National Institute for Applied Statistics Research

Australia

University of Wollongong, Australia

Working Paper

17-11

Extensions in Linear Mixed Models

and Design of Experiments

David Butler, Brian Cullis, and Julian Taylor

Copyright © 2021 by the National Institute for Applied Statistics Research Australia, UOW.
Work in progress, no part of this paper may be reproduced without permission from the Institute.

National Institute for Applied Statistics Research Australia, University of Wollongong,

Wollongong NSW 2522, Australia Phone +61 2 4221 5435, Fax +61 2 4221 4998.
Email: [email protected]
 Statistics for the
Australian Grains Industry

Extensions in Linear Mixed

Models and Design
of Experiments
David Butler1 & Brian Cullis2,3 & Julian Taylor4
1
Principal Biometrician, Plant Science, Agri-Science QLD, DEEDI
2
Professor of Biometry, School of Mathematics & Applied Statistics, Faculty of
Informatics, University of Wollongong
3
CMIS, Environmental Informatics & Food Futures Flagship
4
Research Scientist, School of Food Agriculture & Wine, University of Adelaide

Workshop in conjunction with:

Biometrics by the Blowholes
Presenter Affiliations:
 Contents

Contents 1

B The design and analysis of variety trials using mixtures of composite

and individual plot samples. 3
1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Reducing replication for expensive traits . . . . . . . . . . . . . . . . . . . 10
3 Using mixtures of composite and individual plot samples with multi-phase
experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

Bibliography 30

D OD: An optimal design companion to ASReml 32

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2 Introduction to optimal design . . . . . . . . . . . . . . . . . . . . . . . . . 33
3 The linear mixed model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4 Optimal design criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5 Algorithms for optimal design . . . . . . . . . . . . . . . . . . . . . . . . . 43
6 A design generator: od() . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
7 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
9 The od() class . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Bibliography 65

J Whole Genome Analysis using Linear Mixed Models 69

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2 WGAIM theoretical method . . . . . . . . . . . . . . . . . . . . . . . . . . 72

1
 3 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4 The R package wgaim: A casual walk through . . . . . . . . . . . . . . . . 85
5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

Bibliography 107
Chapter

B
The design and analysis of
variety trials using mixtures of
composite and individual plot
samples.

1 Background
Field trials for evaluating plant varieties are conducted in order to obtain information on a
range of traits including grain yield and grain quality traits, for example as in the National
Variety Trials (NVT) system used for obtaining information on late stage evaluation of
cereal varieties in Australia. Some grain quality traits are costly to measure so that it is not
possible to test all plots individually. It has therefore become common practice to test a
single composite sample for each variety in the trial but this precludes the use of an efficient
statistical analysis. Smith et al. (2011) proposed an approach in which a proportion of
varieties be tested using individual replicate samples. The remaining varieties may either
be tested as composite samples or with reduced replication. This allows application
of efficient mixed model analyses for both the inexpensive traits (using data from the
complete set of individual replicate plots) and expensive traits (using data from a subset
of the plots or a mixture of composite and individual plot samples). In the first chapter
of these notes we present the key results of Smith et al. (2011) beginning with a brief
review of statistical analysis methods for the analysis of single and multi-environment
field trials. We then consider two extensions namely the notion of an embedded field
design and secondly an analysis of composite and individual plot samples. This is initially
done in a single phase experiment. We consider the extension of this idea to multi-phase
experiments in 3

1.1 Statistical analysis methods for field trials

The literature on methods for the analysis of field trials (which includes the specific setting
of variety trials) is quite expansive but the methods can be broadly classified as either
randomisation or model based. In the former the model for the random and residual effects
is determined purely from the block structure whereas in the latter it is either assumed or

3
selected with the objective of providing a good fit to the data. Model based approaches
for the analysis of variety trials aim to account for the effect of spatial heterogeneity on
the prediction of genotype contrasts. Typically the heterogeneity reflects the fact that, in
the absence of design effects (that is, treatment and block effects), data from plots that
are close together (that is, neighbouring plots) are more similar (positively correlated)
than those that are further apart. Numerous authors have proposed analytical methods
to remove the effects of such trend. We use the approach originally proposed by Gilmour
et al. (1997) and extended by Stefanova et al. (2009). These authors assume that field
trials are arranged as rectangular arrays indexed by rows and columns (extensions to
other arrangements are straight-forward). Gilmour et al. (1997) extended the approach
of Cullis & Gleeson (1991) by partitioning spatial variation into two types of smooth trend
(local and global) and extraneous variation. Local trend reflects, for example, small scale
soil depth and fertility fluctuations. Global trend reflects non-stationary trend across the
field. Extraneous variation is often linked to trial management, in particular, procedures
that are aligned with the field rows and columns (for example, the sowing and harvesting
of plots). Certain procedures may result in row and column effects (systematic and/or
random). In the Gilmour et al. (1997) approach global trend and extraneous variation
are accommodated in the mixed model by including appropriate fixed and/or random
effects. Local stationary trend is modelled using a covariance structure for the residuals.
A plausible model that has broad application for two-dimensional (row by column) field
trials is a separable autoregressive process of order 1 (here-after denoted AR1×AR1) as
originally proposed by Cullis & Gleeson (1991) and used by Gilmour et al. (1997) and
Stefanova et al. (2009).
For complete generality we consider a series of t trials (synonymous with environments)
in which a total of m varieties has been grown (without necessarily all varieties being
grown in all trials). First we examine the statistical model used for the analysis of a
single variety trial then build on this to describe the analysis for the complete series of
trials.

1.1.1 Single trial analysis

We consider the analysis of the j th trial (j = 1 . . . t) within the series and assume it
comprises nj plots laid out in a rectangular array of cj columns by rj rows (so that
nj = cj rj ). Let y j be the nj × 1 vector of data, ordered as rows within columns. The
statistical model for y j can be written as

y j = X j τ j + Zg j ug j + Zp j up j + ej (1)

where τ j is a vector of fixed effects with associated design matrix X j (assumed to have
full column rank); ug j is the m × 1 vector of random variety effects with associated design
matrix Zg j ; up j is a vector of random non-genetic (or peripheral) effects with associated
design matrix Zp j and ej is the vector of residuals for the trial. In the simplest case the
vector τ j comprises an overall mean (intercept) for the trial but may include effects to
Table 1: Trial layout information for canola data: numbers of rows, columns and entries
and proportion p of entries with two replicates. Trial mean oil content also presented.
Trial Rows Columns Entries Mean oil p
1 48 6 213 38.2 0.352
2 51 6 232 43.9 0.319
3 52 6 245 40.6 0.273
4 52 6 252 46.0 0.238
5 53 6 254 45.6 0.252
6 49 6 220 38.9 0.336
7 53 6 260 47.5 0.223

accommodate spatial trend (see below). The effects up j include effects for major blocking
factors associated with the trial design and possibly effects to accommodate
 spatial
 trend
(see below). For simplicity we assume independent variety effects with var ug j = σg2j I m
where σg2j is the genetic variance for trial j. Other genetic variance models are possible,
in particular a known relationship matrix may be incorporated (see Oakey et al., 2007).
Note that if fewer than m varieties were grown in the j th trial then Zg j will contain some
zero columns.
In terms of the residuals we follow the spatial modelling approach of Stefanova et al.
(2009) in which a separable autoregressive process of order one (denoted AR1×AR1) is
assumed so that var (ej ) = Rj = σj2 Σcj ⊗ Σrj where Σcj and Σrj are the cj × cj and
rj × rj correlation matrices for the column and row dimensions (so each is a function of
a single autocorrelation parameter, that is, ρcj and ρrj respectively).
The spatial analysis process is sequential, commencing with the fitting of a base-line
model followed by the application of diagnostics to assess model adequacy. In particular
the presence of outliers, non-stationary global trend and extraneous variation is investi-
gated (see Stefanova et al., 2009). If there is evidence of the latter it may be accommodated
in the model by including appropriate effects in either τ j or up j .

1.1.2 MET analysis

We now consider the full series of t trials and let y = (y 01 , y 02 , . . . y 0t )0 denote the vector
of individual plot data combined across trials. The MET model proposed by Smith et al.
(2001) for y can be written as

y = Xτ + Zg ug + Zp up + e (2)

where τ = (τ 01 , τ 02 , . . . τ 0t )0 is a vector of fixed effects with associated design matrix X =


0 0 0 0
diag (X j ); up = (up 1 , up 2 , . . . up t ) with associated design matrix Zp = diag Zp j
and variance matrix var (up ) = Gp ; e = (e01 , e02 , . . . e0t )0 with associated variance matrix
var (e) = R = diag (Rj ).
The vector ug = (ug 01 , ug 02 , . . . ug 0t )0 is the mt × 1 vector
 of variety effects for individual
trials with associated design matrix Zg = diag Zg j . We assume that var (ug ) =
 Table 2: Residual log-likelihoods for MET models fitted to canola data
Variance parameters Residual
Model for Ge genetic total log-likelihood
1. Diagonal 7 37 -1679.89
2. Uniform 2 32 -1179.61
3. FA1 14 46 -1115.38
4. FA2 20 52 -1103.58

Ge ⊗ I m where Ge is the t × t genetic variance matrix across trials (that is, with diagonal
elements given by the genetic variances for each trial and the off-diagonal elements the
genetic covariances between pairs of trials). Here we have assumed independence between
genotypes but as noted for single trial analysis the use of a relationship matrix is also
possible. Various forms for Ge may be used. We follow the approach of Smith et al.
(2001) who advocate the use of factor analytic (FA) models. The FA model for ug is
given by:
ug = (Λ ⊗ I m )f + δ (3)
where Λ is a t × k matrix of trial loadings (k being the number of factors included in the
model), f is an mk × 1 vector of variety scores and δ is the mt × 1 vector of residual
genetic effects. We assume var (f ) = I mk and var (δ) = Ψ ⊗ I m where Ψ = diag (ψj )
where ψj is known as the specific variance for the j th trial. It then follows that

Ge = ΛΛ0 + Ψ

1.2 Analysis of example

Here we consider oil content data (measured using NIR as percentage of whole grain) from
a series of 7 canola breeding trials (data kindly supplied by Canola Breeders Western Aus-
tralia Pty Ltd). The number of entries (synonymous with varieties) grown in individual
trials ranged from 213 to 260 (Table 1) with a total of 260 entries across all trials. Entries
were mainly breeding lines but some commercial varieties were also included. Each trial
was laid out as a rectangular array of plots indexed by rows and columns (Table 1). The
design for each trial was a partially replicated (p−rep) design (Cullis et al., 2006). In these
designs a proportion p of entries is replicated with the remainder having only single plots.
The value of p for these trials ranged from 0.223 to 0.352 (Table 1). Each design was
resolvable in the sense that all replicated entries occurred once in the first block (columns
1-3) and then again in the second block (columns 4-6).
In the Smith et al. (2001) approach for analysis the first step is to determine ap-
propriate spatial models for individual trials. This is most efficiently achieved from a
computational perspective by fitting a simple form for the genetic
 variance matrix in
equation (2). We often use a diagonal form, that is, Ge = diag σg2j since this is anal-
ogous to analysing the data for each trial separately. The spatial models so determined
can then be re-checked once the final genetic model has been fitted. For the canola data
Table 3: Estimates of genetic variance parameters from FA2 model fitted to canola data.
FA parameters Genetic covariance/correlation matrix
Trial λ1 λ2 Ψ 1 2 3 4 5 6 7
1 0.762 0.258 0.296 0.944 0.76 0.71 0.71 0.77 0.79 0.69
2 1.163 0∗ 0.076 0.886 1.429 0.89 0.89 0.89 0.85 0.93
3 0.986 −0.003 0.201 0.750 1.146 1.173 0.84 0.84 0.79 0.87
4 1.463 −0.040 0.395 1.104 1.701 1.442 2.536 0.84 0.79 0.88
5 1.259 0.278 0.214 1.031 1.464 1.240 1.830 1.877 0.88 0.83
6 0.868 0.410 0.072 0.767 1.009 0.854 1.253 1.207 0.993 0.75
7 1.284 −0.272 0.100 0.908 1.493 1.266 1.888 1.541 1.003 1.822

∗
Parameter not estimated but fixed at this value to ensure a unique solution

the base-line model includes random block effects for each trial (included in the vector
up ) with a separate block variance for each trial and a separate spatial covariance model
for the errors for each trial. Following the fit of this base-line model the use of diagnostics
as described in Stefanova et al. (2009) revealed the need to add linear regressions on row
number for the first 3 trials (added to the model as fixed effects in τ ) and random column
effects for trials 2 and 6 (added to the model as extra random effects in up ).
Having determined acceptable spatial models we investigate more appropriate models
for the variety effects across trials. The diagonal model for Ge allows for a separate
genetic variance for each trial but assumes the variety effects in different trials to be
uncorrelated. The simplest correlation model that is often applied (implicitly) to MET
data is the uniform model that assumes a common genetic correlation for all pairs of trials
and a common genetic variance. The residual log-likelihoods from fitting the uniform and
diagonal models are given in Table 2.
We then fitted factor analytic models for the genetic variance structure. The residual
log-likelihood from fitting an FA model with one factor (denoted FA1) to the canola data
is given in Table 2. Since the uniform model is nested within the FA1 model they can
be formally compared using a residual maximum likelihood ratio test (REMLRT). The
FA1 model provided a significantly better (p < 0.001) fit to the data. An FA2 model
was then fitted and a REMLRT revealed this to be significantly better (p < 0.001) than
the FA1 model. Note that in order to ensure uniqueness of the loading matrix in the
FA2 model it is necessary to impose a single constraint (Smith et al., 2001). We chose to
constrain the second element in the second loading to be equal to zero. This is purely for
computational reasons and has no biological basis. If a meaningful interpretation of the
loadings is required we usually rotate the solution. With only 7 trials it is not possible
to fit higher order FA models so we conclude that the FA2 model provides the best fit to
these data.
Estimates of the variance parameters and fixed effects from the FA2 model are given
in Tables 3 and 4. The estimated genetic variances range from 0.94 to 2.54 and the
genetic correlations are very strong, ranging from 0.69 to 0.93 (Table 3). Examination of
Table 4: Estimates of non-genetic fixed effects and variance parameters from FA2 model
fitted to canola data
Fixed effects Variances Autocorrelations
Trial Intercept lin(row) block column residual column row
1 38.2 0.012 0.104 0.326 0.13 0.39
2 43.9 −0.044 0.087 0.306 0.377 0.20 0.45
3 40.6 −0.018 0.082 0.594 0.14 0.59
4 45.9 0.000 1.282 0.35 0.78
5 45.7 0.271 2.217 0.26 0.61
6 38.9 0.000 0.123 0.478 0.27 0.55
7 47.6 0.000 0.707 0.21 0.56

Table 4 shows the existence of substantial non-genetic variation in most of the trials. As
already discussed there was significant linear trend across rows for trials 1-3 and significant
variation between columns for trials 2 and 6. Variation between blocks was large for trial
5. Stationary spatial trend was evident in most trials with the largest autocorrelation
parameters associated with trial 4.
The final step in the analysis was to obtain E-BLUPs of the variety effects for each
trial. Of particular relevance to this paper is the impact of using the Smith et al. (2001)
approach to analysis on these predictions. Figure 1 contains pairwise plots of the predicted
variety effects for trial 5 obtained using three methods, namely (a) variety averages of raw
data (expressed as deviations from the trial mean); (b) E-BLUPs, ũg5 , from model 1 in
Table 2 (diagonal form for Ge ); (c) E-BLUPs, ũg5 , from model 4 in Table 2 (FA2 form
for Ge ). There are large changes in moving from (a) to (b) that are the result of having
modelled spatial variation. Note that method (a) is analogous to the practice of measuring
composite samples for each entry. Then there are also substantial changes in moving from
(b) to (c) that are the result of incorporating information from other (highly correlated)
trials.
 −4 −2 0 2 4 6
6
4
2

raw means
0
−2
−4
2
1

E−BLUPs from
diagonal model
0
−1
−2
4

4
2

E−BLUPs from
0

FA2 model
−2

−2
−4

−4

−4 −2 0 2 4 6 −2 −1 0 1 2 −4 −2 0 2 4

Figure 1: Predicted variety effects for trial 5 obtained using three methods.
2 Reducing replication for expensive traits
There may be traits for which it is prohibitively expensive to take measurements on all
individual replicate plots in a trial. We assume that we are limited to testing sj samples.
Currently it is common practice (for example in NVT) to use sj = m samples where each
sample is a composite of all bj replicate plots for a variety. We propose schemes that
lie between the two extremes of using all composite samples (sj = m) and all individual
plot samples (sj = nj ). In order to describe the schemes we consider one of the examples
presented by Smith et al. (2011).
This example is similar to that of the real breeding program example of Section 1.2
in which we wish to test 90 canola entries for grain yield and oil content. Unlike the
real example, however, we now assume a fully replicated design (with two replicates) for
measuring grain yield making a total of 180 plots. We assume that these are laid out in a
similar manner to the real example with 6 columns and 30 rows and with the first block
comprising columns 1-3 and the second block columns 4-6. In order to test oil content we
assume we are limited to testing 120 samples.

2.1 Two replicate example

In order to reduce the number of samples from sj = 180 to sj = 120 we propose to use
replicate plots of 30 entries only. The remaining 60 entries may either be tested as single
plots or as composite samples formed from the two replicate plots of these entries. These
approaches are discussed in turn in the following sections.

