay, with allele frequencies fixed — so r^2_t = r^2_0 (1-c)^{2t}, decaying twice as fast. The half-life is t_{1/2} = \ln 2 / [-\ln(1-c)] \approx 0.693/c for small c. At the human genome-average rate of roughly 1 cM per Mb:

SeparationcHalf-life of D≈ years (27 y/gen)
Different 0.51 generation27
10 Mb0.16.6 generations180
1 Mb0.0169 generations1,900
100 kb0.001693 generations19,000
10 kb0.00016,900 generations190,000

The generation time is the pinned 27 years (verified facts). The usable literature spans 25–30, so the last column carries about ±10% that the generation counts do not — quote generations when you can.

Read the table twice. LD between unlinked loci is transient — halved every generation, gone in a handful. LD at 10 kb has a half-life on the order of the age of our species — the oldest fossils assigned to Homo sapiens, from Jebel Irhoud, are dated to 315 ± 34 thousand years (Hublin et al. 2017), so 190,000 years is comparable to that, not longer than it. Between those extremes lies the window that makes association mapping possible: at 10–100 kb, LD survives long enough to be reliably present and decays fast enough to localise a signal to something smaller than a chromosome arm.

7. What creates LD

graph LR M["<b>mutation</b><br/>new allele on ONE haplotype<br/>D' = 1 by construction"] --> S["standing LD<br/>in the population"] DR["<b>drift / bottleneck</b><br/>sampling covariance<br/>between haplotypes"] --> S AD["<b>admixture / structure</b><br/>D = m(1−m)·Δp<sub>A</sub>·Δp<sub>B</sub><br/><i>genome-wide, unlinked loci too</i>"] --> S SE["<b>selection</b><br/>hitchhiking, sweeps,<br/>epistatic combinations"] --> S S -->|"× (1−c) every generation"| R["<b>recombination</b><br/>+ time<br/>+ gene conversion"] R --> LE["linkage equilibrium<br/>D = 0"]

is the fundamental source. A new B appears on exactly one , which already carries some allele at every other site — say A. Then p_{AB} = p_B and

D = p_B - p_Ap_B = p_B(1 - p_A) = p_ap_B

Since p_B is tiny, D_{\max} = \min(p_Ap_b, p_ap_B) = p_ap_B, so D' = 1 exactly. But

r^2 = \frac{(p_ap_B)^2}{p_Ap_a\,p_Bp_b} = \frac{p_ap_B}{p_Ap_b} \approx \frac{p_a p_B}{p_A} \to 0

Complete LD and near-zero correlation, simultaneously, from the same table. Every variant enters the population this way. Everything that follows is erosion.

generates LD as a sampling artefact: with finite N, the haplotype counts passed to the next generation are a multinomial draw, and multinomial cells covary. Balancing this against recombination gives Sved's approximation (Sved 1971):

\mathbb{E}[r^2] \approx \frac{1}{1 + 4N_e c}

This one is quoted, not derived here — the derivation tracks along a whole chromosome segment rather than at a single site, and needs machinery this chapter has not built. But its shape is not new to you. It is the same drift-versus-flux equilibrium as the island model of Ch 27 §5: drift builds the association at rate ~1/(2Ne) per generation, a second process tears it down at a constant per-generation rate, and the equilibrium depends only on the product of the two. Set FST ≈ 1/(1 + 4Nm) beside it and the only substitution is c for m — recombination playing the role migration played there. Two different quantities, one balance argument.

The formula says LD extent is inversely proportional to — and, because a correlation at distance c is dominated by events roughly 1/(2c) generations back, that LD at short distances reports on ancient N_e while long-range LD reports on recent N_e. One genome, many time slices.

Bottlenecks are drift concentrated: few surviving haplotypes, LD spikes genome-wide. This is why Finnish, Ashkenazi and Icelandic populations have long haplotypes and are efficient for mapping.

produces LD by mixing, and the algebra is the omitted-variable-bias calculation of S5 §6. Population 1 contributes fraction m, population 2 fraction 1-m; each is internally at equilibrium. Then p_{AB} = m\,p_{A1}p_{B1} + (1-m)p_{A2}p_{B2} while p_A = m\,p_{A1} + (1-m)p_{A2} and similarly for p_B. Subtracting and collecting terms:

D = m(1-m)\,(p_{A1} - p_{A2})(p_{B1} - p_{B2})

The LD is the product of the two frequency differences times the mixing variance. It applies to every pair of loci in the genome, linked or not, which is exactly why unmodelled ancestry produces genome-wide false positives (Ch 28, Ch 51). Continuous population structure does the same thing with a continuum of strata.

Selection creates LD by hitchhiking: a beneficial allele rising in frequency drags its entire chromosomal neighbourhood with it, because the neighbourhood is what it is physically attached to. Epistatic selection — favouring particular combinations — maintains LD against recombination indefinitely.

What destroys it: recombination and time, and nothing else of consequence. (Ch 18) transfers tracts of a few hundred bases and erodes LD at sub-kilobase scales where crossovers are too rare to matter. Recurrent mutation at hypermutable sites places the same allele on different backgrounds, which lowers D' without any recombination at all.

