Problem set 08 — Quantitative genetics
Covers Ch 30–32.
Attempt before revealing. is the one quantity in this book where knowing the definition and understanding it are nearly unrelated skills — problem 5 tests the second.
Problems are roughly in order of difficulty. ★ marks the two worth returning to.
1. Variance decomposition
A population is measured for a . Assume V_I = 0, Cov(G,E) = 0, no G×E.
| Component | Value (trait units²) |
|---|---|
| V_A additive | 42 |
| V_D dominance | 12 |
| V_E environmental | 66 |
(a) Compute V_P, H² and h². (b) State precisely what each means. "How genetic the trait is" is not acceptable for either. (c) Predict the correlations for parent–offspring, full sibs, half sibs, MZ twins. (d) ⚠ Trap. The observed MZ correlation comes back at 0.61. Diagnose it, and say what it implies for every other estimate.
Solution
(a) V_P = 42 + 12 + 66 = 120; V_G = 42 + 12 = 54
H² = V_G/V_P = 54/120 = 0.45 h² = V_A/V_P = 42/120 = 0.35
(b)
- H² = 0.45: 45% of the variance in this population, in this environment, at this time tracks at all, dominance included — the ceiling on prediction from a full genotype.
- h² = 0.35: 35% of the variance is predictable from parental — the fraction carried by , the part that survives . It is the offspring–midparent slope, and the only one of the two that appears in the .
Both describe a population, never a trait and never an individual.
(c) Using Cov = ½·E[k]·V_A + Pr(k=2)·V_D:
| Relationship | Covariance | Correlation |
|---|---|---|
| Parent–offspring | ½(42) = 21 | 21/120 = 0.175 |
| Full sibs | ½(42) + ¼(12) = 24 | 24/120 = 0.200 |
| Half sibs | ¼(42) = 10.5 | 10.5/120 = 0.0875 |
| MZ twins | 42 + 12 = 54 | 54/120 = 0.450 |
Parent–offspring is exactly h²/2, MZ is exactly H², and dominance appears only in the full-sib line — a parent and child cannot share a genotype by descent.
(d) Excess = 0.61 − 0.45 = 0.16. MZ twins share the whole genome, so their correlation is H² + c² — there is no unused genetic term available. Hence ĉ² = 0.16.
Shared environment is not confined to twins, and that is the damage. Full sibs should now correlate 0.20 + 0.16 = 0.36, and the estimator h² = 2t_FS would return 0.72, double the truth. The full-sib estimate is an upper bound, not an estimate — hence animal breeding's preference for half sibs by a common sire, reared by different dams.
2. Parent–offspring regression
Adult height in 1,200 families, female heights multiplied by 1.08 to put both sexes on one scale (as Galton did). σ_P = 7.0 cm. The regression of offspring on midparent has slope 0.62.
(a) Give h², V_A and σ_A. (b) Derive why the midparent slope estimates h² directly while a single-parent slope estimates h²/2. "Because there are two parents" is not the answer. (c) ⚠ Trap. The same sample regressed on fathers only gives slope 0.39. A student doubles it to h² = 0.78 and calls the two estimates inconsistent. Spouses correlate ρ = 0.25 for height. Reconcile them.
Solution
(a) The midparent slope is the , no scaling constant: h² = 0.62.
V_P = 7.0² = 49.0 cm²; V_A = 0.62 × 49.0 = 30.38 cm²; σ_A = √30.38 = 5.51 cm
(b) Both regressions have the same numerator; the factor of two lives entirely in the variance of the predictor.
Cov(P_O, P_parent) = ½V_A — the offspring shares one per locus IBD with this parent. Cov(P_O, MP) = ½[½V_A + ½V_A] = ½V_A — averaging two ½V_A covariances returns ½V_A.
Denominators: Var(parent) = V_P, but Var(MP) = (V_P + V_P)/4 = V_P/2 under random mating:
b_O·parent = (½V_A)/V_P = ½h² and b_O·MP = (½V_A)/(½V_P) = _h_²
Averaging parents adds no genetic information; it halves the noise in the predictor.
(c) Assortative mating breaks the single-parent estimator and leaves the midparent one alone.
