Genetics: Published Articles Ahead of Print, published on March 29, 2010 as 10.1534/genetics.110.

114249

Bayesian inference of genetic parameters based on

conditional decompositions of multivariate normal
distributions
Jon Hallander∗ , Patrik Waldmann§ , Chunkao Wang∗∗ and Mikko J. Sillanpää†,‡

∗
Department of Forest Genetics and Plant Physiology,
SE-90183 Swedish University of Agricultural Sciences, Umeå, Sweden

†
Department of Mathematics and Statistics,
Rolf Nevanlinna Institute,
FIN-00014 University of Helsinki, Helsinki, Finland

‡
Department of Agricultural Sciences,
FIN-00014 University of Helsinki, Helsinki, Finland

§
Thetastats,
SE-22471 Uardavägen 91, Lund, Sweden

∗∗
Department of Animal Science,
Iowa State University, Ames, Iowa, 50011-3150

January 15, 2010

1
Running head: Conditional decompositions
Keywords: WinBUGS, Mixed linear model, Genetic evaluation,
Complex pedigree, Markov chain Monte Carlo

Corresponding author: Jon Hallander

Department of Forest Genetics and Plant Physiology,
SE-90183 Swedish University of Agricultural Sciences, Sweden

Telephone number: +46 (0)90 786 8280

Fax number: +46 (0)90 786 8165
E-mail: [email protected]

2
 ABSTRACT

It is widely recognized that the mixed linear model is an important tool for parameter es-

timation in the analysis of complex pedigrees, which include both pedigree and genomic

information, and where mutually dependent genetic factors are often assumed to follow

multivariate normal distributions of high dimension. We have developed a Bayesian sta-

tistical method based on the decomposition of the multivariate normal prior distribution

into products of conditional univariate distributions. This procedure permits computation-

ally demanding genetic evaluations of complex pedigrees, within the user-friendly computer

package WinBUGS. In order to demonstrate and evaluate the exibility of the method, we

analyzed two example pedigrees: a large non-inbred pedigree of Scots pine (Pinus sylvestris

L.) that includes additive and dominance polygenic relationships; and a simulated pedigree

where genomic relationships have been calculated based on a dense marker map. The anal-

ysis showed that our method was fast and provided accurate estimates, and that it should

therefore be a helpful tool for estimating genetic parameters of complex pedigrees quickly

and reliably.

3
Much eort in genetics has been devoted to revealing the underlying genetic architecture of

quantitative or complex traits. Traditionally, the polygenic model has been used extensively

to estimate genetic variances and breeding values of natural and breeding populations, where

an innite number of genes is assumed to code for the trait of interest (Bulmer 1971; Falconer

and Mackay 1996). The genetic variance of a quantitative trait can be decomposed into

an additive part that corresponds to the eects of individual alleles, and a part that is

non-additive because of interactions between alleles. Attention has generally been focused

on the estimation of additive genetic variance (and heritability), since additive variation

is directly proportional to the response of selection via the breeder's equation (Falconer

and Mackay 1996, chap 11). However, in order to estimate additive genetic variation and

heritability accurately, it can be important to identify potential non-additive sources in

genetic evaluations (Misztal 1997; Ovaskainen et al. 2008; Waldmann et al. 2008), especially

if the pedigree being analyzed contains a large proportion of full-sibs and clones, as these

in particular give rise to non-additive genetic relationships (Lynch and Walsh 1998, pp

145). The polygenic model using pedigree and phenotypic information i.e. the animal model

(Henderson 1984) has been the model of choice for estimating genetic parameters in breeding

and natural populations (Abney et al. 2000; Sorensen and Gianola 2002; O0 Hara et al. 2008).

Recent breakthroughs in molecular techniques have made it possible to create genome-

wide, single nucleotide polymorphism (SNP) maps. These maps have helped to uncover a

vast amount of new loci responsible for trait expression and have provided general insights

into the genetic architecture of quantitative traits (e.g. Valdar et al. 2006; Visscher 2008;

Flint and Mackay 2009). These insights can help when calculating disease risks in humans,

4
when attempting to increase the yield from breeding programs, and when estimating re-

latedness in conservation programs. High density SNPs of many species of importance to

science and agriculture can now be scored quickly and relatively cheaply, for example in mice

(Valdar et al. 2006), chickens (Muir et al. 2008), and dairy cattle (VanRaden et al. 2009).

In the analysis of populations of breeding stock, the inclusion of dense marker data has

improved the predictive ability (i.e. reliability) of genetic evaluations compared to the tra-

ditional phenotype model, both in simulations (Meuwissen et al. 2001; Calus et al. 2008;

Hayes et al. 2009) and when using real data (Legarra et al. 2008; VanRaden et al. 2009;

González-Recio et al. 2009). Meuwissen et al. (2001) suggested that the eect of all markers

should rst be estimated, and then summed, in order to obtain genomic estimated breeding

values (GEBVs). An alternative procedure, where all markers are used to compute the ge-

nomic relationship matrix (in place of the additive polygenic relationship matrix) has also

been suggested (e.g. Villanueva et al. 2005; VanRaden 2008; Hayes et al. 2009); this matrix

is then incorporated into the statistical analysis to estimate GEBVs. A comparison of both

procedures (VanRaden 2008) yielded similar estimates of GEBVs in cases where the eect

of an individual allele was small. In addition, if not all pedigree members have marker in-

formation, a combined relationship matrix derived from both genotyped and ungenotyped

individuals could be computed; this has been shown to increase the accuracy of GEBVs

(Legarra et al. 2009; Misztal et al. 2009). Another plausible option to incorporate marker

information is to use low-density SNP panels within families, and to trace the eect of SNPs

from high-density genotyped ancestors, as suggested by Habier et al. (2009) and Weigel

et al. (2009). However, fast and powerful computer algorithms, which can use the marker

5
information as eciently as possible in the analysis of quantitative traits, are needed in order

to obtain accurate GEBVs from genome-wide marker data.

The present study describes the development of an ecient Bayesian method for incorpo-

rating general relationships into the genetic evaluation procedure. The method is based on

expressing the multivariate normal prior distribution as a product of one-dimensional normal

distributions, each conditioned on the descending variables. When evaluating the genetic pa-

rameters of natural and breeding populations, high dimensional distributions are often used

as prior distributions of various genetic eects, such as the additive polygenic eect (Wang

et al. 1993), multivariate additive polygenic eects (Van Tassell and Van Vleck 1996), and

quantitative trait loci (QTL) eects via the identical-by-decent matrix (Yi and Xu 2000).

