30 — Quantitative traits and variance
Before this: Ch 10 · Ch 26 · Ch 29 · Time: ~45 min
Statistics needed: S2 Distributions · S5 Variance and regression
Everything so far has sorted individuals into classes: affected or not, round or wrinkled, recombinant or parental. Most traits refuse to do that. Height, blood pressure, yield, LDL cholesterol, and every disease risk that isn't a single broken gene form a smear, not a set of bins.
This chapter builds the machinery for the smear. It is the quantitative spine of the book: the variance components defined here are what is a ratio of (Ch 31), what mapping partitions (Ch 32), and what a is trying to predict (Ch 53).
What you'll be able to do
- Derive why many loci of small effect produce an approximately normal distribution, and compute the exact class frequencies for 2 and 3 loci
- Convert a binary disease outcome into an underlying continuous liability, and compute the threshold and the mean liability of affected individuals from a prevalence
- Write the full variance decomposition \mathrm{Var}(P) including the covariance and interaction terms, and say what each requires to vanish
- Define as the variance of — the fitted values of the regression of genotypic value on count — and compute \alpha, V_A and V_D for a one-locus model
- Compute the additive genetic covariance and r_G from per-locus average effects, and explain why a phenotypic correlation near zero can conceal a strongly negative genetic one
- Derive the expected covariance between relatives from allele-sharing probabilities, and explain why parent–offspring and full-sib covariances differ
The core idea
Take a that looks nothing like Mendelian. Now suppose it is controlled by 200 loci, each with two , each shifting the trait by a small amount, inherited independently. An individual's genotypic value is then a sum of 200 nearly-independent random contributions.
A sum of many independent small contributions has one limiting shape — the central limit theorem, S2 §3.
Quantitative genetics is Mendelian genetics plus the central limit theorem. No new inheritance mechanism was ever required. The continuous distribution is what discrete particulate inheritance looks like when you add up enough particles and blur the result with environment.
Everything else in this chapter is bookkeeping on that sum: splitting its variance into the part a parent can transmit, the part that dies with the individual, and the part that isn't genetic at all.
1. The dispute that Fisher settled
For twenty years after Mendel's rediscovery in 1900, two camps were at war. The biometricians (Galton, Pearson, Weldon) measured continuous traits, found smooth distributions and correlations between relatives of about ½ for parent–offspring, and concluded inheritance was blending. The Mendelians (Bateson, de Vries) had discrete factors, clean ratios, and no account of height.
Both were right about their data. The resolution — Fisher's 1918 paper The Correlation between Relatives on the Supposition of Mendelian Inheritance, the paper that coined the word variance — was to show that the biometricians' observations were a consequence of the Mendelians' mechanism.
Take n unlinked loci, p at each, and let each copy of the "+" add a to the trait. Let X_i \in \{0,1,2\} be the count of + alleles at locus i. Under Hardy–Weinberg, X_i \sim \text{Binomial}(2, p), so \mathrm{Var}(X_i) = 2pq. The genotypic value is
where the variance adds because the loci are independent — which requires equilibrium (Ch 29), not merely different .
Statistics: why variances of independent contributions add, and what the cross term 2\,\mathrm{Cov} is when they do not, are covered in S5 §2.
With n loci there are 2n+1 genotypic classes and G is a scaled Binomial(2n, p). The CLT does the rest.
Watch it happen. An F2 from a cross of two pure lines gives p = \tfrac12 at every locus:
1 locus + alleles k: 0 1 2 3 classes
frequency: 1 2 1 / 4
2 loci 0 1 2 3 4 5 classes
1 4 6 4 1 / 16
3 loci 0 1 2 3 4 5 6 7 classes
1 6 15 20 15 6 1 / 64
This is exactly what Nilsson-Ehle saw in wheat kernel colour in 1909: an F2 in which 1/64 of seeds were as white as one grandparent and 1/64 as red as the other, with a graded series between — three loci, behaving perfectly Mendelian, producing something that reads as blending unless you count carefully.