With mates correlated ρ, each parent's becomes informative about the other parent's — Cov(A_dam, P_sire) = h²Cov(P_dam,P_sire) = ρV_A — so
b_O·father = ½h²(1 + ρ)
0.39 = ½ × h² × 1.25 → h² = (2 × 0.39)/1.25 = 0.78/1.25 = 0.624
which matches the midparent 0.62 to rounding. There was never an inconsistency, only an unapplied correction.
Why midparent escapes — numerator and denominator inflate by the same factor:
Cov(P_O, MP) = ¼[V_A + ρV_A + ρV_A + V_A] = ½V_A(1 + ρ) Var(MP) = ¼[V_P + 2ρV_P + V_P] = ½V_P(1 + ρ)
The (1 + ρ) cancels exactly. Assortment still raises V_A itself; the slope faithfully reports that inflated current .
3. Twin study and the equal-environments assumption
A of a physiological trait reports r_MZ = 0.68, r_DZ = 0.42.
(a) Estimate h², c², e² by Falconer's formula. (b) Derive the formula and state when the factor of 2 is exact. (c) A second trait gives r_MZ = 0.80, r_DZ = 0.32. Compute and interpret. (d) ⚠ Trap. State the equal-environments assumption. If MZ pairs share trait-relevant environment more than DZ pairs do, in which direction is h² biased, and by how much per unit of violation? (e) Name a bias running the other way.
Solution
(a) ĥ² = 2(0.68 − 0.42) = 2(0.26) = 0.52
ĉ² = 2r_DZ − r_MZ = 0.84 − 0.68 = 0.16 (check: 0.52 + 0.16 = 0.68 = r_MZ ✓)
ê² = 1 − 0.68 = 0.32 — non-shared environment and measurement error.
(b) With d² = V_D/V_P, and both twin types sharing family environment equally:
r_MZ = h² + d² + c² r_DZ = ½h² + ¼d² + c² r_MZ − r_DZ = ½h² + ¾d²
so h² = 2(r_MZ − r_DZ) − 1.5d². The factor of 2 is exact only when V_D = V_I = 0; otherwise the formula charges all non-additive variance to additivity and is biased upward. If d² = 0.08 here, true h² = 0.52 − 1.5(0.08) = 0.40, not 0.52.
(c) ĥ² = 2(0.48) = 0.96; ĉ² = 0.64 − 0.80 = −0.16
A negative variance component is impossible, so the model has failed and you stop. The usual culprit is non-additive variance, which enters r_MZ at full weight and r_DZ at ¼ or less. Report "h² ≤ 0.96, non-additive component unquantified". A negative ĉ² is the most useful diagnostic the design offers.
(d) EEA: MZ and DZ pairs are equally correlated for trait-relevant environments — not that they are treated identically, only that any environmental sharing is equal across twin type.
Direction: upward, and the trap is the magnitude. If MZ pairs share an extra ε of environmental correlation,
ĥ²_biased = 2[(r_MZ + ε) − r_DZ] = ĥ²_true + 2ε
The bias is twice the violation: with ε = 0.05, ĥ² = 2(0.73 − 0.42) = 0.62 rather than 0.52. Differencing amplifies exactly the quantity the design cannot observe. The critiques with real force: MZ twins are treated more alike, and share a chorion about two-thirds of the time while DZ never do — prenatal, so it touches physical traits too.
(e) Assortative mating biases it downward: correlated mates raise r_DZ, since those twins' parents now have covarying , while r_MZ cannot rise. The biases run both ways with unknown magnitudes, which argues for reporting a range.
4. The breeder's equation
Broiler 6-week body weight: μ = 2,400 g, σ_P = 200 g, h² = 0.40 from a half-sib design. The top 5% are selected as parents, both sexes equally.
(a) Compute i, S, R, and the new mean. (b) Project 10 generations naively. (c) ⚠ Trap. Correct the second-generation response for the fact that truncation changes the variance as well as the mean. (d) Records after 10 generations show ΣS = 3,960 g and ΣR = 1,150 g. Compute realised and account for the shortfall. N_e = 30.