A Bayesian framework is adopted in order to obtain posterior distributions of all unknown

parameters, estimated by using Markov chain Monte Carlo (MCMC) sampling algorithms

in the software package WinBUGS (Lunn et al. 2000 ; 2009 ). By performing prior calcu-

lations in the form of the factorized product of simple univariate conditional distributions,

the computational time of the MCMC estimation procedure is reduced considerably. This

feature permits rapid inference for both the polygenic model and the genomic relationship

model. Moreover, the decomposition allows for inbreeding of varying degree, since the cor-

rect genetic covariance structure can be inferred into the analysis. In the present paper,

we test the method on two previously published pedigree datasets: phenotype data from

a large pedigree of Scots pine, incorporating information on both additive and dominance

genetic relationships (Waldmann et al. 2008); and genomic information obtained from a

genome-wide scan of a simulated animal population (Lund et al. 2009).

6
 METHODS

Statistical model: Following Henderson (1984), we made use of the following linear
mixed eect model under Gaussian assumptions:

y = Xb + Zu + e, (1)

where: y is a vector of size n × 1 containing phenotypic records of a continuous trait

for all members in the population; b is a vector of size p × 1 containing systematic en-

vironmental eects; u is a vector of size n × 1 containing genetic eects that follow a

multivariate normal distribution with zero mean vector, and covariance structure Gσu2 ;

X and Z are known incidence matrices relating phenotypic records to respective loca-
tion parameters included in (1); and e is a vector containing independent residual er-

rors that follow a multivariate normal distribution with zero mean vector, and covari-

ance structure Iσe2 , where I is the identity matrix of order n. Note that records can be

missing for some pedigree members (here, yi =0 NA0 if individual i has a missing record) .

Typically, u contains the additive polygenic eect, although non-additive genetic eects

such as dominance, or QTL eects estimated from marker data, could be included in

the model. To make inferences in model (1), the mixed model equations (MMEs) can

be formed, so that y is associated to Gσu2 . A well-known, ecient Bayesian technique

for the estimation of genetic parameters in the linear mixed eect model is Markov chain

7
Monte Carlo (MCMC) methods (e.g. Wang et al. 1993; Sorensen and Gianola 2002; Bauer

et al. 2009).

The most common distribution used for u in model (1) is the multivariate normal distri-

bution (Lynch and Walsh 1998, pp 194), since u contains n variables (u1 , u2 , . . . , un ), that

on their own are assumed to be normally distributed. The multivariate normal is, therefore,

a natural choice of distribution for u. For example, the traditional polygenic model relies

on normal distribution assumptions for the Mendelian inheritance of genes from parents to

ospring (Bulmer 1971). In addition, when using Bayesian inference to estimate parameters

in model (1), the property of the multivariate normal distribution helps to form conditional

distributions, which are of key importance in MCMC sampling (Sorensen and Gianola 2002;

Rue and Held 2005). The hierarchical structure of model (1) can be usefully interpreted

as a graphical model, which facilitates computations because this representation allows the

joint distribution of genetic eects and other parameters to be broken down into products

of local components (Lauritzen et al. 1990). In the present paper, genetic eects (u) are

assumed to follow a Gaussian distribution with an imposed dependency structure given by

either the pedigree, estimated relatedness from markers; or both. The factorization of the

dependency structure in the graph gives (1) a Markov property (Lauritzen et al. 1990),

which can be successfully utilized in Bayesian MCMC methods. See Rue and Held (2005)

for a comprehensive survey of this topic; and see Steinsland and Jensen (2010) for how to

use the Markov property for making inference in the classical animal model. In addition, the

standard decomposition procedure of the additive polygenic eect (Thomas 1992; Lin 1999;

Waldmann 2009), utilizes the Markov property of the animal model, where an ospring is

8
conditioned on its parents in the analyzed pedigree.

Conditional expressions of multivariate normal distributions: Drawing samples

from multivariate normal distributions quickly becomes computationally demanding as the

number of parameters becomes large, e.g. when performing genetic evaluations of complex

pedigrees. A reasonable alternative, therefore, is to decompose the multivariate normal

distribution into conditionally dependent parts. If we replace the multivariate normal prior,

with the product of these (lower dimensional) distributions, both the mean and the variance

are shifted, for each conditional distribution. Let us assume that we wish to decompose u

into two subsets of column vectors, uT = [uT1 uT2 ], where u1 and u2 are of length l and
n − l, respectively. The mean and variance can be expressed as:
· ¸ · ¸
E (u1 ) V ar (u1 ) Cov (u1 , u2 )
E (u) = and V ar (u) = , (2)
E (u2 ) Cov (u1 , u2 ) T
V ar (u2 )

where V ar (u) = Gσu2 is believed to be a positive denite, symmetric matrix of order q . By

decomposing the multivariate normal distribution (Jensen 1998), we obtain:

E (u2 |u1 ) = E (u2 ) + (u1 − E (u1 ))T V ar (u1 )−1 Cov (u2 , u1 ) , (3)

V ar (u2 |u1 ) = V ar (u2 ) − Cov (u2 , u1 )T V ar (u1 )−1 Cov (u2 , u1 ) . (4)

Hence, the distribution of u1 conditional on u2 is multivariate normal according to

u2 | u1 ∼ M V N (E(u2 | u1 ), V ar(u2 | u1 )). (5)

More generally, for u = {u1 , u2 , u3 , . . . , uN }, where N is the number of partitions of u

9
(typically the number of members in the pedigree, i.e. N = n ), we have
N
Y
p(u1 , u2 , u3 , . . . , uN | σu2 ) = p(ui |ui−1 , . . . , u1 , σu2 ) = p(u1 | σu2 )p(u2 | u1 , σu2 )p(u3 | u2 , u1 , σu2 ) . . .
i=1
(6)

where σu2 is the variance component of u. Our target is to generate u, which is a realization

from a multivariate normal distribution with given mean (vector of zeros) and covariance

structure Gσu2 . The vector u is here generated by element-wise draws, ui : (i = 1, 2, . . . , N ),

from univariate (lower-dimensional) normal distributions conditioned on all the elements

that have been drawn so far, i.e., from p(ui | u1 , u2 , . . . , ui−1 , σu2 ). It should be emphasized

that this sequential strategy is exact and will lead to the correct vector u, drawn from the

full multivariate normal distribution M V N (0, Gσu2 ). First, let us assume individual-wise

partitions for u. Conditional expectation and conditional variance of p(ui |u1 , . . . , ui−1 , σu2 )

are thought of here as the weighted mean and the weighted variance for pedigree member i.