Now add environment. Real are the genotypic classes plus noise, so the observed distribution is a mixture of normals centred on the class values. A useful criterion: an equal-weight mixture of two normals whose means differ by d with common standard deviation \sigma is unimodal when d \le 2\sigma. The class spacing here is a, so once \sigma_E \gtrsim a/2 the bumps merge and the histogram is smooth. With 3 loci and modest environmental variance you can no longer see the seven classes at all — which is precisely why the biometricians never found them.
Two forces conspire, and only one of them involves n. Environmental variance blurs the spacing directly. The loci act more subtly: with a held fixed, adding loci leaves the spacing at a and instead widens the range to 2na and the genetic standard deviation to \sqrt{2npq}\,a, so the steps become a finer and finer texture relative to the spread. Only when the loci divide a fixed total does the spacing itself shrink — the Nilsson-Ehle case, where two pure parental lines pin the difference \Delta between the extremes, so n loci give a = \Delta/2n. Either way, you need very few loci before the discreteness is undetectable.
2. Three kinds of quantitative trait
| Type | Measurement | Examples | Distribution |
|---|---|---|---|
| Continuous | real-valued | height, blood pressure, LDL, expression level | approximately normal, often after transformation |
| Meristic | a count | bristle number, vertebrae, litter size, seed number | discrete but many-valued; Poisson-ish, treated as continuous |
| Threshold | present/absent | type 2 diabetes, schizophrenia, cleft palate, twinning | binary, modelled as a latent continuum |
Threshold traits are the important case for human genetics, because most of what we care about is binary and most of what we can model is continuous.
The liability threshold model
Posit an unobserved continuous variable — liability — that is the sum of all genetic and environmental contributions to risk. Give it the CLT treatment: assume it is normal, and standardise it to L \sim \mathcal{N}(0,1), which costs nothing since the scale is arbitrary. An individual is affected when L exceeds a threshold T.
Statistics: the normal distribution — its density \varphi, its cdf \Phi, and the central-limit argument that licenses assuming it for a latent sum — is covered in S2 §3.
With prevalence K:
The mean liability of affected individuals is the inverse Mills ratio,
Worked, for K = 0.01: T = 2.326, \varphi(2.326) = 0.3989\,e^{-2.706} = 0.02665, so i = 0.02665/0.01 = 2.665. Affected individuals sit on average 2.67 standard deviations up the liability scale.
This is not decoration. It buys three things:
- A single binary outcome becomes a continuous trait, so all the variance machinery below applies unchanged, on the liability scale.
- Dominance appears for free. A trait with a threshold behaves non-additively on the observed scale even when liability is perfectly additive — relatives of severely affected probands have higher risk, recurrence risk falls off faster than \tfrac12 per degree of relationship, and none of that requires interaction between loci.
- becomes scale-dependent, which is why Ch 31 §3 distinguishes heritability on the observed 0/1 scale from heritability on the liability scale, and why performance for a disease is reported as liability-scale R^2 or AUC (Ch 53).
Converting a heritability between the two scales
Scale-dependence is not a nuisance to be gestured at; it is a conversion you have to do before two numbers are comparable. A heritability computed on the observed 0/1 scale — the fraction of the variance of the indicator that is additive-genetic — depends on the prevalence and on the case fraction the study happened to sample. Two case–control studies of the same disease, with identical underlying genetics, report different observed-scale heritabilities purely because one recruited one control per case and the other recruited four.
The liability scale is the one that does not move, and the transformation onto it (Lee et al. 2011) is
with K the population prevalence, T = \Phi^{-1}(1-K) and \varphi(T) exactly as above, and P the fraction of the sample that is cases. It is two corrections multiplied. The factor K(1-K)/\varphi(T)^2 converts a binary-scale variance to a liability-scale one; the second factor K(1-K)/[P(1-P)] undoes the ascertainment, because deliberately over-sampling cases inflates the variance of the indicator without touching the genetics. Draw the sample at random from the population instead and P = K, the second factor is 1, and only the scale change remains.