2.1.1 Using subsets of plots

In this scenario the remaining 60 entries are tested as single plots so that a total of 120
samples is tested and each sample corresponds to a different plot. In terms of the subset
of plots to be tested one approach, here-after termed “subset selection”, would be to
generate an efficient design for grain yield then select a subset of plots for testing oil
content without regard to their spatial location. We will call the overall layout an OT1
design to reflect the fact that it has been optimised with respect to a single trait (usually
grain yield) only. It is usually desirable for field designs to comprise a contiguous set of
plots so that subset selection is not ideal for oil content. As an example we may proceed
as in Figure 2 in which all plots in the first block will be quality tested and then 30 plots in
the second block (corresponding to the pre-chosen set of replicated entries for this trial).
An alternative approach is to restrict the subset of plots for oil content to occupy a
rectangular sub-section of the overall layout (and thence a contiguous set of plots). In
this example in which the full design comprises two replicates we note that the subsetting
of plots creates a replication pattern synonymous with a p−rep design. We may therefore
a priori create an embedded design that is a valid p−rep design in its own right. In
our example we chose to locate the p−rep design within the sub-section corresponding
 1 1 49 38 7 54 12
2 54 67 45 1 25 56
3 85 23 56 51 38 57
4 57 16 7 81 30 67
5 76 25 70 21 23 22
6 22 62 10 45 84 8
7 68 15 43 89 10 83
8 73 84 21 40 68 65
9 47 42 11 70 82 73
10 79 40 33 47 66 2
11 18 66 44 74 59 42
12 14 6 32 36 79 44
13 12 36 71 3 11 6
14 34 28 59 14 37 55
15 55 5 3 17 48 62
row

16 88 8 75 43 5 18
17 89 29 48 60 88 86
18 58 19 86 61 72 29
19 53 35 20 75 19 24
20 69 31 26 78 20 77
21 2 77 60 35 85 26
22 78 13 50 15 33 49
23 87 4 81 76 13 53
24 80 46 72 87 16 28
25 41 37 90 50 39 4
26 9 65 17 46 69 90
27 51 52 39 9 64 80
28 83 64 82 63 34 71
29 30 63 24 32 27 52
30 74 61 27 31 41 58
1 2 3 4 5 6

column

Figure 2: OT1 design for testing grain yield and oil content. Full layout for grain yield
arranged as 30 rows by 6 columns with 2 replicates of 90 entries (block 1 = columns 1-3;
block 2 = columns 4-6). Entry numbers are given for each plot. The subset of plots for
testing oil content (shaded plots) comprises 2 replicates of 30 entries (numbers in bold
face) and single replicates of 60 entries (numbers in plain face).

to the first 20 rows and all 6 columns. This will be called the “embedded area” and the
remaining plots comprise the “non-embedded area”.
In this way we may create a design that is as efficient as possible for both grain yield
and oil content. This can be achieved in a number of ways, two of which include:

1. an efficient p−rep design is generated for the embedded area, then conditional on
this a design is generated for the non-embedded area, or

2. the designs for the embedded and non-embedded areas are generated simultaneously

where in each case the varieties to be replicated in the embedded area are predetermined.
The designs for the simulation study in Section 2.2 were generated under the second
approach using software currently under development for the R statistical environment
(R Development Core Team, 2011), and detailed below.
 It is widely accepted that data from individual field trials often exhibits spatial cor-
relation so may be analysed using models of the form in equation (1). A logical step has
been to accommodate spatial correlation in the design phase (see Martin & Eccleston,
1997; Eccleston & Chan, 1998, for example). We adopt this approach here.
Given a (prespecified) replication scheme, potential extraneous effects and a residual
correlation structure, an interchange algorithm governed the assignment of varieties to
plots in order to optimise a specific design criterion. For this study, the variety effects
were considered fixed and the criterion chosen was A-optimality, where the average pair-
wise variance of variety effects is minimised. Computationally, the A-value is calculated
using the mixed model equations corresponding to the pre-specified form of mixed model
relevant to the design problem. In our example the mixed model comprised fixed effects
for the overall mean and varieties and random effects for blocks, rows, columns and a
factor with two levels indicating whether the plot is in the embedded or non-embedded
area. A separable AR1×AR1 correlation structure was assumed for the residuals. The
autocorrelations were set at 0.3 and 0.6 for the column and row dimensions respectively,
the variance ratios for all random effects were set to 0.1 and without loss of generality
the residual variance was fixed at 1.0. These parameter values reflect our experience in
analysing grain yield data from thousands of trials across a number of crop types. Al-
though the corresponding values for grain quality traits are as yet largely unknown we
note that the designs are fairly robust to mis-specification of the parameters (Eccleston
& Chan, 1998) so that the values given here provide a reasonable starting point. The A-
value is computed from the coefficients matrix for varieties, adjusted for all other effects
by applying the matrix sweep operator (Gentle, 1998) to the mixed model equations.
The design is resolvable for replicated varieties, and the permutation of varieties to
plots was restricted to be within each of the embedded and non-embedded areas, while
respecting this resolvability. The design so constructed ensures that all 30 replicated
varieties in the first 20 rows occurred once in the first block (columns 1-3) and again in
the second block (columns 4-6). Figure 3 shows one realisation of this process. We will
call the overall layout an OT2 design to reflect the fact that it has been optimised with
respect to two traits. Note that the replicated varieties for testing oil content in Figure 2
are identical to those in Figure 3.
In the context of METs, it is important to consider design for the complete series
rather than individual trial designs in isolation. This is particularly so for p−rep designs
(including embedded p−rep designs) in the context of developing a (partial) replication
scheme. To illustrate, we extend our scenario once again to match the real example in
terms of a series of 7 trials. In this setting we wish to construct the embedded p−rep
designs such that entries are as equally replicated across trials as possible. For our scenario
this results in a series of p−rep designs in which entries have a total of either 9 (60
entries) or 10 (30 entries) replicates across all trials. We can achieve a (partially) balanced
allocation of replication levels by regarding the MET as an incomplete block design with
7 replicates (corresponding to trials) of 90 treatments, and two blocks per replicate of
 1 60 66 2 84 18 26
2 63 64 37 56 66 78
3 90 19 14 47 10 76
4 5 50 84 4 48 22
5 48 79 15 5 37 74
6 46 73 30 61 72 23
7 83 23 26 79 52 45
8 29 13 53 3 44 17
9 40 28 52 38 80 62
10 88 36 42 81 20 25
11 72 9 38 43 36 82
12 62 32 20 85 51 27
13 3 11 27 7 58 31
14 71 21 35 87 73 2
15 34 24 4 21 75 33
row

16 54 16 75 86 12 55
17 57 55 39 69 24 41
18 65 69 8 6 57 59
19 7 49 41 67 30 19
20 78 70 68 1 89 77
21 12 1 76 8 88 15
22 59 85 56 42 70 13
23 81 33 45 29 90 65
24 18 17 86 34 28 83
25 80 58 43 35 54 49
26 87 47 25 60 39 9
27 31 6 61 16 14 63
28 10 89 74 53 32 50
29 22 82 51 64 71 40
30 67 44 77 46 68 11
1 2 3 4 5 6

column

Figure 3: OT2 design for testing grain yield and oil content. Full layout for grain yield
arranged as 30 rows by 6 columns with 2 replicates of 90 entries (block 1 = columns 1-3;
block 2 = columns 4-6). Entry numbers are given for each plot. The subset of plots for
testing oil content (shaded plots) comprises 2 replicates of 30 entries (numbers in bold
face) and single replicates of 60 entries (numbers in plain face).

sizes 30 and 60, representing the replicated and unreplicated sets, respectively.
We note that there may be some loss in efficiency for grain yield in using an OT2
design, irrespective of the approach taken for its construction, compared with a design
that is optimal for this trait alone. This will be investigated in the simulation study of
Section 2.2.

2.1.2 Using mixtures of composite and individual plot samples

Recall that we have a scenario in which 30 of the 90 entries are being tested as individual
replicates. Here the remaining 60 entries are tested as composite samples of grain from
their (two) replicate plots. Thus we will be testing the same total number (120) of samples
as in Section 2.1.1 but these are now a mixture of composite and individual plot samples.
In terms of the approaches for testing oil content as described in Section 2.1.1 we would
use individual plot samples for two replicates of 30 entries (so for the layout in Figures
2 and 3 this corresponds to the 60 shaded plots with entry numbers in bold face) but
instead of using individual plot samples for a single replicate of the remaining 60 entries
(corresponding to the shaded plots with entry numbers in plain face) we composite the
grain from these plots with their remaining replicate plot (that is, the non-shaded plots).
The use of a mixture of composite and individual plot samples has important implica-
tions for the statistical analysis. First we consider the analysis of a single trial. Using the
notation of Section 1.1.1 we now assume for ease of explanation that the nj plots com-
prise a fully replicated trial with bj replicates of m varieties (so that cj rj = nj = mbj ).
Let ms denote the number of varieties for which we will use composite samples. It is
assumed that all replicates of each of these ms varieties will be combined to form a com-
posite sample so that the final number of grain samples to be quality tested is given by
sj = (m − ms )bj + ms . Now we consider a transformation of the data vector y j that is
commensurate with a compositing process, namely the averaging of individual replicate
data for a subset of the genotypes. This is a reasonable representation of the physical
process when the sample units (in our case grain samples from individual replicates) enter
into the composite in equal amounts and the physical act of pooling does not affect the
trait of interest (Lancaster & Keller-McNulty, 1998). In this way the experimental units
under consideration are ‘transformed’ from the original nj plots to sj samples. Of these
sj samples, ms are composites of bj plots and the remainder correspond to individual
(original) plots. We denote the sj × nj transformation matrix by D j and the transformed
data by z j = D j y j . The spatial model for the data is then given by

z j = D j X j τ j + D j Zg j ug j + D j Zp j up j + D j ej (4)

where the effects τ j , ug j , up j and ej , design matrices X j , Zg j and Zp j and variance

structures for ug j , up j and ej are as defined in Section 1.1.1. This model involves some
non-standard design matrices.
Finally we note that the MET analysis for data that involve a mixture of composite
and individual plot samples can be obtained as an extension of the single site analysis in
an analogous manner to Sections 1.1.1 and 1.1.2.

2.2 Accuracy of new approaches

The performance of the new approaches was assessed using a simulation study for the
case described in Section 2.1, namely the MET with 7 trials each laid out as 30 rows by 6
columns and subsequent testing of 120 samples from each trial. Two series of field layouts
were used corresponding to OT2 and OT1 designs. The same varieties were replicated in
the two sets of designs. Data were generated according to the final model fitted to the
real example of Section 1.2, that is, with genetic variance parameters as given in Table 3
and fixed effects and non-genetic variance parameters as given in Table 4.
Data were generated for the full trial layouts of 180 plots in each. For each generated
data-set and design type, eight methods were used to obtain predicted variety effects. In
the first seven methods a mixed model analysis was fitted and the E-BLUPs of variety
 1 2 3 4 5 6 7

(a) specific variance (b) first loading (c) second loading

0.20

0.10 0.12 0.14 0.16 0.18 0.20 0.22

S S S

S
M M

0.14
F
S
S M
0.15

M
M M
F F S
S F
S

0.12
S S
M M F
M
0.10

F S S
F S M F M

M S F F
M S F
S
F S M M
M

0.10
F S F
S F M
standard error

M F
0.05

F
M M
F F F

(d) residual variance 0.25 (e) column correlation (f) row correlation
S M M
0.5

0.20
0.4

S
M F S M
0.20

S
M M
M
M M

0.15
0.3

F S
S S S
M
S
S S M
0.15

M M
0.2

S S

0.10
M F M
S S
F S
S
M F M F
M F F
S F F S
F
0.1

F F
0.10

M F F F
M
S S F
F
0.05

F F F

1 2 3 4 5 6 7 1 2 3 4 5 6 7

Trial

Figure 4: Simulation (N=300) standard errors for variance parameter estimates from
models fitted to the full data-set (MF, labelled as “F”), the data corresponding to the
subsetted plots (MS, labelled as “S”) and the data comprising mixtures of composite and
individual plot samples (MM, labelled as “M”) for OT2 designs.

effects for each trial obtained. In the final method no analysis was conducted and the
“predicted” variety effects were taken to be the raw data expressed as deviations from
trial means. The methods were as follows:

MF: The true model was fitted to the full vector of generated data denoted by y =
(y 01 . . . y 07 )0 where y j (j = 1 . . . 7) is the 180 ×1 data vector for the j th trial

MFD: As for MF but a diagonal model for Ge was fitted

MS: The true model was fitted to the sub-vector of y corresponding to the subsetted
plots for each trial (so data from 120 plots only)

MSD: As for MS but a diagonal model for Ge was fitted

MM: The true model was fitted to the vector z = (z 01 . . . z 07 )0 where z j = D j y j corre-
sponds to the mixed sample type data (60 individual plot samples and 60 composite
samples) for the j th trial
Table 5: Correlations between true and predicted entry effects for diagonal models fitted
to the full data-set (MFD), the data corresponding to the subsetted plots (MSD) and the
data comprising mixtures of composite and individual plot samples (MMD). Correlations
given for both OT2 designs (optimised for both grain yield and oil content) and OT1
designs (optimised for grain yield only). Values are averages over all entries in each trial.
MFD MSD MMD
Trial OT2 OT1 OT2 OT1 OT2 OT1
1 0.927 0.926 0.874 0.871 0.912 0.910
2 0.946 0.945 0.900 0.894 0.926 0.927
3 0.924 0.923 0.867 0.857 0.899 0.897
4 0.957 0.957 0.911 0.904 0.927 0.930
5 0.866 0.866 0.780 0.767 0.823 0.825
6 0.924 0.926 0.864 0.859 0.897 0.899
7 0.938 0.937 0.888 0.881 0.918 0.917

MMD: As for MM but a diagonal model for Ge was fitted

MC: A model was fitted to the vector of data corresponding to a full compositing scheme
(that is, 90 composite samples for each trial). In the absence of replication within-
trial spatial variation can not be reliably modelled and the genetic variance model
must be simplified. The ‘best possible’ sub-model of the true model for this scenario
can be written as in equation (2) with the vector τ comprising the trial means, the
vector up omitted and the design matrix for the genetic effects given by an identity
matrix. We may still apply an FA model to the data but the residual genetic effects
(δ in equation (3)) are completely confounded with the plot errors e. Thus we
write ug = (Λ ⊗ I m )f so that Ge = ΛΛ0 ⊗ I 90 and assume var (e) = Ψ ⊗ I 90 =
diag (ψj I 90 ).

MR: As for MC but no model was fitted

The correlations between the true and predicted variety effects were calculated for
each method and design type. For the ith variety in the j th trial this was calculated (for
any given method and design type) as
PN
ũgijk − ũ¯gij.



k=1 ugijk − ūgij.
qP
2 PN
2 (5)
N ¯
k=1 ugijk − ūgij. k=1 ũgijk − ũgij.

where ugijk is the true (generated) effect for entry i in trial j and simulation k (= 1 . . . N )
and ũgijk is the associated predicted effect (E-BLUP from the analysis for the first seven
methods and corrected raw data for the last). The term ūgij. denotes the mean across N
simulations of the true effects for entry i in trial j and ũ¯gij. is the analogous mean for the
predicted effects. In our study N = 300 and corresponds to those simulations in which
the algorithm for fitting the mixed models converged for all methods.
First we consider estimation of the variance parameters for those methods in which
the full model was able to be fitted, namely MF, MS and MM. In terms of bias all of
Table 6: Correlations between true and predicted entry effects for diagonal models fitted to
the data corresponding to the subsetted plots (MSD) and the data comprising mixtures of
composite and individual plot samples (MMD) for OT2 designs only. Values are averages
over all entries, over entries tested as two samples (replicates) and entries tested as single
samples (replicates for MSD and composites for MMD).
MSD MMD
Trial all 2 samples 1 sample all 2 samples 1 sample
1 0.874 0.917 0.853 0.912 0.908 0.914
2 0.900 0.934 0.883 0.926 0.925 0.926
3 0.867 0.912 0.845 0.899 0.900 0.899
4 0.911 0.941 0.897 0.927 0.931 0.925
5 0.780 0.851 0.745 0.823 0.831 0.819
6 0.864 0.912 0.840 0.897 0.903 0.895
7 0.888 0.927 0.868 0.918 0.920 0.916

these methods were very similar and showed no significant bias for any of the parameters.
This was a particularly important finding for MM which involves spatial modelling of
data derived from a mixture of composite and individual plot samples. We believe that
this is the first occurrence in the literature of such an analysis. In terms of the precision
of variance parameter estimates there were some interesting differences between methods.
Figure 4 shows the simulation standard errors of variance parameter estimates for MF,
MS and MM for OT2 designs. Note that the block and column variances have been
omitted from this figure but showed little difference between methods. As expected the
precision was best for MF for all parameters. The standard errors for genetic parameters
(specific variances and loadings) were always lower for the method employing a mixture
of composite and individual samples (MM) compared with the method using a subset of
the plots (MS) whereas the converse was true for the spatial auto-correlation parameters.
The latter may be expected due to spatial trend being “smeared” when grain samples
from different plots are composited.
In terms of correlations between the true and predicted variety effects we first consider
the diagonal models that correspond to separate analyses of individual trials. Table 5
gives the average correlations between true and predicted variety effects for each trial (ie.
averaged over all the entries in a trial) for MFD, MSD and MMD for both design types.
In terms of the effect of design types we immediately see that for the subsetting approach
(MSD) there is a small but consistent increase in the correlations for OT2 compared
with OT1 designs. This may have been expected due to the fact that the embedded
designs for oil content were optimised in the OT2 scenario whereas in OT1 they were
not. Importantly the correlations for the data corresponding to the full two replicate
designs (MFD) were almost identical for the two design types. This implies that there
will be minimal loss in efficiency for grain yield in using an OT2 design. Also there was
no substantial or consistent difference between types when some plots were composited.
We will therefore focus on OT2 designs in what follows.
Table 5 also shows that the correlations for the approach involving composite samples
Table 7: Correlations between true and predicted entry effects for factor analytic models
fitted to the full data-set (MF for both OT2 and OT1 designs), the data corresponding to
the subsetted plots (MS for both OT2 and OT1 designs), the data comprising mixtures
of composite and individual plot samples (MM for both OT2 and OT1 designs) and
the data comprising composite samples alone (MC). Also given are correlations for data
comprising composite samples alone but with no model fitted (MR). Values are averages
over all entries in each trial.
MF MS MM MC MR
Trial OT2 OT1 OT2 OT1 OT2 OT1 OT2 OT2
1 0.939 0.938 0.907 0.904 0.929 0.928 0.836 0.917
2 0.967 0.967 0.955 0.952 0.962 0.961 0.944 0.908
3 0.948 0.948 0.925 0.923 0.937 0.937 0.896 0.890
4 0.967 0.966 0.948 0.944 0.954 0.955 0.906 0.897
5 0.934 0.933 0.914 0.911 0.923 0.923 0.900 0.795
6 0.947 0.947 0.921 0.919 0.935 0.936 0.890 0.878
7 0.959 0.958 0.941 0.938 0.950 0.950 0.926 0.910

(MMD) were always superior to the subsetting approach (MSD). In these diagonal models
the difference can largely be explained by the reduced replication in the latter. Table 6
shows the average correlations for each trial for the two classes of entry, namely the 30
tested as two separate samples (replicates) and the 60 tested as single samples (either
single replicates in the case of MSD or composite of two replicates in the case of MMD).
The correlations for the two replicate entries are lower for MMD compared with MSD. This
may be due to the fact that the spatial correlation parameters are less well estimated with
a mixture of composite and individual plot samples (see Figure 4). The correlations for
the single sample entries are substantially higher for MMD compared with MSD reflecting
the greater replication implicit in the composite samples. The overall result is that the
average correlations for each trial are higher for MMD.
We now consider the correlations between true and predicted variety effects from the
full MET analyses. Table 7 shows that the methods employing reduced replication, ei-
ther with or without composite samples (MM and MS respectively) performed favourably
when compared with the fully replicated method (MF). The MS and MM correlations
were consistently lower than for MF but offset against these losses is the fact that 33 31 %
fewer samples were measured for MS and MM compared with MF so that the associated
reduction in cost must be taken into account. The method based on mixtures of com-
posite and individual plot samples (MM) was substantially better than the method based
on reduced replication alone (MS) and there was no difference between the two types of
design for MM whereas for MS the OT2 designs were superior. The full composite ap-
proach, either with (MC) or without (MR) a mixed model analysis had substantially lower
correlations between true and predicted variety effects compared with all other methods.
We note that the genetic variances for oil content for the trials in this study were
high relative to error and this is reflected in the high correlations between true and
predicted variety effects. In our experience we have found that many quality traits exhibit
Table 8: Simulations based on reduced genetic variances and correlations. Correlations
between true and predicted entry effects for factor analytic models fitted to the full data-
set (MF for both OT2 and OT1 designs), the data corresponding to the subsetted plots
(MS for both OT2 and OT1 designs), the data comprising mixtures of composite and
individual plot samples (MM for both OT2 and OT1 designs) and the data comprising
composite samples alone (MC). Also given are correlations for data comprising composite
samples alone but with no model fitted (MR). Values are averages over all entries in each
trial.
MF MS MM MC MR
Trial OT2 OT1 OT2 OT1 OT2 OT1 OT2 OT2
1 0.815 0.813 0.731 0.721 0.789 0.784 0.605 0.768
2 0.905 0.904 0.865 0.860 0.889 0.888 0.797 0.735
3 0.850 0.848 0.793 0.782 0.822 0.817 0.720 0.708
4 0.905 0.905 0.854 0.847 0.874 0.875 0.743 0.714
5 0.834 0.834 0.783 0.774 0.809 0.805 0.721 0.547
6 0.868 0.871 0.822 0.815 0.847 0.847 0.743 0.682
7 0.904 0.904 0.865 0.860 0.886 0.886 0.838 0.752

lower genetic variance and weaker genetic correlations than observed here. We therefore
undertook a further simulation study in which we generated data based on the same
non-genetic effects (that is, as in Table 4) but altered the genetic parameters so that the
genetic variance for each trial was one quarter of the value used in the first simulation
study. This was achieved by halving the specific variances and reducing the loadings.
This not only reduced the genetic variances but also the genetic correlations, the average
pairwise correlation being 0.69 compared with 0.82 for the first simulation study. The
correlations between true and predicted variety effects from this simulation study for the
full MET analyses are given in Table 8. This table reveals the same patterns as in Table
7 but the differences between methods are greater.
3 Using mixtures of composite and individual plot
samples with multi-phase experiments

3.1 Background
Despite the importance of phenotyping quality traits in plant breeding programs and
associated genomic research, literature on experimental design and statistical analysis for
these traits is scarce. Most quality traits are obtained from multi-phase experiments in
which entries (to represent varieties, breeding lines or other types of genetic material) are
first grown in a field trial then further processed in the laboratory. Smith et al. (2006)
presented a general mixed model approach for the analysis of multi-phase data, with
particular emphasis on quality trait data that are often highly unbalanced and involve
substantial sources of non-genetic variation and correlation. They also detailed is a new
approach for experimental design that employs partial replication in all phases. The
motivation for this was the high cost of obtaining quality trait data and thence the need
to limit the total number of samples tested, but still allow use of the mixed model analysis.
They conducted a simulation study to show that the combined use of the new designs and
mixed model analysis has substantial benefits in terms of the genetic gain from selection.