Solution
(a) x = Φ⁻¹(0.95) = 1.6449
φ(1.6449) = 0.39894 × e^(−1.6449²/2) = 0.39894 × e^(−1.35284) = 0.39894 × 0.25851 = 0.10313
i = 0.10313/0.05 = 2.0626 S = iσ_P = 2.0626 × 200 = 412.5 g (selected parents mean 2,812.5 g) R = h²S = 0.40 × 412.5 = 165.0 g → new mean 2,565 g
Check: R = i h²σ_P = 2.0626 × 0.40 × 200 = 165.0 ✓
(b) 10 × 165.0 = 1,650 g → mean 4,050 g.
(c) Truncation takes a slice off the top of a distribution, and slices have less variance. The variance-reduction coefficient is
k = i(i − x) = 2.06271 × (2.06271 − 1.64485) = 2.06271 × 0.41786 = 0.8619
(i and x carried to five places here; the rounded 2.0626 of part (a) would give 0.8615.)
The trap is reading this as lost . It is not: selecting the tail creates negative between like-signed loci — the Bulmer effect — half of which recombination restores each generation.
V_A = 0.40 × 40,000 = 16,000 g²; residual = 24,000 g².
V_A(1) = V_A(1 − ½kh²) = 16,000 × (1 − 0.5 × 0.8619 × 0.40) = 16,000 × 0.82762 = 13,242 g² V_P(1) = 13,242 + 24,000 = 37,242 → σ_P(1) = 192.98 g; h²(1) = 13,242/37,242 = 0.3556
R(2) = 2.0626 × 0.3556 × 192.98 = 141.5 g
A 14% shortfall in one generation, with no change whatever. Iterating, V_A settles near 12,380 g² (h² ≈ 0.340) by generation five, for a steady R ≈ 134 g: 165.0 + 141.5 + 135.8 + 134.3 + 134.0 + 5 × 133.8 ≈ 1,380 g over ten generations.
(d) h²_realised = ΣR/ΣS = 1,150/3,960 = 0.290 against a 0.40. (ΣS is 3,960, not 10 × 412.5 = 4,125, because σ_P shrinks.)
- Bulmer covers 1,650 → ~1,380 g. Diagnostic: relax selection two generations and V_A recovers as recombination breaks the disequilibrium. Nothing was lost.
- The remaining ~1,380 → 1,150 g (17%) needs allele (diagnostic: reverse selection also fails) or opposing natural selection — real in broilers, where the heaviest birds have reduced fertility and leg pathology, so effective S falls below nominal.
- is not the answer at generation 10. Robertson's limit puts total advance at ≈ 2N_eR₁ = 2(30)(165) = 9,900 g, half-life ≈ 1.4N_e = 42 generations.
Realised heritability is the number to trust — the only one measured on the intervention actually performed.
5. The heritability misconceptions ★
Adult height in a well-nourished European population has h² ≈ 0.8, σ_P ≈ 7 cm. All four claims below are false. For each, say what went wrong and what the correct statement is.
(a) "80% of an individual's height is genes; 20% is environment." (b) "Since height is 80% heritable, environmental intervention can't move it much." (c) "Population X averages 8 cm shorter than Y. Since h² = 0.8, most of that gap must be genetic." (d) "h² = 0.8 is a fact about human height, so it holds in any population."
Solution
(a) partitions variance, not an individual's trait value. Variance is a property of a group; you have one height and there is nothing to decompose. Correctly: of the variance among people in this population, 80% tracks additive genetic differences. Your height is not 80% genes and 20% food — it is 100% both, as a rectangle's area is not 60% length. The killer counterexample: number of fingers is near-perfectly genetically specified and has heritability near zero, because nearly all its variance is industrial accidents.
(b) High heritability says nothing about malleability.
Dutch height. Mean adult male height rose roughly 20 cm since the mid-nineteenth century on nutrition, sanitation and disease burden — 20/7 = 2.9 σ_P. Heritability stayed high throughout, because it measures how well rank within a cohort tracks , and rank is stable while the whole distribution translates.
Phenylketonuria. Heritability near 1, untreated causes severe irreversible disability, completely preventable by a low-phenylalanine diet — an intervention that did not exist in the environments the heritability was estimated over. Heritability is estimated over the environments that happened to be present, and interventions are attempts to leave that range.
(c) Heritability contains no information about between-group differences.