To compute weights for the mean (for individual i), the following general expression can be

used:
i−1
X
W (i) = w(i, j)uj . (7)
j=1

The precalculated weights are then read into WinBUGS together with the data. The code

for the weights in the model only includes one indexed univariate normal distribution with

conditional mean (eqn. 7) and variance (eqn. 4) as a prior for ui . Hence, for every pedi-

gree member i, we have one vector w(i, j) : j = 1, 2, . . . , (i − 1) calculated for the mean

where most of the terms are zero; this feature yields a sparse format which is suitable for

storing the weights. The weights for the mean (w(i, j)) and variance, which specify the con-

10
ditional prior distribution of each individual, need to be calculated only once and are thus

computed outside the MCMC estimation (i.e. before compilation of the code). The order of

the weights is important since the drawing of samples needs to follow the same unique order

throughout the simulation process. Furthermore, it is also possible to use the same principle

to update the parameters in blocks, sampling from multiple multivariate normal distribu-

tions, each of small dimension. When estimating the additive polygenic eect, the approach

proposed here gives identical results to the standard decomposition suggested earlier (e.g.

Thomas 1992; Lin 1999; Waldmann 2009) for non-inbred pedigrees (non-related parents).

Our proposed method could, however, incorporate non-zero covariances between parents,

and inbreeding coecients greater than zero, if complex relationships between relatives aris-

ing from dominance are neglected (e.g. Abney et al. 2000). Two small numerical examples

for computing a realization of additive and dominance polygenic eects, and illustrating the

eect of inbreeding on the additive polygenic covariance structure are given in the Appendix.

Bayesian inference of the factorized model: Following Waldmann (2009), uniform

distributions were assigned as priors to the standard deviations, since this prior distribution

was shown to be a good and robust non-informative choice of prior for variance compo-

nents (Gelman 2006). This prior corresponds to an inverse-χ2 distribution with −1 degrees

of freedom, that is p(σi2 ) ∝ σi−1 , i = u, e. To obtain an upper boundary for the uniform

distributions, we performed a preliminary analysis in which uninformative inverse-gamma

distributions were assigned as priors for the variance components, thereby, obtaining esti-

mates of standard deviations; these were then multiplied by 5 to obtain the upper bounds.

The chosen upper bounds were considerably larger than the upper bounds of the 95% highest

11
probability density (HPD; Box and Tiao 1973) regions obtained in the preliminary analysis.

Note that this procedure is a pragmatic solution and should not be viewed as a strictly

Bayesian solution. A at, noninformative prior was assigned to the systematic environmen-

tal xed eect in both examples, as bj ∼ N (0, 106 ) for systematic eect level j . For u, we used

the common multivariate normal distribution as prior: u ∼ M V N (0, Gσu2 ) (Sorensen and

Gianola 2002), and then used the decomposition of the multivariate normal distribution into
Q
univariate normal distributions, proposed here as p(u|σu2 ) = qi=1 p(ui |u1 , u2 , . . . , ui−1 , σu2 ).

The vector y is assumed to follow a Gaussian distribution; thus the likelihood function is

given as:
n
Y Y ½ ¾
1 (yi − bj − ui )2
p(y|b, u , σe2 ) = p(yi |bj , ui , σe2 ) = √ exp − , (8)
i=1 i∈O
2πσe 2σe2

where bj is the corresponding systematic eect level (covariate) of pedigree member i, con-

nected through X ; and O is the set of members in the pedigree for which phenotypic

records are available. If individual i has a missing record, note that p(yi |bj , ui , σe2 ) = 1. If

conditioning on hyper-parameters is neglected, and the location parameters are believed to

be independent a priori, the joint distribution of all parameters conditional on the data is

proportional to:

p(b, u, σu2 , σe2 |y) ∝ p(b)p(u|σu2 )p(σu2 )p(σe2 )p(y|b, u, σe2 ). (9)

The phenotype model (1) was run in the Bayesian software package WinBUGS (Lunn

et al. 2000) version 1.4.3, which is freely available on //www.mrc-bsu.cam.ac.uk/bugs/.

WinBUGS exploits a graphical modelling technique to translate the supplied prior distribu-

tions of the parameters into corresponding full conditional distributions. The computation

12
of the weights (7) using (3) and (4) was executed in ANSI C. The WinBUGS code is available

in electronic supplement S1, while the computer code used for calculating weights is available

in electronic supplement S2. To check mixing properties of our implementation in WinBUGS

and compare the mixing to alternative implementations, we calculated the eective sample

size (ESS) of the obtained MCMC chains (Kass et al. 1998; Waagepetersen et al. 2008).

ESS can be seen as the number of independent samples from the estimated posterior which

contain equivalent amount of information (i.e. exceed same estimation accuracy) than our

dependent MCMC samples. Low ESS values indicate poor mixing (i.e. high auto-correlation

between consequtive samples) in the MCMC-chain.

Polygenic example: To verify our proposed decomposition method, data acquired from
a 26-year-old eld trial of Scots pine (Pinus sylvestris L.), previously published by Wald-

mann et al. (2008), Finley et al. (2009), Hallander and Waldmann (2009), and Waldmann

(2009), were analyzed to obtain posterior distributions of additive and dominance polygenic

eects for all trees in the pedigree. The pedigree consists of 52 parents crossed according

to a partial diallel design resulting in mixed half-sib and full-sib families totaling 4970 sur-

viving ospring. The parents were assumed to be non-related and non-inbred. In total, 202

families were distributed over approximately 4 ha of forest. The eld trial was subdivided

into 70 square (or nearly square) blocks, which were used in the subsequent evaluations as a

systematic environmental eect. Several traits of interest for breeding purposes were mea-

sured in 1997, although for the present study we chose to analyze only trunk diameter at

breast height (DBH). The mean value of DBH was 114 mm. We made use of the following

covariance structure in the mixed linear model: G1 σu21 = Aσa2 and G2 σu22 = Dσd2 , where A

13
is the additive relationship matrix; σa2 is the variance component of additive genetic eects;

D is the dominance relationship matrix; and σd2 is the variance component of dominance
genetic eects. A and D were calculated using standard equations (Lynch and Walsh 1998,

pp 763 and 768). Uniform densities were assigned as prior distributions for σa and σd as

described in the previous section. See Waldmann et al. (2008), Finley et al. (2009) and

Hallander and Waldmann (2009), for a more detailed analysis of the Scots pine pedigree.

Genomic example: This was a simulated dataset, typical of the data acquired from an
animal breeding protocol, consisting of 5865 pedigree members from seven generations. The

dataset is freely available from the QTLMAS XII workshop webpage:

http://www.computationalgenetics.se/QTLMAS08/QTLMAS/DATA.html (Lund et al. 2009).