3.2 Multi-phase experiments - statistical analysis

The approach we follow for the analysis of multi-phase plant breeding experiments builds
on that of Brien (1983) and Wood et al. (1988). In both of these papers the analysis
of multi-phase experiments is conducted by determining the experimental structure then
including appropriate terms in an ANOVA table. The experiments considered by these
authors are restrictive, however, in the sense that some degree of orthogonality is re-
quired. Brien (1983) discusses orthogonal designs that can be analysed using standard
ANOVA. Wood et al. (1988) consider a class of two-phase designs with non-orthogonal
block structure and provide an ANOVA approach to analysis but comment that a full
analysis involving recovery of information would require a more sophisticated procedure
based on REML estimation of variance parameters. Thus the linear mixed model pro-
vides a natural framework for the analysis of multi-phase experiments. We therefore use
a general linear mixed model that removes any restrictions concerning orthogonality of
block structures is proposed. In simple orthogonal cases the approach provides equivalent
analyses to those proposed in Brien (1983) and Wood et al. (1988).

3.3 Multi-phase experiments - a hypothetical example

Before considering this general mixed model a simple orthogonal two-phase milling ex-
periment that can be analysed using ANOVA is examined. It is assumed that r field
replicates of g entries are grown in a field trial that is designed as an RCB. Grain samples
from each of the rg field plots are split into d smaller samples to be used as replicates
in the laboratory process. Thus there is a total of n = rgd samples to be milled. It
is assumed that rg samples can be processed each day so that the full trial requires d
days. Field plots are randomised to times in the milling process using an RCB design
with days as blocks. A single sample from each of the rg field plots is processed each day
and the plots are allocated completely at random within a day. The data measured for
each sample is the flour yield.
In order to develop the analysis within an ANOVA framework the usual practice of
assuming that block effects are random and treatment effects are fixed is followed. Thus,
for illustrative purposes it is assumed here that entry effects are fixed. As previously noted
the determination of the block structure can be difficult in the context of multi-phase
experiments and the concepts in Brien (1983) may be helpful. The basic principle is to
include terms in the model that capture the randomisation processes used in each phase
of the experiment. In the two-phase quality experiment there are two randomisations,
namely the randomisation of entries to field plots then the randomisation of field plots
to ‘positions’ in the laboratory process. Thus the effects for block factors associated
with each of these randomisations must be included in the analysis. Accordingly and
based on the randomisation processes described above the symbolic model formula for
the hypothetical example can be written as

y ∼ 1 + entry + mrep + frep + frep.plot + mrep.order (6)

where ‘1’ represents an overall mean, entry is a factor with g levels, mrep is a factor (for
replicates in the milling process) with d levels, frep is a factor (for field replicates) with r
levels, plot is a factor (for plots within field replicates) with g levels and order is a factor
(indexing the order of processing of samples within days) with rg levels. Thus the final
term in (6) is the residual term that is also represented generically as units. Note the
convention in the model formula of underlining to indicate those terms that correspond
to fixed effects. The ANOVA table associated with the model in equation (6) is given in
Table 9. A key feature of the analysis is the existence of a residual term for each of the
two phases. The first phase (field plot) residuals are represented in the model by the term
frep.plot and the second phase (laboratory) residuals by mrep.order (or units).

3.4 General linear mixed model for multi-phase experiments

The hypothetical milling example is atypical of trials for measuring quality trait data in
that the data are rarely balanced nor are the designs orthogonal. Typically there is partial
rather than complete replication in terms of both the field and laboratory. This has arisen
from the work of Smith et al. (2006) who extended the idea of Cullis et al. (2006) who
introduced so-called p-rep designs, by considering so-called pq-rep designs for two phase
experiments, pqr-rep designs for three phase experiments and so on. Furthermore, as it
the case with our examples, often only a subset of the entries (and/or plots) from the field
trial is quality tested. Thus in general ANOVA cannot be used but a mixed model analysis
 Table 9: ANOVA table for hypothetical milling example.

Strata/Decomposition df Model term

mean 1 1
mrep d−1 mrep
mrep.order d(rg − 1)
frep r−1 frep
frep.plot r(g − 1)
entry g−1 entry
residual (r − 1)(g − 1) frep.plot
residual (d − 1)(rg − 1) units
total rgd

must be conducted. However, the same general principles are followed, in particular the
inclusion of residual terms for all phases of the experiment.
As with the analysis of field trials, trend in multi-phase trial data can be modelled
in order to improve the response to selection. In multi-phase quality trials the potential
exists to model trend (spatial or temporal) associated with the residuals for any of the
phases. The type of trend modelling depends on the trait and/or measurement process.
We focus on an approach which is consistent with the analysis of the examples, ie flour
yield and amylase in wheat. The modelling approach for field variation is analogous to
that of Gilmour et al. (1997) in that trend (either field or laboratory) is partitioned into
global trend, extraneous variation and local stationary trend. The latter is accommodated
using covariance models. It is important to note that, in the spirit of a randomisation
based analysis, terms in the mixed model that are associated with the block structure are
maintained irrespective of their level of significance. In contrast, model-based terms and
covariance structures are only included if found to be statistically significant.
For simplicity attention is restricted to two-phase experiments but the extension to
more phases is straight-forward. An experiment is considered in which the first phase field
trial consists of a rectangular array indexed by field rows (1 . . . r) and columns (1 . . . c)
making a total of np = rc plots. In the second phase laboratory trial it is assumed that
a total of n samples is tested. Note that, like the field trial, many laboratory trials can
be indexed using a two-dimensional co-ordinate system. For example in a wheat milling
trial the samples may be milled as a set number, s, of samples per day for d days (so
that n = sd). This has a two-dimensional spatial layout of mr rows and mc columns so
often n = mr mc . Thus a rectangular structure for the laboratory process is assumed here.
Extensions to non-rectangular or non-contiguous arrays for either the field or laboratory
layouts are straightforward. The mixed model for the n × 1 data vector y may be written
as in a hierarchical manner by

y|β f = X m τ m + Z f β f + Z pm upm + em
β f = X f τ f + Z g ug + Z pf upf + ef (7)
where τ m and τ f are vectors of fixed effects for the laboratory and field respectively, upm
and upf are vectors of peripheral effects for the milling and field respectively, ug is the
vector of entry effects and ef and em are the vectors of residual effects for the milling and
field respectively.
In the simplest case the fixed effects in (7) comprise a single effect, namely an overall
mean, but may include effects for missing values or covariates to model trend for either
the field and/or the milling phases. The vectors of peripheral effects include effects for
those terms associated with the experimental design in each phase and other effects as
required to model variation. The presence of two residual terms one for the first (field)
phase and one for the second (milling) phase is fundamental. Note that not all field plots
may be quality tested in which case the matrix Z f will contain columns whose elements
are all zero. In the next section we will incorporate this aspect in an explicit manner.
Also note the assumption of random (rather than fixed) entry effects which is consistent
with the applications we consider.
Variance models for the random effects in (7) are chosen from those described for the
analysis of field trials and available in ASReml-R and ASReml.
As an illustration consider the ANOVA for the hypothetical milling experiment of
Section 3.3 but now regard the entry effects as random rather than fixed. The vector ug
in equation (7) corresponds to the term entry in equation (6); ef corresponds to frep.plot;
upm contains mrep effects, upf contains frep effects; em corresponds to mrep.order and τ m
contains a single parameter only, namely an overall mean. In the ANOVA model each set
of random effects is assumed to be independent and each with the default variance matrix
(a scaled identity matrix). Thus the general mixed model of equation (7) encompasses
ANOVA models for multi-phase data (as discussed in Brien, 1983; Wood et al., 1988, for
example) but has much broader application since non-orthogonal designs and unbalanced
data are easily handled and more general covariance structures can be considered. The
latter is particularly important for the modelling of local stationary trend associated with
either the field or laboratory phase and extensions considered in the next section.

3.5 Multi-phase experiments with partial compositing

In this section we present an approach which combines the ideas found in Smith et al.
(2011) and discussed in Section 2.1.2 with the approach presented above for the design
of multi-phase experiments. For purpose of illustration and simplicity we consider the
analysis of a two phase experiment.
The model is given by

y|β f = X m τ m + Z f C f S f β f + Z pm upm + em
β f = X f τ f + Z g ug + Z pf upf + ef (8)

where all of the vectors are as for (7). The matrices C f and S f are compositing and
selection matrices which perform the compositing and selection of field plots. The elements
of S f are either 0 or 1 while the elements of C f are typically 0 or 0.5, but may be any
other real number depending on the compositing regime.
For further details on the model and analysis see Cavanagh et al. (2012).
4 Examples
Both examples are kindly provided by Colin Cavanagh, Food Futures Flagship, CSIRO
and involve phenotyping of lines (and commercial varieties) as part of the Multi-parent
Advanced Generation Inter-cross (MAGIC) project see From mutations to MAGIC: re-
sources for gene discovery & delivery in crop plants (2008) for details.

4.1 Narrabri Milling Experiment

The field trial was sown at Narrabri in northern NSW in 2010 and involved a total of 773
entries (760 experimental lines and 13 commercial varieties) sown in a resolvable p-rep
design. The entire trial comprised 1000 plots arranged in a rectangular arrany with 50
column × 20 rows. Table 10 presents the frequency distribution for lines, commercial
varieties and entries. Note that there was a sowing replacement with one experimental
line sown in four plots instead of two plots. (This was not part of the original design.)

Table 10: Summary of frequency distribution for experiment lines, commercial varieties
and entries in the Narrabri 2010 field trial.

Plots Lines Varieties Entries

1 0 554 554
2 8 205 213
3 4 0 4
4 1 1 2
Total 13 760 773

A total of 480 entries was considered for milling. Of these 350 were selected on the basis
of genetic diversity from the entire trial for those entries which did not get milled from a
companion trial conducted at Yanco in 2009. An additional 130 entries was selected which
were milled from the Yanco last year but maximize diversity of the current set. Table 11
presents a summary of the types of grain samples used in the final milling design. A total
of 509 grain samples was created. These samples were classified according to the following
status:

c1 a composite sample created from equal mixing of two field plots (of the same entry).
There were 89 of these.

c2 a composite sample created from equal mixing of two field plots (of the same entry)
but with sufficient grain to form two milling duplicates. There were 25 of these.

dup a sample of grain from a single field plot with sufficient grain to form two milling
duplicates. These grain samples were taken from field plots with either a single
replicate of an entry (10) or two replicates of an entry (2). For the latter case,
the other plot was not used for milling. This ensured a more even distribution of
replication across entries. There were 12 of these.
Table 11: Summary of type of grain samples used in the milling experiment for the
Narrabri 2010 trial.

Multipliers for
Sample No. Grain Field Entries Mill
Type samples plots samples
Composite (c1) 89 1 2 1 1
Comp-dup (c2) 25 1 2 1 2
Mill-dup (dup) 12 1 1 1 2
Field-rep (frep) 58 1 1 0.5 1
Field-single (sing) 325 1 1 1 1
Total 509 623 480 546

frep a sample of grain from each of the two field plots for each of 29 entries. There were
58 of these.

sing a sample of grain from a single field plot. There were 325 of these, representing
the remaining entries which were not replicated in the field or chosen as a “dup”
sample.

In summary, the milling design involved a total of 480 entries, 509 grain samples from
623 field plots and 546 milling samples. This strategy allowed for a more comprehensive
sampling of the field trial, appropriate levels of direct replication for both phases (namely
6.0 and 7.3% for field and milling respectively). Additionally we have indirect replication
of 28.4% of entries through composited samples. A total of 66 entries had either milling
or direct field replication, and 89 indirect field replication (ie c1 sample type).
The milling experiment design involved two resolvable blocks each of 39 days, with 7
samples per day.
Now comes the challenge to form an analysis!

4.1.1 Decomposition of sources of variation

Table 12 presents a skeletal decomposition of sources of variation for the Narrabri milling
experiment. This is merely a convenient way to present the terms which have been fitted in
the model to account for the randomisation processes, as well as accounting for potential
sources of variation. Note that the design search for the field, milling and plate phases
usually includes all of these terms.

4.2 Yanco amylase experiment

The field trial was sown at Yanco, southern NSW in 2009 and is a companion to the
Narrabri trial. A total of 1077 entries was sown in a p-rep design. The trial comprised
1620 plots arranged in a rectangular array with 20 columns by 81 rows. There were three
 Table 12: Decomposition of sources of variation for Narrabri milling experiment.

Decomposition Model term

mean 1
mrep MRep
mrep.mday MRep : MDay
mrep.mday.mord
frep FRep
fcol FCol
frow FRow
fcol.frow
entry Entry
residual FCol : FRow
residual units

adjoining (physical) bays such that bay 1 spanned all columns and rows 1 to 27, bay 2
spanned all columns and rows 28 to 54 and bay 3 spanned all columns and rows 55 to 81.

Table 13: Summary of frequency distribution for experiment lines, commercial varieties
amd entries in the Yanco 2009 field trial.

Plots Lines Varieties Entries

1 0 556 556
2 0 507 507
3 9 0 9
4 4 0 4
7 1 0 1
Total 14 1063 1077

Table 13 presents the frequency distribution for lines, commercial varieties and entries
in the field trial. Note that the 14 commercial varieties had additional plots, as is often
the case in these trials. The trait we consider is the amylase content of the flour. This is
measured using an ELISA plate in which samples of flour are placed in wells on a 96 well
plate (8 columns by 12 rows). This enzymatic reaction produces amylase which is then
quantified by measuring the absorbance by passing each plate through a scanner.
This experiment is a three phase experiment. Phase 1 is the field trial, phase 2 is the
milling trial and phase 3 is the ELISA plate assay. In considering an approach to the
design, consideration needs to be given to all potential sources of variation in all three
phases. Additionally we need to determine the experimental unit and the observational
unit for this experiment. The experimental unit is the field plot, while the observational
unit is the ELISA plate well. There may be sources of variation due to the scanner
(aligned with the rows and columns of the plate), as well as variation between ELISA
plates, variation due to milling days, milling order, milling order within milling days from
Table 14: Summary of frequency distribution of replication for each phase of the Yanco
amylase experiment

Sample Replication
Phase type 1 2 3 Samples Units
ELISA 488 132 0 620 752 (16)
Milling
Comp 131 0 0 131 131
Other 390 48 1 439 489
Total 521 48 1 570 620
Field
Total 698 570

phase 2, and finally variation between field bays, field columns, field rows and field plots
from phase 1.
A total of 498 entries was selected to be tested, using a similar approach outlined
above (based on maximising genetic diversity). Grain samples were classified as for the
Narrabri milling experiment, except there was no c2 types. (This idea has only recently
been thought of and tested.) A total of 570 grain samples were used for milling. These
comprised 131 c1 samples, 132 frep samples, 49 dup samples and 258 sing samples. This
information is presented in table 14. Of the 49 samples classified as dup, 48 had 2 milling
duplicates, while 1 was triplicated in the milling phase. This occurred during the experi-
ment, as the technician replaced another grain sample with this sample.
The resulting milling experimental design had a total of 620 milling units, comprising
62 days with 10 samples per day. The milling design was a resolvable q-rep design with 2
resolvable MillReps being MillDays 1 to 31 and 32 to 62 respectively.

Table 15: Summary of frequency distribution of milling replicate and ELISA plate of the
Yanco amylase experiment

Plate MillRep-1 MillRep-2

1 94 0
2 94 0
3 94 0
4 78 16
5 9 85
6 1 93
7 0 94
8 16 78

The ELISA phase of the experiment was a r-rep chain-block design with a total of 8
ELISA plates, comprising 96 wells arranged in 8 PlateColumns × 12 PlateRows. A total of
132 of the 620 milling units were duplicated in this phase. An additional 16 blanks were
 Table 16: Decomposition of sources of variation for Yanco amylase experiment.

Decomposition Model term

mean 1
plate Plate
pcol PColumn
prow Prow
plate.pcol Plate : PColumn
plate.prow Plate : PRow
plate.prow.prow
mrep MRep
mrep.mday MRep : MDay
mrep.mday.mord
fbay FBay
fcol FCol
frow FRow
fcol.frow
entry Entry
residual FRol : FRow
residual MRep : MDay : MOrd
residual units

included (two per plate) for quality control purposes. These data were excluded from the
analysis.
Use of a chain-block design was necessary due to the need to commence processing
the flour samples before the full milling experiment was completed. Fortunately, some
flour samples could be frozen which allowed linkage of the milling days with the ELISA
plates. The frequency distribution of MillRep × Plate is presented in table 15, illustrating
the nature of the chain-block design with the linkage between MillDays 1 to 5 (aka MillRep)
and Plates.
The resulting frequency distribution of entries was in the ELISA experiment was
(254, 235, 8, 1) for 1, 2,3 and 4 wells respectively.

4.2.1 Decomposition of sources of variation

Table 16 presents a skeletal decomposition of sources of variation for the Narrabri milling
experiment. This is merely a convenient way to present the terms which have been fitted in
the model to account for the randomisation processes, as well as accounting for potential
sources of variation. Note that the design search for the field, milling and plate phases
usually includes all of these terms.
 Bibliography

Brien, C. J. (1983). Analysis of variance tables based on experimental structure. Bio-

metrics 39 53–59.

Cavanagh, C., Smith, A., Butler, D., Thompson, R. & Cullis, B. (2012). On
the analysis of multi-phase experiments with partial compositing. submitted xx xx.

Cullis, B., Smith, A. & Coombes, N. (2006). On the design of early generation va-
riety trials with correlated data. Journal of Agricultural, Biological and Environmental
Statistics 11 381–393.

Cullis, B. R. & Gleeson, A. C. (1991). Spatial analysis of field experiments - an

extension to two dimensions. Biometrics 47 1449–1460.

Eccleston, J. & Chan, B. (1998). Design algorithms for correlated data. In COMP-
STAT98 Proceedings in Computational Statistics. Heidelberg: Physica-Verlag, 41–52.

From mutations to MAGIC: resources for gene discovery, v. & delivery

in crop plants (2008). Cavanagh, c. and morell, m.k. and mackay, i. and powell, w.
Current Opinion in Plant Biology 11 215–221.

Gentle, J. E. (1998). Numerical linear algebra for applications in statistics. Statistics

and Computing. New York: Springer-Verlag.

Gilmour, A. R., Cullis, B. R. & Verbyla, A. P. (1997). Accounting for natural

and extraneous variation in the analysis of field experiments. Journal of Agricultural,
Biological and Environmental Statistics 2 269–293.

Lancaster, V. A. & Keller-McNulty, S. (1998). A review of composite sampling

methods. Journal of the American Statistical Association 93 1216–1230.

30
Martin, R. J. & Eccleston, J. (1997). Construction of optimal and near-optimal
designs for dependent observations using simulated annealing. Research report 479/97,
Dept. of Probability and Statistics, University of Sheffield.

Oakey, H., Verbyla, A., Cullis, B., Wei, X. & Pitchford, W. (2007). Joint
modelling of additive and non-additive (genetic line) effects in multi-environment trials.
Theoretical and Applied Genetics 114 1319–1332.

R Development Core Team (2011). R: A Language and Environment for Statistical

Computing. R Foundation for Statistical Computing, Vienna, Austria. URL http:
//www.R-project.org/.

Smith, A., Cullis, B. R. & Thompson, R. (2001). Analyzing variety by environ-

ment data using multiplicative mixed models and adjustments for spatial field trend.
Biometrics 57 1138–1147.

Smith, A., Lim, P. & Cullis, B. (2006). On the design of multi-phase experiments for
quality trait data. Journal of Agricultural Science 144 393–409.