The two pots. Take genetically variable seed. Split it at random into two pots.
POT A — full soil POT B — depleted soil
30 cm 26 cm 33 cm 27 cm 12 cm 10 cm 14 cm 12 cm
mean 29 cm mean 12 cm
WITHIN each pot: environment is uniform, so essentially all
variation is genetic → h² ≈ 1 in both pots.
BETWEEN pots: the 17 cm gap is 100% environmental,
by construction — the seed was randomised.
Within-group heritability is as high as it can be; the between-group difference is entirely environmental. Both are true at once, and this is not an edge case: heritability is computed from within-group variance and has no term for between-group means. The reverse seals it — two genetically distinct seed sources on identical soil give a 100% genetic gap that the within-pot heritabilities again say nothing about.
⚠ The sophisticated version also fails. "σ_E = √0.2 × 7 = 3.13 cm, so an 8 cm gap needs mean environments differing by 8/3.13 = 2.56 environmental SDs — implausible." The Dutch 2.9 σ shift happened within one population, entirely environmentally: σ_E, estimated over environments varying within a population, does not bound differences between them. Ch 58 has the history.
(d) Heritability is a property of a population, in an environment, at a time. Hold V_A fixed at 0.8 (units where the original V_P = 1, so V_E = 0.2) and vary only environment:
| Environmental variance | h² = V_A/(V_A + V_E) |
|---|---|
| V_E = 0.2 (original) | 0.8/1.0 = 0.800 |
| V_E doubled to 0.4 | 0.8/1.2 = 0.667 |
| V_E quadrupled to 0.8 | 0.8/1.6 = 0.500 |
| V_E → 0 (uniform environment) | 1.000 |
Not one changed in that table, which is why reported heritabilities differ across countries and decades — and note the perverse corollary: equalising environments raises heritability. A high h² can be an achievement of social policy, not a on it.
6. Liability threshold and sibling recurrence risk
A disease has prevalence K = 2%. Liability L ~ 𝒩(0,1); of liability h²_L = 0.60.
(a) Compute the threshold T and the mean liability of affected individuals, i. (b) Compute sibling recurrence risk and λ_s. (c) ⚠ Trap. A second disease has prevalence 0.5% and the same h²_L = 0.60. Compute λ_s. Why does this make λ_s a poor cross-disease measure of "how genetic" a condition is?
Solution
(a) K = 1 − Φ(T), so T = Φ⁻¹(0.98) = 2.0537
φ(2.0537) = 0.39894 × e^(−2.0537²/2) = 0.39894 × e^(−2.1089) = 0.39894 × 0.12137 = 0.04842
i = φ(T)/K = 0.04842/0.02 = 2.421
Affected individuals average 2.42 SD up the liability scale — well above the threshold of 2.0537.
(b) Falconer's construction: a relative of a proband has expected liability shifted by r·h²_L·i, with r = ½ for full sibs. (The proband sits at i; a fraction h²_L of that deviation is additive genetic; a sib inherits half.)
Shift = 0.5 × 0.60 × 2.421 = 0.7263
K_sib = 1 − Φ(T − 0.7263) = 1 − Φ(2.0537 − 0.7263) = 1 − Φ(1.3274) = 1 − 0.9078 = 0.0922 ≈ 9.2%
λ_s = 0.0922/0.02 = 4.61
A 1-in-11 sibling risk against a 1-in-50 population risk. The naive "sibs share half the genes, so halve something" has no meaning here: the calculation runs through a shift in a latent normal followed by re-integrating the tail, and a 0.73 SD shift multiplies risk by 4.6 because the threshold is crossing the steep part of that tail.
(c) For K = 0.005:
T = Φ⁻¹(0.995) = 2.5758 φ(2.5758) = 0.39894 × e^(−3.3174) = 0.39894 × 0.036248 = 0.014460 i = 0.014460/0.005 = 2.892
Shift = 0.5 × 0.60 × 2.892 = 0.8676 K_sib = 1 − Φ(2.5758 − 0.8676) = 1 − Φ(1.7082) = 1 − 0.9562 = 0.0438 ≈ 4.4% λ_s = 0.0438/0.005 = 8.76
Identical liability heritability, and λ_s nearly doubles — 4.61 versus 8.76 — purely because one disease is four times rarer. Probands come from a more extreme tail (i rises 2.42 → 2.89) and the sib risk is divided by a smaller K.