Worked, for a 50:50 case–control study of a disease with K = 0.01 reporting h^2_{\text{obs}} = 0.10. Reusing T = 2.326 and \varphi(T) = 0.02665 from above,
Just over half the reported figure — and the difference is entirely scale and sampling, with not a single differing between the two numbers.
Two cautions. K is an epidemiological input from outside the study, so an error in the prevalence propagates straight into h^2_l. And the conversion fixes the axis a number is reported on and nothing else: confounding, shared environment and a misspecified threshold model all survive it untouched.
3. The phenotype model, and the terms everyone drops
Start with the decomposition that every textbook writes:
Genotypic value plus environmental deviation. Taking variances:
That covariance term is not a technicality, and dropping it is not conservative — it is an assumption that are distributed independently of environments. In a randomised field trial or a common-garden design that assumption is true by construction. In observational human data it is routinely false:
- Parents transmit and rear the child. Alleles associated with reading ability are carried by parents who buy books — genetic nurture, and it inflates every family-based estimate.
- Niche picking: individuals select environments correlated with their propensities.
- Selective breeding programmes give high-merit animals better feed, so \mathrm{Cov}(G,E) > 0 by management decision.
Allowing effects to be non-additive across the genotype/environment boundary gives
V_{G\times E} is non-zero exactly when reaction norms are non-parallel — when the difference between two genotypes depends on the environment they are in. Phenylketonuria is the cleanest case: the PAH genotype's effect on cognition is severe on a normal diet and near zero on a phenylalanine-restricted one. Not a small correction; a sign that "the effect of the genotype" is not a well-defined quantity without naming the environment.
Both extra terms require designs that break the confound to estimate — randomisation, cross-fostering, within-family comparison, or measured environments. Neither is safely assumed away.
4. Decomposing V_G: the regression that defines additive variance
Genetic variance splits further:
V_A is the single most important quantity in the field and the one most often defined badly. "The variance due to genes acting additively" is wrong, or at least useless — it makes V_A sound like a property of gene action. It is not. Here is the real definition, in the regression language of S5.
The setup
One locus, A_1 and A_2 at frequencies p and q = 1-p. Put the three genotypic values on a scale centred midway between the homozygotes:
| Allele count X | Frequency (HWE) | Genotypic value G | |
|---|---|---|---|
| A_1A_1 | 2 | p^2 | +a |
| A_1A_2 | 1 | 2pq | d |
| A_2A_2 | 0 | q^2 | -a |
d is the dominance parameter: d=0 is pure additivity, d=a is complete dominance of A_1, |d|>a is . Population mean:
The regression
Regress genotypic value on allele count. That is the whole idea. X is a predictor taking values 0, 1, 2; G is the response; fit an ordinary least-squares line through the three points weighted by their genotype frequencies.
so the slope is
\alpha is the average effect of an allele substitution: the expected change in genotypic value from swapping one A_2 for one A_1, averaged over the genetic backgrounds that actually occur at their actual frequencies. (Classically it is derived by uniting an A_1 gamete with a random gamete from the population and taking the deviation from M; that gives average effects \alpha_1 = q\alpha and \alpha_2 = -p\alpha, whose difference is \alpha. Same number, more work.)
Breeding value and the two variance components
The of an individual is its fitted value on that line, as a deviation:
The dominance deviation is the residual, D = G - M - A: what the regression could not capture, which for a single locus is the curvature of three points around a line.
Statistics: the least-squares slope as \mathrm{Cov}(x,y)/\mathrm{Var}(x), and the orthogonality of residuals to fitted values that splits a total sum of squares with no cross term, are covered in S5 §4.
And because residuals are orthogonal to fitted values by construction, \mathrm{Cov}(A,D) = 0 and V_G = V_A + V_D with no cross term. Summing over loci — again requiring equilibrium — gives V_A = \sum_i 2p_iq_i\alpha_i^2. When loci also interact, the residual after fitting all single-locus regressions and their two-locus interaction terms splits into V_I = V_{AA} + V_{AD} + V_{DD} + \dots.
