Lecture 4

BLUP Breeding Values

Guilherme J. M. Rosa
University of Wisconsin-Madison

Mixed Models in Quantitative Genetics

SISG, Seattle
18 – 20 September 2018

Linear Mixed Effects Model

y = Xβ + Zu + e
responses residuals
incidence fixed random
matrices effects effects

⎡u⎤ ⎛ ⎡0⎤ ⎡G 0 ⎤ ⎞
⎢e ⎥ ~ MVN⎜ ⎢ ⎥, ⎢
⎜ 0 0 Σ⎥ ⎟
⎟
⎣ ⎦ ⎝⎣ ⎦ ⎣ ⎦⎠

1
 Estimation of Fixed Effects

y = Xβ + ε
with ε = Zu + e , such that Var[ε] = ZGZ T + Σ

è MLE for β :

βˆ = ( X T V −1X) −1 X T V −1y ~ MVN(β, ( X T V −1X) −1 )

where V = ZGZ T + Σ

Prediction of Random Effects

⎡y ⎤ ⎛ ⎡Xβ⎤ ⎡ V ZG ⎤ ⎞
⎢u⎥ ~ MVN⎜ ⎢ ⎥, ⎢ T
⎜ 0 ⎥ ⎟
⎣ ⎦ ⎝ ⎣ ⎦ ⎣GZ G ⎦ ⎟⎠

E[u | y] = E[u] + Cov[u, y T ]Var −1[y](y − E[y])

= GZT V −1 (y − Xβ) = GZT (ZGZT + Σ) −1 (y − Xβ)
Replacing β by its estimate:

uˆ = GZ T (ZGZ T + Σ) −1 (y − Xβˆ )

2
 Mixed Model Equations

⎡X T Σ −1X X T Σ −1Z ⎤ ⎡βˆ ⎤ ⎡X T Σ −1y ⎤

⎢ T −1 T −1 −1 ⎥ ⎢ ⎥
= ⎢ T −1 ⎥
⎣ Z Σ X Z Σ Z + G ⎦ ⎣uˆ ⎦ ⎣ Z Σ y ⎦

BLUP and BLUE:

uˆ = (Z T Σ −1Z + G −1 ) −1 Z T Σ −1 (y − Xβˆ )

β̂ = {X T [Σ−1 − Σ−1Z(Z T Σ−1Z + G−1 )−1 Z T Σ−1 ]X}−1

× X T [Σ−1 − Σ−1Z(Z T Σ−1Z + G−1 )−1 Z T Σ−1 ]y

Mixed Models in Animal and

Plant Breeding

Animal/plant breeding programs are based on the

principle that phenotypic observations on related
individuals can provide information about their
underlying genotypic values

The additive component of genetic variation is the

primary determinant of the degree to which
offspring resemble their parents, and therefore
this is usually the component of interest in
artificial selection programs

3
 Mixed Models in Animal and
Plant Breeding

Many statistical methods for analysis of genetic

data are specific (or more appropriate) for
phenotypic measurements obtained from planned
experimental designs and with balanced data sets

While such situations may be possible within

laboratory or greenhouse experimental settings,
data from natural populations and agricultural
species are generally highly unbalanced and
fragmented by numerous kinds of relationships

Animal Model
Culling of data to accommodate conventional statistical
techniques (e.g. ANOVA) may introduce bias and/or lead
to a substantial loss of information
The mixed model methodology allows efficient estimation
of genetic parameters (such as variance components and
heritability) and breeding values while accommodating
extended pedigrees, unequal family sizes, overlapping
generations, sex-limited traits, assortative mating, and
natural or artificial selection
To illustrate such application of mixed models in breeding
programs, we consider here the so-called Animal Model in
situations with a single trait and a single observation
(including missing values) per individual

4
 Animal Model
The animal model can be described as:

y = Xβ + Zu + e
y is an (n × 1) vector of observations (phenotypic scores)
β is a (p × 1) vector of fixed effects (e.g. herd-year-
season effects)
u ~ N(0, G) is a (q × 1) vector of breeding values (relative
to all individuals with record or in the pedigree file,
such that q is in general bigger than n)
e ~ N(0, Inσe2) represents residual effects, where σe2 is
the residual variance