So λ_s is not comparable across diseases of different prevalence: "schizophrenia has λ_s ≈ 10 and type 2 diabetes ≈ 3, therefore schizophrenia is more genetic" is unsound without converting both to the liability scale. Same reason performance is reported as liability-scale R² or AUC.
7. Missing heritability, quantified
Estimates for adult height, the best-measured complex trait there is:
| Estimator | h² |
|---|---|
| Twin / | 0.80 |
| Common-SNP GREML (array) | 0.45 |
| Genome-wide-significant SNPs (12,111 SNPs, 5.4M people) | 0.40 |
| Whole-genome-sequence GREML | 0.68 |
(a) Quantify the gaps, absolutely and as fractions of the pedigree estimate. In 2009 about 45 loci explained ~0.05 — what fraction was "missing" then, and now? (b) Give the leading explanation for each gap. There are three, not one. (c) ⚠ Trap. Evaluate "the must be ." (d) Compute the average per-variant effect implied by row 3, at MAF 0.30, in cm. σ_P = 7 cm.
Solution
(a)
| Gap | Absolute | Fraction of 0.80 |
|---|---|---|
| -significant 0.40 → array GREML 0.45 | 0.05 | 6.3% |
| Array GREML 0.45 → WGS GREML 0.68 | 0.23 | 28.8% |
| WGS GREML 0.68 → pedigree 0.80 | 0.12 | 15.0% |
| Pedigree − GWAS-significant | 0.40 | 50.0% |
2009: (0.80 − 0.05)/0.80 = 93.8% missing. Now 0.40/0.80 = 50% missing by the same estimator — and the WGS row shows most of that remainder is untagged, not absent.
(b)
0.40 → 0.45: a power problem, not a variant problem. Both rows share an estimand — additive variance tagged by array SNPs — but row 3 counts only variants clearing p < 5 × 10⁻⁸, and most causal variants sit below it. The gap closes with sample size: height went from 45 loci to 12,111 by adding people (Ch 51).
0.45 → 0.68: rare and low-LD variants arrays cannot tag. h²_SNP is the additive variance of causal variants in LD with the array, so h²_SNP ≤ h² by construction. That the recovered variance concentrates in rare, protein-altering, low-LD variants is a signature of negative selection: larger effects are held at lower frequency, and V_A = 2pqα² penalises rarity twice.
0.68 → 0.80: the pedigree estimate was too high. Shared environment, assortative mating, genetic nurture and non-additive variance all inflate the twin numerator. True additive h² for height is nearer 0.70.
(c) Epistatic variance enters relative covariances with squared coefficients — ¼V_AA between full sibs, ¹⁄₁₆ between half sibs — so it decays fast with relationship distance and is nearly unidentifiable; and at realistic frequencies most epistatic gene action converts into V_A anyway (Ch 30 §4). It is not zero, but the ledger in (b) allocates the whole 0.40 without invoking any interaction.
(d) Mean variance explained per significant variant = 0.40/12,111 = 3.303 × 10⁻⁵ of V_P.
An additive variant explains 2pqβ² in V_P units. At MAF 0.30, 2pq = 2(0.30)(0.70) = 0.42:
β² = 3.303 × 10⁻⁵/0.42 = 7.864 × 10⁻⁵ β = 8.87 × 10⁻³ phenotypic SD per allele = 8.87 × 10⁻³ × 7 cm = 0.062 cm ≈ 0.62 mm
Six tenths of a millimetre per allele, for the most heritable common trait in humans — which is why GWAS needed millions of people, and why common-disease/common-variant was right about "common" and wrong by an order of magnitude about effect size.
8. The Beavis effect ★
An F2 mouse cross with N = 200 is scanned for . A permutation test gives a genome-wide 5% threshold of LOD 3.3.
(a) Why are QTL effect sizes from small studies systematically overestimated, and in which direction? (b) Compute the smallest PVE this experiment can report. What does that mean for a QTL whose true PVE is 4%? (c) Quantify the bias with the conditional-expectation formula. Take SE(δ̂) = 0.20 and threshold c = 0.50, for a low-power case (δ = 0.30) and a high-power case (δ = 0.80). (d) ⚠ Trap. Six QTLs are reported at 12% PVE each in a cross whose is 0.45. What is wrong, what is the fix, and what is not evidence of a false positive?