Smith, A., Thompson, R., Butler, D. & Cullis, B. (2011). The analysis of variety
trials using mixtures of composite and individual plot samples. Journal of the Royal
Statistical Society, Series C 60 437–455.

Stefanova, K., Smith, A. & Cullis, B. (2009). Enhanced diagnostics for the spatial
analysis of field trials. Journal of Agricultural, Biological and Environmental Statistics
14 1–19.

Wood, J. T., Williams, E. R. & Speed, T. P. (1988). Non-orthogonal block structure

in two-phase designs. Australian Journal of Statistics 30A 225–237.
Chapter

D
OD: An optimal design
companion to ASReml

1 Introduction
Experimental designs for plant improvement studies are often selected on an informal
basis, with due consideration to desirable properties such as replication, blocking, orthog-
onality or connectedness. In practice, adjustments are sometimes made to experimental
procedures if an off the shelf design does not quite fit the intended application. Optimal
design allows flexible combinations of the design specification parameters for special cases,
and prespecified residual or treatment correlation structures.
Table 1 defines some notation used here for the design parameters, and the physical
configuration of experiments typically encountered in field, glasshouse or laboratory eval-
uations. Field, or even glasshouse, experiments are typically arranged in a regular row
× column grid of experimental units, with resolvable replicates aligned with either rows
or columns (Figure 1). Field experiments are managed mechanically, with sowing and
harvesting operations along rows, and operations such as plot trimming within columns.

Table 1: Symbols used to describe the configuration of genetic evaluation experiments.

Symbol Definition
v Number of treatments (genotypes) in the experiment
k Number of experimental units per (incomplete) block
s Number of blocks per replicate
r Number of resolvable replicates
ri Number of replicates of treatments i = 1, . . . , v
qr Number of rows of experimental units in the physical layout
qc Number of columns of experimental units in the physical layout
N Total number, qr × qc , of experimental units

Sound experimental design promotes the statistically efficient separation of genetic

and environmental effects in the analysis of phenotypic observations. Commonly, designs
used in plant breeding programs fall within the general class of block designs, and include

32
 CHAPTER 1. OD: AN OPTIMAL DESIGN COMPANION TO ASREML 2

FigureFigure 1.1: Example

1: Example layout of
layout ofaareplicated field trial
replicated fieldwith qr =with
trial 24, qc q=r6,=v =24,
72, qr c==
2, s6,= 9v and k = 8.r = 2,
= 72,
s = 9 Resolvable
and k = replicates are aligned
8. Resolvable with columns.
replicates are aligned with columns.

Column
1 qc
1

k =8
Row

Replicate 1 Replicate 2

qr

1999, Eccleston and Whitaker, 1999, Martin and Eccleston, 1997]. Unreplicated field trials, or those with
restricted
complete replication, follow
or incomplete the same
block, grid patternand
row-column as Figure 1.1, and some
α-designs of the parameters
(Williams & John, k, s,1996).
and r Ex-
mayinclude
ceptions still retainlarge
their meaning in special
scale early cases such asevaluation,
generation augmented or partially
where replicated
methodsdesigns.
based on Yates
(1936a), Federer & Raghavarao (1975a) and Lin & Poushinsky (1983) are in use; more
recently
1.2partially
Introductionreplicated designs
to optimal based on Cullis et al. (2006a) are chosen in this situa-
design
tion. Design methods that extend to spatially correlated data have been of recent interest
(ButlerAn et al., design
optimal 2008;can Cullis et defined
be simply al., 2006a; Williamsdesign
as an experimental et al.,
that2006;
is optimalChan, 1999;to Martin
with respect some &
Eccleston, 1997;
statistical Martin,
criterion, O, say,1986),
based on largely motivated
some measure of the by advances
treatment in fitting
information. theislinear
The goal to max-mixed
modelimise the (treatment)
in this context information
(Cullis etavailable to theGilmour
al., 1998; practitioneretforal.,
the 1997;
statistical model &
Cullis under investiga-1991).
Gleeson,
tion, for a givenresults
Few theoretical number exist,
of experimental units, and sampling
and numerical regime. Optimal
optimization has design
been provides
almosta flexible
exclusively
framework for efficient design in novel experimental situations, such as where no design exists in the
used in the construction of A-optimal designs (Coombes, 2002; Chan, 1999; Eccleston &
literature for the particular combination of treatments, replication and blocking structure, or the un-
Whitaker, 1999; Martin & Eccleston, 1997). Unreplicated field trials, or those with re-
derlying statistical model precludes a standard design. The theory of optimal experimental design is
stricted replication,
credited followandthe
to Kiefer [1959] same grid
is followed patternoverviews
by numerous as Figure 1, and
including the some of Silvey
works by the parameters
[1980],
Atkinson
k, s, and r may and still
Donevretain
[1992], and
their Atkinson
meaninget al. [2007].
in special cases such as augmented or partially
replicatedConsider
designs. the simple linear model
y = Xβ+e

2 where y is a vector of observations,

Introduction X is a N × t design
to optimal matrix (assumed full rank), and e is a N vector
design 2
of residuals assumed normally distributed with mean zero and variance σ . For simplicity, initially let
the t columns
An optimal design of Xcan be t be
continuous
simplyexplanatory
defined variates, and let D denote the
as an experimental designthat
design regionisfrom which with
optimal
the rows of X , x i· , can be drawn. In the discussion that follows, x i· may, depending on context, refer to a
respect to some statistical criterion, O, say, based on some measure of the treatment infor-
row of the design matrix X , or the vector of predictor values at design point i , where there are n distinct
mation. The
design goal is to maximise the (treatment) information available to the practitioner
points.
for the statistical model under investigation, for a given number of experimental units,
The information available to the experimenter is quantified in terms of the variances of the estima-
tors β̂. The variance-covariance matrices of estimable functions of the β̂ are formed from the inverse
information matrix C −1 where

C = σ−2 (X T X ), and
var(β̂) = C −1 = σ2 (X T X )−1
and sampling regime. Optimal design provides a flexible framework for efficient design in
novel experimental situations, such as where no design exists in the literature for the par-
ticular combination of treatments, replication and blocking structure, or the underlying
statistical model precludes a standard design. The theory of optimal experimental design
is credited to Kiefer (1959) and is followed by numerous overviews including the works by
Silvey (1980), Atkinson & Donev (1992), and Atkinson et al. (2007).
Consider the simple linear model

y = Xβ + e

where y is a vector of observations, X is a N × t design matrix (assumed full rank), and

e is a N vector of residuals assumed normally distributed with mean zero and variance
σ 2 . For simplicity, initially let the t columns of X be t continuous explanatory variates,
and let D denote the design region from which the rows of X, xi· , can be drawn. In the
discussion that follows, xi· may, depending on context, refer to a row of the design matrix
X, or the vector of predictor values at design point i, where there are n distinct design
points.
The information available to the experimenter is quantified in terms of the variances
of the estimators β̂. The variance-covariance matrices of estimable functions of the β̂ are
formed from the inverse information matrix C −1 where

C = σ −2 (X T X), and
var(β̂) = C −1 = σ 2 (X T X)−1

Also, the variance of the predicted values

ŷ = X β̂

is
var(ŷ) = σ 2 X(X T X)−1 X T
In general, optimal design seeks to choose an X that minimizes some function of var(β̂),
that is, O = F(C −1 ) in this illustration. Forms for F are usually from the so-called
alphabet series of criteria, some of which are encapsulated by Kiefer (1974) in the Φk
family.
At this point, it is useful to introduce two optimality criteria that illustrate at least
one consequence of the general equivalence theorem outlined below. Of these, a much
studied criterion is D-optimality, where |C −1 | is minimized, or equivalently |C| is max-
imized. D-optimality minimizes the generalized variance of β̂. A criterion concerned
with the variance of the predicted response is G-optimality, which seeks to minimize the
maximum variance of the predicted response over D. That is, G-optimality is defined as
min(max(var(ŷ))).
2.1 Continuous and exact designs
A design is specified by an initially arbitrary measure ξ assigning unit mass over the
region D (Cox & Reid, 2000).
( )
x1 x2 . . . xn
ξ= (1)
w1 w2 . . . wn

where the {xi } are the n distinct design points and {wi } are the associated weights. As
R P
ξ is a measure, D ξ(dx) = 1, and 0 ≤ wi ≤ 1, and i wi = 1. Denoting the information
matrix for β in models such as those in equation (2) by M (ξ), then

Z n
X
M (ξ) = xi· xTi· ξ(dx) = wi M (ξ¯i ) (2)
D i=1

where
N
X
T
X X= xi· xTi· (3)
i=1

and the measure ξ¯i assigns unit mass at design point i.

Mathematically, the problem of finding an optimal design is approached in two ways
(Goos, 2002):

Continuous where the number of observations at a design point may be non-integral.

Design construction can be viewed as purely a numerical optimization problem (Wu,
1978), starting from the general equivalence theorem (below).

Discrete where the design respects the practicality of the integral constraint. A measure
that refers to a design realizable in integers is written as ξN and the weights {wi }
in expression (1) are such that wi N = ri , where ri is the number of replicates at
P
design point i, and ri = N . The summation form of (2)
n
X
M (ξN ) = wi xi· xTi·
i=1

for exact designs becomes the normalised version of (3), that is, M (ξN ) = X T X/N .

Let Ψ(M ) represent a general optimality criterion to be optimized. In the theory of

optimal continuous design, the General Equivalence Theorem (Kiefer & Wolfowitz, 1960)
states

1. the optimal design ξ ∗ minimizes Ψ

2. the minimum of the derivative of Ψ (Ψ0 ) evaluated in ξ ∗ is ≥ 0

3. Ψ0 evaluated in ξ ∗ achieves its minimum at the design points

 Table 2: Optimality criteria
Criteria Definition Computation Interpretation
Pp
tr M −1
 
A i=1 (1/λi ) Minimize the variance of elemen-
Qp tary treatment contrasts.
D i=1 λi |M | Minimize the variance of parame-
ter estimates.
E maxi=1,...,p (1/λi ) Minimize the variance of the least
well determined contrast (of a
set).
G λ Minimize the maximum variance
of predicted values.
V λ Minimize the average variance of
predicted values.

The theorem assumes that Ψ is convex and differentiable, and as a result any minimum
found will be global. D-optimality, where Ψ(M (ξ)) = − log(|M (ξ)|), and Ψ0 = p−var(ŷ),
satisfies these conditions. A direct consequence is that D-optimal designs are also G-
optimal, in that the maximum of the prediction variance over D is minimized (Goos,
2002).
D-optimality and G-optimality are only two of many that have been proposed, others
found useful in practice include:

A-optimality, which minimizes the average of the var(β̂)

E-optimality, which minimizes the variance of the least well estimated contrast aT β
given aT a = 1, and

V-optimality, which minimizes the average prediction variance over D.

For each of these criteria there is an analogue to the General Equivalence Theorem,
compactly described in terms of the general criterion max Φ(M (ξ)), where Φ is a concave
function on the set of information matrices M (ξ). For D-optimality Φ is the concave
function log |M (ξ)|, for example (Atkinson, 1982).
Kiefer (1974) expressed these optimality measures in terms of the sum of the k th powers
of the (non-zero) eigenvalues {λi } of M . Assuming M is of full rank, the general Φk value
is ( λ−k /p)1/k for 0 ≤ k ≤ ∞. Special cases of interest are k = 0, 1, ∞, corresponding
P

to D-, A-, and E-optimality, respectively. Table 2 summarises the common criteria found
in the literature and introduced in this section.
In practice all designs are discrete (or exact), however, the General Equivalence The-
orem is not valid in general for discrete designs. Efficient, though not necessarily optimal
exact designs may sometimes be found by rounding N × wi , from a corresponding optimal
continuous design, to the nearest integer. The optimality of continuous designs can be
proven, whereas a computer search is generally necessary to establish the optimality of a
discrete design. It should be bourne in mind that when comparing exact designs, it is the
relative values of the optimality criterion Ψ that are of importance, rather than how well
ξN approximates ξ ∗ (Atkinson et al., 2007).

2.2 Categorical designs

The simple example introduced in (2) assumed that the design matrix X was of full rank
and the columns of X were continuous variates, such as successive powers in a polynomial
regression, for example. This is an example of a response surface type design, where in
general the objective is to select or exchange the rows of X from the design space D
to optimize Ψ. The emphasis in pioneering work from agriculture such as Fisher (1935)
was on evaluating the effect of qualitative factors on biological response variables, as well
as their inclusion as a means of accounting for extraneous environmental variation. The
theory of optimal design is often presented in terms of response surface designs, however,
the principles and results hold in both settings, though the computational processes may
differ.

2.3 Scope
We only consider discrete, categorical designs where typically, though not exclusively, the
columns of X are dummy variables formed from categorical factors. In this context, a
design can be considered as a permutation P of the rows of X, with D then the set of all
such permutations.
In plant genetic evaluation programs, all treatment comparisons are generally of inter-
est. The A-optimality criterion is equivalent to minimising the average parwise variance
of elementary contrasts and is usually considered the most appropriate in these circum-
stances. We use the A-optimality criterion, or a random effects equivalent, exclusively
here. Exceptions, though not considered here, arise when the goal could be to maximise
genetic gain, for example, in which case alternative criteria may be more appropriate.

3 The linear mixed model

This section revisits the general linear mixed model introduced in Chapter 1, and defines
some elements relevent in the context of design, and the examples that follow. As be-
fore, we begin where a vector of observations, y, is modelled by a linear combination of
explanatory variables. For example, grain yield from plots in a genotype evaluation trial
may be modelled by a linear predictor comprising fixed and random terms representing
management effects, random genetic effects, blocking effects (such as field row or column),
for example, and a possibly correlated residual error structure. If grain samples from these
plots were subsequently milled in the laboratory for flour yield, then the linear predictor
would be extended to include milling effects such as day, trend within day, machine or
operator effects in addition to the environmental effects carried from the field.
 Historically, elements of the random error, and indeed those in all random terms,
were usually considered uncorrelated. This assumption is relaxed here using optimal
design methods, and a range of variance models may be considered for all random terms.
Typically, first order autoregressive models are often assumed for the error structure
arising in a field experiment, though many other models are possible. Gilmour et al.
(1997) discuss a recommended approach to the analysis such data.
In genetic improvement experiments, the genotype effects are often considered as re-
alisations of a random variable, sometimes with a simple variance model that assumes
these effects are independent and identically distributed with a common variance, σg2 , say.
Plant breeders keep crossing records, so the pedigree of an individual and its relatives is
known, and an additive genetic relationship matrix (A), and its inverse can be derived
(see Henderson, 1976). More recently, DNA markers can be used to derive a genetic re-
lationship matrix among individuals that is identical in state; either way, both relax the
iid assumption and allow a more plausible correlation structure among individuals.
With y as the N × 1 vector of observations, the linear mixed model can be written as

y = Xτ + Zu + e (4)

where τ is a t × 1 vector of fixed effects, X is an N × t design matrix which associates

observations with the appropriate combination of fixed effects, u is the b × 1 vector of
random effects, Z is the N × b design matrix which associates observations with the
appropriate combination of random effects, and e is the N × 1 vector of residual errors.
Note that X may not necessarily be of full column rank.
We assume " # " # " #!
u 0 G(γ) 0
∼N , σ2 (5)
e 0 0 R(φ)
where the matrices G and R are functions of parameters γ and φ, respectively. The
parameter σ is an overall scale parameter, and without loss of generality in the design
context we can set σ = 1. We subsequently use κ to denote the vector of variance
parameters in place of γ, which is traditionally reserved for variance ratios.

3.1 Variance structures for the errors: R

The vector e will in some situations be a series of vectors indexed by a factor or factors.
For example, in the classical analysis of a single field experiment R = σ 2 In , however, for
a series of s such experiments (Smith et al., 1998, for example) we may consider R to
have the form ⊕sj=1 σj2 Inj . That is, we can write e = [e01 , e02 , . . . , e0s ]0 and the ej represent
the errors of sections of the data. More generally, Cullis et al. (1998) consider the spatial
analysis of multi-environment trials in which

Rj = Rj (φj )
= σj2 (Σj (ρj ) + ψj I nj ) (6)
and each section represents an independent experiment. This model accounts for between
trial error variance heterogeneity (σj2 ), possibly a different spatial variance model for each
trial (Σj and ρj ) and different measurement errors (ψj ). As in Chapter 1, we assume a
separable first order autoregressive process across the column and row dimensions, so for
experiment j
Σj (ρj ) = Σcj (ρcj ) ⊗ Σrj (ρrj )
where ρcj and ρrj are the scalar autocorrelation parameters. We only consider single
experiments in this Chapter and so can drop the subscript j.

3.2 Variance structures for the random effects: G

The b×1 vector of random effects may be composed of w subvectors u = [uT1 uT2 . . . uTw ]T
where the subvectors ui are of length bi and these subvectors are usually assumed inde-
pendent normally distributed with variance matrices Gi . We can write G as
 
G1 0 ... 0 0
0 G2 ... 0 0
 
 
w

G = ⊕i=1 Gi =  .. .. .. .. ..
 . . . . .

0 0 . . . Gw−1 0 
 

0 0 ... 0 Gw

There is a corresponding partition in Z: Z = [Z 1 Z 2 . . . Z w ]. Each submatrix, Gi ,

is assumed to be the kronecker product of one, two or more component matrices. These
matrices are indexed for each of the factors constituting the term in the linear model. For
example, in a multi-trial analysis the term site:genotype has two factors and so the matrix
Gi is comprised of two component matrices defining the variance structure for each factor
in the term.
Assuming separability, models for the component matrices Gi are defined by

Gi = Gi1 ⊗ Gi2 ⊗ . . . ⊗ Gif

for f component factors. The vector ui is therefore assumed to be the vector represen-
tation of a f -way array. For two factors (f = 2) the vector ui is simply the vec() of a
matrix with rows and columns indexed by the component factors in the term, where vec()
of a matrix is an operator which stacks the columns of its argument into a vector.

3.3 Variance models

Variance models for R and G can be either on the correlation or covariance scale. The
simplest correlation model is the identity model: for example, if the v genotype effects in
a plant variety experiment were considered normally distributed as N (0, σg2 ), then that
component matrix of G would be a scaled identity σg2 I v . We consider other variance
models for random genetic effects below.
3.4 Estimation
The mixed model equations (Robinson, 1991) which are given by
" #" # " #
T −1 T −1 T −1
X R X X R Z τ̂ X R y
−1 −1 −1 = . (7)
T
Z R X Z R Z +GT
ũ Z T R−1 y

These can be written as

C β̃ = W T R−1 y (8)
T
where C = W T R−1 W + G∗ , W = [X Z] , β = τ T uT and


" #
0 0
G∗ = .
0 G−1

Also note that " # " # !

τ̂ − τ 0
β̃ − β = ∼N , C −1 (9)
ũ − u 0

3.5 Random treatments, partitioning and decomposing genetic

effects
In many applications the treatments, usually genetic lines, are considerd as random effects.
In this case we partition u as (uTg , uTb )T where the genotype effects are in ug and other
non-genetic random effects are in ub , and assume:
     
ug 0 σg2 Gg 0 0
 ub  = N  0  ,  0 Gb (κb ) 0 
     

e 0 0 0 R(φ)

where κb and φ are vectors of unknown variance components and correlation parameters.
The expanded mixed model equations are now:
    
Z Tg R−1 Z g + G−1g ũ g Z T
g R −1
y
T −1 T −1   T −1 
X R Z X R X τ̂ = X R y 
 
 g   
−1 −1 −1 −1
T
Zb R Zg T T
Z b R X Z b R Z b + Gb ) ũb Z b R−1 y
T

where Z = [Z g Z b ]. Sweeping out those equations for τ and ub gives

Z Tg SZ g + G−1 ũg = Z Tg Sy


g

C gg ũg = Z Tg Sy

where
" #−1 " #
X T R−1 X ··· XT
S = R−1 − R−1 [XZ b ] ,
Z b R X Z b R Z b + G−1
T −1 T −1
b Z Tb R−1
 var(ũg − ug ) = C gg (10)
and C gg = C −1 gg
gg . This suggests an optimality criterion for design based on some F(C ).

