INFORMATION TO USERS
This reproduction was made from a copy of a manuscript sent to us for publication
and microfilming. While the most advanced technology has been used to pho-
tograph and reproduce this manuscript, the quality of the reproduction is heavily
dependent upon the quality of the material submitted. Pages in any manuscript
may have indistinct print. In all cases the best available copy has been filmed.
The following explanation of techniques is provided to help clarify notations which
may appear on this reproduction.
1. Manuscripts may not always be complete. When it is not possible to obtain
missing pages, a note appears to indicate this.
2. When copyrighted materials are removed from the manuscript, a note ap-
pears to indicate this.
3. Oversize materials (maps, drawings, and charts) are photographed by sec-
tioning the original, beginning at the upper left hand corner and continu-
ing from left to right in equal sections with small overlaps. Each oversize
page is also filmed as one exposure and is available, for an additional
charge, as a standard 35mm slide or in black and white paper format. *
4. Most photographs reproduce acceptably on positive microfilm or micro-
fiche but lack clarity on xerographic copies made from the microfilm. For
an additional charge, all photographs are available in black and white
standard 35mm slide format.*
♦For more information about black and white slides or enlarged paper reproductions,
please contact the Dissertations Customer Services Department.
UMI
Dissertation
Information Service
University Microfilms International
A Bell & Howell Information Company
300 N. Zeeb Road, Ann Arbor, Michigan 48106
8618794
Keele, John William
ESTIMATION OF (CO)VARIANCE COMPONENTS BY WEIGHTED AND
UNWEIGHTED SYMMETRIC DIFFERENCES SQUARED, AND SELECTED
MIVQUE'S: RELATIONSHIPS BETWEEN METHODS AND RELATIVE
EFFICIENCIES
The Ohio State University Ph.D. 1986
University
Microfilms
I n t£m &tl O PI Si 300 N. Zeeb Road, Ann Arbor, Ml 48106
ESTIMATION OF (CO>VARIANCE COMPONENTS BY WEIGHTED AND UNWEIGHTED
SYMMETRIC DIFFERENCES SQUARED, AND SELECTED MIVQUE's: RELATIONSHIPS
BETWEEN METHODS AND RELATIVE EFFICIENCIES
DISSERTATION
Presented in Partial Fulfillment of the Requirements for
the Degree Doctor of Philosophy in the Graduate
School of The Ohio State University
By
John William Keele, B.S., M.S.
* * * * *
The Ohio State University
1986
Dissertation Committee:
Walter R. Harvey Approved by
Francis R. Allaire
Michael E. Davis IjJoJt^ f?. jig /
Adviser
Keith M. Irvin Department of Dairy Science
iry 3c
ACKNOWLEDGMENTS
I would like to thank Walter R. Harvey for his suggestions,
endless support, patience, example and the idea to do this research.
Thanks are due to the other members of my committee, Drs. Frank R.
Allaire, Mike E. Davis, and Keith M. Irvin for their helpful
suggestions and comments. The painstaking task of typing this thesis
was accomplished by Debbie Gallagher to whom I am most grateful. David
and Kathy Keller deserve special thanks for making copies and providing
a place for me to stay while taking my final exam. I am grateful to
the Gilmore family for the Gilmore Award. I owe a debt of gratitude to
the Department of Dairy Science, The Ohio State University, and the
taxpayers of Ohio for my associateship and the privilege to study at
this great University. I would like to thank my wife, Wendy for her
understanding, faith, and willingness to put up with all of this. I
would also like to thank my son Ben for being himself and being a good
kid despite my inexperience as a parent. Most of all, I would like to
thank God who provided the sequence of events that made this work
possible.
11
VITA
April 2, 1957 Born - ChaJlis, Idaho
1979 B.S., University of Idaho,
Moscow, Idaho
1979-1982 Research Assistant, Department of
Animal Science, University of
Idaho, Moscow, Idaho
1982 M.S., University of Idaho,
Moscow, Idaho
1982-1986 Research Associate, Department of
Dairy Science, The Ohio State
University, Columbus, Ohio
PUBLICATIONS
Keele, J.W. 1982. Digestion and metabolism of diets with whole
cottonseed or extended soybeans fed to cows. Unpublished Masters
thesis. University of Idaho, Moscow, Idaho.
Keele, J.W. and R.E. Roffler. 1982. Nutrient digestion in cows fed
rations containing whole cottonseed or extended soybeans. J. Dairy
Sci. (suppl. 1) 65:137 (abstract).
in
TABLE OF CONTENTS
ACKN0WLED3EI1ENTS ii
VITA iii
LIST OF TABLES vi
LIST OF FIGURES vii
INTRODUCTION 1
REVIEW OF LITERATURE 6
VARIANCE COMPONENT ESTIMATION PROBLEM 6
MIVQUE AND REML FOR ANIMAL MODEL 7
NONADDITIVE RELATIONSHIP MATRICES 9
REDUCED ANIMAL MODEL 9
FEASIBILITY OF MIVQUE AND REML 10
METHODS MORE FEASIBLE THAN MIVQUE AND REML 11
COMPUTING A 14
UNWEIGHTED MIVQUE 14
CONCLUSIONS 16
THEORY 17
THE EQUIVALENCE OF MIVQUE AND WEIGHTED SYMMETRIC DIFFERENCES
SQUARED 17
1. Definition of Synmetric Differences Squared .... 20
2. Expectation of Symmetric Differences Squared .... 21
3. Synmetric Differences Squared Equations 23
4. Variance-Covariance Matrix Among Symmetric
Differences Squared 23
5. The Inverse of the Error Variance-Covariance Matrix
of Y 25
6. Weighted Synmetric Differences Squared Equations . . 26
iv
7. The Equivalence of SDS Weighted by the Inverse of
the Error Variance-Covariance Matrix and
MCVQUE(O) 27
8. The Equivalence of SDS Weighted by the Inverse of the
Total Variance-Covariance Matrix and MIVQUE . . 34
CX5MPUTAT00WAL REQUIREMENTS OF WSDS OR MIVQUE(O) 37
METHODOLOGY 49
RESULTS AND DISCUSSION 57
SUMMARY AND CONCLUSIONS 73
LITERATURE CITED 77
APPENDICES 80
A 80
B 84
C 92
v
LIST OF TABLES
TABLE PAGE
1. Storage, multiplications, and additions for each type of
element required to set up the MIVQUE(O) equations when
relationship categories are not considered 38
2. Multiplications, additions and storage needed to compute and
save the matrices needed to set up the WSDS equations when
relationship categories are considered 43
3. Multiplications, additions and storage required to save
matrices needed to set up WSDS and SDS equations for mice data
of Grimes and Harvey (1980) When relationship categories are
considered 46
4. Storage, multiplications, and additions required to compute
each type of element needed to obtain MIVQUE(O) or WSDS
equations for the data of Grimes and Harvey (1980) when
relationship categories are not considered 47
5. Comparison of standard errors of SDS estimates of variance
components obtained by simulation study of Grimes and Harvey
(1980) for a maternal effects model with standard errors
obtained by the numerical method of the current study .... 57
6. Efficiency of SDS estimates of (co) variance components
relative to MIVQUE estimates computed from mating designs of
Thompson (1976) and Eisen (1967), replicated 200 times each . 59
7. Efficiency of WSDS estimates of (co) variance components
relative to MIVQUE estimates computed from mating designs of
Thompson (1976) and Eisen (1967), replicated 200 times each . 62
8. Efficiency of MIVQUE(0,M,E) estimates of (co)variance
components relative to MIVQUE estimates computed from mating
designs of Thompson (1976) and Eisen (1967), replicated 200
times each 66
9. The efficiency of MIVQUE(l) estimates of (co)variance
components relative to MIVQUE estimates computed from 200 sets
of designs described by Thompson (1976) and Eisen (1967) . . 70
vi
LIST OP FIGURES
FIGURE PAGE
1. Two mating designs of Thompson (1976) 50
2. The 18 individuals comprising a set of an Eisen (1967) design 51
VI 1
UtfTRODUCTTON
Variance and covariance components and ratios of these are used to
characterize the type and quantity of genetic and environmental
variation present in a population of animals. In addition to being
unbiased and efficient (small sampling variance), it is desirable that
estimates of (co)variance components be computable with a reasonable
amount of computer time and memory. Even though the amount of computer
time and memory considered to be reasonable has increased with improved
technology, computational feasibility is still an important attribute
of variance component estimation in light of the large amount of data
usually needed to obtain relatively precise estimates. The challenge
is to choose from the computable unbiased methods, the most efficient
method.
The computational requirements for estimating genetic variance
components also depend on the type of family structures that are
present in the population. Populations of farm animals usually contain
many different types of relatives. One way to account for
relationships among animals is to use an animal model. In general,
analysis of variance (ANOVA) procedures are not adequate under the
animal model, although they have been used as ad hoc procedures in the
past (Dickerson, 1942; Hazel et al., 1943; Eisen, 1967).
Henderson (1985a and 1985b) showed how to obtain estimates of
additive and nonadditive genetic variance components by minimum
1
2
variance quadratic unbiased estimates (MIVQUE) (Pao, 1971) for assumed
prior values, and restricted maximum likelihood estimates (REML)
(Patterson and Thompson, 1981), under the animal model. Henderson's
approach for obtaining MIVQUE and REML require the inverse of one
matrix of order n and the g-inverse of another matrix of order n+p,
where n is the number of observations, and p is the number of fixed
effects. REML can be obtained by iterative MIVQUE, or by the
expectation maximization (EM) algorithm (Dempster et al„ 1977). Both
MIVQUE and REML are not computationally feasible for large (greater
than 500-1000 animals) data sets under the animal model.
When the genetic model is additive, the computational requirements
for REML and MIVQUE can be reduced considerably. With the Henderson
approach to obtaining MIVQUE and REML, the mixed model equations (MME)
(Henderson et al„ 1959) and best linear unbiased predictions of random
effects (BLUP) are obtained as intermediate results. Quaas and Pollack
(1980) showed that a reduced animal model (RAM) with MME only for
parents of progeny with records was equivalent to the full animal model
with MME for every animal. Quaas and Pollack (1980) described RAM for
a multitrait model appropriate for beef cattle, and Henderson (1985c)
described RAM for the unitrait case. Hudson and Kennedy (1985) were
able to obtain BLUP for parents and nonparents under RAM on swine data.
In this data set, there were 5.6 times as many animals as there were
parents with progeny records. Therefore, using RAM instead of the
animal model saves computer storage because fewer MME are needed.
However, even under RAM, the number of MME probably were too large to
obtain MIVQUE or REML estimates of variance components because Kennedy
3
et al. (1985) used a simplified model for data in which the animal
model was appropriate. For the swine data set of Hudson and Kennedy
(1985), there were 4,000 to 20,000 MME under RAM for the four breeds
considered. A need exists for relatively efficient variance component
estimation procedures that can be computed under RAM, or the animal
model if there are nonadditive genetic effects for large data sets.
Henderson (1985a) suggested an approximate MIVQUE called diagonal
MIVQUE or Henderson's Simple Method for data sets in which the g-
inverse of the left hand side of MME is too large to compute. The
relative efficiency of this method has not been well characterized
under the animal model, but Henderson (cited by Hudson and Van Vleck,
1982) found this method to be more efficient than method 3 of Henderson
for selected data sets in which method 3 could be used.
Another method of reducing the computations is to assume that all
variances except the error variance are zero prior to applying MIVQUE
(Rao, 1971). These estimates are MIVQUE if unknown variances except
the error variance actually are near zero. This method is known as
MIVQUE(O). However, MIVQUE(O) was found to be inefficient relative to
ANOVA when the variances other than error variance were larger than the
error variance for selected 2 stage nested designs (Brocklebank and
Giesbrecht, 1984).
Grimes and Harvey (1980) extended the symmetric sums of products
(SSP) method of Koch (1968) to the animal model. They used their
method, which they called symmetric differences squared (SDS) to
estimate genetic and environmental (co)variance components for weight
and gain traits of 1,780 mice. While this method required little
4
computer storage, it required a large amount of computer time.
Christian (1980) showed how to reduce the computer time needed to
perform SDS by grouping SDS resulting from pairs of individuals with
the same relationship before obtaining expectations. In addition,
Christian (1980) suggested that SDS could be corrected for
inquadmiss ability (uniformly less efficient than some other method) by
weighting SDS by the inverse of the error variance-covariance matrix
among SDS (WSDS). However, completing the analysis by WSDS as
described by Christian would require computational steps proportional
to n raised to a power of four, where n is the number of observations.
Setting some of the parameters other than the error variance to
zero prior to applying MIVQUE can result in a method that is less
computationally demanding than MIVQUE. For example, when permanent
maternal environmental effects are important, the permanent maternal
environmental and residual variance could be set to nonzero priors and
all other (co)variances set to zero. This method might have a higher
efficiency than MIVQUE(O), without greatly increasing the computational
requirements over that of MIVQUE(O). This method will be referred to
in this paper as MIVQUE(0,M,E).
In addition to reducing computational requirements of variance
component estimation methods, it is convenient to avoid the assignment
of prior values. Rao (1971) suggested that all priors be given a
value of one when nothing is known a priori about the (co)variance
components. This method will be referred to in this paper as
MLVQUE(l).
5
The objectives of this study are to: 1) demonstrate that SDS
weighted by the inverse of the error variance-covariance matrix among
SDS (WSDS) is equivalent to MIVQUE(O); 2) demonstrate that SDS weighted
by the inverse of the total variance-covariance matrix among SDS is
MIVQUE when the priors are correct; 3) present a computationally
feasible algorithm for obtaining WSDS; and 4) evaluate the influence of
the unknown parameters on the efficiency of SDS, WSDS, MIVQUE (0,M,E),
and MIVQUE(l) relative to MIVQUE.
REVIEW OF LITERATURE
VARIANCE COMPONENT ESTIMATION PROBLEM
Methods of estimating genetic and environmental (co)variance
components utilize records and the relationships among individuals who
produced the records. Farm populations have a family structure that
consists of many different types of relatives. One way to account for
these relationships when estimating {co)variance components is to use
the animal model (Henderson and Quaas, 1976; Quaas and Pollack, 1980;
Henderson, 1985a,b,c, ; Hudson and Kennedy, 1985). Analysis of variance
(ANOVA) procedures are not appropriate under the animal model because
the random part of the model can not be written as the sum of sets of
mutually uncorrelated random effects. When the animal model is
appropriate, ANOVA procedures should be weighted by the inverse of the
variance-covariance matrix among the estimated covariances among
relatives to be efficient (Eisen, 1967). This method is usually not
computationally feasible. The methods that can be used to estimate
(co) variance components assuming an animal model are (i) minimum
variance quadratic unbiased estimation (MIVQUE) (Rao, 1971); (ii)
maximum likelihood (ML) (Hartley and Rao, 1967); (iii) restricted
maximum likelihood (REML) (Patterson and Thompson, 1971); (iv)
symmetric differences squared (SDS) (Grimes and Harvey, 1980); and (v)
weighted symmetric differences squared (WSDS) (Christian, 1980). In
addition, various methods which are adaptations of MIVOJJE (Henderson,
6
1984, 1985a) can be applied to an animal model.
MIVQUE AND REML FOR ANIMAL MODEL
Henderson (1985a, b) showed how to obtain MIVQUE and REML
estimates of additive and nonadditive genetic variance components under
an animal model. Henderson first obtained best linear unbiased
A A
predictions (BLUP) for total genotypic merits (m) and residuals (e),
then expressed the quadratics needed for MIVQUE and REML as quadratic
functions of ta and e. This process avoids the need to obtain the
inverse of the n x n variance-covariance matrix among the observations
(V) where n is the number of observations. However, the inverse of the
a
variance-covariance matrix (M) among m is still needed to set up the
mixed model equations (MME) of Henderson et al. (1959). In general,
the inverse of M will be as difficult to obtain as the inverse of V
when there is one record per animal. Once the MME are obtained, m can
be calculated by iteration without finding the g-inverse of C, where C
is the left hand side of MME. Submatrices of the g-inverse of C are
A A
needed however to find the expectations of the quadratics in m and e.
A A
The quadratics in m and e are set equal to their expectations and the
resulting equations are solved to obtain MIVQUE. Computing the g-
inverse of C is costly because this matrix is of order n+p, where p is
the number of fixed constants in the model.
In practice, the parameters are not known; therefore guesses or
estimates by other methods serve as priors for the MIVQUE algorithm.
The closer the priors are to the true parameters the closer the
estimates will be to MIVQUE. However, the closeness of the priors to
the true parameters is unknown with experimental data.