There are 6000 loci evenly distributed over six chromosomes (1000 markers per chromosome),

with 0.1 cM between markers. Forty-eight QTLs, each of small eect, were assumed to code

for the trait of interest. Pedigree and phenotype information were available from the rst four

generations of animals. The animals from generations ve to seven had no given phenotype,

but did have complete marker information. From each generation, 15 males and 150 females

were randomly selected and mated according to a hierarchical mating design, resulting in a

total of 1500 animals being born per generation. Interested readers are invited to consult

Lund et al. (2009) for further details of the dataset. The genomic relationship matrix (or

realized relationship matrix) was computed using the second method proposed by VanRaden

(2008) which follows G = ZDZ0 . In the present paper, Z = M − P, where: M is a matrix

of size n × m containing genotypes at all m loci for all n members of the pedigree; P is

a matrix of the same size containing allele frequencies that dier from 0.5 at all loci; and

14
nally, D is a diagonal matrix of size m × m with diagonal elements 1
m[2pi (1−pi )]
where pi is

the allele frequency at locus i. We analyzed a subset of the complete pedigree consisting of

the rst four generations (1014 animals in total) in order to reduce computational time. See

electronic Table S1 for original identication numbers of pedigree members in the analyzed

sub-population.

RESULTS

Polygenic example: To facilitate a comparison between the results obtained from our
proposed method and the results reported by Waldmann et al. (2008), we performed the same

procedure as Waldmann et al. (2008) when analyzing the Scots pine pedigree. One MCMC

was run for a total of 225,000 iterations, from which the rst 25,000 iterations were omitted

(burn-in) from the analysis, and every 10th iteration was saved (thinned), resulting in an ef-

fective sample of 20,000 iterations. Table 1 shows the results of the analysis using our method,

together with the results from Waldmann et al. (2008) for the trait DBH. Our posterior

point-estimates and their 95 % HPD regions closely agree with those obtained by Waldmann

et al. (2008), although slightly dierent degrees of freedom for the inverse-χ2 distributions

used as prior to the variance components (−1 in our method while Waldmann et al. (2008)

used −2 in theirs) could cause some dierences in the respective posterior distributions. We

believe, however, that the priors have little inuence on the parameter estimates obtained

from the analyzed data, mainly due to both the large size of the pedigree and the similar

15
parameter estimates obtained by Waldmann et al. (2008) and Finley et al. (2009) using

dierent priors.

The additive and dominance covariance structures resulted in 9,940 and 61,247 non-

zero weights, respectively. For variance components and heritability, 95% HPD intervals

were estimated using the R library, 'boa' (Smith 2007). In WinBUGS, each scan of the

MCMC took 0.4241 seconds on an AMD Opteron(tm) Dual Core Processor (2.39 GHz) with

1 GB of RAM. The corresponding average time, for each MCMC scan, using the method in

Waldmann et al. (2008) was 1.840 seconds. In each iteration in the MCMC procedure, all

non-zero weights for means need to be multiplied by the corresponding genetic parameter

of the preceding pedigree members (a matrix-vector multiplication: see equation 7). The

actual number of non-zero weights depends on the covariance structure, as more non-zero

relationships will result in a higher number of non-zero weights. For the Scots pine data,

little computational time was needed to obtain reliable posterior estimates due to the leve

l of sparseness of the additive and dominance polygenic relationship matrices. Consequently,

the total computational time is greatly reduced in the current example because of the few

non-zero weights. Hence, our proposed method seems to be very benecial for analyzing

polygenic data.

The slightly lower eective sample size (ESS; sample size adjusted for auto-correlation)

obtained by our method reects the fact that single-site Gibbs sampling is performed in

WinBUGS, while the hybrid Gibbs sampler, proposed by Waldmann et al. (2008) also

applies block updating of parameters, which is known to improve the mixing of the MCMC

chain. In addition, the transformation of genetic covariance structures applied in Waldmann

16
 et al. (2008) improves the ESS further. However, in this case, the marginally lower ESS of

the current method is well compensated for by the improved speed. Furthermore, we tested

two additional updating options (block hybrid and conjugated multivariate) in OpenBUGS

version 3.0.3 (Thomas et al. 2006), to see whether the ESS was improved. Only the ESS of

dominance genetic eects was improved while the ESS of the other parameters was unaected

or even decreased (results not shown). On the other hand, the computing time increased

markedly: we therefore believe that the standard updating (multivariate forward) option in

WinBUGS/OpenBUGS gives acceptable mixing of the MCMC. Furthermore, we investigated

whether changing the sequential order of conditioning would generate a dierent number of

weights and, thereby, a dierence in computational time. However, the number of weights

remained the same, regardless of the sequence, and we therefore can conclude that the order

is unimportant from a computational perspective.

Genomic example: First, a purely additive polygenic model was used to obtain initial
values of the variance components for all 225,000 iterations. A preliminary analysis was then

conducted in order to estimate the standard deviations of the variance components, which

were used to set the upper limit of the uniform prior distributions. The MCMC procedure was

run for 23,000 scans in total, from which the rst 3,000 scans were discarded to give 20,000

saved iterations. The heritability estimates obtained are shown in Table 2. The posterior

point-estimates from our analyzed subset agree closely with the true corresponding parameter

values given in Lund et al. (2009), both for posterior mean and mode. However, the additive

polygenic model resulted in heritability point-estimate being too large although with both

models, true value is included within the estimated 95% HPD regions. In addition, in Table 2,

17
it is important to note that 95% HPD region obtained by a genomic model is clearly more

narrow than the one given by a polygenic model, suggesting that the inclusion of the genomic

relationship matrix improves the estimation accuracy of heritability. Similar conclusions have

been drawn by, for example, Meuwissen et al. 2001, Villanueva et al. (2005) and Hayes

et al. (2009). For a thorough explanation of improved accuracy, see Xu (2006).

The computational eort in the genomic example was unfortunately massive, due to the

large number of non-zero weights (in total: 513,591). On the same computer as that used for

the polygenic example, each scan took 39.40 seconds. Initially, we truncated small elements in

the genomic relationship matrix in order to reduce the number of non-zero elements, but this

modication eectively prevented convergence of the MCMC chain. Truncating small values

in the resulting weight matrix, instead of the genomic matrix itself, might be a more suitable

procedure because the individual weights for computing the variance of each genetic parameter

(i.e. each normal distribution) are correctly calculated given the genomic covariance matrix,

and only weights used for computing the mean of each normal distribution are aected.