The Matrix A
The matrix G describing the covariances among the
random effects (here the breeding values) follows
from standard results for the covariances between
relatives
It can be shown that the additive genetic covariance
between two relatives i and i’ is given by 2θ ii ' σ a2 ,
where θii' is the coefficient of coancestry between
2
individuals i and i’, and σ a is the additive genetic
variance in the base population
Hence, under the animal model, G = Aσ a2 , where A
is the additive genetic (or numerator) relationship
matrix, having elements given by a ii' = 2θii'

5
 The Matrix A

For each animal i in the pedigree (i = 1, 2,…,n), going from

older to younger animals, compute aii and aij (j = 1, 2,…,i-1)
as follows:
If both parents (s and d) of animal i are known:
aij = aji = (ajs + ajd)/2 and aii = 1 + asd/2
If only one parent (e.g. d) of animal i is known:
aij = aji = ajd/2 and aii = 1
If parents unknown:
aij = aji = 0 and aii = 1

Animal Sire Dam

Example 1 - -
2 - -
1 2 3 1 2
4 1 -
5 4 3
4 3 6 5 2

5 6

pedigree matrix A

6
 Animal Model

In general, in animal/plant breeding interest is

on prediction of breeding values (for selection
of superior individuals), and on estimation of
variance components and functions thereof,
such as heritability
The fixed effects are, in some sense, nuisance
factors with no central interest in terms of
inferences, but which need to be taken into
account (i.e., they need to be corrected for
when inferring breeding values)

Animal Model
Since under the animal model G −1 = A −1σ a−2 and
R −1 = I n σ e−2 , the mixed model equations can be
expressed as:
# T &# & # XTy &
% X X XTZ (% β̂ (=% (
%$ Z T X Z T Z + λA −1 ('$ û (' %$ Z T y
% (
'
σ e2 1− h 2
where λ = = 2 , such that:
σ a2 h
−1
" % " T % " T %
$ β̂ ' = $ X X XTZ ' $ X y '
$# û '& $# Z T X Z T Z + λA −1 '& $ ZT y '
# &

7
Conditional on the variance components ratio λ, the
BLUP of the breeding values are given then by:

uˆ = (Z T Z + λA −1 ) −1 Z T (y − Xβˆ )
These are generally referred to as Estimated Breeding
Values (EBV)

Alternatively, some breeders associations express

their results as Predicted Transmitting Abilities (PTA)
(or Estimated Transmitting Abilities (ETA) or Expected
Progeny Difference (EPD)), which are equal to half the
EBV, representing the portion of an animal’s breeding
values that is passed to its offspring

The amount of information contained in an animal’s

genetic evaluation depends on the availability of
its own record, as well as how many (and how close)
relatives it has with phenotypic information

As a measure of amount of information in livestock

genetic evaluations, EBVs are typically reported
with its associated accuracies

Accuracy of predictions is defined as the

correlation between true and estimated breeding
values, i.e., ri = ρ(û i , u i )

Instead of accuracy, some livestock species

genetic evaluations use reliability, which is the
squared correlation of accuracy (ri2 )

8
 Prediction Accuracy
The calculation of ρ(û i , u i ) requires the diagonal
elements of the inverse of the MME coefficient
matrix, represented as:
−1
# T & # ββ &
X X XTZ ( =% C Cβu (
C =% T
%$ Z X Z T Z + λA −1 (' %$ C uβ C uu ('

It can be shown that the prediction error variance of

EBV û i is given by:

PEV = Var (û i − u i ) = c iuu σ e2

uu
where c i is the i-th diagonal element of C uu , relative
to animal i.