The assumption above that var(ug ) = σg2 I v may be plausible in certain cases where
genotypes are from diverse genetic backgrounds, and the degree of relatedness among
individuals is insignificant. Typically, in genetic improvement programs there is shared
ancestry among individuals and the simple model of independent genotype effects is un-
realistic.
A first step in decomposing the genetic effects is to distinguish between additive (that
is, the combined effects of alleles at two or more gene loci are equal to the sum of their
individual effects) and non-additive effects. Assuming additive and non-additive effects
are independent, we can decompose ug as

ug = ua + uā (11)

with var(ug ) = σa2 A + σā2 I v , where ua is the vector of additive genetic effects, and uā is
the vector of non-additive genetic effects, both associated with the design matrix Z g
The matrix A is a known (fixed) additive genetic relationship matrix derived from
ancestral records, however, from the mixed model equations (7) it is actually A−1 that
is required. Numerous methods exist, with many dating back to Henderson (1976), to
generate A−1 directly from pedigree records. The decomposition in (11) can be extended
to include other genetic relationships, such as dominance, but this will not be considered
any further here.
Often dim(A) > v, that is there are more individuals in the pedigree than will actually
appear in the design. On that basis, the (complete) vector of genotype effects, uG , say,
could be partitioned as uTG = (uTg , uTḡ )T , where uḡ represent those related genotypes that
do not appear in the design, but, may be used in calculating O in certain situations.
Butler et al. (2012) discuss this further, but it will not be expanded on in this document.
The implications for design of recognising a genetic covariance structure among treat-
ments are at least twofold:

1. The optimum allocation of genotypes to experimental units will be affected by the

relatedness of neighbours,

2. The optimal choice of test genotypes to replicate in series of p-rep designs may be
better determined.

4 Optimal design criteria

Table 2 lists several optimality criteria that have found popular use. As previously noted,
the A-optimality criterion is equivalent to minimising the average pairwise variance of all
elementary treatment contrasts; it is usually considered the most appropriate in circum-
stances such as comparative genetic evaluation experiments where all treatments are of
equal interest.

4.1 Fixed effects

Assuming genotypes are fixed effects, consider the mixed model equations in (7) reordered
(that is, permute the columns of W and the corresponding elements of β), so that the
information matrix C in (8) and its generalised inverse are partitioned as
" # " #
gg go
C gg C go C C
C= ; C− = ; (12)
C og C oo C og C oo

where C gg is the v × v coefficients matrix for genotypes, and C oo is the coefficients matrix
for all other fixed and random effects. If τ g is the v × 1 (sub)vector of fixed genotype
effects in τ , from (9) we can write

var(τ̂ g − τ g ) = C gg (13)

The A-optimality criterion minimises the average variance of the elementary contrasts
(τi − τj ), and
var(τ̂i − τ̂j ) = vij = cii + cjj − 2cij
where C gg = {cij }. Denoting the A-value of a design by A = v̄, it can be shown
Raghavarao (1971), for example, that
 
2 gg 1 T gg
A= tr [C ] − 1v C 1v (14)
v−1 v

and a design is said to be A-optimal if the A-value is a minimum for all designs under
consideration
Aopt = minD {Ai }.
If C − is a Moore-Penrose generalised inverse, ij cij = 0 and the A-value of a design is
P

just
A ∝ tr [C gg ]

4.2 Random effects

If genotypes are considered as random effects, possibly with a known relationship matrix
(that is, var(ug ) = σg2 A), an A-value analagous to (14) can be derived from (7) in a
similar manner to the above. Bueno Filho & Gilmour (2003), for example, use this
direct random effects equivalent of A-optimality in assessing competing designs for various
treatment correlation structures. In such cases, that is, models that include a known
genetic relationship matrix, an alternative criterion can be derived from considering the
so-called coefficient of determination, or the squared correlation between ug and ũg . It
turns out that this leads to proposing the average prediction error variance (PEV) as an
appropriate criterion, that is
O ∝ tr [C gg ]
using (10).

5 Algorithms for optimal design

Algorithms for optimal design have historically been considered in the context of designs
with quanitative factors, for both continuous and exact cases. Early work in this area
pre-dates the widespread use of high speed computing, and borrows from generic methods
for function minimisation. In adapting these early methods to categorical designs, the
objective function is replaced by the chosen criterion, such as A- or D-optimality, or PEV,
for example, and the interchange algorithms are recast in a permutation framework.
Even for small designs it is generally impossible to enumerate all solutions in D, and
an efficient sampling strategy directed by an optimisation algorithm is necessary. Design
search methods share a common set of features in exploring the design space D, and these
include:

1. A suitable starting design

2. An optimality criterion O

3. An interchange policy to move to a neighbouring configuration

4. Rules to assess the effect of a treatment interchange on O

5. An acceptance policy for new configurations

6. A stopping rule to terminate the search

Figure 2 illustrates the simplified structure of a design optimization algorithm. The

key features underlying this in the context of the computing paradigm discussed below
include:

• Specifying the linear model and the pre-specified random components.

• Choosing an appropriate criterion O.

• A mechanism to identify the target treatment factor and an interchange policy.

• An efficient optimization strategy to search the design space.

 • Specifying the linear model and the pre-specified random components.

• Choosing an appropriate criterion O.

• A mechanism to identify the target treatment factor and an interchange policy.

• An efficient optimization strategy to search the design space.

Figure 2: A simplified design optimization procedure.

Figure 1.2: A simplified design optimization procedure.

Begin

initialise components

O0 ← initial criterion

p← sample(D)

evaluate O

Y Y
O < O0 reject swap?
N N
O0 ← O

N
stop?

Accept p,O0

5.1 Search
Search algorithms
algorithms

The reject
The reject and and
stopstop decision boxes
decision boxes inin
Figure 1.2 together
Figure with the
2 together sampling
with of Dtheencapsulate
of D encapsulate
the sampling search
algorithm in the optimization process. Rejection rules may be probability based, in which case inferior
the search algorithm in the optimization process. Rejection rules may be probability
solutions will sometimes be accepted in order to widen the sampling of D. Stopping rules may be based
based, on
in iteration
which count,
case ainferior
minimumsolutions
threshold forwill
a change in the criterion,
sometimes or an analagous
be accepted physical
in order toprocess.
widen the
Two search algorithms have been implemented in od(), and these are described below.
sampling of D. Stopping rules may be based on iteration count, a minimum threshold for
a change in the criterion, or an analagous physical process. Two search algorithms have
Simulated annealing
been implemented in od(), and these are described below.
Simulated annealing is a stochastic search method introduced by Metropolis et al. [1953]. Kirkpatrick
et al. [1983] proposed an algorithm based on the analogy between the annealing of solids and solving
5.1.1 Simulated
combinatorial annealing
optimisation problems. Using this analogy, the Boltzmann probability
[−E /k t ]
Simulated annealing is a stochastic search P (Emethod
)∝e introduced by Metropolis et al. (1.15)
(1953).
Kirkpatrick et al.
determines the (1983) proposed
probability distributionan algorithm
of system energiesbased on the
E for a given analogyt . The
temperature between
constantthe
k an-
is Boltzmann’s constant. The energy E corresponds to the value of the objective function and the states
nealingofof solids and solving combinatorial optimisation problems. Using this analogy,
the solid correspond to feasible solutions.
the Boltzmann probability
P (E) ∝ e[−E/kt] (15)
determines the probability distribution of system energies E for a given temperature t.
The constant k is Boltzmann’s constant. The energy E corresponds to the value of the
objective function and the states of the solid correspond to feasible solutions.
Within a sequence of iterations, a new solution E2 is accepted unconditionally if ∆E =
E2 − E1 is negative, otherwise E2 is accepted according to Metropolis’s criterion based
on the Boltzmann probability. That is, a random number δ in [0, 1] is generated and E2
accepted if δ ≤ exp(−∆E /t).
The minimum energy state corresponds to the optimum solution and several cooling
schedules have been proposed in the literature Osman (1991); Lundy (1985).
 Simulated annealing has been used in optimal design problems (Chan, 1999; Elliot
et al., 1999) but Coombes (2002) found its performance inferior to tabu search based
methods.

5.1.2 Tabu search

The tabu algorithm Glover (1989, 1990) is an iterative procedure characterised by flexible
memory structures and local search strategies, with the ability to escape local minima.
A basic tabu algorithm involves three main strategies: forbidding, freeing and short-term
strategy.
The forbidding strategy guards against cycling problems by preventing solutions vis-
ited within the last Ts iterations from being revisited. Ts is normally called the tabu list
size and the tabu list is a FIFO queue. The freeing strategy is used in conjunction with an
aspiration criterion to decide what (if any) candidates exit the tabu list. The short-term
strategy manages the interaction between the two previous strategies. For example, a lo-
cal descent strategy could be used to evaluate admissible solutions in the neighbourhood
of the current solution.
The work of Coombes (2002) suggests that reactive tabu search (Battiti & Tecchiolli,
1994) could be more efficient as the optimisation loop in design search problems. Reactive
tabu modifies the basic strategy to include a tabu list of dynamic length, determined by
a self tuning heuristic.
od() implements the tabu search strategy operating with a memory resident search
list (Woodruff & Zemel, 1993). A unique key is generated for each design permutation
vector pi , and stored in a rapid access memory structure. The design key is calculated
as a 32 bit checksum using the computing industry standard Cyclic Redundancy Check
(CRC) algorithm (Press et al., 1996) with polynomial 0x04C11DB7.

6 A design generator: od()

Many of the concepts in the previous sections of this chapter are being incorporated into
an R package od(), where designs are specified using symbolic model formulae in the
spirit of ASReml. We begin with some brief computing details specific to od() that put
the function arguments in context.

6.1 Computational background

Let W = [X Z g Z b ] and permute the columns of W such that W = [W 1 W 2 ] and
W 1 = Z g (or that partition of Z g conformal with the subset of ug forming O).
Given invariant design matrices for non-genetic effects, our goal is to minimise O
through an optimal allocation of genotypes to the experimental units. This is conveniently
achieved with a permutation matrix P operating on the rows of W 1 . Initially P = I N .
Consider a perturbation function S() operating on the rows of P , such that P ∗ = S(P )
is a new permutation. A new design is produced as W ∗1 = P ∗ W 1 . Computationally, a
design is represented in a permutation vector p that indexes the rows of P , and a new
design p∗ is generated as p∗ = s(p), where s() is the corresponding vector perturbation
function. A straightforward interchange function for s() is typically chosen which swaps
two elements of p(rows of P ), subject to any resolvability constraints. Given an initial
p0 = [1, 2, . . . , n], the algorithm minimises O through repeated applications of s(p) under
the supervision of an optimisation strategy such as TABU search or simulated annealing
(see Coombes, 2002). At the time of writing the TABU implementation was unreliable
and simulated annealing is used in the examples that follow.
The optimization proceeds as shown in Figure 2, where a move to an new (neighbour-
ing) design is achieved using the permutation strategy describe above, and O is calculated
as:

1. Form the mixed model equations

W T R−1 W + G∗ where G∗ = G−1 ⊕ 0

2. Absorb the equations for the effects in W 2 to form C gg

3. Form C gg and calculate O

6.2 The od()function

Table 3 outlines the arguments to od() and the corresponding components of the linear
model or methodology to which they refer (where applicable). Section 9 reproduces the
R help pages for the od() package.

Table 3: Arguments to od()

Argument Value Model/method context
fixed ∼ fixed model terms X
random ∼ random model terms Z
rcov ∼ formula specifying the error terms R
permute ∼ formula specifying term(s) to permute W1
swap ∼ formula controlling legal treatment exchanges S()
Gstart list assigning values of pre-specified variance parameters G
Rstart list assigning values of pre-specified variance parameters R
ginverse list specifying inverse relationship matrices A−1
method c(’avalue’,’PEV’) O
search c(’anneal’, ’tabu’, ’random’)
data dataframe containing initial configuration

Variance models for random terms are specified as special model functions in the same
manner as ASReml-R. Currently these models are limited to id(), idv(), ar1(), ped() and
ide() while the package is being developed. A data frame in which to resolve the named
terms in the model formulae, and containing the initial design configuration, must be
provided; this is in the usual observation × variate array with model terms declared as
factors or variables as appropriate.
The function returns an object of class od with the following components:

• A dataframe with the permute terms in design order

• The permutation vector p.

• The optimality criterion O.

• List(s) of the pre-specified variance components.

• The function call.

Methods for summary() and plot() exist, though currently have limited capability.

7 Examples
CHAPTER 1. OD: AN OPTIMAL DESIGN COMPANION TO ASREML 13
7.1 A randomised block
1.7 Examples
We begin by illustrating some features of od() with the simplest and arguably most
commonly usedblock
A randomised block design. Consider the following experimental layout:
We begin by illustrating some features of od() with the simplest and arguably most commonly used block
design. Consider the following experimental layout:

I Block II
Configuration parameters (Table 1)
1 A A Configuration parameters (Table 1.1)
v =5
v=5
B B k =5 k=5
s =1
r =2
s=1
Plot

C C qr = 5 r=2
qc = 2
q =5
N = 10 r
D D qc = 2
N = 10
5 E E

> ## Initial configuration

> rb <- data.frame(Trt=rep(LETTERS[1:5],2),
R> ## >> Initial configuration
Block=factor(rep(1:2,each=5)),
Column=factor(rep(1:2,each=5)),
R> rb > <- data.frame(Trt=rep(LETTERS[1:5],2),
Row=factor(rep(1:5,2))) Block=factor(rep(1:2,each=5)),
> ## generate a design
+ > Column=factor(rep(1:2,each=5)), Row=factor(rep(1:5,2)))
rb.od <- od(fixed= ~Trt, random= ~Block, permute= ~Trt,
R> ## generate aswap= ~Block, maxit=100, data=rb)
design
> names(rb.od)
R> rb.od <- od(fixed= ~Trt, random= ~Block, permute= ~Trt, swap= ~Block,
+ [1] "call"
maxit=100, "design"
data=rb) "criterion" "permutation" "Gstart"
[6] "Rstart"
R> names(rb.od)
Before considering the returned object rb.od, there are some implicit and explicit actions of the call that
warrant explanation:
[1] "call" "design" "criterion" "permutation" "Gstart"
Block appears in the random formula but no variance model or initial values have been given.
[6] •"Rstart"
The default model in this case is a scaled identity, with a default scale factor, that is, the variance
component for blocks, of 0.1.

• The rcov model formula is not specified, so the default error model, N (0, I 10 ) has been used, re-
membering that σ2 is fixed at 1.0.

• The swap model formula enforces the resolvability constraints of the design by restricting treat-
ment exchanges to occur only within levels of the Block factor.
Before considering the returned object rb.od, there are some implicit and explicit actions
of the call that warrant explanation:

• Block appears in the random formula but no variance model or initial values have
been given. The default model in this case is a scaled identity, with a default scale
factor, that is, the variance component for blocks, of 0.1.

• The rcov model formula is not specified, so the default error model, N (0, I 10 ) has
been used, remembering that σ 2 is fixed at 1.0.

• The swap model formula enforces the resolvability constraints of the design by re-
stricting treatment exchanges to occur only within levels of the Block factor.

The optimized design is returned in the design component of the returned object, and
the corresponding permutation vector p is returned in the permutation component.

rb.od$permutation

[1] 1 2 3 4 5 6 7 8 9 10

rb.od$design

Trt Block Column Row units

1 A 1 1 1 1
2 B 1 1 2 2
3 C 1 1 3 3
4 D 1 1 4 4
5 E 1 1 5 5
6 A 2 2 1 6
7 B 2 2 2 7
8 C 2 2 3 8
9 D 2 2 4 9
10 E 2 2 5 10

The (apparently) optimized design is just a copy of the initial dataframe, a result that
surprises new users. This example resolves into complete blocks, has no incomplete block-
ing, extraneous effects or correlated error structure to drive treatment exchange. Under
this model all configurations are optimal and manual randomization of treatments is the
only guard against systematic effects.
To illustrate this point, an autoregressive error structure is added to the model and
initial values are set using the od.init() function.

R> rb.init <- od.init(Block=0.1, "Column:Row|Column"=0.2, "Column:Row|Row"=0.4)

R> rbsp.od <- od(fixed= ~Trt, random= ~ Block, rcov= ~ar1(Column):ar1(Row),
+ permute= ~Trt, swap= ~Block, Rstart=rb.init, maxit=100, data=rb)
R> rbsp.od$design
 Trt Block Column Row
1 A 1 1 1
2 C 1 1 2
3 E 1 1 3
4 B 1 1 4
5 D 1 1 5
6 E 2 2 1
7 A 2 2 2
8 D 2 2 3
9 C 2 2 4
10 B 2 2 5

The initial A-value is 0.9320, while that of the optimized design above is 0.6300.
The final configuration exhibits some classic properties of a spatial design based on an
autoregressive process, such as diagonal self-neighbours and no self-adjacencies; that is
the design is binary with respect to the Row factor:

R> table(rbsp.od$design$Trt,rbsp.od$design$Row)

1 2 3 4 5
A 1 1 0 0 0
B 0 0 0 1 1
C 0 1 0 1 0
D 0 0 1 0 1
E 1 0 1 0 0

In this small example, the binary configuration will most likely always be found, but as
N gets larger there is no guarantee that a binary design will result. However, if the initial
design is binary then the optimal design will also be binary, subject to an appropriate
model.

7.2 A glasshouse example

This example is a glasshouse experiment from the national LMA screening program (Mrva
& Mares, 2001). Briefly, this is stage one of a multi-stage process where (wheat) plants are
grown in pots in a glasshouse prior to being transferred to other laboratory procedures.
Two replicates of the treatments are arranged in pots in a 3 × 4 grid within trays across
glasshouse benches, where trays and benches are contiguous. The centre two pots in each
tray are left empty (blank), and are indicated by cross-hatching in the figure below. The
main purpose of this example is to further illustrate the action of the swap argument.

R> ## LMA glasshouse winter 2011

R> ## design factors, no treats
R> ## 448 unique treats (+extra stds/rep)
 A glasshouse example

This example is a glasshouse experiment from the national LMA screening program [Mrva and Mares,
2001]. Briefly, this is stage one of a multi-stage process where (wheat) plants are grown in pots in a
glasshouse prior to being transferred to other laboratory procedures. Two replicates of the treatments
are arranged in pots in a 3 × 4 grid within trays across glasshouse benches, where trays and benches are
contiguous. The centre two pots in each tray are left empty (blank), and are indicated by cross-hatching
in the figure below. The main purpose of this example is to further illustrate the action of the swap
argument.
1 2 3 4 5 6 7 8

1 2 3 4 5
1
2 1
v = v448
= 448

ay
Tr
I
3 r = r2= 2
4
2 k =ks12= 12
= 46
Replicate

Bench
s = q46
r = 12
3 q c = 92
qr = 12
II

4
qc = 92

> ## LMA glasshouse winter 2011

> ## design factors, no treats
> ## 448 unique treats (+extra stds/rep)
R>## At> ##
10 At 10 tray,
per per tray,
46 46 trays/rep, 460
trays/rep, 460 units/rep
units/rep
> ## 92 cols X 12 rows
R>## 92> cols X 12 rows
> Column <- rep(1:92,rep(12,92))
R>
> Row <- rep(1:12,92)
R>Column <- rep(1:92,rep(12,92))
> Replicate <- rep(rep(1:2,c(6,6)),92)
> Tray <- matrix(rep(rep(1:92,rep(3,92))),nrow=23,byrow=TRUE)
R>Row <- rep(1:12,92)
> Tray <- as.vector(t(cbind(Tray,Tray,Tray,Tray)))
R>Replicate <- rep(rep(1:2,c(6,6)),92)
> ## Blank - control in centre two cells of each tray
> cc <- sort(c(seq(from=2,by=4,length.out=23),seq(from=3,by=4,length.out=23)))
R>Tray ><-
rr matrix(rep(rep(1:92,rep(3,92))),nrow=23,byrow=TRUE)
<- c(2,5,8,11)
R>Tray ><- as.vector(t(cbind(Tray,Tray,Tray,Tray)))
Blank <- rep(0,92*12)
> Blank[(is.element(Column,cc) & is.element(Row,rr))] <- 1
R>## Blank - <-
> Treat control in centre two cells of each tray
rep(0,92*12)
R> >
cc <- sort(c(seq(from=2,by=4,length.out=23),seq(from=3,by=4,length.out=23)))
> ghex <- data.frame(Column=factor(Column),
R>rr <-> c(2,5,8,11) Row=factor(Row),
R>Blank <- rep(0,92*12) Replicate=factor(Replicate),
>
> Tray=factor(Tray),
R>Blank[(is.element(Column,cc)
> & is.element(Row,rr))] <- 1
Blank=factor(Blank),
R>Treat> <- rep(0,92*12) Treat=Treat)
> ghex[1:15,]
R>
R> ghex <- data.frame(Column=factor(Column),
Column Row Replicate Tray Blank Treat Row=factor(Row),
1 1 1 1 1 0 0
+ Replicate=factor(Replicate), Tray=factor(Tray), Blank=factor(Blank),
+ Treat=Treat)
R> ghex[1:15,]

Column Row Replicate Tray Blank Treat

1 1 1 1 1 0 0
2 1 2 1 1 0 0
3 1 3 1 1 0 0
4 1 4 1 2 0 0
5 1 5 1 2 0 0
6 1 6 1 2 0 0
7 1 7 2 3 0 0
8 1 8 2 3 0 0
9 1 9 2 3 0 0
10 1 10 2 4 0 0
11 1 11 2 4 0 0
12 1 12 2 4 0 0
13 2 1 1 1 0 0
 14 2 2 1 1 1 0
15 2 3 1 1 0 0

At this point the design factors exist but the treatments are yet to be allocated; this
process is not relevent to the discussion and has been omitted for brevity. Note that
the factor Blank identifies two zones: the empty pot positions, and the rest. The call to
generate a design using the complete dataframe gh is:

R> gh.od <- od(fixed = ~ Entry, random = ~ Replicate+Tray+Column+Row,

+ rcov = ~ ar1(Column):ar1(Row), permute = ~ Entry, swap = ~ Replicate:Blank,
+ Gstart=gh.init, Rstart=gh.init, maxit=5, data=gh)

where the main purpose of the spatial error structure is to promote some neighbour bal-
ance properties in the configuration. The swap argument in this example limits treatment
exchanges to be within the intersection of the levels of the nominated factors (Repli-
cate:Blank), thus maintaining resolvability and banning swaps between zones.