Conversely, if elements in the genomic covariance structure are truncated before the

decomposition procedure, then weights for computing the means and the variances of the

normal distributions of all parameters are aected. By truncating small elements in the weight

matrix, the computational time could potentially be greatly reduced, although the accuracy

of estimated posterior and 95% HPD regions would be negatively aected. However, some

preliminary experiments on truncating weights have given promising results (i.e. the obtained

posterior estimates were only slightly aected), although we chose to include the full weight

matrix when analyzing the simulated data (other results not shown).

18
 DISCUSSION

Decomposition of high dimensional multivariate normal distributions provides a very exible

Bayesian method that allows us to make inferences in linear mixed eect models with a large

number of genetic parameters. For example, the proposed method can be used for the following:

variance component based linkage and association mapping methods for the estimation of QTL

eects; estimating non-additive genetic eects, such as dominance and epistasis; estimation

of genomic breeding values; estimation of maternal and permanent environmental eects,

which are important in breeding evaluations. The approach was implemented in the user

friendly computer software WinBUGS, and was shown to be fast and accurate on both real

and simulated data. By using this approach, researchers will be able to perform advanced

genetic evaluations of complex traits and pedigrees without possessing advanced knowledge

in animal models and programming.

Recently, several studies have shown that the accuracy of genetic evaluations can be in-

creased by incorporating the genomic relationship matrix (Villanueva et al. 2005; Misztal et

al. 2009; Hayes et al. 2009). For large, complex pedigrees with a high number of polymor-

phic markers, the resulting genomic relationship matrix will probably be dense, i.e. most

pedigree members will have non-zero, pair-wise estimated relatedness. A general problem

with this approach is that when making inferences in animal models, most currently avail-

able methods rely on either sparse solvers (e.g. Schaeer and Kennedy 1986; Johnson and

19
Thompson 1995; Waldmann et al. 2008) or on ecient graph model techniques (Wilkinson

and Yeung 2004; Rue and Held 2005; Steinsland and Jensen 2010). Compared to standard,

non-sparse methods, these methods will not result in the same reduction in computational

time when incorporating sparse covariance structures (i.e. when using pedigree information

only). Unfortunately, even though our proposed method can handle dense genetic covari-

ance structures, as demonstrated in the genomic example, the computational time required

is massive. One way to overcome this hurdle would be to truncate small elements in the

weight matrix obtained by our approach, thereby obtaining a good approximation of the

genetic covariance matrix . An alternative method would be to apply transformation on

the covariance matrix to facilitate the inference of parameter estimation as, for example,

suggested by Mrode and Thompson 1989, Wilkinson and Yeung (2004) and Waldmann et al.

(2008). As a result, estimating parameters using the linear mixed model does not depend

on the sparsity of the genetic covariance structure.

A good mixing property of the MCMC method is very important in Bayesian analysis in

order to obtain reliable posterior estimates, especially if parameters are highly correlated in

the model. In Gibbs sampling, the updater samples from the fully conditional posterior dis-

tribution, which is proportional to the likelihood function and the prior distribution through

Bayes theorem. Our proposed method samples from the multivariate normal distribution for

the prior, but samples from the likelihood function are taken for one parameter at a time,

which introduces dependencies to the posterior (i.e. introduces higher correlation between

parameters in the posterior as these are drawn element-wise). Thus, the mixing property

of our algorithm does not match the mixing properties achieved with block updating of pa-

20
rameters, where sampling from both prior distribution and the likelihood are performed in

a block (García-Cortés and Sorensen 1996; Roberts and Sahu 1997). On the other hand, for

element-wise or single-site updating (without decomposition), sampling from both likelihood

and prior are made for each parameter, which introduces heavy dependencies to the posterior

distribution, resulting in poor mixing properties of the MCMC chain (Sorensen and Gianola

2002). Hence, our method should result in better mixing than single site updating but result

in less eective mixing than block updating of parameters. This insight is conrmed, to some

extent, empirically when the mixing property in our approach was compared to the mixing

property of the standard single-site sampler (Sorensen and Gianola 2002) implemented in C

(results not shown). However, this comparison should be interpreted bearing in mind that

WinBUGS uses an expert system that attempts to utilize the most appropriate sampling

scheme for each stochastic node (Lunn et al. 2000, pp 328).

If a better mixing property is warranted, one plausible alternative might be to com-

bine our suggested decomposition approach with block updating of parameters into a hy-

brid sampler. A similar approach (i.e. combining single-site and block sampling) was im-

plemented by Waldmann et al. (2008), which resulted in better mixing of the MCMC

chain than was obtained in pure single-site updating. However, it would not be possible

to implement the combined sampling approach in WinBUGS, as the block updating re-

quires a large equation system to be solved during each iteration in the MCMC procedure.

If only single-site sampling is possible (due to computational limitations), the parameters

can be randomly updated in each MCMC step and not in the same sequential order, as

suggested by Levine and Casella (2006). An additional alternative to improve mixing is

21
 to apply transformation of the location parameters in the model (Vine et al. 1996;

Waldmann et al. 2008).

The lack of freely available computer packages designed for the genetic evaluation of com-

plex pedigrees using a Bayesian framework has, unfortunately, prevented more regular use

of these models. The graphical model representation within the Bayesian software package

WinBUGS is very well suited for decomposition of joint distributions into products of local

components, i.e. parent and ospring nodes in a directed acyclical graph (DAG) (Lunn et al.

2000 ; 2009 ). WinBUGS also applies an intelligent, automatic approach to the choice of up-

daters needed in the implementation of the MCMC procedure. Both Damgaard (2007) and

Waldmann (2009) successfully executed the animal model in WinBUGS and produced re-

sults that show how evaluating the genetics of complex pedigrees can be performed smoothly

without the need of expensive hardware and software. As an extension to these studies, we

have shown how general relationship structures can be decomposed and hence be eciently

implemented in WinBUGS.

One important improvement oered by our proposed method, compared to the standard

factorization method of additive polygenic eects (e.g. Thomas 1992; Lin 1999; Waldmann

2009), is the ability to obtain realizations from the correct additive covariance structure of

inbred populations. If such populations are analyzed with the standard factorization model,

additive variance, and consequently heritability, can be overestimated, depending on the

level of inbreeding and the size of the pedigree. Problems with handling inbred pedigrees

arise with the standard model because the covariances between parents are assumed to be

zero. Since it is not uncommon to have some degree of inbreeding in both breeding and

22
natural populations of animals and plants, the standard factorization of additive polygenic

relationships can be erroneous. Dierences in the estimated posterior of the polygenic vari-

ance components, obtained by the standard factorization model and our approach, need

further verication in extensive computer simulations. It should be noted that the non-

additive genetic relationships, introduced by inbreeding, can have a considerable inuence

on the genetic variances ( Harris 1964; Abney et al. 2000).