8
REML can be obtained by iterative MIVQUE when only estimates in
the parameter space are allowed. The solution from the k 01 iterate of
iterative MIVQUE is obtained by applying MIVQUE using as priors the
solution from the (k-l) t " iterate. This process is continued until
prespecified convergence criteria are met. Estimates from the
literature, ANOVA estimates, MIVQUE(O) estimates, MIVQUE(l) estimates,
or guesses can be used as priors for the first iterate.
REML can also be obtained by applying an expectation maximization
(EM) (Dempster et al., 1977) algorithm. The EM algorithm results in
much simpler calculations at each iterate of REML than iterative
MIVQUE. The quadratics in m and e are equated to their expectations
under the pretext that the prior (co) variances are correct.
Advantages of the EM algorithm are that if prior values are
assigned from within the parameter space, estimates will not converge
outside the parameter space, and at each iterate the likelihood is
guaranteed to increase. A disadvantage of the EM algorithm is that it
converges very slowly. This can be a real problem if each iterate
requires 20 minutes to an hour of computer time. Iterative MIVQUE
converges more rapidly than EM, however, estimates outside the
parameter space occur, and computational requirements per iterate can
be much greater than EM algorithm. If the computational requirements
of an iterate of iterative MIVQUE are not too much greater than that of
an iterate of the EM algorithm, it is this author's opinion that
iterative MIVQUE should be the method chosen because iterative MIVQUE
usually converges in 3 to 5 iterations whereas the EM algorithm can
take well over 50 iterations to converge.
9
Swallow and Monahan (1984) compared MIVQUE with ANOVA estimates
used as priors (MIVQUE(A)) to REML by iterative MIVQUE for 10 different
number patterns for a random one-way classification using simulation.
They found that the efficiencies of MIVQUE(A) and REML were of such a
similar magnitude that the extra iterations required to complete REML
for convergence were not justified.
NONADDITIVE RELATIONSHIP MATRICES
Henderson (1985a) showed how the matrices of coefficients of
relationship for dominance (D), additive by additive epistasis (AA),
additive by dominance epistasis (AD), etc. could be obtained from
Wright's numerator relationship matrix (A) using results due to
Cockerham (1954) when no inbreeding has occurred. When inbreeding has
occurred, the nonadditive genetic effects are difficult to interpret
(Cockerham, 1954); therefore an additive genetic model is usually
assumed as an approximation when inbreeding exists. When an additive
genetic model is assumed, the computational requirements of MIVQUE or
REML can be reduced. The greatest reduction comes about because of the
structure of the inverse of A, and because the inverse of A can be
obtained directly without first finding A (Henderson, 1973; Quaas,
1976) .
REDUCED ANIMAL MODEL
Quaas and Pollack (1980) showed that a reduced animal model (RAM)
with equations in the MME only for parents who have progeny records is
equivalent to a full animal model with equations in the MME for all
animals. Nonparent BLUPs can be obtained from the BLUPs of their
parents, the herd-year-season effects, and the residuals. Their model
10
was a multiple trait model which included both direct and maternal
genetic effects and a correlation between direct and maternal genetic
effects. Henderson (1985c) described the RAM for single traits.
Utilizing RAM greatly reduces the size of the MME which reduces the
computer storage required to obtain BLUP. For example, Hudson and
Kennedy (1985) obtained BLUP's for swine using Ontario Record of
Performance data of Yorkshire, Landrace, Hampshire, and Duroc breeds.
In this data set, there were 5.6 times as many total animals as there
were parents and ancestors with progeny with records. However, the
computation of MIVQUE or REML estimates of (co)variance components
using quadratics in BLUP still is not feasible because the g-inverse of
the left-hand side of the MME is needed for these calculations; and
this matrix can be quite large even under RAM.
In practice, BLUP's are computed by substituting estimates of
variance components for the corresponding parameters in the MME. These
estimates can come from either the same data set or from an independent
data set. When estimates of variance components are used in the MME in
place of true parameters, the resulting estimates of genetic values are
not actually BLUP but are probably close approximations if the
estimates of variance components are precise. Also, if the estimates
of variance components used with BLUP are obtained from the same data
set as BLUP then the BLUP estimates obtained are biased, but this bias
is likely to be very slight if the data set is large.
FEASIBILITY OF REDUCED ANIMAL MODEL
Kennedy et al. (1985) estimated the variance components used by
Hudson and Kennedy (1985) using an approximate REML procedure
11
originally due to Henderson (cited by Hudson and Van Vleck, 1982).
Kennedy et al. (1985) used the same, or some of the same, data to
estimate variance components as Hudson and Kennedy (1985) used to
obtain BLUP. Kennedy et al. (1985) used a model in which the random
effects were sires and litters. The sires and litters were assumed to
have constant variances, and the correlations among sires, among
litters, and between sires and litters were assumed to be zero. If RAM
was appropriate for obtaining BLUP, why did these authors not use RAM
to estimate variance components? The answer is that it is not feasible
to use REML or MIVQUE to obtain estimates of variance components under
RAM when the MME contain 4,000 to 20,000 equations as was the case in
these studies. It is not known whether it is better to use an
approximate genetic model and an efficient method such as REML as was
done in this study, or to use a more accurate genetic model such as RAM
and a less efficient method such as diagonal MIVQUE, SDS, or WSDS When
the more efficient method cannot feasibly be used with RAM. In the
remainder of this paper, it will be assumed the animal model or RAM is
appropriate, and methods that are more computationally feasible than
REML or MIVQUE will be discussed.
METHODS MORE FEASIBLE THAN MIVQUE AND REML
Henderson (1985a) proposed a method called diagonal MIVQUE when
the number of MME is too large to feasibly obtain the g-inverse of the
coefficient matrix (left hand side of MME). For this method,
approximate solutions for genetic random effects (g) are obtained by
dividing the right hand sides of MME by the corresponding diagonal
element of the coefficient matrix. Quadratics in terms of approximate
12
g (i.e., (g)) and approximate e (i.e., (e)) are then computed. These
quadratics are then equated to their expectations, and the resulting
equations are solved for estimates of variance components. The
expectations of these quadratics are easier to compute than
expectations of quadratics in BLUP for g and e because the inverse of
the diagonal elements of the coefficient matrix are used in place of
the g-inverse of the coefficient matrix.
The inverse of V is proportional to I n for most models if all
variances and covariances except the residual variance are assumed to
be zero for the purpose of simplifying calculations. If MIVQUE is
applied in this case the resulting estimates are MIVQUE if all
(co) variances except the error variance are close to zero relative to
the error variance. This method was suggested by Rao (1971) to help
improve computational feasibility. However, these estimates have been
shown to be poor relative to ANOVA when variances other than the error
variance are as large as eight times the size of the error variance for
some hierarchial designs (Brockelbank and Giesbrecht, 1985). The
extent to which this is true for the animal model is not known.
Grimes and Harvey (1980) extended the use of symmetric sums of
products (SSP) of Koch (1967, 1968) to the animal model. They called
their method symmetric differences squared (SDS). They compared the
standard errors of SDS and ANOVA for a random paternal half sib
analysis using simulated data and found that SDS was slightly less
efficient than ANOVA for this case. Grimes and Harvey (1980) also
used simulation of a maternal effects model to evaluate the standard
errors of SDS estimates of variances due to direct genie effects,
13
maternal genie effects, permanent maternal environmental effects, and
residual effects, and the covariance between direct and maternal genie
effects computed from data simulated using mating designs A and B
described by Thompson (1976). They found that the standard errors of
the estimates were quite large even for data sets of 1,600 animals
resulting from simulating 200 A or B sets. Grimes and Harvey (1980)
demonstrated the computational feasibility of SDS by using it to
estimate the same (co)variances as in their maternal effects simulation
described above from the weight and gain traits of 1,780 mice in a
random mating control population. This procedure required very little
computer storage but used considerable computer time because of the
calculation of 1,583,310 symmetric differences squared and their
expectations .
Christian (1980) showed how the computer time requirements of SDS
can be reduced by summing groups of symmetric differences squared that
have the same expectation and then computing the expectation of the sum
of each group of SDS. These sums and their expectations are equated,
and the resulting equations solved to obtain estimates of (co) variance
components. Christian (1980) proved that this method is equivalent to
SDS. Computer time was less because fewer multiplication operations
were needed. The number of addition operations required for this
method was not different from SDS. In addition, Christian (1980)
suggested that SDS weighted by the inverse of the error variance-
covariance matrix among SDS (WSDS) would correct for the
inquadmissibility (inefficiency) of SDS. However, to complete the
analysis by this method as Christian described, it would require
14
computational steps proportional to n raised to a power of four. Where
n is the number of observations.
COMPUTING A
The numerator relationship matrix (A) is needed for SDS, WSDS,
REML, MIVQUE and diagonal MIVQUE under the additive and nonadditive
genetic animal model. The computation and storage of this matrix can be
costly. Hudson et al. (1982) described a computer algorithm that can
compute and use A by storing only the diagonal elements and the nonzero
off diagonal elements of A. This algorithm requires storage for three
vectors of order n and two vectors of order equal to the number of
nonzero elements of A. They found that only 6.6 to 22.6% of the
n(n+l)/2 elements of the upper triangle of A were nonzero for samples
from five populations of sires.
UNWEIGHTED MIVQUE
In addition to computational feasibility, it would be convenient
if methods of estimating (co)variance components did not depend on the
assignment of prior values. Methods such as REML and ML do not depend
on the assignment of prior values. Such methods are iterative and can
require large amounts of computer time when more than 500 to 1,000
parents are included in a data set especially when the animal model is
applied. MIVQUE is not computationally feasible under the animal model
with large data sets, and it depends on prior values for its optimal or
near-optimal properties.
Rao (1971) suggested that MIVQUE could be applied with all prior
(co)variances assigned a value of one including the residual variance
so that the experimenter would not have to assign prior values. This
15
method will be abbreviated MIVQUE(l) in the remainder of this paper.
However, this method would not be computationally feasible for the
animal model with large data sets.
Brocklebank and Giesbrecht (1984) compared MIVQUE(l) to ANOVA for
4,225 different parameter combinations involving 15 different 2 stage
nested designs. For all but one of the 15 designs, MIVQUE(l) resulted
in lower standard errors than did ANOVA for more than one-half of the
parameter combinations.
CONCLUSIONS
REML or MIVQUE with reasonable priors yield efficient estimates of
(co)variance components. However, when the animal model is appropri-
ate, estimates by REML and MIVQUE require too much computer time for
large data sets with greater than 500 to 1,000 parents. For additive
genetic models, Quaas and Pollack (1980) described a reduced animal
model equivalent to the animal model that has MME only for parents of
progeny with records. This reduces the size of the MME needed for
obtaining BLUP estimates of genie values and for obtaining genie (co)-
variance components (Henderson, 1985a, b, c). However, with today's
computers it appears not to be feasible to obtain REML or MIVQUE under
even the reduced animal model as evidenced by the choices made in the
studies of Hudson and Kennedy (1985) and Kennedy et al. (1985). There-
fore, the breeder is forced to choose an approximate model and use an
efficient method, such as REML as it appears that Kennedy et al. (1985)
did, or choose the animal model or RAM and a less efficient, more
computationally feasible method for estimating variance components.
Hopefully, in the future with improving computer technology and innova-
tive computing shortcuts, this choice will not have to be made.
There is a need to find a computationally feasible algorithm for
computing WSDS, and to compare the efficiency of WSDS with other
computationally feasible methods such as MIVQUE(O), SDS, and diagonal
MIVQUE under the animal model.
16
THEORY
THE EQUIVALENCE OF MIVQUE AND WEIGHTED SYMMETRIC DIFFERENCES SQUARED
It is shown in this section that Symmetric Differences Squared
(SDS) (Grimes and Harvey, 1980) weighted by the inverse of the
variance-covariance matrix among squared differences (Y) is MIVQUE
(Rao, 1971) when the true parameters are used to construct the Var(Y).
The idea of weighting SDS to obtain more efficient estimates of
variance components was suggested by other workers (Forthofer and Koch,
1974? Christian, 1980).
Forthofer and Koch (1974) suggested that Symmetric Sums of
Products (SSP) (Koch, 1967) might be weighted in some way to obtain
more efficient estimates of variance components. Grimes and Harvey
(1980) showed that SDS estimates are equivalent to SSP estimates (Koch,
1967) for the one-way classification random model. Christian (1980)
suggested that SDS could be corrected for inquadmissibility
(inefficiency) by weighting by the inverse of the error variance-
covariance matrix of Y. While MIVQUE is appropriate for both mixed and
random models, SDS was defined only for the random model (Grimes and
Harvey, 1980). Therefore, only a random model need be considered when
showing the equivalence of MIVQUE and weighted SDS.
The linear model for the phenotypes or observations to be assumed
is
c
y = l n P + E u ± + e [1]
i=l
17
18
where:
V is a fixed constant,
n is the total number of observations,
y is an n x 1 vector of observations,
l n is an n x 1 vector of l's,
c is the number of random sources of variation not including e,
and
u^ and e are n x 1 norma] random vectors.
The first and second moments of the random effects are:
E(e) = EU^) = 0,
Var(u ± ) = V ± o\,
Var(e) = I n eg,
Covfu^e') = 0,
CovfUjyUj) = 0, if ij*j, and
Var(y) = Z V ± o? + I n a| , [2]
i=l
where V^ and l n are known matrices. For example, V^ might be Wright's
numerator relationship matrix and V"2 might be the dominance
relationship matrix. a? and ag represent unknown variance and
covariance components. It should be noted that elements within a set
of random effects can be correlated in many different ways. However,
correlations among random sets of effects are assumed to be zero.
First, the equivalence of MIVQUE and weighted SDS will be shown
for a case where the observations are uncorrelated. When all variances
and covariances except the residual error variance are assumed to be
zero, the observations are uncorrelated. In the second part of the
19
proof, it is shown that weighted SDS and MIVQUE are equivalent for
cases in which correlations exist among observations in y. The vector
of observations y is transformed to a random vector y* whose elements
are uncorrelated. It is shown that weighted SDS on y* results in the
same estimates of variance components as weighted SDS on y. Then,
applying the first part of the proof for uncorrelated observations in
y, weighted SDS on y* is equivalent to MIVQUE on y* which is equivalent
to MIVQUE on y.
The proof of the equivalence of weighted SDS and MIVQUE can be
divided into the following steps.
1. The vector of symmetric differences squared Y is defined using
slightly different notation than that of Grimes and Harvey
(1980). Y is defined as a vector of squares (HY^) minus two
times a vector of crossproducts (Y2).
2. The expectation of Y is given. These expectations are needed
to set up the SDS and weighted SDS equations.
3. The equations to solve to obtain estimates of variance
components by SDS (unweighted) are given.
4. The variance-covariance matrix for Y is given. The inverse
of this matrix is needed to obtain estimates by weighted SDS.
5. The inverse of the error variance-covariance matrix among
differences squared in Y is given. Note that this is
proportional to the inverse of the variance-covariance matrix
of Y if there are no correlations among observations in y.
6. The equations to solve to obtain estimates of variance
components by weighted SDS are given.
20
7. It will be shown that SDS weighted by the inverse of the error
variance-covariance matrix of Y (WSDS) is equivalent to
MIVQUE(O).
8. Using the results from step 7, it will be shown that SDS
weighted by the inverse of the total variance-covariance
matrix is equivalent to MIVQUE.
1. Definition of Symmetric Differences Squared
The q = n(n-l)/2 x 1 vector of symmetric differences squared is
formed by taking all possible unique differences among observations.
These difference are then squared. Define these differences squared as
y= C(y 1 -y 2 ) 2 (y 1 -y 3 ) 2 ...(y 1 -y n ) 2 (y 2 -y3) 2 ...(y 2 -y n ) 2 ...(y n . 1 -y n ) 2 ]'
which may be written as:
y = t(y k - yi ) 2 ] C33
for k = 1, 2, ... n-1; 1 = k + 1, k + 2 , ... n.
Expand the squares in [3] and express Y as a vector of squares (HYj)
minus two times a vector of crossproducts Y 2 in [4J.
Y = [y 2 - 2y kYl + y 2 ]
= Cy 2 + yj] - 2[y kyi ]
= BYj - 2Y 2 [4]
where:
*1 = Cyf y 2 , ... y 2 ^, and [5]
Y 2 = Cy 2 y 2 y^ ... y^ y 2 y 3 ... Y^y^' C6]
H is a q x n matrix containing zeros and ones, which specifies the
elements of Y^ to be incorporated into Y. It is defined as
21
H =
1
1
1
1
1
1
1
1
[7]
... 1 1
There are three useful matrix identities involving H that will be
needed in the proof that follows. These are
[8]
[9]
ffl n = 21 q<
H'l,
(n-l)l n# and
H'H = (n-2)^ + l n l^ . [10]
Identity [8] can be explained by noting that every row of H has two
ones and n-2 zeros. Likewise, identity [9] can be explained by noting
that every column of H has n-1 ones and (n-l)(n-2)/2 zeros. The
diagonals of [10] are all n-1 because each column of H has n-1 ones and
(n-l)(n-2)/2 zeros. Each of the off-diagonal elements of [10] is one
because each pair of columns of H has a one in the same row only one
time.
