3
$\begingroup$

I am trying to model the phenotype of a trait as $X = G + E$, where $G$ and $E$ are the genetic and environmental effects. (I'll ignore the distinction between broad-sense and narrow-sense heritability in this question.) Wikipedia tells me that $$h^2 = \frac{\mathrm{Var}(G)}{\mathrm{Var}(X)}$$ The same page also gives the breeder's equation $R = h^2S$ but also says "Note that heritability in the [breeder's equation] is equal to the ratio [${\mathrm {Var}}(G)/{\mathrm {Var}}(X)$] only if the genotype and the environmental noise follow Gaussian distributions." So let $G \sim N(0, \sigma_G^2)$ and $E \sim N(0, \sigma_E^2)$ be distributed according to normal distributions. Then $X \sim N(0, \sigma_G^2 + \sigma_E^2)$ so $$h^2 = \frac{\mathrm{Var}(G)}{\mathrm{Var}(X)} = \frac{\sigma_G^2}{\sigma_G^2 + \sigma_E^2}$$

Now I want to perform a selection, choosing a parent with $X=a$. We can expect the child to have the trait $\mathbb E[G \mid X=a] + E'$ where $E' \sim N(0, \sigma_E^2)$ is some other environmental effect. (I don't want to bother with recombination and multiple parents, so we can just think of it as cloning the parent: given that $X=a$, we expect the parent's genetic part to have been $\mathbb E[G \mid X=a]$.)

To find $\mathbb E[G \mid X=a]$, note that $\mathbb E[X \mid X=a] = a$ and by linearity of expectation, $\mathbb E[X \mid X=a] = \mathbb E[G \mid X=a] + \mathbb E[E \mid X=a]$. To get the terms to cancel, we want to write $E$ as $\alpha E$ where $\alpha E$ has the same distribution as $G$. Since $\alpha E \sim N(0, \alpha^2 \sigma_E^2)$, we want $\alpha^2 \sigma_E^2 = \sigma_G^2$, so $\alpha = \sigma_G/\sigma_E$. Thus we have $$\begin{align}a &= \mathbb E[G \mid X=a] + \frac1\alpha\mathbb E[\alpha E \mid X=a] \\ &= \mathbb E[G \mid X=a] + \frac1\alpha\mathbb E[G \mid X=a] \\ &= \left(1+\frac1\alpha\right)\mathbb E[G \mid X=a]\end{align}$$ Hence $$\mathbb E[G \mid X=a] = \frac{a}{1+\frac1\alpha} = \frac{a}{1+\frac{\sigma_E}{\sigma_G}}$$

Since the parent was selected at $X=a$ and the child has mean $\frac{a}{1+\frac{\sigma_E}{\sigma_G}}$, we have $$h^2=\frac{R}{S} = \frac{1}{1+\frac{\sigma_E}{\sigma_G}} = \frac{\sigma_G}{\sigma_G+\sigma_E}$$

Now my problem is that the two expressions for $h^2$ don't match. They look pretty similar, but one of them has all the terms squared while the other one doesn't. What's going on/where did I go wrong?

ETA: To give a concrete example, suppose 80% of variance in the trait is explained by genes. Then the breeder's equation predicts the child to have $0.8a$. But computing $\mathbb E[G \mid X=a]$ gives $a/(1 + \sqrt{0.8^{-1} - 1}) = (2/3)a$. So the two approaches predict different things (although if it was 50% they would give the same prediction).

$\endgroup$
1
$\begingroup$

I found my mistake, which is where I said "To get the terms to cancel, we want to write $E$ as $\alpha E$ where $\alpha E$ has the same distribution as $G$." What we actually want is to find a constant $\alpha$ such that $\mathbb E[\alpha E\mid X=a] = \mathbb E[G\mid X=a]$. In my question, I assumed that this was the same thing as finding $\alpha$ such that $\mathbb E[\alpha E] = \mathbb E[G]$; hence I was trying to find $\alpha$ to make the distributions of $\alpha E$ and $G$ the same (the means are both 0, so I just needed to get the variances to equal each other). But actually, conditioning on $X=a$ shifts the distributions so the $\alpha$ I found in my question doesn't work.

Since my previous strategy doesn't work, I ended up skipping the step of finding $\alpha$ entirely. I realized that the random vector $(G, X)=(G,G+E)$ is a bivariate normal, because we can write $$\begin{pmatrix}G \\ G+E\end{pmatrix} = \begin{pmatrix}\sigma_G & 0\\ \sigma_G & \sigma_E\end{pmatrix}\begin{pmatrix}Z_1 \\ Z_2\end{pmatrix}$$ where $Z_1,Z_2$ are standard normal random variables.

Now Wikipedia gives the conditional distribution for bivariate normals, which in our case is $G\mid X=a$. We only need the mean, which is $\mu_G + \frac{\sigma_G}{\sigma_X}\rho (a - \mu_X)$. Both $\mu_G$ and $\mu_X$ are zero, and $\rho$ is the correlation which is $$\rho = \rho_{G,X} = \frac{\mathrm{cov}(G,X)}{\sigma_G \sigma_X} = \frac{\mathrm{cov}(G,G) + \mathrm{cov}(G,E)}{\sigma_G \sigma_X} = \frac{\sigma_G^2}{\sigma_G \sigma_X} = \frac{\sigma_G}{\sigma_X}$$ Thus we end up with $$\mu_G + \frac{\sigma_G}{\sigma_X}\rho (a - \mu_X) = \frac{\sigma_G^2}{\sigma_X^2}a$$ which is exactly what we want.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.