Three consequences that follow immediately from the formulas, and that people get wrong:
V_A is a property of a population, not of a locus. \alpha contains p. The same allele with the same biochemical effect contributes different additive variance in two populations with different frequencies, and contributes zero when fixed. There is no frequency-free "effect size" of a locus in this framework.
A locus with complete dominance still contributes mostly additive variance if the allele is rare. Set d = a: V_D/V_A = 2pqd^2/\alpha^2, which for small p is \approx 2p a^2 / 4a^2 = p/2 \to 0. At p=0.1 the is 5.6% of the additive variance despite gene action being completely . Dominance in the mechanism does not imply dominance variance in the population.
Additive variance is maximised at intermediate frequencies. With d=0, V_A = 2pq\,a^2, peaking at p = \tfrac12. A large-effect allele at p = 0.001 contributes 0.002a^2 against 0.5a^2 — a 250-fold difference — which is most of the reason finds common variants of tiny effect and misses rare variants of large effect (Ch 51).
5. Why V_A is the part that matters
Parents transmit , not . dismantles every genotype in the parent and reassembles new ones at random in the offspring. The additive component is the part of the that survives that dismantling; the dominance component is a property of a pairing that is destroyed and re-drawn every generation.
That is the entire justification for the privileged status of V_A, and it explains the whole downstream structure of the field:
- The response to selection is R = h^2 S with h^2 = V_A/V_P, not V_G/V_P (Ch 31). Selecting parents on phenotype changes the offspring mean only through the transmissible part.
- Animal and plant breeders estimate estimated for selection candidates, and V_D is a nuisance term — except in crossbreeding programmes, where the point is precisely to generate favourable genotype combinations that will not breed true.
- Dominance and contribute to resemblance between relatives only insofar as the relatives share whole genotypes or whole combinations — which is why §7's coefficients differ between the components.
6. Genetic correlation and pleiotropy
Two traits measured on the same individuals each have a . The additive genetic covariance is, summing over loci under equilibrium,
A locus contributes to r_G only if it has a non-zero average effect on both traits — — or if it is in LD with a locus affecting the other trait. The two sources are not distinguishable from the correlation alone, which is a recurring problem in interpretation (Ch 52).
r_G is not r_P, and the environmental correlation can run the other way. Milk yield and fertility in dairy cattle are negatively genetically correlated while better-managed herds score higher on both, so the phenotypic correlation understates the genetic conflict — which is how decades of selection for yield quietly degraded fertility. Selecting on one trait always drags correlated traits along, at a rate set by r_G.
7. Resemblance between relatives — where the coefficients come from
This is the bridge to estimating anything, because V_A and V_D are not observable; the covariances between relatives are.
Set up the accounting at a single locus. The of an individual is the sum of two independent allelic contributions, so each carries variance V_A/2. Relatives covary exactly to the extent that they carry the same allele copies — copies identical by descent (IBD), inherited from a common ancestor.
Let k be the number of alleles a pair shares IBD at a locus (k \in \{0,1,2\}). Then:
- Each IBD-shared allele contributes V_A/2 to the covariance, so the additive contribution is \mathbb{E}[k]\cdot V_A/2.
- Dominance deviations are properties of the whole , so they covary only when the pair has the same genotype IBD — both alleles shared. Contribution: \Pr(k=2)\cdot V_D.
Now just count, assuming random mating and no inbreeding.
Parent–offspring. The offspring received exactly one allele from this parent, and it is IBD with certainty; the other came from an unrelated parent. So k=1 always, \mathbb{E}[k]=1, \Pr(k=2)=0:
Note what the zero means: a parent and child cannot share a genotype IBD, so no amount of contributes to parent–offspring resemblance.
Full sibs. From the mother, both sibs receive one of her two alleles; they receive the same one with probability ½. Independently, the same for the father. So k \sim \text{Binomial}(2, \tfrac12): \Pr(0)=\tfrac14, \Pr(1)=\tfrac12, \Pr(2)=\tfrac14, giving \mathbb{E}[k]=1 and \Pr(k=2)=\tfrac14:
Half sibs. Sharing only one parent, k=1 with probability ½ and 0 otherwise: \mathbb{E}[k]=\tfrac12, \Pr(k=2)=0, so \mathrm{Cov}(\text{HS}) = \tfrac14 V_A.