2. Expectation of Symmetric Differences Squared
The expected value of Y (E(Y)) is needed to set up the SDS and
weighted SDS equations. E{Y) is written as a function of unknown
variance components and the elements of V^ as follows:
E{Y) = Edffj^ - 2Y 2 ) - HE(Y X ) - 2E(Y 2 ). [Ill
The expected value of Yj^ is
E(Y X ) = l n p 2 + X x a 2 [12]
where:
x l -
o = to 2. & 2 ' '
< V 11>1 < v ll>2
< V 22>1 < v 22>2
o c e J ,
< v ll>c
< v 22>c
< v nn>l < v nn>2 • • • < v ™)
nn'c
, and
,th
* v kk^i ^ s the diagonal element of V^,
The expected value of Y 2 is
where
x 2 =
< V 12>1
< V 13>1
< v m> 1
< V 23>1
E(Y 2 ) = l qU 2 + X 2 a 2 /
< v 12>2
< v 13>2
< v ln> 2
(v 23 ) 2
< v 12>c
< v 13>c
■
< v ln>c
< v 2n>c
■
< v 2n> 1 < v 2n> 2
( v 2n>c °
< v n-l,n>l < v n~l,n>2 • - * < v n-l,n>c °
(v }c] ) i is the element in the k th row and 1 th column of V^.
It should be noted that when there is no inbreeding, X^ is
x l _ ^c+l-
22
[13]
[14]
[15]
, and [16]
[17]
Substitute [12] and [15] into [11] and the E(Y) can be written as:
E(Y) = Hl n u 2 + HX^ 2 - 21 q v 2 - 2X30 2 . [18]
Substitute [8] into [18] and E(Y) can be written as:
E{Y) = 21qM 2 + HX 2 a 2 - 21 v 2 - 2X 2 a 2 , which, after
combining terms and rearrangement is
23
E(Y) = {HX X - 2X2)o 2 , or
= Xo 2 ,
where
X=HX 1 - 2X 2 . [19]
3. Symmetric Differences Squared Equations
Estimates of variance and covariance components can be obtained by
solving
X'Xg 2 = X'Y. [20]
It should be emphasized that [20] are the same equations as used by
Grimes and Harvey (1980) to obtain estimates of variance components.
4. The Variance - Oovariance Matrix Among Symmetric Differences
Squared
The inverse of the variance-covariance matrix of Y is needed to
obtain estimates of variance components by weighted SDS. Define Var(Y)
as
Var(Y) = Cov(Y,Y'),
= [Cov«y k - yi ) 2 f (y m - y ) 2 >], or
= 2 ^ v km " v lm " v ko + v lo> 2 3< C213
for k = 1, 2, . .. n - 1; 1 = k + 1, k + 2, ...n,
and
m = 1/ 2, ... n — 1; o = m + 1, m + 2, ... n,
where Vj an is the element in the "k^ row and the m column
of Var(y) = V.
The inverse of [21] is difficult to represent algebraically because V
can be any positive definite nonsingular matrix. Fortunately, this
inverse is not needed to accomplish the objectives of this paper.
24
First, the inverse of [21] will be obtained for the special case when V
is proportional to I_, which can be written as:
VaI n-
It will then be shown that even when Var(y) is complicated, y can be
transformed to a y* with a variance proportional to I n , and that
estimates of variance and covariance components by MIVQUE and weighted
SDS are identical whether the analyses are completed on y or y*.
The variance of Y can be rewritten as
Var(Y) - VardHi - 2Y 2 )
= CovCHYl - 2Y 2 ,YjH' - 2Y 2 )
= CovtHYpYjH' ) - 2Cov(Hy 1# Y 2 ) - 2Cov{Y 2 ,YjH' ) + 4Cov(Y 2 ,Y 2 ), or
= H VarfY-^H' - 2 H CovfY-^Y^) - 2Cov(Y 2 ,Yi)H' + 4Var(Y 2 ). [22]
When all variance and covariance components except o~ are assumed to be
zero, the second moments involving Y^ and Y 2 are
VarCY^ = 2 a^, [23]
Var(Y 2 ) = a*I q , [24]
Cov(Y lf Y 2 ) = 0^ q , and [25]
Cov(Y 2 ,Yi) = pgn. [26]
Substitute [23], [24], [25], and [26] into [22] and the Var(Y) is
Var(Y) = 2 a^HH' + 4 o*I q ,
= 2 a^(HH' + 2I q ), or
= 2 o^W,
when Oq is the only nonzero prior value, where
W = {HH 1 + 2l q ). [27]
It should be noted that SDS weighted by the inverse of W is equivalent
to SDS weighted by the inverse of Var(Y) if
25
W a Var(Y) .
It should be emphasized that
Var(Y) a W
whenever
Var(y) ot^.
This is true if all variance components except o ^ are assumed to be
zero, or if y is a random vector with variance proportional to I n .
5. The Inverse of the Error Variance - Oovariance Matrix of Y
The inverse of W can be written as:
VT 1 = (BH' + Zlg)" 1 . [28]
Now, let us apply equation [17] of Henderson and Searle (1981) to find
the inverse of the sum of 2 matrices. Their equation [17] is
(A + UBUT 1 = A" 1 - A^UCB" 1 + U'A^lD^U'A" 1 [29]
If we let A = 2I a
U = H, and
B = Ij^, and substitute into [29],
W _1 can be written as
WT 1 = [^I q - ^Hdn + ^H'H)" 1 ^ |]. [30]
After rearrangement, [30] is
W" 1 = ^ [I q - H^ + H'Hr^-H']. [31]
Substitute [10] into [31], and observe that
(2^ + H'H)" 1 = (21^ + (n-2)l n + l^) -1 , or after combining terms,
= <"*n + Vn> _1 - ^
Now we can apply equation [3] of Henderson and Searle (1981) to
26
evaluate [32]. Their equation [3] is
b
(A + buv') -1 = A -1 A" 1 uv , A" 1 . [33]
1 + bv'A~i
We let
A=nl n ,
u = 1 n'
v ' = *n
b= 1,
and substitute into Z331 to obtain
n
which after simplification is
n " 2n
Substitute [34] into [31] to obtain
= ; ^-^Vn'- ^
^ = I Cl q " n" H(I n - In VA> H '3' ****
simplifies after rearrangement to
W-l = i [I q - i BH' + L l q l q ]. [35]
6. Weighted Symmetric Differences Squared Equations
Estimates of variance components by weighted SDS can be obtained
by solving
X'W^Xa 2 = X'W^Y. [36]
The solution to [36] yields the same estimates that Christian (1980)
27
suggested after correcting SDS for inguadmissibility.
7. The Equivalence of SDS Weighted by the Inverse of the Error
Variance-Covariance Matrix and MIVQUE(O)
It will now be shown that the solution to [36] is equivalent to
MIVQUE if all variances and covariances except cjg are zero, or if
Varfyjai^. it will first be shown that
X'frtc = CtrfMVjMVj)], [37]
for
(i, j = 1, 2, ... c, c+1),
where
V c+1 = V
m - *n - \ v;- ^BJ
and
CtrtMV^MVj)]
is a (c + 1) x (c + 1) matrix that contains the left hand side of
equations used by Rao (1971) to obtain MIVQUE(O) estimates of variance
and covariance components. It will then be shown that
X'W^Y = [y'M^My], [39]
for
(i = 1, 2, ... c, c+1),
where (y'MVjMy] is a (c + 1) x 1 vector that is the right hand side of
the equations used by Rao (1971) to obtain MIVQUE(O) estimates of
variance components. Let estimates of variance components by MIVQUE(O)
and SDS weighted by W -4- be denoted by o and o , respectively,
where
28
a 2 = (x'lrto^x'W^Y, C4o]
and
o 2 = Ctr(HV i MVj)]" 1 [y'MV i My]. [41]
Therefore, if equalities [37] and [39] are true then [37] and [39] can
be substituted into [40] to obtain
o 2 = [tr(MV i MVj)]"' 1 [y , MV i My3,
which after applying [41] is
"2 * 2
Therefore, if equalities [37] and [39] are true and if Var(y)oti n ,
estimates of variance components by MIVQUE and SDS weighted by W" 1 are
equivalent.
If [35] is substituted into [36], the left hand sides of [36] are
X'\r l X = ^ (XjH' - 2X£)[l q - ^ HH' + -- lgl^fHX-L - 2 X 2 ),
which after expanding through the center parentheses can be written as:
= ^ [(X[H' - 2X^(HX 1 - 2X 2 )
(XjH' - 2X 2 )HH'(HX 1 - 2X 2 )
+ -- (X^ - 2X^1 l^fHX-L - 2X 2 )]. [42]
rr ^ ^
The first term within the brackets of [42] is
(X^H" - 2X3) (HX 1 - 2X 2 ) = XJH'HX-l - ^H)^ - ^{H'Xg + 4X3X3,
which after substituting [10] for H'H simplifies to
{n^JXjX-L + X^lJXj - 2X2HX-L - 2X}H'X 2 + 4X!X 2 . [43]
29
It should be noted that [43] is the same as the left hand side of
the SDS equations used by Grimes and Harvey (1980) to obtain estimates
of variance and oovariance components.
The second term within the brackets of [42] is l/n multiplied by
(XjH* - 2X£)HH'(HX 1 - 2X 2 ) = XJH'HH'IWl - 2XjH , HH , X 2
- 2X 2 HH'HX 1 + 4X2HH'X 2 ,
which after substituting [10] for IfH simplifies to
(n^^XjX-L + On-^JXj^lj^ - 2(n-2)XjH'X 2
- 4X^1^X2 - 2(n-2)X 2 HX 1 - 4X 2 l q I ] £C 1 + 4X 2 EH'X 2 . [44]
The third term within the brackets of [42] is 2/n 2 multiplied by
(XJH* - 2X 2 )l q l q (HX 1 - 2X 2 ) = X^B'1 C ^^K 1 - 2XjH'l q l q X 2 " 2X 2 1 q 1 q HX l
+ 4X21^2,
which after substituting [9] for H'l q simplifies to
(XiH' - 2X 2 )l q l q (HX 1 - 2X 2 ) = (n-D^lnl^C-L - 2(n-l)Xjl n l q X 2
- 2(n-l)X 2 l q l' r X 1 + 4X21^2- [45]
We can substitute [43], [44], and [45] into [42] to obtain
X'W" 3 * = - {(n^X-JX,^ + Xjljjl^ - 2X 2 HX 1 - 2XjH'X 2 + 4X3X3
- - [(n^^X-^ + On^X-J^l,^ - 2(n-2)XjH'X 2
~ 4X i 1 n 1 q K 2 - 2(n-2)X 2 HX 1 - W^Ri + 4X2HH'X 2 ]
+ -- [(n-1) 2 X^l^ - 2(n-l)Xil n l q X 2
n
- 2{n-l) X^lgl,^! + 4X2^1^3]}, [46]
or
30
x'w^x = ( - n ~ 2 - } - x^ +y x^ifo - 2 - x^ - 2 - x>px 2 + 2X3X3
+ ^ X i W2 + ~ n 2 ^Wl " I X 2 HH,X 2 + ^2 Wft' C473
which after further rearrangement is
X'vrtc = (X^ + 2X2X 2 )
- - {X 2 H + Xi)(H'X 2 + X x )
+ ^ < X i X n + 2^ V <Wl + 21 <?2> • C48]
n^
The following matrix equalities are useful:
[tr^V^] - Xj^ + 2X3X3, [49]
for i, j = 1, 2, ... c + 1,
^1^ = 1& + 21^X3, [50]
and
^j 1 ^ = 1^1 + 21 q*2- C 51 ^
Equalities [49] through [51] are proven in Appendix A. Substitute [49]
through [51] into [48] to obtain
X'W _1 X = [trtViVj)] - I [tray-jV^)] + -- [trd^l^Vjln)], which
after rotating traces is
= [tr(V iVj )]- -^[trtVil^Vj)]- ^[tr(l n l^V iVj )
■*-i [tr(l n l^ Vi l n l n V 5 )], or
31
X'WTlx = {trCViVj - i V^l^Vj - i l^V^ + i- l^V^l^] >,
= (trCVi - i l^i^ViXVj - i VA v j^ >-
Finally, substitute [38] into [523 to obtain
X'VT^X = trf^MVj), which proves [37].
Substitute [35] into the right hand sides of [36] to obtain
X . W -1 Y = \_ {x ^. _ 2X^Cl q - i HH 1 + -- lgl^KHY! - 2Y 2 ) f
which after expanding through the center parentheses becomes
X'W~ 1 Y= - C(X]H' - 2X^)(HY 1 - 2Y 2 )
- - (XjH 1 - 2X 2 )HH'(HY 1 - 2Y 2 )
+ -- {X^H' - 2X 2 )1 C J^(HY 1 - 2Y 2 )]. [53]
rr
The first terra within the brackets of [53] is
(XJH'- 2X 2 )(HY 1 - 2Y 2 ) = XjH'HYj^ - 2X2HYJ - 2XjH'Y 2 + 4X 2 Y 2 ,
which after substituting [10] for H'H simplifies to
(n-2)XjSr 1 + X^l^ - 2X3^ - 2XjH'Y 2 + 4X^2- [54]
It should be noted that [54] is the same as the right hand side of the
SDS equations used by Grimes and Harvey (1980) to obtain estimates of
32
variance and covariance components.
1
The second terra within the brackets of [53] is - multiplied by
n
(XJH' - 2X£)HH'(HY 1 - 2Y 2 ) = XjH'HH'Hr^ - 2XjH'HH'Y 2
- 2X 2 HH'HY 1 + 4X£EH'Y 2 ,
which after substituting [10] for H'H simplifies to
(n-2) 2 X{Y 1 + (3n-4)Xil n l^Y - 2(n-2)X|H'Y 2
- 4Xj_l n l^Y 2 - 2(n-2)X 2 HY 1 - 4X 2 l q l^Y 1 + 4X2HH'Y 2 [55]
2
The third term within the brackets of [53] is — multiplied by
n 2
(XjH'- 2X2)1^^ - 2Y 2 ) = XiH'lgl^ - 2X^1^2 - 2X^1^^
which after substituting [9] for H'l q simplifies to
(n-l) 2 Xil n l nYl - 2(n-l)Xil n l4Y 2
- 2(n-l)X 2 l q l^Y 1 + 43^1 q l q y 2 . [56]
We can substitute [54] through [56] into [53] to obtain (after
simplification)
X' W -1 Y = i™ } - Xfr + ^- Xil n 17! - I X^ - I X^'Y 2 + 2X^2
+ -^ 2 X i 1 n 1 o Y 2 + 2 ~2 W^l " \ X ^ H ' Y 2 + ^ W?* ^57]
which after further rearrangement is
33
X'W^Y = (X^ + 2X^Y 2 )
- - (X^H + X[)(H'Y2 + Y x )
n
+ -~ (Xil n + 2X^1 Jd^ + 21^Y 2 ). [58]
rr
The following matrix equalities are useful:
[y' Vi y] = X£Y X + 2X2'Y 2 , [59]
yy'l n = H'Y 2 + Y v [60]
and
l^yy'l n =i;Y 1 + 2^Y 2 . [61]
Proofs of [59] through [61] are in Appendix A.
We can substitute [59] through [61] into [58] to obtain
X-^Y = [y'V iY - \ l n V iW 'l n + lg^&flj'
= [y'V iy - ly'ViVJy - ^y'l n l^V iy + ^fl^^Vtfl,
= [y'Vi(y - \ i^y) - i y'iniAVity - \ 1^)],
= [(y'v i -iy'l n l^V i )(y-il n l^y)],
= [(y' - i y^l^Cy - i l^y)],
= Ey' t^ - I VA)Vid n - I l n l n )y] C62]
Finally, substitute [38] into [62] to obtain
X'W^Y = [y'MlTjMy],
34
to prove equality [39].
Therefore, it is shown that the equations for computing estimates of
variance and covariance components by MIVQUE are equivalent to SDS
equations weighted by the inverse of Var(Y) if the Var(y) al n . Since
the equations are equivalent then so are the estimates of variance
components.