8. Blocks, hotspots, and the array that made GWAS possible

Crossovers are not uniformly placed. In humans, roughly 80% of recombination occurs in 10–20% of the sequence, concentrated into hotspots a couple of kilobases wide. Their positions are set largely by , whose zinc-finger array binds a sequence motif and deposits histone marks that recruit the double-strand breaks initiating meiotic recombination. PRDM9 is highly polymorphic, so hotspot locations differ between individuals — and the gene is non-functional in dogs, which consequently have hotspots at instead.

Punctate recombination produces punctate LD: long stretches with almost no historical recombination, separated by short intervals across which D' collapses. Those stretches are . Within a block, the number of distinct common is far smaller than 2^k for k sites — typically three to five account for the great majority of . Gabriel and colleagues' 2002 survey found block lengths approximately exponentially distributed with a mean of about 22 kb in European and Asian samples and 11 kb in Yoruba samples.

That factor of two is the Sved formula in action. African populations never passed through the out-of-Africa bottleneck, so their effective size over the relevant window is roughly twice that of non-Africans; \mathbb{E}[r^2] \approx 1/(1+4N_ec) then predicts LD extending about half as far. Equivalently: with a larger N_e the genealogy of an African sample coalesces further back, so more recombination has accumulated along its branches — not because the population is older in generations, but because its lineages take longer to find a common ancestor. Shorter African LD blocks are a demographic prediction, not a curiosity.

The engineering consequence: if a block contains 40 common variants but only four common haplotypes, you do not need 40 assays. two or three well-chosen , and the rest are predicted at r^2 \ge 0.8. This collapses ~10 million common human variants to a few hundred thousand array positions — around 500,000 for European ancestry, more for African, where blocks are shorter and tags cover less. Arrays at that density are cheap enough to run on hundreds of thousands of people, which is the entire economic basis of .

The International HapMap Project (from 2002) existed to build that map. I and II genotyped over 3.1 million SNPs in 270 individuals from four population samples; Phase III extended to about 1.6 million SNPs in 1,184 individuals from 11 populations. Its output — which variants tag which — is what array designers used. It was later superseded as a reference panel by 1000 Genomes (3,202 high- samples), and TOPMed, but the conceptual contribution stands.

One caveat: the block model is an idealisation. LD decay is not a step function, block boundaries depend on the algorithm and the sample, and hotspot positions vary between individuals because PRDM9 do.

9. Long haplotypes as a signature of recent selection

Combine §6 and §7 and you get one of the sharpest inferences in population genetics.

A new starts at frequency 1/(2N) on one long, intact . Under alone, reaching high frequency takes on the order of 4N_e generations — tens of thousands. Over that time, a 1 Mb haplotype survives with probability (0.99)^{40{,}000} \approx e^{-400}: zero. Common alleles are normally old, and old alleles normally sit on short haplotypes.

Now let selection drive the allele up in, say, 300 generations. The probability that a 1 Mb haplotype around it survives intact is (0.99)^{300} \approx e^{-3} \approx 0.05 — small per , but the allele is now carried by a large number of chromosomes, so long unbroken copies are still abundant. The observable signature is a violation of the normal relationship: high frequency together with an unusually long shared haplotype.

LCT is the canonical case (Ch 00). The European lactase-persistence allele — a single base change in an inside 13 of the neighbouring gene MCM6, about 14 kb upstream of LCT — reaches roughly 90% frequency in Scandinavia while still sitting on a shared haplotype of order 1 Mb. From the table in §6, a 1 Mb haplotype has a half-life of ~69 generations, so the allele cannot be more than a few thousand years old; independent estimates put it at roughly 5,000–10,000 years, coincident with cattle domestication. East African pastoralists carry different causal variants on their own long haplotypes — convergent evolution, dated the same way.

The statistics that formalise this — EHH, iHS, XP-EHH — are simply "how far does homozygosity extend from this allele, relative to what its frequency predicts". They are developed in Ch 33.

10. Three consequences you will meet again

A hit is a tag, not a cause. The reported variant is the one that happened to be on the array and happened to have the largest test statistic. Dozens of variants in the same block have nearly the same r^2 with the causal site and therefore nearly the same p-value; which one wins is close to noise. (Ch 52) attacks this with , huge samples that resolve small r^2 differences, and multi-ancestry data — African LD blocks being shorter, adding African samples narrows the credible set faster than adding more of the same ancestry.

do not transfer. The estimated tag effect is \beta_{\text{tag}} = \beta_{\text{causal}} \cdot \operatorname{Cov}(G_c,G_t)/\operatorname{Var}(G_t) — a quantity that depends on the LD structure of the discovery population. Applied where r differs, the variance explained scales as r^2_{\text{new}}/r^2_{\text{discovery}}, and if the sign of r flips, so does the contribution. Add differing frequencies and the score degrades sharply out of ancestry, which is the central technical problem of Ch 53.