Solution
(a) An unconditional estimate is unbiased: E[δ̂] = δ. But you never report one. You report δ̂ given that it cleared significance, and estimates that fluctuated up clear it while those that fluctuated down never appear. The published value is the mean of a distribution truncated from below, so the bias is always upward, and larger the lower the power.
(b) χ² = 2·ln(10)·LOD ≈ 4.605 × 3.3 = 15.20. For small PVE the non-centrality is χ² ≈ N × PVE, so
PVE_min = 15.20/200 = 7.6%
Nothing below 7.6% can appear in this experiment at all. A QTL with true PVE 4% cannot be reported honestly even in principle: on the rare occasions it clears the threshold it must be reported at ≥ 7.6%, an inflation of 7.6/4.0 = 1.90×, at least 90%, built in before a single mouse was phenotyped. Putting a 4% QTL at the threshold needs N = 15.20/0.04 = 380 animals, and more for decent power.
(c) E[δ̂ | δ̂ > c] = δ + σ_δ̂·λ((c − δ)/σ_δ̂), λ(x) = φ(x)/(1 − Φ(x)).
Low power, δ = 0.30: x = (0.50 − 0.30)/0.20 = 1.00; power = 1 − Φ(1.00) = 0.159
λ(1.00) = 0.24197/0.15866 = 1.5251 E[δ̂ | sig] = 0.30 + 0.20 × 1.5251 = 0.605 → 2.02× the truth, +102%
High power, δ = 0.80: x = (0.50 − 0.80)/0.20 = −1.50; power = 1 − Φ(−1.50) = 0.933
λ(−1.50) = φ(1.5)/Φ(1.5) = 0.12952/0.93319 = 0.1388 E[δ̂ | sig] = 0.80 + 0.20 × 0.1388 = 0.828 → 1.035×, +3.5%
The bias is governed entirely by power, matching Beavis's simulations: about double at ~100 progeny, modest at ~500, largely gone at ~1,000. In the limit c ≫ δ, λ(x) → x and the expression collapses to E[δ̂ | sig] ≈ c — the reported effect is pinned near the detection threshold, almost independently of the truth, exactly as the 7.6% floor says.
(d) 6 × 12% = 72% of the variance explained in a cross whose total genetic contribution is 45%. The reported QTLs explain more variance than exists — arithmetic, not biology, and routine in the early QTL literature.
Fixes, in order: estimate effects in an independent sample (discovery tells you where, never how big); failing that, invert the formula in (c) as a shrinkage correction; and report the detection floor beside every effect.
⚠ The corollary that gets misapplied in reverse: a QTL whose effect shrinks on replication is not thereby a false positive. Shrinkage is the expected behaviour of a real locus. Evidence against a locus is failure to reach significance in a replication powered for the replicated effect size — not a smaller number. The identical structure governs discovery betas and (Ch 53).
Where you went wrong
| Error pattern | What to re-read |
|---|---|
| Used H² where the needs h² | Problem 1, Ch 31 §1 |
| Doubled a single-parent slope without checking assortative mating | Problem 2(c) — the slope is ½h²(1 + ρ) |
| Reported Falconer's h² without computing c² | Problem 3(c) — a negative ĉ² means stop |
| Assumed an EEA violation biases h² by the size of the violation | Problem 3(d) — it biases by twice it |
| Reapplied the original h² in generation 2 of a selection programme | Problem 4(c) — Bulmer effect |
| Read h² = 0.8 as a statement about an individual | Problem 5(a) — variance decomposes, individuals do not |
| Inferred a between-group cause from a within-group | Problem 5(c) — the two pots |
| Treated h² as a constant of the trait | Problem 5(d) — halve V_E and it rises |
| Compared λ_s across diseases of different prevalence | Problem 6(c) — λ_s is a function of K |
| Blamed the missing-heritability gap on | Problem 7(c) — three gaps, three causes |
| Quoted a discovery effect size, or called a shrinking replicate a false positive | Problem 8 |