8. The Equivalence of SDS Weighted by the Inverse of the Total
Variance-Covariance Matrix and MIVQUE
Now it will be shown that MIVQUE and weighted SDS are equivalent
even when Var{y) is not proportional to I n . One strategy is to obtain
a transformation of y, say Ty such that Var(Ty) = I n , and then show
that weighted SDS completed on Ty is equivalent to MIVQUE. This is the
procedure to be followed here. Let V - LDL', where L'I» = I n and D is a
diagonal matrix of positive real numbers. Then, the inverse of V is
V 1 = UT 1 !.* . [63]
If we obtain the transformation of y — > Ty, and let
T = D" 1 / 2 !/ , [64]
V~ can be written as a function of T as follows:
TT 1 = LD- 1 / 2 D~ 1 / 2 L' = T'T. [65]
The variance of Ty is
Var(Ty) = TVT',
= D^^L'UXj'LD" 1 / 2 = 3^, or
= I ojTVjT' + ogpr 1 . [66]
There are two important points to be made here. First, Var(Ty) aI n so
that the results given previously on the equivalence of weighted SDS
and MIVQUE are valid when the analysis is on Ty. Second, the Var(Ty)
35
[66] is written in terms of the same unknown parameters as the Var(y)
[2],
Let us denote the vector of transformed observations as
y* = Ty= CyJ y*E ... y*] ' , and [67]
the vector of symmetric differences squared as
Y* = C(y£ - yf) 2 ]. for [68]
(k = 1, 2, ... n - 1; 1 = k + 1, k + 2, ... n).
The columns of Xf are now the diagonals of TVjT" and D" 1 instead of the
diagonals of V^ and I n . Also, the columns of X£ are the q off diagonal
elements of TVjT' and D -i instead of the off diagonal elements of V^
and I n . Hence,
Var(Y*) = 2W* = <HH' + 21 ), and [69]
t-1 ^
1 . 1 .2
W*-= 2 -Cl q - n -HH' +n --l qV . [70]
W*" 1 [70] is identical to W" 1 [35] and W* [69] is identical to W [27].
The weighted SDS equations, using y are
X^VarfY)]"^ 2 = X^VartY)]" 1 ^ [71]
and the weighted SDS equations using y* are
X*'W* _1 X* cr^X*^*" 1 ^. [72]
The solutions to equations [71] and [72] are identical.
Now let
M * = *n " Tl^T'TV" 1 ^,
which after substituting [65] for T'T is
D^ - n^l^r 1 !^)- 1 ^ . [73]
By equality [37] the left hand side of weighted SDS equations using y*
is
36
x *i W *-l x * _ [tr(M*TV,T'M*TV^T')],
x J
which after rotation of traces is
and since
then
= [tr(T , M*TV i T'M*TVj);j, [74]
t*m*t = T'djj - Ti n u; 1 ir 1 i n )- 1 i n T , )T,
= T'T - T , Tl ri (i; i V r " 1 :L n )~ 1 i; i T'T, or
= V 1 - V^l^yy-l^-lyy-l = P , [75]
Xt'W'hc* = CtrtPViEVj)], [76]
for i, j = 1/ 2, . .. c + 1,
which is the left hand side of the equations used to estimate MIVQUE
(Rao, 1971) for model [1]. Now by equality [39] the right hand side of
the weighted SDS equations using y* is
xa.^-ly* = [y'T'MVTVjT'^Ty],
which after substituting in [75] for T'M*T is
X^^-ly* = [y'FVjPy],
which is the right hand side of the MIVQUE equations (Rao, 1971). It
has therefore been shown by these derivations that the weighting of the
symmetric differences squared by the inverse of the total variance -
covariance matrix to estimate the variance components yields estimates
that are minimum variance unbiased estimates if the prior values used
are near the true parameters. It must be emphasized that transformation
of y to y* was simply a device used to prove the above, which would not
be recommended as a computational approach in practice.
37
COMPUTATIONAL REQUIREMENTS OF WSDS OR MIVQUE(O)
The computational requirements to complete analysis by MIVQUE(O)
or WSDS will now be discussed. The left and right-hand side of the
MIVQUE(O) or WSDS equations are as follows:
trtvfjtrCV^) . . . trtV^) tr^)
tr(V§) . . . tr(V 2 V c ) tr(V 2 )
[trfMV.jMVj)] =
symmetric
2
n
symmetric
tr(v£) tr(V c )
n.
1 nV 2 V c l n l^V 2 l n
■ *
n
1
+ —
n 2
1*V„1„
n c n
n
IWVn l^Valn . . . 1^1^ n]
[77]
and
[y'MV-jMy] = Cy'V^y y'V 2 y . . . y'V^ y'yiT
- ^V^y l n V 2 y . . . 1^ ny] '
+ Y^Wn W„ • • ' ^Vn ^' . [78]
The multiplications, additions, and storage needed to obtain and save
the elements in the right hand sides of [77] and[78] are in table 1.
38
Table 1. Storage, multiplications and additions for each type
of element required to set up the MIVQUE(O) or WSDS equations
when relationship categories are not considered.
Type of
element
Number of
elements of
this type 3
c(c+l)/2
Total multiplications
for all elements of
of this type*
n(n+l)c(c+l)/4
Total additions
for all elements
of this type
trfVjVj)
n(n+l)c(c+l)/4
tr(V ± )
c
nc
n
1
n
WVn
c(c+l)/2
nc(c+l)/2
cn 2 +nc(c+l)/2
Wn
c
cn(n+l)/2
y'V iy
c
cn(n+l)/2
cn(n+l)/2
^y
c
nc
n(n+l)c
y'y
1
n
n
1
1
n
a c is the number of sets of random effects other than error
"n is the total number of observations.
Computing [tr(MV^MV^)] and [y'MVjMy] from the elements in Table 1
and solving for the estimates of variance components also requires
computer time and storage, but these requirements are small relative to
the requirements needed to obtain the elements represented in Table 1.
Additional computer resources are needed to compute and save V^
matrices and store y.
For animal breeding problems, the elements of V^ are usually
coefficients of relationship which specify the fraction of the i
variance or covariance component's contribution to the covariance
39
between two members of y. The most widely used V^ matrix in animal
breeding problems is Wrights numerator relationship matrix (A) . If
inbreeding has occured, a strictly additive genetic model is usually
invoked as an approximation because the nonadditive effects under
inbreeding are difficult to interpret (Cockerham, 1954) . If no
inbreeding has occured, nonadditive effects such as dominance and
epistasis can also be considered. Maternal effects and covariance
between direct and maternal effects can be important in animals.
Matrices of dominance and epistatic coefficients of relationship can be
constructed from Wright's numerator relationship matrix (A) if there is
no inbreeding (Henderson, 1985) . A can be saved for later use by
storing only its p nonzero elements, where p £ n(n+l)/2 and p is the
number of nonzero elements in the upper triangle of A (Hudson et al . ,
1982) . Three vectors of order n and one vector of order p in addition
to the p vector containing the nonzero elements of A are needed to
locate nonzero elements in A. This process can save considerable
computer storage. Hudson et al. (1982) found that only 6 to 22.6% of
the n(n+l)/2 upper triangular elements of A were nonzero for five
breeds of dairy sires.
The multipications required to complete WSDS or MIVQUE(O)
equations can be reduced if the related pairs of individuals having
records in y can be classified into a relatively small number of
relationship categories. For example, with the mice data of Grimes and
Harvey (1980), there were 107,398 related pairs of individuals
distributed among only 39 relationship categories. Christian (1980)
showed how to take advantage of these relationship categories to reduce
40
the computer time required to compute SDS by reducing the number of
multiplications needed. It will now be shown how to reduce the number
of multiplications required to do WSDS or MIVQUE(O) when there are a
relatively small number of relationship categories.
If we can assume that no inbreeding has occured, X^ = l n *c+l ^^
can be substituted into [47] and [57] to obtain (after simplification)
i -1 _ i 2 • i 2
X'W X = (n-l)l c+1 l^. +1 - - X^l q l^. +1 - - 1 C+1 1^X 2
+ 2X 2 X 2 - I X^'^ + '2 X 2Vq*2' ™
and
X'WT l Y = (y'y - ny 2 )l c+1 + 2X 2 Y 2 - 2yX 2 Hy
+ 2y 2 X*,l q [80]
for the left and right-hand sides of the WSDS equations. For purposes
of comparison, it seems appropriate to give the SDS equations in this
same notation. Substitute [17] into [43] and [54] to obtain
XX = 2n(n-l)I c+1 l ( L +1 - 4Xp q li +1 - 41 c+1 lgX 2
+ 4X 2 X 2 , and [81]
X'Y = 2n(y'y - ny 2 )l c+1 - 2X3^ + 4X 2 Y 2 [82]
for the left and right hand sides of the SDS equations.
We can let
X 2 = ZU, [83]
where D is an r x (c+1) matrix whose i th column contains the
coefficients for the i™ variance component among the r relationship
categories, and Z is a q x r matrix of zeros and ones that specifies
the relationship category to which the pair of individuals belong whose
41
records are multiplied together to make an element of Y2.
We can also let
B = Z'Z, [84]
where B is a diagonal matrix whose m tn diagonal element, for m = 1, 2,
... r, is the number of pairs of individuals related by the m
relationship category. This is the same as the W matrix defined by
Christian (1980).
Now we can let
N = H'Z, [85]
or
N =
n ll n 12 * * * n lr
n 21 n 22 • * • n 2r
[86]
n nl "n2 * • • "nr
where n^ is the number of individuals in the entire sample related to
individual k by the relationship of the m 1 -" relationship category, for
k = 1, 2, . .. n, and m = 1, 2, ...r.
Hence,
N'N =
z n
l kl
? n kl n k2
? n kl n k2 Z n i
k2
J "k^kr
k
? n k2 n kr
We can let
? n kl n kr ? n k2 n kr • • • ? n kr
Q = z'l„#
[87]
[88]
where Q is an r x 1 vector whose i th element is the same as the i
ith
42
diagonal element of B. Substitute [83] , [84] , [85] , and [88] into [79]
and [80] to obtain
X'frtc = (n-l)l c+1 li +1 - 1 U'Ql^i - I 1 C+1 Q'U
2 4
+ 20'BU - - O'N'NU + -- U'QQ'U, [89]
n n 2
and
X'tf^Y = (y'y - ny 2 )l c+1 + 2U'Z , Y 2 - 2yU'N'y + 2y 2 U'Q, [90]
for the left and right hand sides of the WSDS equations.
Likewise, substitute [83] , [84] , [85] and [88] into [81] and [82] to
obtain
X'X = 2n(n-l)l c+1 l^ +1 - 40'Ql^ - 41 C+1 Q'D + 40'BU, [91]
and
X'Y = 2n(y'y - ny 2 )l c+1 - 20*11^ + 4U , Z , Y 2 , [92]
for the left and right hand sides of the SDS equations. Matrices on
the right hand side of [89] , [90] , [91] and [92] that can be stored
economically are in table 2. Also in table 2 are the storage
requirements for the matrices, and the number of multiplications and
additions required to compute the matrices. It is assumed that square
symmetric matrices will be half stored. Further computations required
to obtain estimates of variance components by WSDS or SDS are small
relative to the computations needed to construct some of the matrices
in table 2, therefore, these are not given.
The r diagonal elements of B can be stored as Q so that B does not
need to be stored separately. It is clear that estimates of variance
components by SDS will take less time to compute than estimates of
43
variance components by WSDS because N'N is needed for WSDS but not for
SDS. A computational example for obtaining SDS and WSDS by [89]
through [92] is presented in Appendix C
Table 2. Multiplications, additions and storage
needed to compute and save the matrices needed to set
up the WSDS equations when relationship categories are
considered.
Matrix
Storage 3 ' b
Multipl icat ions c
Additions
U
re
re
Q
r
n(n-l)/2
N'N
r(r+l)/2
nr(r+l)/2
n 2 +nr(r+l)/2
Z'Y 2
r
n(n-
-D/2
n(n-l)/2
N'y
r
nr
nr
N'Yl
r
nr
nr
n
1
n
y'y
1
n
n
y
1
1
n
a r is the number of relationship categories.
^c is the number of random effects other than residual.
c n is the total number of observations.
The computational requirements for completing WSDS using the mice
data of Grimes and Harvey (1980) will now be discussed. Arithmetic
steps and storage required to compute WSDS by [89] and [90] and by [77]
and [78] will be compared. For the mice data analyzed by Grimes and
Harvey (1980), O, Q, r, c, n, and p are as follows:
a
ii
M| I-
Q3COOCOmU1^iNibitb*iiMi*>*iliOJWWli)NWWIOMMWW(OHHI-'MHHHMM
oowoo>o*»aiajjiMioi-'HooooajfeHoa)itii[»wioMi-'oouiiMJWWPi-'00
0>^<M^(D>bMCOOi^OaiOGO>MOOOOiM^aM»>tiObO^O^OiM>OtOOIOOIOO
oooc^ooooooooooooooooooooooooooooooooooo
ooooooooooooooooooooooooooooooooooooooo
*»
*»
45
Q = [4,739 10,215, 536 1,367 556 936 40 4,419 256 14,862 19,203 743 747
1,280 304 621 302 6,323 264 56 107 176 4,603 2,546 9,214 127 821 56
491 347 2,546 190 9,475 144 170 5,578 1,458 61 1,519]',
r = 39,
c = 4,
n = 1,780, and
p = 109,178.
The first three columns of U were obtained directly from the last three
columns, excluding the first row, of table 5 of Grimes and Harvey
(1980) . The columns of U are the coefficients for the variance and
covariance components for the covariances among individuals related by
the relationship categories. The first column contains the
coefficients for the direct additive genetic variance (a^) , the second
for maternal additive genetic variance (a^L) , the third for direct-
maternal additive covariance {Oggn,) t the fourth for permanent maternal
environmental variance (a^) , and the fifth for residual variance (a^) .
Q was obtained from the first column, excluding the first row, of Table
5 of Grimes and Harvey (1980) .
The basic matrices needed to construct WSDS and SDS equations are
given in table 3 with their storage requirements and the numbers of
multiplications and additions required to obtain them for the mice data
of Grimes and Harvey (1980) .
A small amount of additional computer resources is needed to
compute the WSDS or SDS equations from the matrices in table 3 and to
solve for estimates of variance components.
46
A large amount of computer time and storage is needed to compute
and store A. Grimes and Harvey (1980) traced pedigrees back through 2
generations to compute A. Assuming that their A is essentially
correct, the procedure of Hudson et al. (1982) would require storage
for 223,696 numbers. In addition, computer time proportional to
n 2 = 3,168,400 would be needed to compute A. The elements of the other
V^ can be obtained from A. In order to determine how much is gained by
taking advantage of relationship categories by using [89] and [90] , we
need to evaluate computer requirements when we do not take advantage of
relationship categories as in [77] and [78] .
Table 3. Numbers of multiplications,
additions and stored real numbers required to
save matrices needed to set up WSDS and SDS
equations for mice data of Grimes and Harvey
(1980) .
Matrix
D
Storage
156
Mu ] t ipl ica t ions
Additions
156
Q
39
1,583,310
N'N
780
1,388,400
4,556,800
Z'Y 2
39
1,583,310
1,583,310
N'y
39
69,420
69,420
N , Y 1
39
69,420
69,420
n
1
1,780
yty
1
1,780
1,780
y
1
1,780
47
The storage, multiplications, and additions required to obtain the
matrices needed for MIVQUE(O) or WSDS if we do not take advantage of
relationship categories for the mice data of Grimes and Harvey (1980)
are presented in table 4.
Table 4. Numbers of stored real numbers,
multiplications, and additions required to
compute each type of element needed to obtain
the MIVQUE(O) or WSDS equations for the data of
Grimes and Harvey (1980) if relationship
categories are not considered.
Type of
Element
tr(V ± Vj
Storage
) 1015, 85C
Multiplications
1,900
Additions
15,850,900
tr(V ± )
4
7,120
n
1
1,780
Wj 1 !!
10
17,800
12,691,400
Wn
4
1
6,340,360
Y'Vjy
4
6,340,360
6,340,360
^y
4
7,120
12,680,720
y'y
1
1,780
1,780
y
1
1
1,780
Comparison of tables 3 and 4 demonstrates that consideration of
relationship categories increases the requirement for computer storage
but reduces the requirement for multiplication and addition operations.
Therefore, consideration of relationship categories by using
expressions [89] and [90] instead of expressions [77] and C78] to set
48
up WSDS equations will reduce computer time required substantially and
increase computer storage requirement slightly when the number of
relationship categories is small. Computer resource requirements for
obtaining A are the same whether or not relationship categories are
considered.