Prediction Accuracy

The PEV can be interpreted as the fraction of

additive genetic variance not accounted for by
the prediction

Therefore, PEV can be expressed also as:

PEV = (1 − ri2 )σ a2
uu 2 2 2
such that c i σ e = (1 − ri )σ a , from which the
reliability is obtained as:

ri2 = 1 − c iuu σ e2 / σ a2 = 1 − λc iuu

9
 Animal Model
herd 1

herd 2
⎡ u1 ⎤
⎢ ⎥
⎡310⎤ ⎡1 0⎤ ⎡1 0 0 0 0⎤ ⎢u 2 ⎥ ⎡ e1 ⎤
⎢270⎥ = ⎢1 0⎥ ⎡ h1 ⎤ + ⎢0 0 1 0 0⎥ ⎢ u ⎥ + ⎢e ⎥
⎢ ⎥ ⎢ ⎥ ⎢h ⎥ ⎢ ⎥⎢ 3 ⎥ ⎢ 3 ⎥
⎢⎣350⎥⎦ ⎢⎣0 1⎥⎦ ⎣ 2⎦
⎢⎣0 0 0 1 0⎥⎦ ⎢u 4 ⎥ ⎢⎣e 4 ⎥⎦
⎢⎣ u 5 ⎥⎦

y = X β + Z u + e

Animal Model
2
Breeding values: u ~ N (0, Aσ u ) , with

⎡ 1 0 0.5 0.5 0.25 ⎤

⎢ 0 1 0 0.5 0 ⎥⎥
⎢
A = ⎢ 0.5 0 1 0.25 0.5 ⎥
⎢ ⎥
⎢ 0.5 0.5 0.25 1 0.125⎥
⎢⎣0.25 0 0.5 0.125 1 ⎥⎦

−1
" % " T % " T %
$ β̂ ' = $ X X XTZ ' $ X y '
$# û '& $# Z T X Z T Z + λA −1 '& $ ZT y '
# &
σ e2 1− h 2
λ= = 2
σ 2u h

10
 R Code
animal model
y<-matrix(c(310,270,350),nrow=3)
X<-matrix(c(1,1,0,0,0,1),nrow=3) toy example
Z<-matrix(c(1,0,0,0,0,0,0,1,0,0,0,0,0,1,0),nrow=3, byrow = TRUE)
A<-matrix(c(1,0,0.5,0.5,0.25,
0,1,0,0.5,0,
0.5,0,1,0.25,0.5,
0.5,0.5,0.25,1,0.125,
0.25,0,0.5,0.125,1),nrow=5)

h2<-1/3 # heritability ⎧ ĥ1 = 290

a=(1-h2)/h2 ⎪
⎪ ĥ 2 = 348
# crossproducts ⎪ û = 4.0
XX<-crossprod(X,X)
1 ⎪ 1
XZ<-t(X) %*% Z
h2 = →α=2 ⎨ û 2 = 0.0
ZX<-t(Z) %*% X
ZZ<-crossprod(Z,Z)+a*solve(A) 3 ⎪ û = −4.0
⎪ 3
# mixed model equations ⎪ û 4 = 2.0
# coefficient matrix and right hand side ⎪
C<-rbind(cbind(XX,XZ),cbind(ZX,ZZ)) ⎩ û 5 = −2.0
rhs<-rbind(t(X) %*% y,t(Z) %*% y)

#solution
theta.hat <- solve(C) %*% rhs

Animal Model
The animal model can be extended to model multiple
(correlated) traits, multiple random effects (such as
maternal effects and common environmental effects),
repeated records (e.g. test day models), and so on
Example (Mrode 1996, pp74-76): Weaning weight (kg)
of piglets, progeny of three sows mated to two boars:

11
A linear model with the (fixed) effect of sex, and the
(random) effects of common environment (related to
each litter) and breeding values can be expressed as X:

y = Xβ + Zu + Wc + e
Weight Residual
Sex Breeding Common
values environment

Assuming that σ 2u = 20 , σ c2 = 15 and σ e2 = 65 , the MME

are as follows:
# T &# β̂ & # X T y &
% X X XTZ XTW (% ( % (
% Z T X Z T Z + A −1λ1 T
Z W (% û ( = % Z T y (
% (% ( % (
%$ W X
T
WTZ W T W + Iλ 2 ('% ĉ ( % W T y ('
$ ' $
σ e2 σe2
where λ1 = 2
= 3.25 and λ 2 = 2
= 4.3!
σu σc

The BLUEs and BLUPs

(inverting the numerator
relationship matrix) are:

Mrode example

12