We have, in the present study, demonstrated the benet of decomposing the multivariate

normal distribution, often used as prior for genetic eects in the standard, linear mixed eect

model. To our knowledge, this procedure of decomposing the prior distribution has not been

utilized before, in the context of parameter estimation in genetics. The decomposion approach

was put forward and excecuted successfully in WinBUGS by Vines et al. (1996); they utilized

a random eect model to analyze clinical data in the context of epidemiology. However,

they did not include covariance between random eects, which makes the decomposition

procedure more complex. In general, there also exist alternative procedures for ecient

implementing the prior distribution, which deserves more attention. One such alternative,

which is likely to be equally ecient than the approach presented here, is obtained by

multiplying vector of univariate normal distributed variables by a Cholesky factor (square

root) of the original covariance matrix (Golub and Van Loan 1996). As in our approach,

(Cholesky) weights can be calculated once, prior to the WinBUGS analysis. Both procedures

involves one matrix-vector multiplication each iteration in the MCMC process and are,

therefore, likely to be computational equally time consuming for analysis of large pedigrees.

23
 ACKNOWLEDGMENTS

This work was supported by funding from: the Research School of Tree Breeding and Forest

Genetics for JH and PW; Föreningen svensk skogsträdsförädling for CW and PW; and re-

search grants from the Academy of Finland and the University of Helsinki's Research Funds

for MJS. The Scots pine data used in the polygenic example were provided by the Swedish

Forest Research Institute, Skogforsk. In addition, we wish to thank the associate editor and

three anonymous reviewers for comments that improved the manuscript.

LITERATURE CITED

Abney, M., M. S. McPeek and C. Ober, 2000 Estimation of variance components of quanti-

tative traits in inbred populations. Am. J. Hum. Genet. 66: 629-650.

Bauer, A. M., T. C. Reetz, F. Hoti, W. -D. Schuh, J. Léon et al., 2009 Bayesian prediction

of breeding values by accounting for genotype-by-environment interaction in self-pollinating

crops. Genet. Res. 91: 193-207.

Box, G. E. P., and G. C. Tiao, 1973 Bayesian Inference in Statistical Analysis. Wiley, New

York.

Bulmer, M. G., 1971 Eect of selection on genetic variability. Am. Nat. 105: 201-210.

Calus, M. P. L., T. H. E. Meuwissen, A. P. W. de Roos and R. F. Veerkamp, 2008 Accuracy

24
of genomic selection using dierent methods to dene haplotypes. Genetics 178: 553-561.

Damgaard, L. H., 2007 How to use Winbugs to draw inferences in animal models. J. Anim.

Sci. 85: 1363-1368.

Falconer, D. S., and T. F. C. Mackay, 1996 Introduction to Quantitative Genetics. Longman,

New York.

Finley, A. O., S. Banerjee, P. Waldmann and T. Ericsson, 2009 Hierarchical spatial modeling

of additive and dominance genetic variance for large spatial trial datasets. Biometrics 65:

441-451.

Flint, J., and T. F. C. Mackay, 2009 Genetic architectures of quantitative traits in ies, mice

and human. Genome Res. 19: 723-733.

García-Cortés, L. A., and D. Sorensen, 1996 On a multivariate implementation of the Gibbs

sampler. Genet. Sel. Evol. 28: 121-126.

Gelman, A., 2006 Prior distributions for variance parameters in hierarchical models. Bayesian

Anal. 1: 515-534.

Golub, G. H., and C. F. Van Loan, 1996 Matrix Computations, 3rd Edition. Johns Hopkins

University Press, Boltimore.

González-Recio, O., D. Gianola, G. J. Rosa, K. A. Weigel and A. Kranis, 2009 Genome-

assisted prediction of a quantitative trait measured in parents and progeny: application to

food conversion rate in chickens. Genet. Sel. Evol. 41: 3.

Habier, D., R. L. Fernando, and J. C. M. Dekkers, 2009 Genomic selection using low-density

marker panels. Genetics 182: 343-353.

Hallander, J., and P. Waldmann, 2009 Optimum contribution selection in large general tree

25
breeding populations with an application to Scots pine. Theor. Appl. Genet. 118: 1133-

1142.

Harris, D. L., 1964 Genotypic covariances between inbred relatives. Genetics 50: 1319-1348.

Hayes, B. J., P. M. Vissher and M. E. Goddard, 2009 Increased accuracy of articial selection

by using the realized relationship matrix. Genet. Res. 91: 47-60.

Henderson, C. R., 1984 Applications of Linear Models in Animal Breeding. University of

Guelph, Guelph, ON, Canada.

Jensen, D. R., 1998 Multivariate normal distribution. In: Armitage P, Colton T, editors.

Encyclopedia of Biostatistics. Chichester, Wiley. pp. 2906-2911.

Johnson, D. L., and R. Thompson, 1995 Restricted maximum likelihood estimation of vari-

ance components for univariate animal models using sparse matrix techniques and average

information. J. Dairy Sci. 87: 449-456.

Kass, R. E., B. P. Carlin, A. Gelman, and R. Neal, 1998. Markov Chain Monte Carlo in

practice: a roundtable discussion. Am. Stat. 52: 93-100.

Lauritzen, S. L., A. P. Dawid, B. N. Larsen and H.-G. Leimer, 1990 Independence properties

of directed Markov elds. Networks 20: 491-505.

Legarra, A., I. Aguilar and I. Misztal, 2009 A relationship matrix including full pedigree and

genomic information. J. Dairy Sci. 92: 4656-4663.

Legarra, A., C. Robert-Granie, E. Manfredi and J. M. Elsen, 2008 Performance of genomic

selection in mice. Genetics 180: 611-618.

Levine, R. A., and G. Casella, 2006 Optimizing random scan Gibbs samplers. J. Mult. Anal.

97: 2071-2100.

26
Lin, S., 1999 Monte Carlo Bayesian methods for quantitative traits. Comp. Stat. Data

Anal. 31: 89-108.

Lund, M. S., G. Sahana, D. J. de Koning, G. Su and Ö. Carlborg, 2009 Comparison of

analyses of the QTLMAS XII common dataset. I: Genomic selection. BMC Proc. 3: S1.

Lunn, D. J., A. Thomas, N. Best and D. Spiegelhalter, 2000 WinBUGS - a Bayesian mod-

elling framework: concepts, structure, and extensibility. Stat. Comp. 10: 325-337.

Lunn, D. J., D. Spiegelhalter, A. Thomas and N. Best, 2009 The BUGS project: evolution,

critique and future directions. Stat. Med. 28: 3049-3067.