MZ twins. k=2 always: \mathrm{Cov} = V_A + V_D + V_I = V_G.
Writing r = \mathbb{E}[k]/2 (the coefficient of relationship, twice the kinship coefficient) and u = \Pr(k=2):
| Relationship | r | u | Expected covariance |
|---|---|---|---|
| MZ twins | 1 | 1 | V_A + V_D + V_I |
| Parent–offspring | ½ | 0 | \tfrac12 V_A\;(+\tfrac14 V_{AA} + \dots) |
| Full sibs | ½ | ¼ | \tfrac12 V_A + \tfrac14 V_D\;(+\tfrac14 V_{AA} + \dots) |
| Half sibs | ¼ | 0 | \tfrac14 V_A\;(+\tfrac1{16}V_{AA} + \dots) |
| Grandparent–grandchild | ¼ | 0 | \tfrac14 V_A\;(+\tfrac1{16}V_{AA} + \dots) |
| Aunt/uncle–niece/nephew | ¼ | 0 | \tfrac14 V_A\;(+\tfrac1{16}V_{AA} + \dots) |
| First cousins | ⅛ | 0 | \tfrac18 V_A\;(+\tfrac1{64}V_{AA} + \dots) |
The general form is \mathrm{Cov} = rV_A + uV_D + r^2V_{AA} + ruV_{AD} + u^2V_{DD} + \dots. Epistatic terms enter with squared coefficients, so they decay fast with relationship distance: V_{AA} contributes \tfrac14 to full sibs but only \tfrac1{16} to half sibs. This is why epistatic variance is nearly impossible to identify separately in practice, and why a great deal of epistatic gene action shows up statistically as V_A. Note also that parent–offspring and full sibs have the same r and therefore the same \tfrac14 V_{AA}: differencing the two covariances cancels the additive-by-additive term exactly, leaving \tfrac14 V_D (plus the much smaller V_{AD} and V_{DD} terms) — which is what makes that contrast a usable handle on dominance at all.
Three things this table earns you, all cashed in next chapter:
- Parent–offspring and full sibs share the same V_A coefficient but differ in V_D. Comparing them separates the two components. Comparing half sibs (pure V_A, and shared environment usually smaller) with full sibs does it more cleanly.
- Every entry is a testable prediction. Observed sib correlations that exceed \tfrac12 V_A/V_P by more than \tfrac14 V_D/V_P indicate shared environment, assortative mating, or both — the covariance table is the null.
- Regressions of offspring on parent estimate h^2 directly, because the slope is \mathrm{Cov}(P,O)/\mathrm{Var}(P) = \tfrac12 V_A / V_P = \tfrac12 h^2.
The assumptions doing the work: random mating (assortative mating inflates the additive covariances between all relatives), no inbreeding (which makes k=2 possible where the table says it isn't), equilibrium, no shared environment beyond what is modelled, and no \mathrm{Cov}(G,E). Every one of them fails somewhere in human data, and each failure has a signature. That is Chapter 31's subject.