METHODOLOGY
It was established in the current study that estimates of
(co) variance components by SDS weighted by the inverse of the error
variance-covariance matrix (WSDS) are MIVQUE if all {co) variances
except Og are near zero relative to o|. since SDS and WSDS are more
computationally feasible than MIVQUE and REML under the animal model,
it would be of interest to know how efficient SDS and WSDS estimates
are relative to MIVQUE as a ?/°g departs from zero. Other approximate
MIVQUE methods that reduce the computational requirements, or remove
the arbitrariness of assigning priors without markedly reducing
efficiency would also be of interest. Permanent maternal environmental
effects (variance = c*) will usually account for a portion of the
phenotypic variance of measurements taken on young animals. MIVQUE
with all priors set to zero except a ^, and a^ (MIVQUE (0, M, E,)) is
easier to compute than MIVQUE because the Var(y) under the assumption
that all variances except a^ and a| are zero can be easily inverted
using partitioned matrix techniques (Henderson and Searle, 1981) . To
avoid the arbitrariness of assigning prior values, Rao (1971)
recommended assigning priors: o* = 1, for all i, and a~ = 1. This
method will be referred to in this paper as unweighted MIVQUE or
MIVQUE (1).
A numerical study was conducted to evaluate the influence of the
unknown (co) variances on the efficiency of SDS, WSDS, MIVQUE (0, M, E) ,
49
50
and MIVQUE(l) relative to MIVQUE {true parameters used as priors) under
the animal model.
Two mating designs (Figure 1) described by Thompson (1976) and
three mating designs described by Eisen (1967) (figure 2) were chosen
for this evaluation. Thompson's (1976) designs were chosen because
they are appropriate for species with one or two offspring, whereas
Eisen' s (1967) designs were chosen because they are appropriate for
litter bearing species. For the Thompson designs, a random sire (S) is
mated to two random dams (Dj and D2) to produce a male and a female
offspring from each mating. In design A, a female progeny from one dam
and a male progeny from the other dam are bred to random mates (M^ and
M2) to produce two progeny each. In design B, the female progeny from
both dams are bred to random males to produce two progeny each. This
results in eight progeny to form a set of a Thompson design.
Figure 1. Two mating designs of
Thompson (1976) . In design A,
!>! and P3 are males and P 2 and
P 4 are females. In design B, P 2
and P 3 are males and P-^ and P 4
are females.
51
For the Eisen designs, s sets are sampled from initial random
matings. The 18 individuals that comprise a set are shown in figure 2.
S^ and S 2 are male parents, D±, D 2 , ... Dg are female parents, and 0^,
2 , ... Og are offspring. Each set contains three unrelated families.
In design I, Sj and S 2 are a full sib family, D^, D 2 , D3 and D4 are a
full-sib family, and D5, Dg, D7, and Dg are a half sib family. In
design II, S^, Dj, and D 2 are a full sib family; S 2 , D3, and D4 are a
full-sib family; and D 5 , Dg, D 7 , and D Q are a full sib family. Design
III is the same as design II with the exception that D5, Dg, Dy, and Dg
are a half-sib family instead of a full-sib family.
Figure 2. The 18 individuals comprising a set of an Eisen (1967)
design. Si and S 2 are male parents, D 1# D 2 , ... Dg are female
parents, arid O^, 2 , ... Og are offspring.
The model assumed in this study for the phenotype of the i th
individual from the jt dam is
*ij = v+ 9i + 9j + ™j + e ij' t 93 J
where p is a fixed unknown constant,
g^ is the direct genie effect of the 1 th individual,
g^ 1 is the maternal genie effect of the j^ 1 dam,
and
mj is the permanent maternal environmental effect of the j dam,
e^ is the temporary environmental effect,
E<e i; j) = E(mj) = E(<£j) = E(g ± ) = 0,
52
Var{ gi ) = q|, Var{<^) = o^, VarOrij) = oj, Cov( gi , gj) = ° ggm and
Var(e i j k ) = a|, and
Cov{gj,mj) = Oov(e 1 , e^) = Cov(g^, irij) = Cov(mj, ey) = Cov(g?J, e^) =
Since individuals of different sets are unrelated, it is convenient to
define the variance of terms in the formulas for sampling variance as a
function of the number of sets (s) . Let y be the vector of phenotypes
for one set. The variance of y can be written
Var(y) = V = V X a2 + v 2 a ggm + V 3 a^ + V 4 a* + i n c|, [94 ]
where n is the number of individuals in a set (n = 8 for Thompson
designs, n = 18 for Eisen designs); and V lr V 2 , V 3 , and V 4 are n x n
matrices of coefficients which when multiplied by °|, Oggu,/ c*™* and cf2
specify the contribution of a|, o , a^ , and ojj- to the variance among
individuals in a set. V^, V 2 , V3, and V 4 for the Eisen and Thompson
designs are in Appendix B for the individuals in a set. The
coefficients in V^, for i = 1, 2, ... 4, were obtained from table 1
of Thompson (1976) and table 1 of Eisen (1967) . For the V^ matrixes in
Appendix B, y was sorted for the Thompson designs as Pp P 2 , ... Pg
(figure 1) and for the Eisen designs as S 1# S 2 , Dj, D 2 , ... Dg, 0-^, 2 ,
. . . 8 (Figure 2) .
Relative efficiency in this study was computed as:
/ Var(a?)
Relative Efficiency = / , [95]
\ / Var(af)
where a? is an SDS, WSDS, MIVQUE (0, M, E) , or MIVQUE(l) estimate of
the i^ 1 variance component and h ? is the MIVQUE estimate of the i
variance component.
53
The sampling variances of 3 and g 2 can be written as a f line t ion
of the number of sets (s) and the V^ matrices defined previously. The
sampling variance among MIVQUE estimates of the variance components for
data resulting from s sets is
Varfa 2 ) = 2Cs[tr(\T 1 V i <Nr 1 V j )] - 2(l^V _1 l n )" :l [l^ ;L V i \r 1 V j \r 1 l n ]
+ dAV" 1 ^) " 2 Il^V^lnl [l^VjV 1 ^] 3 _1 [96]
for i, j = 1, 2, ... 5, where
V 5 = In-
The sampling variance among SDS estimates of (co) variance components is
Var(a 2 ) = (X'X) -1 Var(X'Y) (X'X)" 1 ,
[97]
where for s sets
Var(X'Y) = BstsCs^trtV^nljIjV 2 :^ + (l n Vl n ) 2 )l c+1 l c+1
- sn[ [tr (FiV 2 ) ] 1^ +1 + l c+1 [tr (FjV 2 ) ] 3
+ sn[ [tr (V^ 2 ) ] 1^ +1 + l c+1 [tr (VjV 2 ) ] 3
+ flA^i^n^c+l + W^j^nJ
- HAWiYlnll^! - l c+1 [lAWjVl n ]
+ [trfFiVFjV)]
- ItrtFiWjV)] - [tr^VFjV)]
+ [tr(V i W j V)]3,
[98]
where
P i =
Z [v n ) ±
Z (v 21 ) ±
? < v nl)i
[99]
for 1 = 1, 2,
n, and
54
X'X = 2s£sn 2 l c+1 l£ +1 - [lAVil n ]l' +1 - l^fl'Vjl^ + ttrtV^j)]). [100]
The sampling variance among WSDS estimates of (co) variance components
is
Var(o 2 ) = (X'W^X)" 1 VarfX'W^Y) (X'vrtt)" 1 , [101]
where for s sets
VartX'W^Y) = 2ts[tr(W i Wj)] - - [l^WjWjln]
1
+ -- f2 l'vi fi'V'W'i i + 2ri'wi l ri'wi 1
o a n v - L n ia n v n vv i J 'n J ZL - L n vv i- l n J L - L n vv n- l n J
+ ri'w-vi i ri'vi ] + fi'V'i i [l'wvi ]
LJ -n vv i VA n J t-'rrj-'-n- 1 L4 -n Y i J -n J L - l n vv 3 v - L n J
2
- -- ri'vi fri'vvi i ri'vi 1 + ri'vi lri'w-i n
o l± n va n l LX n vv i a n J ^n^-'-n- 1 lx n v i A n J lJ, n vv n J, n J J
n J
+ S ^A^n) 2 tlA V iV tl n Vjl n ] 3 . ^02]
and
X^X = sttr^Vj)] - 2 ElnVjVjlJ + i- [1^1^ [l'Vjl^ . [103]
The sampling variance among MIVQUE (0, M, E) or MIVQUE(l) estimates of
(co) variance components is as follows:
Var(a 2 ) = (IHS)"" 1 Var(RHS) (ms)" 1 , [104]
where
a 2 = (LHSJ^RHS. [105]
55
For s sets,
Var(RHS) - 2{s[tr{GV i GVj)]
+ 2[^v- 1 v i Gi n 3Ei r ;v^ 1 v j Gi n ]
+ Cl^GViGl^El^Vj^ 1 !^ }
+ Ci^v i Gi n ]Ci-v^ 1 v j ^i n ]>
+ ^A^ln)" 4 ^^) 2 ^^^ 1 ^]:^^^ 1 !^}, [106]
where
V a = o%y 4 + a^, for MIVQUE (0, M, E), and
V a ~ v l + v 2 + V 3 + V 4 + I n f for M*VQUE{1), and
G = V^W^ 1 ; and
+ (^^V'^^^i^yciA^vs; 1 ^:. C1073
Two hundred sets were considered resulting in 1,600 individuals
for Thompson (1976) designs and 3,600 individuals for Eisen designs.
Each parameter other than the error variance was set to two levels
- direct genie at 15 or 30, maternal genie at 15 or 30, correlation
between direct and maternal genie at -.1 or -.3, and permanent maternal
environment at 20 or 50. The phenotypic variance was held constant at
100. Taking all possible combinations of the four parameters at two
levels each results in 16 parameter set combinations. The error
variance was computed as the difference between 100 and the sum of all
56
other components. However, for the parameter set in which ag = 30,
0q m = 30, a ggm = -.1, and ojf, = 50, the sum of these components is 104
>100. Therefore, for this set, a^ = 27.8, a| m = 27.8, 0g qm = -.1, ajj,
= 46.3, and c| = 3.70 to hold the phenotypic variance constant and to
o 7 7
keep a* positive. In addition, the parameter set a g = 9.8, a gm = 9.8,
a = -1.96, and cr^ = 21.56 was used to make a direct comparison of
this numerical procedure with the simulation study of Grimes and Harvey
(1980).
RESULTS AND DISCUSSION
Standard errors of SDS estimates of (co) variance components
computed as the square root of the result of [97] agree quite well with
the standard errors obtained in the simulation study of Grimes and
Harvey (1980) (table 5) .
Table 5. Comparison of standard errors of SDS estimates of variance
components obtained by simulation study of Grimes and Harvey (1980) for
a maternal effects model with standard errors obtained by the numerical
method of the present study.
Item
Mating'
Design
Parameters
a a 2 a 2
ggm gm m
o2
e
Parameter values
Current study A
Grimes and Harvey (1980)
Current study B
Grimes and Harvey (1980)
Parameter values
9.8 -1.96 9.8 9.8 21.56
Standard errors
9.3 9.2 15.9 9.8 5.1
9.4 10.1 17.1 10.3 5.0
12.1
12.8
6.2
6.8
7.9
7.9
4.9
5.0
6.2
6.4
a The mating designs were due to Thompson (1976) , of which 200
replicates were used.
°qr a qm' a m' a™ 3 o 2 are variances due to direct genie, maternal
genie, permanent maternal environmental, and environmental effects,
and a nam is the covariance between direct and maternal genie effects.
The efficiencies [95] of SDS and WSDS estimates of (co) variance
components relative to MIVQUE are in tables 6 and 7. In tables 6 4 and
7, the efficiencies corresponding to each parameter set are ranked by
o^ from high to low. Efficiency of SDS and WSDS relative to MIVQUE
seems to follow for the most part the same ranking as °| (table 6 and
57
58
7) . This result is to be expected for the relative efficiency of WSDS
because the assumptions underlying the MIVQUE(O) property of WSDS are
violated to a greater and greater extent as cf| becomes smaller. This
is because °?/ °g increases as cr^ decreases in this study.
Discrepancies between the rankings of efficiency of WSDS and a| are
due to certain combinations of parameters having more effect on
efficiency than other combinations. The efficiencies of SDS and WSDS
relative to MIVQUE were quite sensitive to decreases in cr| for all
designs except Thompson B. For the Thompson B design, efficiency of
SDS and WSDS varied little over all parameter sets studied. Comparison
of tables 6 and 7 showed that WSDS is more efficient than SDS for
almost all parameter sets, estimates of variance components, and mating
designs. Exceptions to this result occur when °^ is estimated from
Eisen designs I, II, and III or when a| ± s estimated from Eisen design
III.
59
Table 6. Efficiency 3 of SDS estimates of (co) variance components
relative to MIVQUE estimates computed from mating designs of Thompsons
and Eisen" replicated 200 times each.
Parameters
•5
"2
°g
Efficiency
°ggm °gm
of
;2
a m
Rank d
°1
°%
gm
r ggm
4
Thompson A Design
1
59.00
15
15
-.3
20
.66
.67
.68
.71
.66
2
53.00
15
15
-.1
20
.66
.66
.68
.72
.66
3
47.72
15
30
-.3
20
.65
.64
.66
.71
.64
3
47.72
30
15
-.3
20
.68
.67
.68
.72
.67
5
39.24
15
30
-.1
20
.65
.63
.66
.71
.63
5
39.24
30
15
-.1
20
.67
.66
.68
.72
.65
7
38.00
30
30
-.3
20
.66
.64
.66
.71
.64
8
29.00
15
15
-.3
50
.62
.58
.61
.68
.59
9
26.00
30
30
-.1
20
.65
.62
.65
.71
.62
10
23.00
15
15
-.1
50
.60
.55
.60
.67
.57
11
17.72
15
30
-.3
50
.57
.50
.56
.65
.54
11
17.72
30
15
-.3
50
.61
.56
.60
.68
.58
13
9.24
15
30
-.1
50
.51
.44
.52
.63
.48
13
9.24
30
15
-.1
50
.57
.51
.57
.66
.54
15
8.00
30
30
-.3
50
.56
.49
.55
.65
.52
16
3.70
27.8
27.8
-.1
46.3
.52
.45
.53
.64
.49
Thompson B Design
1
59.00
15
15
-.3
20
.80
.89
.88
.95
.82
2
53.00
15
15
-.1
20
.80
.90
.89
.95
.81
3
47.72
15
30
-.3
20
.80
.89
.89
.95
.81
3
47.72
30
15
-.3
20
.82
.91
.88
.94
.83
5
39.24
15
30
-.1
20
.80
.90
.89
.95
.80
5
39.24
30
15
-.1
20
.82
.91
.89
.94
.82
7
38.00
30
30
-.3
20
.82
.91
.89
.95
.83
8
29.00
15
15
-.3
50
.82
.88
.82
.93
.83
9
26.00
30
30
-.1
20
.80
.91
.89
.95
.80
10
23.00
15
15
-.1
50
.82
.89
.81
.93
.82
11
17.72
15
30
-.3
50
.82
.88
.80
.93
.82
11
17.72
30
15
-.3
50
.84
.90
.81
.93
.84
13
9.24
15
30
-.1
50
.81
.88
.77
.93
.81
13
9.24
30
15
-.1
50
.83
.90
.80
.93
.83
15
8.00
30
30
-.3
50
.83
.89
.79
.93
.83
16
3.70
27.8
27.8
-.1
46.3
.81
.90
.79
.93
.81
60
Table 6 continued
Parameters
Efficiency
of
Rank d
n 2
°e
~2
a g
«2
a gm
ggm
°l
Z 2
a g
a ggm
n 2
a gm
n 2
CT m
Jl
Eisen
I Design
1
59.00
15
15
-.3
20
.53
.72
.61
.67
.65
2
53.00
15
15
-.1
20
.53
.71
.62
.67
.62
3
47.72
15
30
-.3
20
.51
.71
.64
.69
.57
3
47.72
30
15
-.3
20
.56
.72
.62
.67
.63
5
39.24
15
30
-.1
20
.49
.69
.64
.69
.53
5
39.24
30
15
-.1
20
.55
.71
.62
.67
.58
7
38.00
30
30
-.3
20
.53
.71
.64
.68
.55
8
29.00
15
15
-.3
15
.46
.64
.56
.64
.46
9
26.00
30
30
-.1
20
.49
.68
.64
.67
.48
10
23.00
15
15
-.1
50
.44
.62
.56
.64
.42
11
17.72
15
30
-.3
50
.40
.58
.55
.63
.37
11
17.72
30
15
-.3
50
.46
.63
.56
.63
.43
13
9.24
15
30
-.1
50
.35
.52
.54
.62
.31
13
9.24
30
15
-.1
50
.42
.58
.56
.63
.37
15
8.00
30
30
-.3
50
.40
.56
.55
.62
.35
16
3.70
27.8
27.8
-.1
46.3
.36
.53
.55
.62
.30
Eisen
II Design
1
59.00
15
15
-.3
20
.62
.76
.72
.75
.67
2
53.00
15
15
-.1
20
.61
.76
.73
.76
.64
3
47.72
15
30
-.3
20
.59
.74
.72
.78
.60
3
47.72
30
15
-.3
20
.66
.78
.72
.75
.66
5
39.24
15
30
-.1
20
.56
.73
.73
.79
.56
5
39.24
30
15
-.1
20
.64
.78
.72
.76
.62
7
38.00
30
30
-.3
20
.62
.75
.72
.78
.60
8
29.00
15
15
-.3
50
.54
.69
.66
.79
.50
9
26.00
30
30
-.1
20
.58
.74
.72
.78
.54
10
23.00
15
15
-.1
50
.51
.67
.66
.80
.47
11
17.72
15
30
-.3
50
.47
.62
.65
.80
.42
11
17.72
30
15
-.3
50
.55
.69
.66
.78
.50
13
9.24
15
30
-.1
50
.41
.57
.65
.81
.36
13
9.24
30
15
-.1
50
.50
.66
.65
.79
.44
15
8.00
30
30.