Lynch, M., and B. Walsh, 1998 Genetics and Analysis of Quantitative Traits. Sinauer Asso-

ciates, Sunderland, MA.

Meuwissen, T. H. E., B. J. Hayes and M. E. Goddard, 2001 Prediction of total genetic value

using genome-wide dense marker maps. Genetics 157: 1819-1829.

Misztal, I., 1997 Estimation of variance components with large-scale dominance models. J.

Dairy Sci. 80: 965-974.

Misztal, I., A. Legarra and I. Aguilar, 2009 Computing procedures for genetic evaluation

including phenotypic, full pedigree, and genomic information. J. Dairy Sci. 92: 4648-4655.

Mrode, R., and R. Thompson, 1989 An alternative algorithm for incorporating the relation-

ships between animals in estimating variance components. J. Anim. Breed. Genet. 106:

89-95.

Muir, W. M., K. S. G. Wong, Y. Zhang, J. Wang, M. A. M. Groenen et al., 2008 Genome-

wide assessment of worldwide chicken SNP genetic diversity indicates signicant absence of

rare alleles in commercial breeds. Proc. Natl. Acad. Sci. USA 105: 17312-17317.

27
O0 Hara R. B., J. M. Cano, O. Ovaskainen, C. Teplitsky, and J. S. Alho, 2008 Bayesian ap-

proaches in evolutionary quantitative genetics. J. Evol. Biol. 21: 949-957.

Ovaskainen, O., J. M. Cano and J. Merilä, 2008 A Bayesian framework for comparative

quantitative genetics. Proc. Roy. Soc. B 275: 669-678.

Roberts, G. O., and S. K. Sahu, 1997 Updating schemes, correlation structure, blocking and

parameterization for the Gibbs sampler. J. R. Stat. Soc. Ser. B 170: 419-431.

Rue, H., and L. Held, 2005 Gaussian Markov Random Fields: Theory and Applications.

Chapman and Hall, London.

Schaeer, L. R., and B. W. Kennedy, 1986 Computing solutions to mixed model equations.

J. Dairy Sci. 69: 575-579.

Smith, B. J., 2007 boa: an R package for MCMC output convergence assessment and poste-

rior inference. J. Stat. Soft. 21: 1-37.

Sorensen, D., and D. Gianola, 2002 Likelihood, Bayesian and MCMC Methods in Quantita-

tive Genetics. Springer-Verlag, New York.

Steinsland, I., and H. Jensen, 2010 Utilising Gaussian Markov Random Field properties of

Bayesian animal models. Biometrics in press.

Thomas, A., R. B. O0 Hara, U. Ligges and S. Sturtz, 2006 Making BUGS Open. R News 6:

12-17.

Thomas, D. C. 1992 Fitting genetic data using Gibbs sampling - an application to nevus

counts in 38 Utah kindreds. Cytogenet. Cell Genet. 59: 228-230.

Waagepetersen, R., N. Ibánêz-Escriche, D. Sorensen, 2008 A comparison of strategies for

Markov chain Monte Carlo computation in quantitative genetics. Genet. Sel. Evol. 40:

28
161-176.

Valdar, W., L. C. Solberg, D. Gauguier, S. Burnett, P. Klenerman et al., 2006 Genome-wide

genetic association of complex traits in heterogeneous stock mice. Nat. Genet. 38: 879-887.

Waldmann, P. 2009 Easy and exible Bayesian inference of quantitative genetic parameters.

Evolution 63: 1640-1643.

Waldmann, P., J. Hallander, F. Hoti and M. J. Sillanpää, 2008 Ecient Markov Chain Monte

Carlo implementation of Bayesian inference of additive and dominance genetic variances in

non-inbred pedigrees. Genetics 179: 1101-1112.

Wang, C. S., J. J. Rutledge and D. Gianola, 1993 Marginal inference about variance com-

ponents in a mixed linear model using Gibbs sampling. Genet. Sel. Evol. 21: 41-62.

VanRaden, P. M., 2008 Ecient methods to compute genomic predictions. J. Dairy Sci. 91:

4414-4423.

VanRaden, P. M., C. P. Van Tassell, G. R. Wiggans, T. S. Sonstegard, R. D. Schnabel et

al., 2009 Reliability of genomic predictions for North American Holstein bulls. J. Dairy Sci.

92: 16-24.
Van Tassell, C. P., and L. D. Van Vleck, 1996 Multiple-trait Gibbs sampler for animal mod-

els: exible programs for Bayesian and likelihood-based (covariance) component inference.

J. Anim. Sci. 74: 2586-2597.

Weigel, K. A., G. de los Campos, O. González-Recio, H. Naya, X. L. Wu et al., 2009 Pre-

dictive ability of direct genomic values for lifetime net merit of Holstein sires using selected

subsets of single nucleotide polymorphism markers. J. Dairy Sci. 92: 5248-5257.

Villanueva, B., R. Pong-Wong, J. Fernández and M. A. Toro, 2005 Benets from marker-

29
assisted selection under an additive polygenic genetic model. J. Anim. Sci. 83: 1747-1752.

Wilkinson, D. J., and S. K. H. Yeung, 2004 A sparse matrix approach to Bayesian compu-

tation in large linear models. Comp. Stat. Data Anal. 44: 493-516.

Vines, S. K., W. R. Gilks and P. Wild, 1996 Fitting Bayesian multiple random eects models.

Stat. Comp. 6: 337-346.

Visscher, P. M., 2008 Sizing up human height variation. Nat. Genet. 40: 489-490.

Xu, S., 2006 Separating nurture from nature in estimating heritability. Heredity 97: 256-

257.

Yi, N., and S. Xu, 2000 Bayesian mapping of quantitative trait loci under the identity-by-

descent-based variance component model. Genetics 156: 411-422.

30
 APPENDIX

Calculation of weights for dominance relationships: We would like to calculate

conditional prior distributions for dominance genetic eects from the pedigree given in Table

3. The corresponding dominance polygenic relationship matrix, D, can be obtained as

described in, for example, Waldmann et al. (2008) as

 
1 0 0 0
 0 1 0 0 
D=  0 0 1 0.25
.

0 0 0.25 1

For individual 1: E(d1 | σd2 ) = 0, V ar(d1 ) = σd2 and d1 | σd2 ∼ N (0, σd2 ).