Common misconceptions
| What people think | What's actually true |
|---|---|
| Continuous traits need a non-Mendelian mechanism | They need only many Mendelian loci plus environmental noise. Fisher 1918 showed the biometricians' correlations are derivable from particulate inheritance |
| Additive variance means "loci with no dominance" | It is the variance of the fitted values from regressing genotypic value on count. A completely locus contributes almost pure V_A when the allele is rare |
| A locus has an effect size, full stop | \alpha = a + d(q-p) depends on , and V_A = 2pq\alpha^2 depends on it twice more. Effect sizes and variance contributions are population-specific |
| V_G is what parents transmit | Parents transmit alleles. V_D and V_I live in combinations that destroys. Only V_A passes through |
| \mathrm{Cov}(G,E) = 0 is a harmless simplification | It is a design property, true in randomised experiments and generally false in human observational data. Genetic nurture and niche picking make it positive, inflating everything estimated from families |
| Full sibs and parent–offspring are equally informative | Both carry \tfrac12 V_A, but full sibs additionally carry \tfrac14 V_D and usually a large shared-environment term. They estimate different things |
| A binary disease trait can't be quantitative | The makes it one. The observed scale is binary; the latent scale is normal, and all the variance machinery works there |
| in the mechanism means epistatic variance in the population | Interaction at the gene-action level converts largely into V_A at realistic allele frequencies. Statistical variance components are not statements about biochemistry |
Worked example
One locus, p = 0.6, q = 0.4, a = 10, d = 4 (partial dominance). Compute everything.
Mean. $M = a(p-q) + 2pqd = 10(0.2) + 0.48(4) = 2 + 1.92 = 3.92$. Check directly: 0.36(10) + 0.48(4) + 0.16(-10) = 3.6 + 1.92 - 1.6 = 3.92. ✓
Total genotypic variance. \mathbb{E}[G^2] = 0.36(100) + 0.48(16) + 0.16(100) = 36 + 7.68 + 16 = 59.68, so V_G = 59.68 - 3.92^2 = 59.68 - 15.3664 = 44.3136.
The regression, explicitly. X \in \{2,1,0\} with frequencies 0.36, 0.48, 0.16.
Formula check: \alpha = a + d(q-p) = 10 + 4(-0.2) = 9.20. ✓
, A = \alpha(X - 2p) = 9.2(X - 1.2):
| X | freq | G | A | D = G - M - A | |
|---|---|---|---|---|---|
| A_1A_1 | 2 | 0.36 | 10 | 9.2(0.8) = 7.36 | 10 - 3.92 - 7.36 = -1.28 |
| A_1A_2 | 1 | 0.48 | 4 | 9.2(-0.2) = -1.84 | 4 - 3.92 + 1.84 = 1.92 |
| A_2A_2 | 0 | 0.16 | −10 | 9.2(-1.2) = -11.04 | -10 - 3.92 + 11.04 = -2.88 |
Means are zero as required: 0.36(7.36) + 0.48(-1.84) + 0.16(-11.04) = 2.6496 - 0.8832 - 1.7664 = 0. ✓ Dominance deviations match -2q^2d = -1.28, $2pqd = 1.92$, -2p^2d = -2.88. ✓
Variance components.
Formula: 2pq\alpha^2 = 0.48(84.64) = 40.6272. ✓
Formula: (2pqd)^2 = (1.92)^2 = 3.6864. ✓
Scale up. Suppose the trait is governed by many such loci contributing in total V_A = 40, V_D = 4, with V_E = 56 and no covariance or interaction terms, so V_P = 100. Then h^2 = V_A/V_P = 0.40 and the predicted correlations between relatives are:
| Relationship | Covariance | Correlation |
|---|---|---|
| Parent–offspring | \tfrac12(40) = 20 | 0.20 |
| Full sibs | \tfrac12(40) + \tfrac14(4) = 21 | 0.21 |
| Half sibs | \tfrac14(40) = 10 | 0.10 |
| MZ twins | 40 + 4 = 44 | 0.44 |
If the observed full-sib correlation came back at 0.35, the excess of 0.14 over the predicted 0.21 is not extra — the maximum dominance contribution here is 0.01. It is shared environment, assortative mating, or \mathrm{Cov}(G,E). The covariance table is what lets you say that.
Connections
- Back to: Ch 10 — the particulate inheritance this chapter sums over; Ch 26 — HWE supplies \mathrm{Var}(X) = 2pq, the denominator of every formula here; Ch 29 — equilibrium is what lets variances add across loci; Ch 09 — is why are not transmitted
- Forward to: Ch 31 turns the relative-covariance table into estimates and the ; Ch 32 finds the individual loci; Ch 51 estimates \alpha for millions of variants at once; Ch 53 sums estimated and lives or dies on the