-.3
50
.48
.62
.64
.79
.42
16
3.70
27.8
27.8
-.1
46.3
.44
.59
.65
.80
.37
61
Table 6 continued
Parameters
~2
°9
Efficiency
a ggm a gm
of
"75
Rank d
4
4
2
a gm
r ggm
2
- CT m
~2
Eisen
III
Design
1
59.00
15
15
-.3
20
.67
.87
.90
.91
.71
2
53.00
15
15
-.1
20
.66
.87
.89
.91
.72
3
47.72
15
30
-.3
20
.65
.85
.87
.91
.66
3
47.72
30
15
-.3
20
.70
.88
.88
.90
.70
5
39.24
15
30
-.1
20
.63
.84
.85
.89
.62
5
39.24
30
15
-.1
20
.68
.87
.86
.88
.66
7
38.00
30
30
-.3
20
.67
.84
.84
.89
.65
8
29.00
15
15
-.3
50
.61
.83
.85
.89
.57
9
26.00
30
30
-.1
20
.63
.82
.81
.85
.58
10
23.00
15
15
-.1
50
.59
.82
.83
.88
.53
11
17.72
15
30
-.3
50
.56
.76
.78
.85
.48
11
17.72
30
15
-.3
50
.62
.83
.81
.86
.54
13
9.24
15
30
-.1
50
.50
.72
.74
.82
.41
13
9.24
30
15
-.1
50
.57
.80
.77
.83
.48
15
8.00
30
30
-.3
50
.56
.75
.75
.82
.46
16
3.70
27.8
27.8
-.1
46.
3 .51
.72
.73
.80
.41
a Efficiency is the standard error of the MIVQUE estimate of
(co) variance component divided by the standard error of the SDS
estimate.
" Mating designs are defined in the text.
c Parameters are defined as follows:
a£, aJL,, a£. and o£ are variances
rcu-aiicLCia iuc uciiucu ao j.uj.iuwo. a' am' m' c ""- i u Q «j.«= »uj.j.uin-ti3
due to direct additive genetic, maternal additive genetic, permanent
maternal environmental, and residual error effects; and r_ gm and aggm
are the correlation and covariance between direct and maternal
additive genetic effects.
Parameter sets are ranked by a£, from high to low.
62
Table 7. Efficiency 3 of WSDS estimates of (co) variance components
relative to MIVQUE estimates computed from mating designs of Thompson*-*
and Eisen" replicated 200 times each.
Parameters 1 -
Efficiency
of
Rank d
°l
•I
a 2
u gm
ggm
m
J2~
°g
Design
a ggm
A 2
a gm
°m
T 2 ~
°e_
Thompson A
1
59.00
15
15
-.3
20
.96
.95
.95
.97
.97
2
53.00
15
15
-.1
20
.96
.93
.94
.95
.96
3
47.72
15
30
-.3
20
.94
.91
.91
.94
.95
3
47.72
30
15
-.3
20
.93
.92
.93
.95
.94
5
39.24
15
30
-.1
20
.92
.89
.89
.91
.93
5
39.24
30
15
-.1
20
.92
.90
.91
.93
.92
7
38.00
30
30
-.3
20
.91
.89
.90
.93
.92
8
29.00
15
15
-.3
50
.87
.80
.83
.89
.87
9
26.00
30
30
-.1
20
.88
.84
.85
.89
.88
10
23.00
15
15
-.1
50
.84
.76
.80
.87
.85
11
17.72
15
30
-.3
50
.80
.70
.75
.84
.80
11
17.72
30
15
-.3
50
.83
.75
.80
.87
.83
13
9.24
15
30
-.1
50
.72
.61
.69
.80
.72
13
9.24
30
15
-.1
50
.76
.69
.75
.84
.77
15
8.00
30
30
-.3
50
.76
.67
.73
.83
.76
16
3.70
27.8
27.8
-.1
46.3
1 .71
.62
.69
.80
.71
Thompson B
Design
1
59.00
15
15
-.3
20
.99
.99
.97
.99
.99
2
53.00
15
15
-.1
20
.98
.99
.97
.99
.98
3
47.72
15
30
-.3
20
.97
.99
.96
.99
.97
3
47.72
30
15
-.3
20
.97
.98
.97
.99
.97
5
39.24
15
30
-.1
20
.96
.99
.95
.98
.96
5
39.24
30
15
-.1
20
.96
.98
.97
.99
.96
7
38.00
30
30
-.3
20
.96
.98
.96
.98
.96
8
29.00
15
15
-.3
50
.99
.95
.87
.96
.99
9
26.00
30
30
-.1
20
.93
.98
.95
.98
.93
10
23.00
15
15
-.1
50
.98
.96
.86
.95
.98
11
17.72
15
30
-.3
50
.98
.95
.83
.95
.98
11
17.72
30
15
-.3
50
.98
.95
.87
.96
.98
13
9.24
15
30
-.1
50
.96
.95
.80
.95
.96
13
9.24
30
15
-.1
50
.96
.95
.85
.95
.96
15
8.00
30
30
-.3
50
.97
.95
.83
.95
.97
16
3.70
27.8
27.8
-.1
46.3
1 .95
.95
.82
.95
.95
63
Table 7 continued
a2
Parameters
cj2 G 2
r
o2
Efficiency
of
Rank^
52
a
a2
— 2 —
IT
e
g
gm
ggm
m
g
ggm
gm
m
e
Eisen
I Design
1
59.00
15
15
-.3
20
.92
.94
.92
.82
.80
2
53.00
15
15
-.1
20
.90
.93
.90
.80
.76
3
47.72
15
30
-.3
20
.87
.89
.85
.74
.68
3
47.72
30
15
-.3
20
.87
.90
.87
.78
.75
5
39.24
15
30
-.1
20
.82
.86
.80
.69
.62
5
39.24
30
15
-.1
20
.82
.88
.83
.74
.69
7
38.00
30
30
-.3
20
.82
.85
.81
.70
.65
8
29.00
15
15
-.3
15
.84
.86
.85
.64
.52
9
26.00
30
30
-.1
20
.74
.80
.74
.63
.55
10
23.00
15
15
-.1
50
.80
.83
.82
.61
.47
11
17.72
15
30
-.3
50
.74
.75
.75
.56
.41
11
17.72
30
15
-.3
50
.76
.80
.80
.60
.48
13
9.24
15
30
-.1
50
.64
.67
.69
.53
.34
13
9.24
30
15
-.1
50
.68
.74
.75
.56
.40
15
8.00
30
30
-.3
50
.66
.69
.71
.54
.38
16
3.70
27.8
27.8
-.1
46.3
.59
.64
.68
.52
.33
Eisen
II Design
1
59.00
15
15
-.3
20
.94
.95
.94
.88
.91
2
53.00
15
15
-.1
20
.91
.94
.92
.86
.88
3
47.72
15
30
-.3
20
.88
.90
.88
.80
.84
3
47.72
30
15
-.3
20
.90
.93
.90
.83
.88
5
39.24
15
30
-.1
20
.84
.88
.86
.78
.79
5
39.24
30
15
-.1
20
.86
.91
.88
.80
.84
7
38.00
30
30
-.3
20
.86
.88
.85
.77
.82
8
29.00
15
15
-.3
50
.83
.85
.87
.76
.71
9
26.00
30
30
-.1
20
.79
.84
.82
.73
.74
10
23.00
15
15
-.1
50
.79
.82
.85
.74
.67
11
17.72
15
30
-.3
50
.73
.74
.80
.69
.61
11
17.72
30
15
-.3
50
.79
.81
.84
.72
.68
13
9.24
15
30
-.1
50
.64
.67
.77
.67
.52
13
9.24
30
15
-.1
50
.71
.76
.80
.69
.61
15
8.00
30
30.
-.3
50
.69
.71
.78
.67
.59
16
3.70
27.8
27.8
-.1
46.3
.63
.67
.76
.66
.53
64
Table 7 continued
o2
Parameters
a 2 cr2
T
' o2
Efficiency
of
Rank^
52"
o2
32
a2
i\ni v\
e
g
gm
L ggm
m
g
ggm gm
m
e
Eisen
III Design
1
59.00
15
15
-.3
20
.95
.97 .95
.87
.83
2
53.00
15
15
-.1
20
.93
.97 .94
.85
.80
3
47.72
15
30
-.3
20
.91
.94 .90
.80
.74
3
47.72
30
15
-.3
20
.92
.95 .93
.82
.78
5
39.24
15
30
-.1
20
.88
.92 .87
.77
.67
5
39.24
30
15
-.1
20
.89
.94 .90
.79
.72
7
38.00
30
30
-.3
20
.89
.92 .88
.76
.70
8
29.00
15
15
-.3
15
.86
.91 .88
.72
.58
9
26.00
30
30
-.1
20
.83
.89 .84
.70
.60
10
23.00
15
15
-.1
50
.83
.89 .87
.70
.53
11
17.72
15
30
-.3
50
.78
.83 .80
.65
.48
11
17.72
30
15
-.3
50
.82
.88 .85
.68
.53
13
9.24
15
30
-.1
50
.69
.77 .76
.61
.40
13
9.24
30
15
-.1
50
.76
.85 .81
.63
.46
15
8.00
30
30
-.3
50
.74
.80 .77
.61
.45
16
3.70
27.8
27.8
-.1
46.3
.68
.77 .75
.59
.40
a Efficiency is the standard error of the MIVQUE estimate of
(co) variance component divided by the standard error of the WSDS
estimate.
Mating designs are defined in the text.
°' °gm' °m'
£, and a \ are variances
c Parameters are defined as follows: o _ _ m , e —
due to direct additive genetic, maternal additive genetic, permanent
maternal environmental, and residual error effects; and r nrm and
are the correlation and covariance between direct
additive genetic effects.
d Parameter sets are ranked by a 2 ,, from high to low.
and maternal
65
The efficiencies of MIVQUE (0, M, E) estimates of (co)variance
components relative to MIVQUE are in table 8. In table 8, efficiencies
corresponding to each parameter set are ranked by cr m + a* from high to
low. The efficiency of MIVQUE (0, M, E) appears to follow the same
ranking with some exceptions as o^ + a | {table 8). Unlike SDS and
WSDS, the efficiency of MIVQUE (0, M, E) relative to MIVQUE remained
relatively high (greater than .80} over all parameter sets studied. In
practice, this is an upper bound on the efficiency that could be
attained for MIVQUE (0, M, E) because estimates for <?2 and a| would
need to be used in place of the true parameters.
66
Table 8. Relative efficiency 3 of MIVQUE (0, M, E) estimates of
(co) variance components relative to MIVQUE estimates computed from
mating designs of Thompson 13 (1976) and Eisen b (1967) replicated 200
times each.
2 2
m e
Parameters
2 2
Ot. OX™.
g u gm
r ggm
r, 2
°m
Efficiency
of
Rank d
^ 2
a g
°ggm
n 2
CT gm
-2
a m
-2
Thompson A
Design
1
79.00
15
15
-.3
20
.99
.99
.99
.99
.99
1
79.00
15
15
-.3
50
.99
.99
.99
.99
.99
3
73.00
15
15
-.1
20
.98
.99
.98
.98
.99
3
73.00
15
15
-.1
50
.99
.99
.99
.99
.99
5
67.72
15
30
-.3
20
.98
.98
.98
.98
.98
5
67.72
15
30
-.3
50
.99
.98
.98
.98
.99
5
67.72
30
15
-.3
20
.97
.98
.98
.98
.97
5
67.72
30
15
-.3
50
.97
.98
.98
.98
.97
9
59.24
15
30
-.1
20
.97
.97
.96
.96
.97
9
59.24
15
30
-.1
50
.99
.98
.98
.97
.99
9
59.24
30
15
-.1
20
.96
.97
.97
.97
.96
9
59.24
30
15
-.1
50
.95
.95
.97
.97
.95
13
58.00
30
30
-.3
20
.96
.98
.97
.97
.96
13
58.00
30
30
-.3
50
.96
.96
.97
.97
.96
15
50.00
27.8
27.8
-.1
46.:
3 .91
.92
.95
.95
.95
16
46.00
30
30
-.1
20
.95
.96
.96
.95
.91
Thompson B
Design
1
79.00
15
15
-.3
20
.99
.99
.99
.99
.99
1
79.00
15
15
-.3
50
.99
.99
.99
.99
.99
3
73.00
15
15
-.1
20
.98
.99
.99
.99
.98
3
73.00
15
15
-.1
50
.98
.98
.98
.98
.98
5
67.72
15
30
-.3
20
.97
.99
.99
.99
.97
5
67.72
15
30
-.3
50
.98
.98
.98
.97
.98
5
67.72
30
15
-.3
20
.97
.98
.99
.98
.97
5
67.72
30
15
-.3
50
.98
.97
.97
.97
.98
9
59.24
15
30
-.1
20
.96
.98
.99
.97
.96
9
59.24
15
30
-.1
50
.96
.96
.96
.95
.96
9
59.24
30
15
-.1
20
.96
.97
.98
.97
.96
9
59.24
30
15
-.1
50
.96
.94
.93
.94
.96
13
58.00
30
30
-.3
20
.96
.98
.99
.97
.96
13
58.00
30
30
-.3
50
.97
.95
.94
.94
.97
15
50.00
27.8
27.8
-.1
46.:
3 .95
.92
.90
.91
.95
16
46.00
30
30
-.1
20
.93
.95
.97
.95
.93
67
Table 8 continued.
Parameters
Efficiency
of
Rank d
2 2
•§
gm
r ggm
a 2
m
~2
°g
a ggm
-2
"sT
~w
Eisen
I Design
1
79.00
15
15
-.3
20
.95
.96
.95
.95
.96
1
79.00
15
15
-.3
50
.96
.97
.96
.96
.96
3
73.00
15
15
-.1
20
.93
.95
.93
.93
.95
3
73.00
15
15
-.1
50
.94
.96
.94
.95
.95
5
67.72
15
30
-.3
20
.91
.92
.89
.89
.92
5
67.72
15
30
-.3
50
.92
.94
.91
.91
.92
5
67.72
30
15
-.3
20
.90
.93
.91
.92
.92
5
67.72
30
15
-.3
50
.92
.95
.93
.94
.92
9
59.24
15
30
-.1
20
.87
.90
.85
.86
.89
9
59.24
15
30
-.1
50
.91
.94
.90
.89
.90
9
59.24
30
15
-.1
20
.87
.91
.87
.89
.89
9
59.24
30
15
-.1
50
.90
.92
.91
.91
.89
13
58.00
30
30
-.3
20
.88
.89
.86
.87
.88
13
58.00
30
30
-.3
50
.91
.93
.91
.89
.90
15
50.00
27.8
27.8
-.1
46.3
.88
.90
.88
.86
.87
16
46.00
30
30
-.1
20
.81
.85
.81
.81
.82
Eisen
II Design
1
79.00
15
15
-.3
20
.97
.98
.97
.96
.98
1
79.00
15
15
-.3
50
.97
.98
.98
.98
.98
3
73.00
15
15
-.1
20
.95
.97
.96
.95
.96
3
73.00
15
15
-.1
50
.96
.98
.97
.97
.96
5
67.72
15
30
-.3
20
.93
.94
.93
.92
.95
5
67.72
15
30
-.3
50
.94
.96
.96
.96
.95
5
67.72
30
15
-.3
20
.94
.96
.94
.93
.95
5
67.72
30
15
-.3
50
.94
.96
.96
.95
.94
9
59.24
15
30
-.1
20
.89
.93
.91
.91
.91
9
59.24
15
30
-.1
50
.92
.95
.96
.95
.93
9
59.24
30
15
-.1
20
.90
.94
.93
.91
.92
9
59.24
30
15
-.1
50
.90
.93
.93
.93
.90
13
58.00
30
30
-.3
20
.91
.93
.91
.90
.92
13
58.00
30
30
-.3
50
.91
.93
.94
.93
.91
15
50.00
27.8
27.8
-.1
46.3
.88
.90
.92
.91
.88
16
46.00
30
30
-.1
20
.85
.90
.90
.88
.87
68
Table 8 continued.