R> table(ghex$Replicate,ghex$Blank)

0 1
1 460 92
2 460 92

Perhaps not the most efficient solution given the wasted legal swaps within the empty
zones, but nontheless effective. An alternative strategy would have been to omit the blank
units from the dataframe and, if considered desirable, use a power model for the spatial
component (should one have been implemented).

7.3 Partially replicated designs

As the name suggests, p-rep designs (Cullis et al., 2006b) replicate a proportion p of the
test treatments in resolvable replicates with the remaining treatments appearing once
only. The figure below illustrates this property.

R> id <-rep(LETTERS[1:7],c(2,2,2,1,1,1,1))
R> prep <- data.frame(Trt=as.vector(
+ matrix(id,nrow=5,byrow=TRUE))[c(sample(1:5),sample(6:10))],
+ Block=factor(rep(1:2,each=5)), Column=factor(rep(1:2,each=5)),
+ Row=factor(rep(1:5,2)))
R> prep

Trt Block Column Row

1 A 1 1 1
2 B 1 1 2
 CHAPTER 1. OD: AN OPTIMAL DESIGN COMPANION TO ASREML 17

Column
1 2 3 4 5 6
1 A

B
F
C v = 213
v = 213
75 × 2 replicates
75 × 2 replicates
138 × 1 replicate
A 138 × 1 replicate
r =2
Row

Replicate 1 Replicate 2 q r =r48

=2
q c =q6
r = 48
N =q288
c = 6
E
N = 288
B

F
D

30 G

> id <-rep(LETTERS[1:7],c(2,2,2,1,1,1,1))
> prep <- data.frame(Trt=as.vector(
> matrix(id,nrow=5,byrow=TRUE))[c(sample(1:5),sample(6:10))],
> Block=factor(rep(1:2,each=5)),
3 C> 1 1 3 Column=factor(rep(1:2,each=5)),
4 F> 1 1 4 Row=factor(rep(1:5,2)))
5 D> prep 1 1 5
6 E 2 2 1
7 A Trt 2 Block Column
2 2 Row
1 A 1 1 1
8 C 2 2 3
2 B 1 1 2
9 G3 C2 1 2 41 3
10 B4 F2 1 2 51 4
5 D 1 1 5
6 E 2 2 1
R> prep.init <- od.init(Column=0.4, Row=0.2, "Column:Row|Column"=0.1,
7 A 2 2 2
+ "Column:Row|Row"=0.4)
8 C 2 2 3
R> prep.od
9 <-G od(fixed=
2 ~Trt,
2 4 random= ~ Column+Row, rcov= ~ar1(Column):ar1(Row),
+ permute=
10 B ~Trt,2 swap=2~Block,
5 Gstart=prep.init, Rstart=prep.init,
+ maxit=50, data=prep)
R> summary(prep.od)$incidence
> prep.init <- od.init(Column=0.4, Row=0.2, "Column:Row|Column"=0.1,
"Column:Row|Row"=0.4)
$Column> prep.od <- od(fixed= ~Trt, random= ~ Column+Row, rcov= ~ar1(Column):ar1(Row),
permute= ~Trt, swap= ~Block, Gstart=prep.init,
Rstart=prep.init, maxit=50, data=prep)
1 2 > summary(prep.od)$incidence
A 1 1
B 1 1 $Column
C 1 1
D 1 0 1 2
E 0 1A 1 1
B 1 1
F 1 0C 1 1
D 1 0
Figure 3: Frequency distribution of A = {aij }, i > j on the correlation scale for the 402
individuals in the CBWA pedigree.

40

30
Percent of Total

20

10

0.0 0.2 0.4 0.6 0.8

Correlation

G 0 1

$Row

1 2 3 4 5
A 1 0 1 0 0
B 1 0 0 0 1
C 0 0 1 0 1
D 0 1 0 0 0
E 0 0 0 1 0
F 0 0 0 1 0
G 0 1 0 0 0

The initial A-value for the starting design was 1.614993, and the optimal A-value at ter-
mination was 1.415362 for an approximate 15% gain. The optimal design is binary with
respect to Row although treatment C is a self-neighbour in the initial random configura-
tion; as noted previously this is not a guaranteed outcome.
The figure above is a real example from the canola breeding program of Canola Breed-
ers WA (CBWA) and we use it to illustrate the use of a numerator relationship matrix (A)
in design. That is, we assume that var(ug ) = σg2 A. The 217 entries in the experiment are
from a pedigree containing 482 individuals, which was subsequently pruned to 402, that
is the 217 trial entries and 185 progenetors. Some indication of the degree of relatedness
in the pedigree is shown in Figure 3 where the relative frequencies of the elements in the
lower triangle of A on the correlation scale are given.
The A−1 matrix is required by the mixed model equations, and like ASReml this is
given to od() in 3 column sparse form:
R> pinv[1:10,]

Row Column Ainverse

1 1 1 0.7897590
2 2 2 2.3440514
3 3 3 0.9543473
4 4 4 1.3209545
5 5 5 0.9997506
6 6 5 0.4995005
7 6 6 0.9997506
8 7 7 1.8990912
9 8 7 0.9990010
10 8 8 2.2989313

R> str(canola.df)

'data.frame': 288 obs. of 4 variables:

$ column: Factor w/ 6 levels "1","2","3","4",..: 1 1 1 1 1 1 1 1 1 1 ...
$ row : Factor w/ 48 levels "1","2","3","4",..: 1 2 3 4 5 6 7 8 9 10 ...
$ block : Factor w/ 2 levels "1","2": 1 1 1 1 1 1 1 1 1 1 ...
$ geno : Factor w/ 213 levels "ICPBER-0001",..: 21 83 193 13 137 70 22 47 168 ...

The mechanism for including a known relationship matrix in od() parallels that for AS-
Reml, that is, the inverse matrix is given by the ginv argument and is linked to the
relevent model term with the ped() special function. By default, od() will generate a
design optimal for those entries in the dataframe.

R> canola.G <- od.init("ped(geno)" = 1.0, block=0.3, column=0.6, row=0.4 )

R> canola.R <- od.init("column:row|column"=0.2, "column:row|row"=0.5)
R> canola.od <- od(random = ~ ped(geno)+block+column+row,
+ rcov = ~ ar1(column):ar1(row),permute = ~ ped(geno), swap = ~ block,
+ Gstart=Ginit, Rstart=Rinit, ginv=list(geno=pinv), optimize='data',
+ search='anneal',method='PEV',maxit=50, data=canola.df)

A term for non-additive genetic effects, specified by the ide() special function, has
been omitted from the above model, and as a consequence we are assuming (perhaps un-
realistically) that the genetic effects are exclusively additive. For this example, there was
a gain of approximately 7.0% in the average prediction error variance when the optimal
configuration ignoring the known genetic relationship matrix was further optimized using
the pedigree information.

7.4 Embedded p-rep designs

The context for this type of design is introduced in Chapter 1 and Smith et al. (2011).
Briefly, some quality characteristics of grain samples from field experiments are typically
expensive (in economic or throughput terms) to determine. The concept of an embedded
 CHAPTER 1. OD: AN OPTIMAL DESIGN COMPANION TO ASREML

Joinly optimise both areas using a common spatial model and including model term(s) d
experimental structure. Treatment interchanges that violate the embedded layout are d

Figure 4: An embedded p-rep design.

Figure 1.4: An embedded p-rep design.
Column
1 2 3 4 5 6

1 A
Full design
Full design
E v = 90
vr == 90
2
B
r=2
F
q = 30
C qr r= 30
qq c==66
c

A Shadedarea
Shaded area
Row

vv==90 90
30
30 x 22 ++6060
x x 1x= 1120
= plots
120
20 E plots
r =2
r=2
B
q = 20
qr r= 20
F qq c==66
c
D

30 G

The third option is illustrated here. The full design is a complete block configuration, an
bedded design is a contiguous array of plots in the first two thirds of the area (Figure 1.4). The
are identified by the levels of the factor Prep, and the integrity of the embedded design is
through
design for the
this search
purposebywithin
declaring theexperiment
a larger swap boundaries asthe
mitigates thehigh
intersection of Prep and Replicat
cost by sampling
fewer plots while maintaining statistical rigour.
> Ginit
Given <- od.init("~Row"=0.1,
a pre-specified "~Column"=0.1,
replication scheme, there "~Rep"=0.1,
are several approaches "~Prep"=0.1)
to sampling
from> the
Rinit <- od.init("~Column:Row|Column"=0.3,"~Column:Row|Row"=0.6)
full design to reduce laboratory costs, including:
> prep.od <- od(fixed = ~ID,
>
Sample random
from the full design by = ~ Prep+Rep+Column+Row,
choosing one complete replicate and subsetting treat-
> ments from a secondrcov
replicate. This leads to a spatially disjoint scheme. = ~ ID,
= ~ ar1(Column):ar1(Row),permute
> swap = ~ Prep:Rep, Gstart=Ginit, Rstart=Rinit,
A>conditional approach whereby an optimal
data=site, p-rep design and optimise the full design
maxit=40)
conditional on fixing the p-rep design.
An example outcome is shown in Figure 1.7, where the shaded areas correspond to the replic
Joinly optimise both areas using a common spatial model and including model
types.
term(s) defining the experimental structure. Treatment interchanges that violate
the embedded layout are disallowed.
Figure 1.5: An example embedded p-rep outcome.

The third option is illustrated here. The full design Column

is a complete block configuration,
1 2 3 4 5 6
and the embedded design is a contiguous array of plots in the first two thirds of the area
1
(Figure 4). The two areas are identified by the levels of the factor Prep, and the integrity
of the embedded design is preserved through the search by declaring the swap boundaries
as the intersection of Prep and Replicate
Row
 > data=site, maxit=40)

An example outcome is shown in Figure 1.7, where the shaded areas correspond to the replicated geno-
types.

Figure
Figure1.5:
5: An
An example embeddedp-rep
example embedded p-rep outcome.
outcome.

Column
1 2 3 4 5 6

Row

20

30

R> Ginit <- od.init("~Row"=0.1, "~Column"=0.1, "~Rep"=0.1, "~Prep"=0.1)

R> Rinit <- od.init("~Column:Row|Column"=0.3,"~Column:Row|Row"=0.6)
R> prep.od <- od(fixed = ~ID, random = ~ Prep+Rep+Column+Row,
+ rcov = ~ ar1(Column):ar1(Row), permute = ~ ID, swap = ~ Prep:Rep,
+ Gstart=Ginit, Rstart=Rinit, data=site, maxit=40)

An example outcome is shown in Figure 5, where the shaded areas correspond to the
replicated genotypes.

7.5 Chain block designs

The chain block design was introduced by Youden & Connor (1953), motivated by appli-
cations in the physical sciences where measurements are often made with high precision,
and replication can be reduced. They have since proved useful in sequential applications
where subjects or samples arrive chronologically, possibly from a prior phase process.
The real example below is from a two phase process (actually three counting the original
field experiment) where flour samples from milled grain were to be scored for viscosity
measurements using a rapid viscosity analyzer (RVA). The full experiment is too large
to be considered in depth here, so for illustration we initially consider a small synthetic
example.

R> x <- 1:4

R> cb <- data.frame(Trt = as.vector(rbind(
 the physical sciences where measurements are often made with high precision, and replication can be
reduced. They have since proved useful in sequential applications where subjects or samples arrive
chronologically, possibly from a prior phase process. The real example below is from a two phase process
(actually three counting the original field experiment) where flour samples from milled grain were to be
scored for viscosity measurements using a rapid viscosity analyzer (RVA). The full experiment is too large
to be considered in depth here, so for illustration we initially consider a small synthetic example.

Block
1 2 3 4
I II III IV

1 A B C D
A ... D replicated groups
A ... D replicated groups
B C D A a ... p unreplicated treatments
a ... p unreplicated treatments
gg= 1= 1
a e i m vv= 20= 20
Time

kk= 6= 6
b f j n s =4
Ns== 24 4
c g o
N = 24
k

6 d h l p

> x <- 1:4

> cb <- data.frame(Trt = as.vector(rbind(
+ >matrix(LETTERS[c(x,x%%4+1)],nrow=2,byrow=TRUE),
matrix(LETTERS[c(x,x%%4+1)],nrow=2,byrow=TRUE),
> matrix(letters[1:16],nrow=4,byrow=FALSE))),
+ >matrix(letters[1:16],nrow=4,byrow=FALSE))),
Block = factor(rep(1:4,each=6)), Block = factor(rep(1:4,each=6)),
+ >Time = factor(rep(1:6,4)), resolvable = rep(c(rep(TRUE,2),rep(FALSE,4)),4))
Time = factor(rep(1:6,4)),
> resolvable = rep(c(rep(TRUE,2),rep(FALSE,4)),4))
R> cb[1:10,]
> cb[1:10,]

Trt Trt
BlockBlockTime
Timeresolvable
resolvable
1 1 A A 1 1 1 1 TRUE
TRUE
2 B 1 2 TRUE
2 3 B a 1 1 23 TRUE
FALSE
3 4 a b 1 1 34 FALSE
FALSE
5 c 1 5 FALSE
4 6 b d 1 1 46 FALSE
FALSE
5 7 c B 1 2 51 FALSE
TRUE
8 C 2 2 TRUE
6 d 1 6 FALSE
9 e 2 3 FALSE
7 10B f 2 2 1 4 TRUE
FALSE
8 C 2 2 TRUE
9 The column
e labelled
2 resolvable
3 is in parallel with the Trt factor and is used by od() to modify the be-
FALSE
haviour of the swap argument. If both swap units have treatments labelled resolvable=FALSE then the
10constraint
f 2
imposed by4 swap is ignored,
FALSE and those treatments can move throughout the design. Thus, in

The column labelled resolvable is in parallel with the Trt factor and is used by od() to
modify the behaviour of the swap argument. If both swap units have treatments labelled
resolvable=FALSE then the constraint imposed by swap is ignored, and those treatments
can move throughout the design. Thus, in the configuration above treatments A-D re-
main within their assigned blocks, while treatments a-p can be exchanged throughout the
design.

R> cb.init <- od.init(Block=0.4, Time=0.1, "Block:Time|Time"=0.3)

R> cb.od <- od(fixed= ~Trt, random= ~ Block+Time, rcov=~Block:ar1(Time),
+ permute= ~Trt, swap= ~ Block, Gstart=cb.init, Rstart=cb.init,
+ maxit = 100, data=cb)
R> matrix(cb.od$design$Trt,nrow=6)
 [,1] [,2] [,3] [,4]
[1,] "j" "m" "c" "a"
[2,] "A" "B" "C" "D"
[3,] "b" "o" "k" "h"
[4,] "l" "e" "g" "f"
[5,] "B" "C" "D" "A"
[6,] "d" "p" "i" "n"

The two phase example introduced above has chain block parameters v = 620, k = 16,
g = (2, 3) for s = 46 blocks (days), where the last block has ony 15 units. The first phase
milling process spanned 62 days, and it was a requirement that the RVA assessment begin
as soon as possible after milling commenced, running in parallel with the milling process
thus restricting the abiliy to allocate treatments throughout the design.
It is possible that experimental units from the milling phase will carry effects from
milling sources of variation into the RVA phase of testing. These potential effects, such
as milling day or mill order must be kept in sync with treatment exchanges throughout
the optimization. The pipe symbol ”—” can be used in the permute formula to specify
the factors additional to the treatment factor that are to be permuted. This design took
some considerable time to run and the od() call is reproduced below to illustrate the form
of the permute argument.

R> load("rva.RData")
R> rva$MillR <- factor(paste(as.character(rva$MillDay),
+ as.character(rva$MillOrd)))
R> str(rva)

'data.frame': 735 obs. of 16 variables:

$ Day : Factor w/ 46 levels "1","2","3","4",..: 1 1 1 1 1 1 1 1 1 1 ...
$ Seq : Factor w/ 16 levels "1","2","3","4",..: 1 2 3 4 5 6 7 8 9 ...
$ MillRep : Factor w/ 2 levels "1","2": 1 1 1 1 1 1 1 1 1 1 ...
$ MillDay : Factor w/ 62 levels "1","2","3","4",..: 2 1 1 2 1 1 1 1 1 1 ...
$ MillOrd : Factor w/ 10 levels "1","2","3","4",..: 1 3 5 2 6 7 8 4 1 9 ...
$ Bay : int 1 3 3 2 1 1 3 1 1 2 ...
$ Row : int 6 71 70 29 16 26 79 13 1 53 ...
$ Col : int 13 1 3 14 5 13 5 11 2 17 ...
$ ID : Factor w/ 508 levels "Alsen","Baxter",..: 342 125 156 276 ...
$ frep : int 1 1 1 1 1 1 1 1 1 1 ...
$ trialUnitID : int 1517 610 771 1621 879 1537 942 1362 621 1888 ...
$ compTrialUnitID: int NA NA NA NA NA 1992 NA NA NA NA ...
$ repDesc : Factor w/ 4 levels "Comp","field rep",..: 4 2 4 4 4 1 4 4 4 ...
$ slDup : logi TRUE FALSE FALSE TRUE FALSE FALSE ...
$ MillR : Factor w/ 620 levels "1 1","1 10","1 2",..: 111 4
6 113 7 8 ...

R> rva.G <- od.init(Day=0.9,Seq=0.6, MillRep=0.1,MillDay=0.9,

+ MillOrd=0.4,MillR=1.0)
R> rva.od <- od(fixed = ~ ID, random = ~ Day + Seq + MillRep + MillDay
+ + MillOrd + MillR, permute=~ID|MillRep+MillDay+MillOrd+MillR,
+ swap=~Day, Gstart=rva.G, maxit=50, data=rva)
7.6 Partial compositing
The rationale behind partially compositing grain samples as a means of recovering (statis-
tical) information, or indeed in practice as a means of gaining information on genotypes
(otherwise lost) with insufficient individual plot samples is detailed elsewhere in these
notes. For example, in Figure 4, the 30 replicated genotypes in the shaded area are tested
in duplicate while the remaining 60 genotypes are tested as composite samples. That is,
a mixed sample from the single plot within the embedded area, and the one outside is
prepared for further testing.
As noted earlier, the analysis of these data in ASReml requires special design matri-
ces which can be constructed using the and() model function. At the time of writing
this model function is not implemted in od(), consequently designs that recognise this
compositing scheme cannot be constructed strictly in accordance with the appropriate
model.