For individual 2: E(d2 | d1 , σd2 ) = E(d2 ) + [d1 ][ σ12 ][0] = 0, V ar(d2 | d1 , σd2 ) = V ar(d2 ) −
d

[0][ σ12 ][0] and d2 | d1 , σd2 ∼ N (0, σd2 ).

d

" 1
#· ¸
σd2
0 0
For individual 3: E(d3 | d1 , d2 , σd2 ) = E(d3 ) + [d1 d2 ] = 0, V ar(d3 |
0 σ12 0
" 1
#· ¸ d

σd2
0 0
d1 , d2 , σd2 ) = V ar(d3 ) − [0 0] 1 = σd2 and d3 | d1 , d2 , σd2 ∼ N (0, σd2 ).
0 σ 2 0
d

Finally, for individual 4:  

1
σd2
0 0  0 
 1
0 
E(d4 | d1 , d2 , d3 , σd2 ) = E(d4 ) + [d1 d2 d3 ]  0 σd2   0  = 14 d3 , V ar(d4 |
1 2
0 0 σ2 1 σ
4 d
 1
 d

σd2
0 0 0
1 2  0 1 
d1 , d2 , d3 , σd2 ) = V ar(d4 ) − [0 0 σ ]
4 d  σd2
0  0  = σd2 − 16
1 2
σd = 15 σ 2 and
16 d
1 2
0 0 1
σd2
σ
4 d
d4 | d1 , d2 , d3 , σd2 ∼ N ( 14 d3 , 15 σ 2 ).
16 d

31
For individual 4, both mean and variance are shifted (reduced) because individuals 3 and

4 are full sibs. The following weights for reduction in mean are, consequently, obtained

for individual 4: w(4, 1) = 0, w(4, 2) = 0 and w(4, 3) = 0.25. Using equation (7), we

obtain W (4) = 0.25d3 . Hence, instead of drawing d from M V N (0, Dσd2 ), we make use of

the following univariate normal distributions: d1 | σd2 ∼ N (0, σd2 ), d2 | d1 , σd2 ∼ N (0, σd2 ),

d3 | d1 , d2 , σd2 ∼ N (0, σd2 ) and d4 | d1 , d2 , d3 , σd2 ∼ N ( 41 d3 , 15 σ 2 ).

16 d

Eect of inbreeding on calculation of additive weights: In Table 4, an example

pedigree is shown that includes a loop; this will cause pedigree member 6 to be inbred. The

corresponding additive relationship matrix is

 
1 0 0 0.5 0 0.25
 0 1 0 0.5 0.5 0.5 
 
 0 0 1 0 0.5 0.25 
A=  .
0.5 0.5 0 1 0.25 0.625 
 
 0 0.5 0.5 0.25 1 0.625 
0.25 0.5 0.25 0.625 0.625 1.125

When obtaining realization from additive polygenic eects of pedigree members 1 to 5, our

proposed method and the standard factorization method (e.g., Lin 1999), give exactly the

same mean and variance used in the normal univariate distributions as a1 , a2 , a3 ∼ N (0, σa2 ),
2 2
2 σa 2 σa
a4 ∼ N ( a1 +a
2
, 2 ) and a5 ∼ N ( a3 +a
2
, 2 ). However, for pedigree member 6, the standard
2
5 σa
factorization method yields a6 ∼ N ( a4 +a
2
, 2 ). Our proposed method, on the other hand,

gives

E(a6 | a1 , a2 , a3 , a4 , a5 , σa2 ) = E(a6 ) + [a1 a2 a3 a4 a5 ]

32
   −1  
1 0 0 0.5 0 0.25
 0 1 0 0.5 0.5    0.5 
  2   2
 0 0 1 0 0.5  σa   0.25  σa = a4
+ a5
,
     2 2
 0.5 0.5 0 1 0.25    0.625 
0 0.5 0.5 0.25 1 0.625

V ar(a6 | a1 , a2 , a3 , a4 , a5 , σa2 ) = V ar(a6 ) − [0.25 0.5 0.25 0.625 0.625]

 −1  
1 0 0 0.5 0 0.25
 0 1 0 0.5 0.5   
   0.5  2 3σ2
 0 0 1 0 0.5   0.25 
 
 σa = 8 .
a

 0.5 0.5 0 1 0.25   0.625 
0 0.5 0.5 0.25 1 0.625

Using our proposed method, for individual 6, we make use of the following normal uni-
2
5 3σa
variate distribution: a6 | a1 , a2 , a3 , a4 , a5 , σa2 ∼ N ( a4 +a
2
, 8 ). Consequently, the weight for

the variance component diers between our method (3/8) and the standard method (1/2);

this will cause a6 to be sampled from an incorrect distribution if the standard method is

applied to the current pedigree.

33
 TABLE 1
Summary statistics including posterior estimates (mode, mean and median) and effective
effective sample size (ESS) obtained from the WinBUGS analysis in the polygenic model
example for additive genetic variance ( ), dominance genetic variance ( ), residual
variance ( ), heritability ( ) and dominance proportion ( ). The results of the
decomposition approach are denoted DEC while the results of Waldmann et al. (2008) are
denoted HYB including both MCMC estimates and restricted maximum likelihood (REML)
estimates
Mode Mean Median 95 % HPD region ESS REML

Paramet DEC HYB DEC HYB DEC HYB DEC HYB DEC HYB HYB
er

53.16 54.70 63.29 62.52 60.49 59.62 [29.47, [27.67, 376.4 417.2 55.95
103.1] 103.7]

77.56 82.88 84.69 88.41 82.5 85.71 [39.86, [39.70, 399.2 456.8 83.06
136.9] 142.2]

726.4 722.2 724.3 721.7 725.0 722.6 [670.7, [665.3, 733.2 756.1 728.5
778.3] 776.8]

0.0617 0.0630 0.0724 0.0714 0.0694 0.0685 [0.0340, [0.0327, 372.2 420.9 0.0645
0.1158] 0.1170

0.0894 0.0939 0.0970 0.1014 0.0946 0.1 [0.0463, [0.0500, 394.5 447.5 0.0957
0.1567] 0.1616]
 TABLE 2
Posterior estimates obtained from the WinBUGS analysis in the genomic model example for
heritability ( ) using a model including genomic relationship matrix ( ) and a model
including additive polygenic relationship matrix ( ). The true value of for the entire
pedigree should be 0.3 (Lund et al. 2009)
Model Mode Mean Median 95 % HPD region
0.3040 0.2980 0.2997 [0.2553, 0.3384]
0.3376 0.3418 0.3397 [0.2191, 0.4691]
 TABLE 3
Example pedigree for sampling of dominance polygenic effects where a 0 indicate that an
individual has an unknown father or mother
Individual Father Mother
1 0 0
2 0 0
3 1 2
4 1 2
 TABLE 4
Example pedigree for sampling of additive polygenic effects where a 0 indicate that an
individual has an unknown father or mother

Individual Father Mother

1 0 0
2 0 0
3 0 0
4 1 2
5 3 2
6 4 5