Parameters
Efficiency
of
Rank d
m e
°i
°jm
r ggm
°2
m
^2
a g
CT ggm
°gm
a m
°e
Eisen
III
Design
1
79.00
15
15
-.3
20
.98
.98
.98
.97
.98
1
79.00
15
15
-.3
50
.98
.98
.99
.98
.98
3
73.00
15
15
-.1
20
.96
.98
.97
.96
.96
3
73.00
15
15
-.1
50
.97
.98
.98
.98
.97
5
67.72
15
30
-.3
20
.95
.96
.94
.94
.94
5
67.72
15
30
-.3
50
.96
.96
.96
.97
.97
5
67.72
30
15
-.3
20
.95
.97
.96
.95
.95
5
67.72
30
15
-.3
50
.95
.96
.97
.96
.95
9
59.24
15
30
-.1
20
.93
.95
.93
.93
.92
9
59.24
15
30
-.1
50
.94
.96
.96
.96
.94
9
59.24
30
15
-.1
20
.93
.95
.95
.93
.92
9
59.24
30
15
-.1
50
.91
.92
.94
.93
.92
13
58.00
30
30
-.3
20
.93
.94
.93
.92
.93
13
58.00
30
30
-.3
50
.92
.94
.96
.94
.93
15
50.00
27.8
27.8
-.1
46.
3 .88
.88
.90
.90
.88
16
46.00
30
30
-.1
20
.89
.92
.91
.90
.89
a Relative efficiency is the standard error of the MIVQUE estimate of
( co ) variance component divided by the standard error of the MIVQUE
(0, M, E) estimate.
b Thompson and Eisen Mating designs are defined in the text.
c Parameters are defined as follows: cig, a 2 ™, a^, and CT | are variances
due to direct additive genetic, maternal additive genetic, permanent
maternal environmental, and residual error effects; and *g and a (
are the correlation and covariance between direct and mater
additive genetic effects.
d Parameter sets are ranked by o| + o 2 , from high to low.
69
The efficiencies of MIVQUE{1) estimates of (co) variance components
relative to MIVQUE are presented in table 9. Like MIVQUE (0, M, E) ,
the efficiency of MIVQUE (1) relative to MIVQUE remained relatively high
(greater than .8) over all parameter sets studied. The efficiencies of
MIVQUE (0, M, E) and MIVQUE (1) are similar with MIVQUE (0, M, E) being
higher for some parameter sets and MIVQUE (1) being higher for others.
70
Table 9. The efficiency 9 of MIVQUE(l) estimates of (co) variance
components relative to MIVQUE estimates computed from 200 sets of
designs described by Thompson 13 (1976) and Eisen b (1967).
Parameters
.c
Efficiency of
o 2
r\
a 2
<£
o 2
GZL
"2
%
%2
_g
ggm
gm
m
u e
g
ggm
gm
m
e
Thompson A
15
-.1
15
20
53
.93
.89
.89
.92
.94
50
23
.97
.95
.95
.96
.97
30
20
39.24
.97
.93
.94
.95
.97
50
9.24
.91
.89
.93
.95
.91
-.3
15
20
59
.91
.87
.87
.90
.93
50
29
.97
.94
.94
.95
.97
30
20
47.72
.95
.90
.91
.93
.95
50
17.72
.95
.92
.94
.95
.95
30
-.1
15
20
39.24
.96
.93
.93
.94
.97
50
9.24
.96
.95
.96
.96
.96
30
20
26
.99
.97
.97
.97
.99
27.8
-.1
27.8
46.3
3.70
.92
.91
.95
.96
.91
30
-.3
15
20
47.72
.94
.90
.89
.92
.95
50
17.72
.98
.95
.96
.96
.97
30
20
38
.97
.93
.93
.94
.97
50
8
.94
.92
.95
.96
.94
Thompson B
15
-.1
15
20
53
.90
.95
.95
.95
.90
50
23
.93
.92
.92
.93
.92
30
20
39.24
.94
.96
.95
.96
.94
50
9.24
.96
.93
.88
.93
.96
-.3
15
20
59
.87
.95
.94
.94
.89
50
29
.91
.92
.91
.92
.90
30
20
47.72
.91
.95
.94
.94
.91
50
17.72
.94
.92
.89
.92
.93
30
-.1
15
20
39.24
.93
.97
.97
.97
.93
50
9.24
.95
.94
.92
.94
.95
30
20
26
.96
.97
.97
.97
.96
27.8
-.1
27.8
46.3
3.70
.97
.94
.91
.95
.97
30
-.3
15
20
47.72
.90
.95
.96
.96
.90
50
17.72
.92
.93
.93
.93
.92
30
20
38
.93
.96
.95
.96
.92
50
8
.95
.92
.90
.93
.94
71
Table 9 continued.
Parameters
,c
Efficiency
of
°1
r ggm
gin
°l
•2
"2
°9
°ggm
5 gm
"2
"2
e
Eisen
I
15
-.1
15
20
53
.84
.91
.88
.89
.87
50
23
.91
.90
.84
.89
.90
30
20
39.24
.91
.93
.90
.93
.91
50
9.24
.88
.80
.80
.89
.85
-.3
15
20
59
.81
.89
.85
.86
.85
50
29
.89
.90
.82
.87
.88
30
20
47.72
.87
.91
.87
.90
.88
50
17.72
.90
.85
.80
.90
.87
30
-.1
15
20
39.24
.92
.95
.92
.93
.93
50
9.24
.91
.89
.86
.91
.89
30
20
26
.96
.95
.93
.96
.95
27.8
-.1
27.8
46.3
3.70
.86
.81
.84
.91
.83
30
-.3
15
20
47.72
.88
.93
.88
.90
.89
50
17.72
.92
.90
.84
.90
.90
30
20
38
.92
.93
.90
.93
.92
50
8
Eisen
.88
II
.82
.81
.89
.84
15
-.1
15
20
53
.88
.92
.93
.95
.90
50
23
.92
.93
.92
.94
.92
30
20
39.24
.93
.95
.93
.95
.93
50
9.24
.88
.85
.89
.92
.86
-.3
15
20
59
.85
.91
.91
.94
.88
50
29
.90
.93
.91
.93
.90
30
20
47.72
.89
.93
.91
.94
.90
50
17.72
.90
.89
.88
.91
.89
30
-.1
15
20
39.24
.93
.96
.95
.97
.94
50
9.24
.91
.92
.94
.96
.90
30
20
26
.96
.97
.95
.97
.96
27.8
-.1
27.8
46.3
3.70
.87
.86
.92
.94
.86
30
-.3
15
20
47.72
.90
.94
.93
.96
.91
50
17.72
.92
.93
.93
.95
.91
30
20
38
.93
.95
.93
.96
.93
50
8
.89
.87
.90
.93
.87
72
Table 9 continued.
______
________
________
Parameters
r a 2
ggm gm
c
a 2
m
°I
a 2
9
Efficiency
a a
ggm gm
of
'§
*2
m
~2
a e
Eisen
III
15
-.1
15
20
53
.88
.94
.90
.93
.90
50
23
.92
.91
.90
.94
.91
30
20
39.24
.93
.94
.93
.95
.93
50
9.24
.88
.82
.86
.92
.87
-.3
15
20
59
.86
.92
.88
.91
.87
50
29
.90
.90
.89
.93
.90
30
20
47.72
.89
.93
.90
.94
.90
50
17.72
.90
.85
.87
.92
.89
30
-.1
15
20
39.24
.94
.96
.93
.96
.94
50
9.24
.92
.90
.91
.95
.90
30
20
26
.97
.96
.95
.97
.96
27.8
-.1
27.8
46.3
3.70
.89
.85
.89
.94
.87
30
-.3
15
20
47.72
.91
.95
.90
.94
.91
50
17.72
.92
.91
.90
.95
.91
30
20
38
.93
.94
.92
.96
.92
50
8
.90
.85
.87
.93
.87
a Efficiency is the standard error of the MIVQUE estimate of
(co) variance component divided by the standad error of the
MIVQUE (1) estimate.
b Thompson and Eisen mating designs are described in the text.
,2 f
2rni
G |m' a m'
Parameter sets are defined as follows: a
variances due to direct genie, maternal "genie,
maternal environmental, and residual effects, and r„
are the correlation and covariance between direct a
genie effects.
and o| are
permanent
i and a ggm
1 maternal
SUMMARY AND CONCLUSIONS
In papulations of farm animals, there are many different kinds of
relatives. In general, analysis of variance (ANOVA) type methods are
not adequate to estimate (co) variance components from data in which the
animal model is needed because of the complex covariance structure
among the observations. Minimum variance quadratic unbiased estimates
(MIVQUE) (Rao, 1971) and restricted maximum likelihood estimates (REML)
(Patterson and Thompson, 1971) have been extended by Henderson (1985a
and b) to an animal model which includes both additive and nonadditive
genetic effects. With Henderson's (1985a and b) approach, the mixed
model equations (MME) and best linear unbiased predictions (HLUP) for
genetic merits are obtained as intermediate results.
Quaas and Pollack (1980) described a reduced animal model (RAM)
with MME only for parents that have progeny records equivalent to the
animal model with MME for all animals when the genetic model is
strictly additive. Using RAM instead of the animal model can save a
great deal of computer storage when analyzing performance records from
some species. Hudson and Kennedy (1985) found that there were 5.6
times as many total animals as there were parents and ancestors that
had progeny with records in Ontario record of performance swine data.
REML and MIVQUE under RAM or the animal model are not
computationally feasible with 1986 computers when data sets are large.
Therefore, under the animal model or RAM with large data sets, methods
73
74
more computationally feasible than REML or MIVQUE are needed to obtain
estimates of (co) variance components.
Symmetric differences squared (SDS) (Grimes and Harvey, 1980),
weighted symmetric differences squared (WSDS) (Christian, 1980),
MIVQUE(O) (Rao, 1971), and diagonal MIVQUE (Henderson) are
computationally feasible methods of estimating (co) variance components
under RAM or the animal model.
Estimates of (co) variance components by symmetric differences
squared (SDS) weighted by the inverse of the error variance-covariance
matrix among SDS (WSDS) are minimum variance quadratic unbiased
estimates (MIVQUE) when all parameters other than the error variance
are near zero. Likewise, estimates of (co) variance components by SDS
weighted by the inverse of the total variance-covariance matrix among
SDS are MIVQUE for the prior parameters chosen.
The algorithm presented in this paper for obtaining WSDS greatly
reduces the computations needed to complete the analysis compared to
the WSDS methodology described by Christian (1980). The number of
multiplication steps required can be further reduced if the
relationships between pairs of animals can be classified into a
relatively small number (5 to 100) of relationship categories. This
algorithm follows the same approach for reducing the number of
multiplications required as Christian (1980) used to reduce the number
of multiplications needed to complete SDS. However, WSDS requires more
multiplication and addition steps to complete than SDS.
Matrices containing coefficients of relationship for additive and
nonadditive genetic effects are needed to apply WSDS or MIVQUE(O) to
75
models in which these effects are important. Relationship matrices for
dominance and epistatic effects can be computed from the numerator
relationship matrix (A) when there is no inbreeding (Henderson, 1985a).
Normally, an additive genetic model is assumed as an approximation when
there is inbreeding because the nonadditive effects are difficult to
interpret under inbreeding (Cockerham, 1954).
For large data sets under the animal model, the numerator
relationship matrix (A) is too large to half store in computer core if
all n(n+l)/2 elements are stored. However, A can be stored and used by
storing only the p < n(n+l)/2 nonzero elements (Hudson et al., 1982).
This process requires the storage of three vectors of order n and two
vectors of order p. Hudson et al. (1982) found that only 6 to 22.6% of
the n(n+l)/2 elements of the upper triangle of A were nonzero for five
dairy sire populations. The computational steps required to obtain A
are proportional to n 2 even with the storage saving procedures of
Hudson et al. (1982).
The sampling standard errors computed from 200 sets of designs A
and B of Thompson (1976) agree with the standard errors obtained by
simulation of the same designs by Grimes and Harvey (1980). These
results substantiate the validity of the numerical and simulation
methods used in these two studies.
WSDS was more efficient than SDS for most of the design-parameter
set combinations studied. Because WSDS was less than 75% efficient
relative to MIVQUE (parameters known a priori) for many of the design-
parameter set combinations studied, there is a need for computationally
feasible methods that are more efficient than WSDS.
76
MIVQUE with all priors set to zero except for the permanent
maternal environmental and residual variance, which were set to their
true values (MIVQUE(0,M,E)), and MIVQUE with all priors including the
error variance set to one (MIVQUE(l)) were greater than 80% efficient
relative to MIVQUE for all design-parameter combinations studied, and
were greater than 90% efficient for over half of the design-parameter
combinations studied. For MIVQUE(0,M,E), the efficiencies obtained in
this study are an upper bound because in practice the priors chosen for
permanent maternal environmental and residual variances would have to
be estimates instead of the true parameters. MIVQUE(0, M,E) is much
easier to compute than MIVQUE or MIVQUE(l) because the inverse of the
variance-covariance matrix among observations is easy to compute when
all parameters except the permanent maternal environmental and residual
variances are assumed to be zero. MIVQUE(l) offers no computational
advantage over MIVQUE, but it frees the experimenter from the
responsibility of assigning prior values.
A method not considered in the current study which offers poten-
tial as far as computational feasibility and efficiency relative to
MIVQUE is diagonal MIVQUE (Henderson, 1985a). Numerical studies
similar to the current one are needed to study the efficiency of
diagonal MIVQUE (Henderson, 1985a) relative to MIVQUE under the animal
or reduced animal model. Henderson (1984) states that diagonal MIVQUE
is more efficient than MIVQUE(O) when the unknown parameters deviate
greatly from zero. Diagonal MIVQUE is only slightly more
computationally burdensome to complete than MIVQUE(O) under the animal
model.
LITERATURE CITED
Brocklebank, J. and F.G. Giesbrecht. 1984. Estimating variance
components using alternative MINQE's in selected unbalanced
designs. In: Experimental Design, Statistical Models, and Genetic
Statistics. Vol. 50. Edited by Klaus Hinkelman. pp 177-211.
Christian, L.E. 1980. Modification and simplification of symmetric
differences squared procedure for estimation of genetic variances
and covarainces. Fh.D. Dissertation. The Ohio State University,
Columbus.
Cockerham, C.C. 1954. An extension of the concept of partitioning
hereditary variance for analysis of covariances among relatives
when epistasis is present. Genetics 39:859.
Dempster, A.P., N.M. Laird, and D.B. Rubin. 1977. Maximum likelihood
from incomplete data via the E.M. algorithm. J. of the Royal
Statistical Society 39:1.
Dickerson, G.E. 1942. Experimental design for testing inbred lines of
swine. J. Anim. Sci. 1:326.
Eisen, E.J. 1967. Mating designs for estimating direct and maternal
genetic variances and direct-maternal genetic covariances. Can. J.
Genet. Cytol. 9:13.
Forthofer, R.N. and G.G. Koch. 1974. An extension of the symmetric sum
approach to the estimation of variance components. Biometrische
Zeitschrift 16:3.
Grimes, L.W. and W.R. Harvey. 1980. Estimation of genetic variances and
covariances using symmetric differences squared. J. Anim. Sci.
50:634.
Hartley, H.O. and J.N.K. Rao. 1967. Maximum likelihood estimation for
the mixed analysis of variance model. Biometrics 54:93.
Hazel, L.N., M.L. Baker, and C.F. Reinmuller. 1943. Genetic and
environmental correlations between the growth rates of pigs at
different ages. J. Anim. Sci. 2:118.
Henderson, C.R. 1973. Sire evaluation and genetic trends. In: Proc. of
the Animal Breeding and Genetics Symposium in Honor of Dr. Jay L.
Lush, pp 10-41. ASAS and ADSA, Champaign, IL.