8 Summary
Optimal design with a flexible computing strategy is a powerful tool for design generation,
particularly in novel situations, or those with awkward experimental restrictions. On the
positive side, the advantages of the technology include:

• Flexibility

– in terms of the underlying linear model, and

– in terms of the experimental configuration, given an appropriate computing
strategy.

• The model based approach strongly links design and analysis. The model formulae
from od() provide both an accurate record of the properties of the design, and the
means of analysis.

• The computational approach using direct methods promotes this flexibility, and
allows computing efficiencies such as exploiting sparsity.

There are, however, unanswered questions and software support issues the include:

• The methodology requires the specification of a linear model, and pre-specified

variance parameters. The principal question that arises is that of robustness

– in terms of the choice of linear mixed model, and

– sensitivity to variance parameter estimates.
– Some work on robustness appears in Chan (1999), where it was found that
the choice of parameter value within a model class was less important than
the choice of variance model family. However, with the range of design models
now possible, the issue of robustness warrants further investigation.
• The choice of optimality criterion is not necessarily resolved.

• The construction of an initial design. Software to at least partially automate this

step is still under construction. It is not possible to anticipate all cases but the ability
to generate binary designs for quasi regular configurations is a current deficiency.
9 The od() class

od Optimal design

Description

Generate optimal categorical designs under a general linear mixed model.

Usage

od(fixed = ~1, random = ~NULL, rcov = ~NULL, permute = ~NULL,

swap = ~NULL, ginv = list(), Gstart = list(), Rstart = list(),
data = sys.parent(), search = c("anneal","tabu", "random"),
method = c("avalue","PEV"), maxit = 1, trace = FALSE, debug = FALSE)

Arguments

fixed a formula specifying the fixed model terms.

random a formula specifying the random model terms.
rcov a model formula specifying the error structure.
permute a formula containing a single term specifying the factor in the fixed
or random set that is to be permuted. The "|" operator can be used
to associate this factor with other terms in the model that should be
permuted in parallel.
swap a formula nominating which factor(s) define legal treatment exchanges
during the optimisation process. Rows of W = cbind(X,Z) are only
permuted within levels of the swap factor (or the intersection of levels
if swap is of the form ’~ A:B’).
ginv a list object associating a known inverse treatment relationship matrix
with a random term in the model. The inverse is given in sparse form
(row, column, value) as lower triangle row-wise.
Gstart a list object containing initial values for the random components, typ-
ically generated from a call to od.init.
Rstart a list object containing initial values for the residual components, typ-
ically generated from a call to od.init.
data a dataframe in which to resolve the terms in the model formulae.
 search a character string specifying the search strategy. The default, "anneal",
specifies classical simulated annealing, "tabu" initiates a robust tabu
search (Taillard, 1991) while "random" simply exchanges treatments
in an undirected manner.
method the optimality criterion. "avalue" is typically used when treatment
effects are considered as fixed, and optimizes the average pairwise vari-
ance of elementary contrasts. "PEV" may be more appropriate for ran-
dom effects, particularly when a known treatment covariance matrix
is present, and optimizes the trace of the inverse coefficient matrix for
treatments.
maxit the number of cooling loops for simulated annealing, or tabu loops if
the search method is tabu, otherwise the number of random exchanges.
trace if TRUE then the optimization progress is reported. Currently FALSE
does nothing.
debug if TRUE various R structures are returned for debug purposes. If nu-
meric (1 or 2), intermediate structures and computations are printed
from the numerical functions.

Details

The mixed model equations are formed from the model given by the fixed, random and
rcov formulae, and the variance parameters pre-specified in Gstart and Rstart. The
objective function is calculated for the model (treatment) term given in the permute
formula. The objective function (avalue or PEV) is optimized under the supervision
of the search strategy for successive exchanges of pairs of treatments, subject to the
conditions set by the terms in swap, until a stopping condition is satisfied. Optimiza-
tion proceeds until maxit is exceeded (anneal, tabu and random), or the temperature
change is lower than the specified minimum for simulated annealing (see od.options).
The cooling schedule for simulated annealing can be either geometric or lundy, and
is set along with other annealing options in od.options.
Special functions may be used in model formulae; currently the only recognised func-
tions are id, ar1, ped and ide.

Value

A list object of class od with the following components:

call the od() function call.

design the dataframe with the factor nominated in the permute formula in
design order.
criterion the final value of the optimality criterion.
 permutation the permutation, or design order, of the rows of data.
Gstart initial values
Rstart initial values

od.init initial values for od()

Description

creates a list object containing initial values for function od.

Usage

od.init(...)

Arguments

... a series of comma separated name = value pairs, where name is a char-
acter string matching a term in an od model, and value is the associ-
ated numeric starting value. If name contains special characters then
it must be quoted. The "|" operator identifies individual components
of direct product structures.

Value

A list object named by the arguments containing starting values for od.

od.options Settings for od()

Description

Allows various options that affect the behaviour of od to be set or examined.

Usage

od.options(...)
Arguments

... a series of comma separated name = value pairs, where name is a char-
acter string matching one of the options below, and value is the al-
lowable setting.

Details

If called with no arguments then a list of current option values is returned. The list
of options is held in an environment .ODenv in the od database.

Value

A list of current options and their values if called as od.options(), otherwise invisible().

Options used in OD

cool the cooling schedule for simulated annealing, recognised values are "lundy" (de-
fault), or "geometric".
cycles if search is anneal, this is the number of random interchanges to evaluate
in exploring the neighbourhood of a solution within a temperature cycle, de-
fault=1000.
alpha temperature reduction coefficient used in geometric cooling, default=0.7.
beta temperature reduction coefficient used in lundy cooling, default=0.7.
prob transition probability, default=0.95.
tstop temperature at which to terminate the search (simulated annealing), default =
1e-5. Optimization terminates when the temperature falls below tstop, or maxit
is exceeded.

Examples

## Get all options

od.options()

## Set the simulated annealing cooling schedule

od.options(cool="geometric")
 Bibliography

Atkinson, A. C. (1982). Developments in the design of experiments. International

Statistical Review 50 161–177.

Atkinson, A. C. & Donev, A. N. (1992). Optimum Experimental Designs. Oxford:

Clarendon Press.

Atkinson, A. C., Donev, A. N. & Tobias, R. D. (2007). Optimum Experimental

Designs, with SAS. Oxford University Press.

Battiti, R. & Tecchiolli, G. (1994). The reactive tabu search. ORSA Journal on
Computing 6 126–140.

Bueno Filho, J. S. S. & Gilmour, S. G. (2003). Planning incomplete block experi-

ments when treatments are genetically related. Biometrics 59 375–381.

Butler, D. G., Eccleston, J. A. & Cullis, B. R. (2008). On an approximate opti-

mality criterion for the design of field experiments under spatial dependence. Australian
and New Zealand Journal of Statistics 50 295–307.

Butler, D. G., Smith, A. B. & Cullis, B. R. (2012). On the design of experiments

where treatments are correlated. Biometrics submitted.

Chan, B. S. P. (1999). The design of field experiments when the data are spatially
correlated. Phd thesis, Department of Mathematics, University of Queensland.

Coombes, N. (2002). The reactive tabu search for efficient correlated experimental de-
signs. Phd thesis, Liverpool John Moores University.

Cox, D. R. & Reid, N. (2000). The Theory of the Design of Experiments. Boca Raton:
Chapman and Hall.

65
Cullis, B. R. & Gleeson, A. C. (1991). Spatial analysis of field experiments - an
extension to two dimensions. Biometrics 47 1449–1460.

Cullis, B. R., Gogel, B. J., Verbyla, A. P. & Thompson, R. (1998). Spatial

analysis of multi-environment early generation variety trials. Journal of the Royal
Statistical Society, Series B 39 1–38.

Cullis, B. R., Smith, A. B. & Coombes, N. E. (2006a). On the design of early

generation variety trials with correlated data. Journal of Agricultural, Biological, and
Environmental Statistics 11 381–393.

Cullis, B. R., Smith, A. B. & Coombes, N. E. (2006b). On the design of early

generation variety trials with correlated data. Journal of Agricultural, Biological and
Environmental Statistics 11 381–393.

Eccleston, J. A. & Whitaker, D. (1999). On the design of optimal change-over

experiments through multi-objective simulated annealing. Statistics and Computing 9
37–42.

Elliot, L. J., Eccleston, J. A. & Martin, R. J. (1999). An algorithm for the

design of factorial experiments when the data are correlated. Statistics and Computing
9 195–201.

Federer, W. T. & Raghavarao, D. (1975a). On augmented designs. Biometrics 31

29–35.

Fisher, R. A. (1935). The Design of Experiments. Edinburgh: Oliver and Boyd.

Gilmour, A. R., Cullis, B. R. & Verbyla, A. P. (1997). Accounting for natural

and extraneous variation in the analysis of field experiments. Journal of Agricultural,
Biological, and Environmental Statistics 2 269–273.

Glover, F. (1989). Tabu search - Part I. ORSA Journal on Computing 1 190–206.

Glover, F. (1990). Tabu search - Part II. ORSA Journal on Computing 2 4–32.

Goos, P. (2002). The optimal design of blocked and split-plot experiments. In P. Bickel,
P. Diggle, S. Fienberg, K. Krickeberg, I. Olkin, N. Wermuth & S. Zeger, eds., Lecture
Notes in Statistics, 164. Springer.

Henderson, C. R. (1976). A simple method for computing the inverse of a numerator

relationship matrix used in prediction of breeding values. Biometrics 32 69–83.

Kiefer, J. (1959). Optimum experimental designs. Journal of the Royal Statistical

Society, Series B 21 272–319.

Kiefer, J. (1974). General equivalence theory for optimum designs (approximate the-
ory). Annals of Statistics 2 849–879.
Kiefer, J. & Wolfowitz, J. (1960). The equivalence of two extremum problems.
Canadian Journal of Mathematics 12 363–366.

Kirkpatrick, S., Gelatt, C. D. & Vecchi, M. P. (1983). Optimisation by simulated

annealing. Science 220 671–680.

Lin, C. S. & Poushinsky, G. (1983). A modified augmented design for an early stage
of plant selection involving a large number of test lines without replication. Biometrics
39 553–561.

Lundy, M. (1985). Applications of the annealing algorithm to combinatorial problems

in statistics. Biometrika 72 191–198.

Martin, R. J. (1986). On the design of experiments under spatial correlation. Biometrika

73 247–277.

Martin, R. J. & Eccleston, J. A. (1997). Construction of optimal and near optimal

designs for dependent observations using simulated annealing. Research Report 479/97,
Department of Probability and Statistics.

Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H. &

Teller, E. (1953). Equations of state calculations by fast computing machine. Journal
of Chemical Physics 21 1087–1091.

Mrva, K. & Mares, D. J. (2001). Induction of late maturity α−amylase in wheat by

cool temperature. Australian Journal of Agricultural Research 52 477–484.

Osman, I. H. (1991). Metastrategy simulated annealing and tabu search algorithms for
combinatorial optimisation problems. Phd thesis, University of London, Imperial Col-
lege, UK.

Press, W. H., Teukolsky, S. A., Vetterling, W. T. & Flannery, B. P. (1996).

Numerical recipes in C. Cambridge, UK: Cambridge University Press.

Raghavarao, D. (1971). Constructions and combinatorial problems in design of exper-

iments. New York: John Wiley and Sons.

Robinson, G. K. (1991). That blup is a good thing: The estimation of random effects.
Statistical Science 6 15–51.

Silvey, S. D. (1980). Optimal Design. London: Chapman and Hall.

Smith, A. B., Cullis, B. R., Gilmour, A. R. & Thompson, R. (1998). Multiplica-

tive models for interaction in spatial mixed model analyses of multi-environment trial
data. In Proceedings of the International Biometrics Conference. Capetown.

Smith, A. B., Thompson, R., Butler, D. G. & Cullis, B. R. (2011). The analysis
of variety trials using mixtures of composite and individual plot samples. Journal of
the Royal Statistical Society, Series C 60 437–455.
Williams, E. R. & John, J. A. (1996). Row-column factorial designs for use in
agricultural field trials. Journal of the Royal Statistical Society. Series C (Applied
Statistics) 45 39–46.

Williams, E. R., John, J. A. & Whitaker, D. R. (2006). Construction of resolvable

spatial row-column designs. Biometrics 62 103–108.

Woodruff, D. L. & Zemel, E. (1993). Hashing vectors for tabu search. Annals of
Operations Research 41 123–137.

Wu, C. F. (1978). Some algorithmic aspects of the theory of optimal designs. Annals of
Statistics 6 1286–1301.

Yates, F. (1936a). A new method of arranging variety trials involving a large number
of varieties. Journal of Agricultural Science (Cambridge) 26 424–455.

Youden, W. J. & Connor, W. S. (1953). The chain block design. Biometrics 9

127–140.
Chapter

J
Whole Genome Analysis using
Linear Mixed Models

1 Introduction
1.1 Background & Computational history
Whole genome analysis is receiving wide attention in the statistical genetics community.
In the context of plant breeding experiments the focus is on quantitative trait loci (QTL)
which attempt to explain the link between a trait of interest and the underlying genetics
of the plant. Many approaches of QTL analysis are available such as marker regression
methods (Hayley & Knott, 1992; Martinez & Curnow, 1992) and interval mapping (Zeng,
1994; Whittaker et al., 1996). These methods are common place in QTL software and are
available for use in R packages such as Karl Bromans qtl package (Broman & Wu, 2010).
This particular suite of software is also complemented with a book (Broman & Sen, 2009)
which has been favourably reviewed (Zhou, 2010).
There has also been some focus on the use of numerical integration techniques for the
analysis of QTL. Xu (2003) and Zhang et al. (2008) suggest the use of Bayesian variable
shrinkage and utilise Markov chain Monte Carlo (MCMC) to perform the analysis. An
MCMC approach is also adopted in the R package qtlbim (Yandell et al., 2005). The
package builds on the qtl package and the Bayesian paradigm allows an extensible list of
trait types to be analysed. The package also makes use of the new model selection tech-
nique, the Deviance Information Criterion (Shriner & Yi, 2009), to aid in identifying the
correct QTL model. Similarly, a non-MCMC approach is adopted in the BayesQTLBIC
package (Ball, 2010) where the QTL analysis involves the use of the Bayesian Information
Criterion (Schwarz, 1978) as a QTL model selection tool.
Unfortunately many of the aforementioned methods and their software lack the ability
to account for complex extraneous variation usually associated with plant or animal based
QTL studies. Limited covariate additions are possible in R package qtlbim and through
the inventive on-line GridQTL software which uses the ideas of Seaton et al. (2002). Kang

69
et al. (2008) uses linear mixed models in the R package EMMA but it does not allow for
extraneous random effects and possible complex variance structures that may be needed
to capture environmental processes, such as spatial layouts, existing in the experiment.

1.2 WGAIM and software package

In this workshop we discuss the whole genome average interval mapping (WGAIM) ap-
proach of Verbyla et al. (2007) and its related software, the R package wgaim. The
package is available from the Comprehensive R Archive Network (CRAN) at http:
//CRAN.R-project.org/package=wgaim. This approach allows the simultaneous mod-
elling of genetic and non-genetic variation through extensions of the linear mixed model.
The extended model allows complex extraneous variation to be captured as well as si-
multaneously incorporating a whole genome analysis to detection and selection of QTL
using a linkage map. The underlying linear mixed modelling analysis is achieved com-
putationally using the R package ASReml-R. The simulation results given in Section 3
and in Verbyla et al. (2007) show that WGAIM is a powerful tool for QTL detection and
outperforms more rudimentary methods such as composite interval mapping. As it incor-
porates the whole genome into the analysis it eliminates the necessity for piecemeal model
fitting along the genome which in turn avoids the use of model selection criteria to control
the number of false positive QTL. In wgaim the false positives are controlled naturally
by assuming a background level of QTL variation through a single variance component
associated with a contiguous set of QTL across the whole genome. This parameter can
then be tested to determine the presence of QTL somewhere on the genome. As a result,
a less cumbersome approach to detecting and selecting QTL is ensured.

1.3 Software Prerequisites

The WGAIM method uses an extension of interval mapping to perform its analysis. For
convenience and flexibility, the wgaim package provides the ability to convert genetic data
objects created in the qtl package to objects for further use in wgaim. The converted
objects retain a similar structure to objects created in qtl and therefore can still be
used with functions within the package. Users of wgaim need to be aware that it is
a software package intended for the analysis and summary of QTL and currently only
contains minimal tools for exploratory linkage map manipulation. Much of the exploratory
work can be handled with functions supplied in the qtl package and users should consult
its documentation if required. In addition, the interval mapping approach of Verbyla et al.
(2007) and its implementation in wgaim is also restricted to populations with only two
distinct genotypes. Some of these populations include, double haploid (DH), back-crosses
and recombinant inbred lines (RIL). To ensure this rule is adhered to, error trapping has
been placed in the appropriate functions of wgaim.
Throughout the WGAIM procedure the underlying linear mixed model analysis is
achieved using the highly flexible R software package ASReml-R, built as a front end
wrapper for the more sophisticated stand alone version, ASReml (Gilmour et al., 2009).
This software allows the user the ability to flexibly model spatial or environmental vari-
ation as well as possible variation that may arise from additional components associated
with the experimental design. It uses an average information algorithm developed in
Gilmour et al. (1995) that allows efficient computing of residual maximum likelihood
(REML) (Patterson & Thompson, 1971) estimates for the variance parameters. The use
of REML estimation in the linear mixed model context becomes increasingly necessary in
situations where the data is unbalanced. Much of its sophistication has been influenced
from its common use in the analysis of crop variety trials (Smith et al., 2001, 2005, 2006)
where complex additional components such as spatial correlation structures or multiplica-
tive factor analytic models need to be incorporated into the mixed model. If available, the
software also allows complex pedigree information to be included (Oakey et al., 2006).
Many of these additional flexibilities in ASReml have also established it as a valuable
software tool in the livestock industries. In more recent years it has been used as a core
engine for more complex genetic analyses as in Gilmour (2007), Verbyla et al. (2007) and
Huang & George (2009). If you are affiliated with an academic institution, the stand
alone software and the R package ASReml-R Discovery is now freely available through
http://www.vsni.co.uk.
2 WGAIM theoretical method
The WGAIM approach is a forward selection method that uses a whole genome approach
to genetic analysis at each step. Following Verbyla et al. (2007), initially a working
model is developed that assumes a QTL in every interval. Thus for a given set of trait
observations y = (y1 , . . . , yn ) consider the model

y = Xτ + Z e ue + Z g g + e, (1)

where τ is a t length vector of fixed effects with an associated n × t explanatory design

matrix X and ue is a b × 1 length vector of random effects with an associated n × b design
matrix Z e . Typically, the distribution of ue ∼ N (0, σ 2 G(ϕ)) and is assumed mutually
independent to the residual vector e ∼ N (0, σ 2 R(φ)) with ϕ and φ being vectors of
variance ratios.
The vector g in (1) represents a r length vector of genotypic random effects with its
associated design matrix Z g . Let m be the total number of markers, c be the number
of chromosomes, mk the number of markers on chromosome k, (k = 1, . . . , c), and qi,k:j
represent the parental allele type for line i in interval j on chromosome k. In WGAIM,
qi,k:j = ±1, reflecting two possible genotypes AA, BB for DH and RIL and AB, BB for
back-cross populations. The ith genetic component of this model is then given by
c m
X Xk −1

gi = qi,k:j ak:j + pi ,
k=1 j=1

where ak:j is QTL effect size assumed to have distribution ak:j ∼ N (0, σ 2 γa ) and pi ∼
N (0, σ 2 γp ) represents a polygenic or residual genetic effect not captured by the QTL
effects.
As in interval mapping the vector of QTL allele types are replaced by the expectation
of the QTL genotype given the flanking markers. Let mk:j be the jth marker on the kth
chromosome then the vector of genotypic effects is
c m
X X k −1

g= (mk:j λk:j,j + mk:j,j+1 λk:j+1,j )ak:j + p

k=1 j=1

= M Λa + p, (2)

where λk:j,j and λk:j+1,j are complicated expressions based on recombination fractions
between the marker and the QTL in the jth interval (see equation (5) and (6) on page
100 of Verbyla et al. (2007)). These parameters require estimation. Verbyla et al. (2007)
suggest applying a parameter reduction technique to produces a vector of genotypic effects
of the form
c m
X Xk −1

g= (mk:j + mk:j,j+1 )λk:j ak:j + p

k=1 j=1

= M ΛE a + p, (3)
where λk:j = θk:j,j+1 /2dk:j,j+1 (1 − θk:j,j+1 ) and θk:j,j+1 , dk:j,j+1 are the the known recom-
bination fraction and Haldane’s genetic distance between marker j and j + 1 respectively
on the kth chromosome. Let M E = M ΛE then M E is an r × (m − c) fully specified
known matrix of pseudo-markers spanning the whole genome. A more detailed overview
of this decomposition and its derivation can be found in Verbyla et al. (2007). The full
working statistical model for analysis is then

y = Xτ + Z e ue + Z g M E a + Z g p + e. (4)