Statistics: that a fitted coefficient is a covariance divided by the predictor's variance, and therefore a property of the sample it was estimated in, is covered in S5 §4.

is structure used as a codec. You 500,000 markers; you want tens of millions. Model the target as a mosaic of segments copied from a reference panel of sequenced haplotypes, with switches between reference haplotypes occurring at a rate proportional to genetic distance. That is a hidden Markov model whose hidden state is which reference haplotype am I copying now and whose emissions are the observed genotypes; forward– backward gives a posterior over the untyped alleles. The reason it works is precisely §6: within a few tens of kilobases, chromosomes really are near-copies of a small number of ancestral haplotypes. The reason it fails is also §6: with a mismatched reference panel, the target's haplotypes are not in the panel and the copying model has nothing good to copy.


Common misconceptions

What people thinkWhat's actually true
LD means the loci are linkedLD is a population-level correlation. and structure create it between loci on different ; old linked loci can be at D = 0
D' = 1 means the marker is a good proxyD' = 1 only says one is missing. A rare on a common background gives D' = 1 with r^2 \approx 0.05 — a 20× sample-size penalty
r^2 = 1 is achievable for any tightly linked pairr^2 = 1 requires equal allele frequencies. A common SNP can never perfectly tag a rare one
The is the causal variantIt is the best-correlated genotyped marker. Which member of the block tops the list is largely sampling noise
LD extends further in African genomes because of more diversityThe opposite: larger N_e, no out-of-Africa bottleneck, deeper coalescence and so more accumulated recombination → blocks about half as long (~11 kb vs ~22 kb)
Recombination creates LDRecombination destroys it, at rate c per generation. , , admixture and selection create it
are discrete objects with real boundariesA useful idealisation over punctate recombination. Boundaries are algorithm- and sample-dependent, and hotspot positions vary between individuals via
A high r^2 means the two variants do the same thing biologicallyIt means they are statistically correlated in that population. One may be functional and the other silent — that is the entire premise of tag-SNP genotyping

Worked example

Two SNPs, rs1 (A/a) and rs2 (B/b), 250 kb apart. 200 phased (100 individuals):

              rs2 = B     rs2 = b     total
   rs1 = A       90          10        100
   rs1 = a       30          70        100
   total        120          80        200

Step 1 — frequencies. p_{AB} = 90/200 = 0.45, p_{Ab} = 0.05, p_{aB} = 0.15, p_{ab} = 0.35. p_A = 100/200 = 0.50, p_a = 0.50, p_B = 120/200 = 0.60, p_b = 0.40.

Step 2 — D. D = p_{AB} - p_Ap_B = 0.45 - (0.50)(0.60) = 0.45 - 0.30 = 0.15. Check by cross-product: (0.45)(0.35) - (0.05)(0.15) = 0.1575 - 0.0075 = 0.15. ✓

Step 3 — D'. D > 0, so D_{\max} = \min(p_Ap_b,\, p_ap_B) = \min(0.20,\, 0.30) = 0.20. D' = 0.15/0.20 = 0.75. Substantial historical recombination — no haplotype is absent.

Step 4 — r². p_Ap_ap_Bp_b = (0.50)(0.50)(0.60)(0.40) = 0.06. r^2 = D^2 / 0.06 = 0.0225/0.06 = 0.375, so r = 0.612.

Note D' = 0.75 against r^2 = 0.375. The pair looks strongly associated historically and predicts each other only moderately.

Step 5 — the cost of tagging. If rs2 is causal and you only rs1: N_{\text{tag}} = N_{\text{causal}}/r^2 = N/0.375 = 2.67N. A study powered at 10,000 samples on the causal variant needs 26,700 at the tag. This pair would fail a standard r^2 \ge 0.8 array-design filter.

Step 6 — a rare third variant. Add rs3, C at p_C = 0.05, all 10 copies on the A background: p_{AC} = 0.05, p_{aC} = 0. D = 0.05 - (0.50)(0.05) = 0.025. D_{\max} = \min(p_Ap_c,\, p_ap_C) = \min(0.475,\, 0.025) = 0.025. D' = 0.025/0.025 = \mathbf{1.00}. r^2 = (0.025)^2 / [(0.5)(0.5)(0.05)(0.95)] = 0.000625/0.011875 = \mathbf{0.053}. Perfect D', and a 19-fold sample-size penalty. Two statistics, same table, opposite verdicts.

Step 7 — decay. At ~1 cM/Mb, 250 kb gives c = 0.0025. After 100 generations: D_{100} = 0.15 \times (0.9975)^{100}. \ln(0.9975) = -0.0025031; \times 100 = -0.25031; e^{-0.25031} = 0.7785. D_{100} = 0.15 \times 0.7785 = 0.1168. r^2_{100} = r^2_0(1-c)^{200} = 0.375 \times (0.7785)^2 = 0.375 \times 0.6061 = 0.227. Half-life of D (and of r, since the allele frequencies are fixed): 0.693/0.0025 = 277 generations ≈ 7,800 years. Because r^2 decays twice as fast, the study's power halves in ~139 generations ≈ 3,900 years.

Connections