77
78
Henderson, CR. 1984. Applications of Linear Models in Animal Breeding.
University of Guelph. p. 173.
Henderson, C.R. 1985a. Best linear unbiased prediction of nonadditive
genetic merits in noninbred populations. J. Anim. Sci. 60:111-117.
Henderson, C.R. 1985b. MIVQUE and REML estimation of additive and
nonadditive genetic variance. J. Anim. Sci. 61:113.
Henderson, CR. 1985c. Equivalent linear models to reduce computations.
J. Dairy Sci. 68:2267.
Henderson, C.R., O. Kempthorne, S.R. Searle, and CM. Von Krosigk.
1959. The estimation of environmental and genetic trends from
records subject to culling. Biometrics 15:192.
Henderson, C.R. and R.L. Quaas. 1976. Multiple trait evaluation using
relatives' records. J. Anim. Sci. 43:1188.
Henderson, H.V. and S.R. Searle. 1981. On deriving the inverse of a sum
of matrices. SIAM Review 23:53.
Hudson, G.F.S., R.L. Quaas, and L.D. Van Vleck. 1982. Computer
algorithm for the recursive method of calculating large numerator
relationship matrices. J. Dairy Sci. 65:2018.
Hudson, G.F.S. and B.W. Kennedy. 1985. Genetic evaluation of swine for
growth rate and backfat thickness. J. Anim. Sci. 61:83.
Hudson, G.F.S. and L.D. Van Vleck. 1982. Estimation of components of
variance by method 3 and Henderson's new method. J. Dairy Sci.
65:435.
Kennedy, B.W., K. Johansson, and G.F.S. Hudson. 1985. Heritabilities
and genetic correlations for backfat and age at 90 kg in
performance-tested pigs. J. Anim. Sci. 61:78.
Koch, G.G. 1967. A general approach to the estimation of variance
components. Technometrics 9:93-118.
Koch, G.G. 1968. Some further remarks concerning 'a general approach to
the estimation of variance components'. Technometrics 10:551.
Patterson, H.D. and R. Thompson. 1971. Recovery of interblock
information when block sizes are unequal. Biometrics 58:545-555.
Quaas, R.L. 1976. Computing the diagonal elements and inverse of a
large numerator relationship matrix. Biometrics 32:949.
Quaas, R.L. and EUT. Pollak. 1980. Mixed model methodology for farm and
ranch beef cattle testing programs. J. Anim. Sci. 51:1277-.
79
Rao, C.R. 1971. Minimum variance quadratic unbiased estimation of
variance components. J. Mult. Anal. 1:445.
Searle, S.R. 1958. Sampling variances of estimates of components of
variance. Ann. Math. Stat. 29:167.
Swallow, W.H. and J.F. Monahan. 1984. Monte Carlo comparison of ANOVA,
MIVQUE, REML, and ML estimates of variance components.
Technomentrics 26:47.
Thompson, R. 1976. The estimation of maternal genetic variances.
Biometrics 32:903.
1.
APPENDIX A
Prove TtrfViVj)] = XjXj + 2X3X2-
Since V^ and V^ are sytrmetric,
[49]
n-1 n
[tr^n = C j (v kk ) ± (v Kk ) :J 3 + 2C z z (v kx ) i (v kl ) j 3,
which can be written as the sum of two row-column multiplications
as follows:
CtrfViVj)] = C(v u ) i (v 22 ) i ... (v^Ji]
(v n)j
(v 22 } j
4 nn' 3
+2 Cv 12 ) i {v 13 ) i ... (v ln ) i (v 23 ) i ... (v n _ 1/n ) ± ]
(v 12)j
< v 13>j
< v ln>j
< v 23>j
^n-l^j
2.
which after substituting in [14] and [16] is
CtrtVjVj)] = X^X X + 2X3X3, which proves [49].
Prove CVjljj] = H'Xj + Xj_.
[V^l n 3 can be expressed in sunmation notation as:
[50]
80
81
[Vjl^ =
I (v lk>j
I (V2k)j
* (v rik>j
= H'
< v 12>j
< v 13>j
(v 23>j
1 < v lk J 1
k#> 2k D
^J^^
(v ll>j
(v 22>j
(v ) •
{v ll>j
< v 22>j
( v nn>j
or
[51]
(v n-l,n>j
which after substituting in [143 and [16] is
[Vj^] = H'X 2 + X lf which proves [50].
Prove [l^Vjln] = 1^ + 21^.
The left side of equality [51] can be written
which after substituting [50] in [Vjjl n ] is
^m 1 = ^^2 + ^v
which after substituting the transponse of [9] for l^H' is
= 1^ + 21iX 2 ' which proves [51].
Prove [y^y] = X]Y ± + 2X^2 [59]
The left side of [59] is
[y'V iy ] = [ E (v^)^] + 2[ I _Z (v kl ) iykyi ],
which can be written as the sum of 2 row column multiplications as
follows :
82
[ y ' Vi y] = C(v 11 ) i (v 22 ) i ... (v^i] yj
■
+ 2C(v 12 ) i (v 13 ) i ... (v ln ) i (v 23 ) i ... (v n _ 1/n ) ± ]
V1V2
^3
yiy n
y n _xy n
5.
which after substituting in [51, [6], [14], and [16] is
[y'Vjy] = XJY-l + 2X 2 Y 2 , vMch proves [59].
Prove yy'ln = H'Y 2 + Y ± .
The left hand side of [60] is
[60]
tyy'i^ =
1 yivk
k
*i Yiyk
yf
S v 2Vk
k
—
+
yi
*
•
•
1 VnVk
k
•
^n
or
83
= H'
^1^2
ny3
rf
•
•
+
*
Yl^n
•
^2*3
*
, which after
6.
substituting in [53 and [6] is
= H*Y 2 + Yj_, which proves [60].
Prove l^yy'ln = Vi + 21^.
Substitute [60] into [61] for yy' 1^ to obtain
= ^'*2 + *n Y l'
which after substituting the transpose of [9] for l^H is
= l^i l + 21 'Y 2 , which proves [61].
[61]
APPENDIX B
The multiplier coefficients that specify the contribution of <Jg, a
ggm'
r gm'
and
o£ to the covariances among the 18 individuals generated
by a replicate of Eisen design I are defined in matrix form as V-^, V 2 ,
V3, and V4. These matrices are as follows:
V, =
V 2 = T
16 8
4
4
8
8
8
8
4
4
16
8
8
4
4
4
4
8
8
16 8
8
8
8
4
4
4
16
8
8
4
8
4
4
16
8
4
4
8
4
16
4
4
4
8
16
4
4
4
8
2
2
2
1
16
4
4
2
8
2
2
16
4
2
2
8
2
16
16
2
2
2
8
symmetric
16
6
16
4
4
16
4
4
6
16
2
2
4
4
16
2
2
4
4
5
16
4
4
2
2
3
3
16
4
4
2
2
3
3
5
16
4 4
1
1
1
1
1
1
1
1
4
1
1
1
1
1
1
1
1
4 4
4
4
5
3
3
3
4
4
4
3
5
3
3
4
4
3
3
5
3
4
3
3
3
5
4
5
1
1
1
1
4
1
5
1
1
4
1
1
5
1
4
4
1
1
1
5
syimetric
4
2
4
2
2
4
2
2
2
4
4
1
4
1
1
4
1
1
1
4
84
85
4 4
4
4 4
4
4
2
2
2
2
4
4
4
2
2
2
2
4
4
2
2
2
2
4
2
2
2
2
4
2
1
V 3 = 4
4
2
4
2
4
2
symmetric
4
2
4
2
2
4
2
2
2
4
4
1
4
1
1
4
1
1
1
4
and
v 4 =
1
1
1 1
1
1
1
1
1
1
1
1
1
1
me
trie
1
1
1
1
1
1
1
1
1
1
The multiplier coefficients that specify the contribution of Cg/ o ggm ,
a*, and a^ to the covariances among the 18 individuals generated fcy
a replicate of Eisen design II are defined in matrix form as V 1# V2#
V 3 , and V4. These matrices are as follows:
86
8
4 4
2
2
4
4
4
4
8
4
4
4
4
2
2
4
4
8 4
4
2
2
2
2
2
8
2
4
2
2
2
2
8
4
2
2
4
2
2
2
8
2
2
2
4
2
2
8
4
4
4
4
2
2
2
1
V l = 8
8
4
4
2
4
2
2
8
4
2
2
4
2
8
2
2
2
4
symmetric
8
3
8
2
2
2
2
1
1
1
1
2
2
2
2
8
3
8
2
2
8
2
2
3
8
1
1
1
1
8
1
1
1
1
3
8
4
4 4
3
3
1
1
1
1
4
4
4
1
1
3
3
1
1
4 4
5
3
1
1
1
1
4
3
5
1
1
1
1
4
4
1
1
5
3
1
1
4
1
1
3
5
1
1
4
4
4
4
5
3
3
3
1
v 2 = ;
4
4
4
3
5
3
3
4
4
3
3
5
3
4
3
3
3
5
syrrmetric
4
2
4
2
2
2
2
1
1
1
1
4
2
4
4
2
4
1
1
2
2
4
1
1
2
2
2
4
87
2
2 2
1
1
2
2
2
1
1
2 2
1
1
2
1
1
2
2
1
1
2
1
1
2
2
2
2
1
1
1
1
1
2
2
2
1
1
1
1
v 3 = 5
2
2
1
1
1
1
2
1
1
1
1
symmetric
2
1
2
2
1
2
2
1
2
1
1
2
1
1
1
2
and
v 4 =
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
me
tri
c
1
1
1
1
1
1
1
1
1
The multiplier coefficients that specify the contribution of Og, a gqm'
a^*, and a^ to the covariances among the 18 individuals generated by
a replicate of Eisen design III are defined in matrix form as V-^, V2,
V3, and V4. These matrices are as follows:
88
16 8 8
4
4
8
8
8
8
16
8
8
8
8
4
4
8
8
16 8
8
4
4
4
4
4
16
4
8
4
4
4
4
16
8
4
4
8
4
4
4
16
4
4
4
8
4
4
16
4
4
4
8
2
2
2
*-fe
16
4
4
2
8
2
2
16
4
2
2
8
2
16
2
2
2
8
symmetric
16
6
16
4
4
16
4
4
6
16
2
2
4
4
16
2
2
4
4
5
16
4
4
2
2
1
1
16
4
4
2
2
1
1
5
16
4 4 4
3
3
1
1
1
1
4
4
4
1
1
3
3
1
1
4 4
5
3
1
1
1
1
4
3
5
1
1
1
1
4
4
1
1
5
3
1
1
4
1
1
3
5
1
1
4
5
1
1
1
1
V 2 = 4
4
1
5
1
1
4
1
1
5
1
4
1
1
1
5
symmetric
4
2
4
2
2
4
2
2
2
4
1
1
4
1
1
1
4
1
1
1
1
4
1
1
1
1
1
4
89
V3 = i
4
4
2
2
4
4
4
2
2
4
4
2
2
4
2
2
4
4
2
2
4
2
2
4
2
4
2
4
2
4
2
metri
c
4
2
4
4
2
4
4
1
4
1
1
4
1
1
1
4
and
V 4 =
10 11
10
1
1
1 1
1
1
1
1
1
1
1
synmetric
1
1
1
1
1
1
1
1
1
The multiplier coef ficients that specify the contribution of o ~, a qqm /
a**, and a^ to the covariances among the eight individuals generated
by a replicate of Thompson design A are defined in matrix form as V-^ f
v 2' v 3* an< ^ v 4* Th ese matrices are as follows:
90
and
16 8 4 4
8
8
2
2
16 4 4
4
4
2
2
1
16 8
2
2
4
4
v l
~ 16
16
2
16
2
8
8
1
8
1
symmetric
16
1
16
1
8
16
8 8
2
2
2
2
8
2
2
2
2
1
8 8
6
6
V 2
= 8
8
8
8
10
1
10
1
symmetric
8
1
8
1
8
8
2 2
2
1
2 2
1
1
V 3
J.
2
1
1
«J
2
2
2
symmetric
2
2
2
2
v„ =
110
10
1 1
1
symmetric
1
1
1
1
1
1
The multiplier coefficients that specify the contribution of a~,
ggm'
2
a gm'
4
and o£ to the covariances among the eight individuals
generated by a replicate of Thompson design B are defined in matrix
form as V^, V^, Vg, and V^. These matrices are as follows:
91
and
16 8 4 4
8
8
2
2
16 4 4
4
4
2
2
1
16 8
2
2
4
4
v l
~ 16
16
2
16
2
8
8
1
8
1
symmetric
16
1
16
1
8
16
4 4
5
5
1
1
4
3
3
1
1
1
4 4
1
1
3
3
V 2
4
1
1
5
5
4
4
4
1
1
symmetric
4
1
4
1
4
4
4 4
2
2
4
2
2
1
4 4
2
2
V 3
A.
~ 4
4
4
4
2
1
2
1
symmetric
4
1
4
1
4
4
v 4 =
110
10
1 1
1
symmetric
1
1
1
1
1
1
APPENDIX C
Computational example follows for symmetric differences squared
(SDS) and SDS weighted by the inverse of the error variance-covariance
matrix among SDS when relationship categories are considered.
Suppose that the observations generated from one replicate of
design A of Thompson (1976) are
y = [25 13 37 11 7 4 9 21]' .
Assume that Var(y) is the same as assumed in Appendix B. Hence,
l^y = 127, and
y - 15.875 .
D contains the first 4 columns of table 1 of Grimes and Harvey (1980),
excluding the 9th row, plus a column of zeros for the fifth column.
Hence,
4
2
8 16 16 16
8 20 8
u =
(1/16)
8 4
4 12 8
2 4
4 4
12
Q is obtained from the 5th column of table 1 of Grimes and Harvey
(1980), excluding the 9th row. Hence,
q= [4 4422242 4]' .
The 4 in the first element of Q, means that there are four pairs of
individuals in a set of this design related as paternal half-sibs. N
92
93
is obtained from figure 1 and table 1 of Grimes and Harvey (1980).
Hence,
N' =
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
2
1
1
2
1
1
2
1
1
2
2
2
2
2
1
1
2
2
2
2
A 2 in the first row and column of N', means that there are two animals
related to P^ as paternal half-sibs. Therefore,
N'N =
16 8 8
16 8
8
4
4
4
6
4
4
4
6
synmetric
4
4
4
2
6
8
8
4
4
4
16
4
4
4
2
4
6
8
8
4
4
4
8
4
16
The vectors of squares (Y^) and crossproducts (Y2) among observations
(y) are
Y ± = [625 169 1369 121 49 16 81 441]\ and
Y 2 = [325 925 275 175 100 225 525 481 143 91 52 117 273 407
259 148 333 777 77 44 99 231 28 63 147 36 84 189] ' .
We can let
B = Diag(Q)
94
The incidence matrix that specifies the relationship category to which
the elements of Y2 belong is
Z =
1
1
1
,0
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
Z does not need to be stored, but is given here for completeness.
Therefore,
Z'Y 2 = [1824 528 949 330 275 1110 1140 143 330]' .
Applying equations [89] and [90],
X'W" 1 * =
5.23828 6.10156 6.46875 6.375
11.0781 10.1875 8.75
11.25 11
symmetric 12
5.1875
4.375
5.5
6
7
and
X'W^Y - [594.633 641.391 708.313 736.75 854.875]'
95
64.5625 57.6250
70
75
83
63.2500
70
70
70
84
88
88
symmetric
96
96
112
Therefore,
(X , W 1 X)~ 1 X , lT 1 Y= [-195.7 148.6-142.9 72.3 219.5]' ,
which are the WSDS estimates of the variance due to direct genie
effects, the covariance between direct and maternal genie effects, the
variance due to maternal genie effects, the variance due to permanent
maternal environmental effects, and the variance due to residual
effects. Applying equations [91] and [92],
X'X =
and
X'Y = [10131.8 9166.5 10588 11732 13678]' .
Therefore,
(X'X)~ 1 X'Y= [-232.3 538.6 -905.4 503.1 237.8]* ,
which are the SDS estimates of the variance due to direct genie
effects, the covariance between direct and maternal genie effects, the
variance due to maternal genie effects, the variance due to permanent
maternal environmental effects, and variance due to residual effects.
Note for this example that the number of relationship categories
was 9 for only 8 animals. Normally in practice the number of
relationship categories would be quite small (less than 50) and the
nutiber of animals could be in the thousands. Note that two estimates
of variance components are negative. This can happen with high
probability for an example this small.
Live Music Archive
Librivox Free Audio
Metropolitan Museum
Cleveland Museum of Art
Internet Arcade
Console Living Room
Books to Borrow
Open Library
TV News
Understanding 9/11