After the fitting of (4) the simple hypothesis H0 : γa = 0 is tested based on the
statistic −2 log Ψ = −2(log L − log L0 ) where L and L0 is the residual likelihood of the
working model (4) with and without the random regression QTL effects, Z a a. Stram &
Lee (1994) suggest that under H0 , −2 log Ψ is distributed as the mixture 21 (χ20 + χ21 ) due
to the necessity of testing whether the variance ratio is on the boundary on the parameter
space.
If γa is found to be significant a putative QTL is determined using an outlier detection
method based on the alternative outlier model (AOM) for linear mixed models from Gogel
(1997) and formalised in Gogel et al. (2001). Verbyla et al. (2007) uses the AOM to develop
a score statistic for each of the chromosomes. For example, for the kth chromosome let
ak0 = ak + δ k where δ k is a vector of random effects such that δ k ∼ N (0, σ 2 γa,k I mk −1 ).
The full outlier model is

y = Xτ + Z e ue + Z g M E a + Z g,k M E,k δ k + Z g p + e, (5)

where Z g,k is the matrix Z g appropriately subsetted to chromosome k. The REML score
is then derived for γa,k and evaluated at γa,k = 0, namely
 
1 1 T
Uk (0) = − tr(C k,k ) − 2 2 ãk ãk , (6)
2 σ γa
where C k,k = Z g,k M E P M E Z g,k with P = H −1 − H −1 X(X T H −1 X)−1 X T H −1 , H =
σ 2 (R+ZGZ T +γa Z g M E M TE Z Ta +γp Z p Z Tp and best linear unbiased predictors (BLUPS)
ãk = γa M TE Z Tg,k P y. This score has mean zero and this will occur exactly when the terms
in the parentheses of (6) are equal. Scores that depart from zero suggest a departure from
γa,k = 0. A simple statistic that reflects this departure can be based on the “outlier”
statistic
Pmk −1 2
2 ãTk ãk j=1 ãk:j
tk = 2 2 = Pmk −1 .
σ γa tr(C k,k ) j=1 var(ã k:j )
This statistic can therefore be calculated from the BLUPS of the QTL sizes and their
prediction error variances arising from the working model. In most cases mixed model
software, including ASReml-R used in wgaim, provide the ability to extract these com-
ponents for this use.
In a similar manner to the above once the chromosome with the largest outlier statistic
is identified, the individual intervals within that chromosome are checked. For example
if the largest t2k is from the kth chromosome, a similar derivation can be followed for the
outlier statistic of the jth interval, namely

ã2k:j
t2k:j = . (7)
var(ãk:j )

A putative QTL is then determined by choosing the largest t2k:j within that chromosome.
It must be stated at this point that although (5) is formulated to derive the theory for
QTL outlier detection there is no requirement to fit this model as the chromosome and
interval outlier statistics only contain components obtainable from a fit of the working
model proposed in (4). Thus there is only a minimal computational cost to determine an
appropriate QTL interval using this method.
Once a QTL interval is selected it is moved into the fixed effects of the working model
(4) and the process is repeated until γa is not significant. After the selection process is
complete the selected QTL intervals appear as fixed effects and the final model is

y = Xτ + Z a,s M E,s as + Z e ue + Z g p + e, (8)

where Z a,s and M E,s contain the the appropriate columns of Z g and M E for the selected
QTL. This complete approach is known as the WGAIM algorithm.

2.1 Marker vs Interval

The WGAIM method derived in the previous section uses a whole genome extension
of interval mapping. The matrix Λ in (2) can be viewed as a mapping matrix that
appropriately maps the marker scores to midpoint pseudo-interval scores. In fact, the
genetic model proposed in (2) can be written as an approximate marker QTL regression
model

g = M aM + p (9)

where the marker QTL sizes are aM = Λa. This suggests the marker QTL sizes have an
assumed distribution of the form aM ∼ N (0, σ 2 γa ΛΛT ) and are correlated. Therefore an
analysis assuming the genetic QTL model (9) with independent marker QTL effects will
be less efficient than the interval mapping equivalent (2).

2.2 Recent advances: WGAIM v1.0+

WGAIM is always being developed to improve its efficiency and stability as well as advance
its capabilities. Recent research in Verbyla & Taylor (2011) has shown the WGAIM
method can be improved through a slight alteration is the selection process as well as
using a random effects formulation for the selected QTL. Verbyla & Taylor (2011) also
show that its possible to efficiently include high dimensional genetic components in the
linear mixed model formulation. These points are discussed below.
2.2.1 The outlier statistics

There is a short relevant point in Verbyla & Taylor (2011) concerning the use of the outlier
statistics in the WGAIM algorithm. After much scrutiny it was found that the use of
the chromosome statistic was flawed for small linkage groups. Consider the scenario of
two markers on a linkage group k. After converting the marker information to a single
interval the chromosome and interval outlier statistics have the property t2k = t2k:1 .
Thus, the chromosome statistic, in this instance, is based on the information contained
in one interval. This interval statistic, in some circumstances, may bias the choice toward
chromosome k and the selection of its only interval. It was discovered after many simula-
tions that a better choice would be to only use the interval outlier statistic to guide the
selection process.

2.2.2 A random effects formulation

It is well known that there is (selection) bias involved in moving the selected QTL to
the fixed effects (Beavis, 1994, 1998). Xu (2003) provides a theoretical justification while
Melchinger et al. (1998) also conclude that sizes are inflated. There is a parallel in general
plant breeding analysis where genetic effects are assumed to be random rather than fixed.
This reduces the bias through shrinkage and provides a more realistic estimate of the size
of a genetic effect. Reducing the bias in QTL analysis would be desirable.
2
In a random effects formulation we assume as ∼ N (0, σas I) so the selected QTL effects
are assumed to be random and have a different variance to the unselected effects. This
makes sense as selected QTL effects exhibit variation from zero because they are QTL.
Thus random version of WGAIM mimics a mixture distribution, in which QTL come from
one distribution and non-QTL come from another distribution. The two distributions
differ in their variances and not their means.

2.2.3 High dimensional analysis

A major issue concerns the possible high-dimension of Mλ in (4) at any stage of the
selection. For ease of presentation we consider the first step of selection. The matrix Mλ
is of size r × m − c, and m − c may be very large, larger than r. If r > m − c there is
no need to invoke dimension reduction. If r < m − c, it is possible to carry out efficient
computations not in dimension m − c but in dimension r which may be much smaller,
and subsequently to recover the m − c effects necessary for selection.
For generality, suppose thata ∼ N (0, σa2 Ga ). In WGAIM, Ga = I m−c . Taylor &
Verbyla (2011b) consider an alternative approach based on penalized likelihoods and in
that case at stage k of the iterative process involved, Ga = W k where W k is a diagonal
matrix of known weights. In these cases, Ga is a known matrix and the results of this
section will hold for any method in which this is the case or can be constructed to be so.
Note that while WGAIM involved pseudo-markers for intervals, these can be replaced by
marker scores.
 Thus suppose for the moment that no QTL have been selected (s = 0 in the discussion
above). The variance model for the random effects M E a is of the form

σa2 M E Ga M TE (10)

which is of size r × r. The matrix M E Ga M TE is known. Then we can re-formulate the

working mixed model (4) as follows. Let

Z λ = (M E Ga M TE )1/2

denote the matrix square root; this can be found using the singular value decomposition
(Golub & van Loan, 1996). If we define an r × 1 vector a∗ as

a∗ ∼ N (0, σa2 I r )

the working model term for interval effects Z g M E a can be replaced

Z g Z λ a∗

and the same variance model (10) is obtained. Thus we can estimate r effects rather than
m − c. Most importantly for QTL analysis however, the m − c effects can be recovered.
Note firstly that the BLUP for a∗ is given by

ã∗ = σa2 (M E Ga M TE )1/2 Z Tg P y

The BLUP of a is given by

ã = σa2 Ga M TE Z Tg P y
so that we can write
ã = Ga M TE (M E Ga M TE )−1/2 ã∗ (11)
thereby recovering the estimated size of potential QTL in each interval.
To calculate the outlier statistic given by (7), we need the variance of the best linear
unbiased predictor (BLUP) of a. If ã denotes the BLUP of a, the variance matrix is

var(ã) = Ga M TE (M E Ga M TE )−1/2 (ã∗ )(M E Ga M TE )−1/2 ME Ga (12)

and so the variance matrix of the BLUPs of a∗ is required, again a lower dimensional
calculation after fitting the reduced dimensional model. The diagonal elements of (ã) can
be used together with (11) to calculate the outlier statistics (7).

2.2.4 Links to genomic prediction

There are important consequences for genomic prediction and selection (Meuwissen et al.,
2001) in the development of the high dimensional theory derived above.
Note, the BLUP for a∗ can be rewritten as

ã∗ = M E ã

so that for genomic prediction, we need only form ã∗ . The individual BLUPs ã obtained
through back transformation are available but not required.
2.3 Final assessment of significance of QTL
Once all the QTL are selected they are summarized differently depending on the formu-
lation.

2.3.1 Fixed effects formulation

The summary of the selected QTL in this formulation can proceed using a Wald statistic.
Let âjk be the fixed effect estimate of the QTL ajk with variance var(âjk ) = σP2 EV,jk . The
wald statistic for this QTL is then given by
ãjk
zjk =
σP EV,jk
Assuming Z ∼ N (0, 1) then a p-value is calculated in the usual manner using
ãjk

Pr Z < − |âjk < 0
σP EV,jk
and
ãjk

1 − Pr Z < − |âjk > 0
σP EV,jk

2.3.2 Random effects formulation

In this formulation, the size of the QTL effect is a best linear unbiased prediction (BLUP).
It is no longer appropriate to test the hypothesis that the effect is zero in order to assess
its significance. Tests of hypotheses pertain to unknown parameters, and random effects
involve distributions of effects.
To provide a measure of the strength of a QTL, the conditional distribution of the
true (random) QTL effect ajk say, given the data is used. That is under the normality
assumptions for a linear mixed model,

ajk |y 2 ∼ N (ãjk , σP2 EV,jk )

where y 2 is the component of the data free of fixed effects (Verbyla, 1990). The mean
of this conditional distribution is the BLUP of ajk , that is the estimated size of the
QTL ãjk , and the variance is the prediction error variance (PEV) of ajk . Thus the proper
assessment of the impact of the QTL involves determining how far the distribution is from
zero. Clearly, distributions that are far from zero indicate a strong QTL effect whereas
distributions which include zero will indicate a weaker QTL. This can be quantified by
calculating a probability somewhat like a p-value, but for which values close to 1 indicate
the QTL is strong. Thus for estimated effects that are positive, the probability that the
true (random) QTL effect is greater than zero is calculated, while for estimated effects
that are negative the probability that the true (random) QTL effect is less than zero is
calculated.
Thus,  
ãjk
Pr(ajk < 0|y 2 ) = Pr Z < −
σP EV,jk
and  
ãjk
Pr(ajk > 0|y 2 ) = 1 − Pr Z < −
σP EV,jk
These probabilities can be found approximately by substituting the estimated σP EV,jk
from the analysis. While LOD scores are not appropriate for WGAIM, the statistic zjk
can be used to calculate a LOD score if this is deemed necessary.

2.4 Current research

Currently only QTL analysis of univariate traits are possible with the accompanying
software, wgaim. However, several extension are being developed and itemized here.

• A multivariate version of the WGAIM algorithm has been researched and prelimi-
nary papers have been submitted (Verbyla & Cullis, 2010; Verbyla et al., 2010). The
software implementation of this multivariate approach is currently being tested.
• An implementation of WGAIM for populations with multiple parents is also being
researched. This has shown to be useful in determining QTL for emerging complex
breeding populations such as the Multi-parent Advanced Genetic Inter-Cross wheat
population being developed in CSIRO, Plant Industry.
• Lastly, research is also currently being conducted to determine the inclusion of
higher order effects such as epistatic interactions into the WGAIM approach.

We are hopeful that all these new approaches will be implemented in future releases
of wgaim.

2.5 Summary
To summarise the WGAIM algorithm provides an efficient method for whole genome
interval or marker QTL analysis. As the algorithm is developed through the extension
of a flexible linear mixed model it provides an appropriate mechanism for more difficult
analysis of experiments that usually occur in plant or animal breeding trials.
3 Simulations
In this section the performance of the fixed and random WGAIM methodologies is exam-
ined through simulation studies. Much of the information contained here can be found in
more detail in Verbyla et al. (2007) and Verbyla & Taylor (2011). The simulation studies
were constructed in a similar manner to the multiple chromosome QTL simulation studies
in Broman & Speed (2002).

3.1 Details
The number of genetic lines and hence the population size considered is l = 100, 200 and
400. These values were chosen to reflect the size of population common in crop trials.
In all simulations, 2 replicates of each line were generated. This differs from Broman &
Speed (2002) who consider a single replicate for each line. However, our simulations are
constructed to ensure that our results are comparable to those of Broman & Speed (2002).
The simple model for the simulated yij , i = 1, . . . , l; j = 1, 2 is given by

yij = µ + gi + eij

where eij ∼ N (0, σ 2 ) are independent random variables. The genetic line effects gi are
given by
XS
gi = qis as + pi
s=1

where there are S QTL, and qis for QTL s is either −1 or 1 depending on the parental
allele that line i carries. The polygenic effects pi ∼ N (0, σg2 ) are independent and also
independent of eij . In all simulations, σ 2 = 1, σg2 = 0.5 and as = 0.378.
Genetic lines are assumed to have a total of 9 chromosomes. The underlying chromo-
somal structures are 100cM in length, with 11 equally spaced markers. Broman & Speed
(2002) locate the QTL at markers, but in our study, the QTL are located at midpoints
of intervals in configurations to be discussed below. This is the more likely situation in
practice. The genetic lines are simulated on the basis of this underlying structure, and
hence each line in the simulation has scores on the 11 markers for each chromosome and
on the QTL.
For each population size (l = 100, 200, 400), a linkage map was constructed on the
basis of the simulated marker scores, and also including the simulated QTL scores to
accurately place the QTL on the estimated linkage map. This was required to confirm
the detection or non-detection of a QTL. Note that while the underlying structure has
chromosomes of length 100cM and equally spaced markers, the estimated linkage map for
a chromosome varied in length from 80 to 120cM, with marker spacing varying from 4cM
to 25cM. The qtl package in R was used to develop the linkage map and then to check
the order of the markers and QTL. The final linkage map was used for all simulations for
a particular population size. Thus map uncertainty was included in the simulation study.
Table 1: Rate of correct identification of each QTL and the total mean number out of 7.

Interval
Size Method C1.4 C1.8 C2.4 C2.8 C3.6 C4.4 C5.1 Total
100 WGAIM 0.470 0.445 0.468 0.722 0.679 0.603 0.564 3.951
CIM7 0.132 0.056 0.208 0.223 0.335 0.241 0.181 1.376
200 WGAIM 0.940 0.952 0.623 0.733 0.790 0.787 0.773 5.598
CIM7 0.706 0.810 0.234 0.330 0.584 0.555 0.614 3.833
400 WGAIM 0.961 0.968 0.937 0.918 0.980 0.994 0.997 6.755
CIM7 0.900 0.946 0.919 0.896 0.990 0.989 0.998 6.638

In this simulation S = 7 QTL are strategically placed on the linkage map. The
locations of the QTL are denoted by Ck.j where Ck is chromosome k and j is interval j
on that chromosome. QTL are located at the midpoints of the intervals C1.4, C1.8, C2.4,
C2.8, C3.6, C4.4, C5.1, unlike Broman & Speed (2002) who assumed that each QTL was
located at the left-hand marker of each interval. The favorable QTL allele is coded as 1 for
all but the second QTL C1.8, for which it is −1. Thus the two QTL on chromosome 1 are
in repulsion while the two QTL on chromosome 2 are in coupling. The proportion of the
total line mean variance due to the 7 QTL was 7×0.3782 /(7×0.3782 +1) = 1/(1+1) = 0.50,
again matching Broman & Speed (2002)). There are additional terms in this calculation
for the QTL in coupling and repulsion due to linkage, but these additional terms cancel.

Table 2: Mean number of linked extraneous QTL on chromosomes 1 to 5 and unlinked

extraneous QTL on chromosomes 6 to 9.

Linked Unlinked
Size Method C1 C2 C3 C4 C5 Total C6-C9
100 WGAIM 0.227 0.203 0.150 0.100 0.069 0.749 0.199
CIM7 0.038 0.092 0.040 0.023 0.005 0.198 0.005
200 WGAIM 0.098 0.145 0.105 0.144 0.049 0.541 0.090
CIM7 0.043 0.258 0.044 0.105 0.018 0.468 0.001
400 WGAIM 0.032 0.029 0.029 0.009 0.008 0.107 0.025
CIM7 0.038 0.025 0.009 0.009 0.001 0.082 0.003

3.2 Fixed WGAIM vs CIM

In this section the methods of QTL analysis chosen for comparison are fixed WGAIM,
and CIM with 7 co-factors (CIM7), the latter chosen by forward selection. Broman &
Speed (2002) found that CIM7 performed well in their simulation study, and as the correct
number of cofactors are included our comparison is again with the best CIM approach.
Direct comparison between WGAIM and CIM is difficult and thorough details of how this
Table 3: Rate of identification of the QTL in coupling and repulsion. Identification is
presented by two by two tables, with, for example, C2.4 on the left and C2.8 on the top
of each two by two table. D denotes correctly detected or identified and D̄ means not
identified.

Repulsion Coupling
WGAIM CIM7 WGAIM CIM7
Pop. size D̄ D D̄ D D̄ D D̄ D
100 D̄ 0.418 0.112 0.830 0.038 0.057 0.475 0.574 0.218
D 0.137 0.333 0.114 0.018 0.221 0.247 0.203 0.005
200 D̄ 0.017 0.043 0.095 0.199 0.060 0.317 0.468 0.298
D 0.031 0.909 0.095 0.611 0.207 0.416 0.202 0.032
400 D̄ 0.015 0.024 0.018 0.082 0.006 0.057 0.007 0.074
D 0.017 0.944 0.036 0.864 0.076 0.861 0.097 0.822

is achieved for this simulation can be found in Verbyla et al. (2007). For each population
size 2000 simulations were carried out.
The correct detection rates for individual QTL are presented in Table 1. CIM7 per-
forms quite badly for population size 100. This is particularly true for the QTL in
coupling. The performance of both methods improves as the population size increases.
WGAIM again has the best overall detection rates.
There are 7 QTL in the simulations. The total number of correctly identified QTL
for the 3 methods and 3 population sizes are also given in Table 1. The results found
in our study with regard to CIM are consistent with those found by Broman & Speed
(2002). All three methods improve their power as the population size increases. The mean
number of correctly identified QTL using WGAIM is always higher than for CIM7, often
considerably so. This shows that WGAIM is much more powerful than CIM at finding
true QTL.

Table 4: Proportion of the 200 simulations in which the QTL was detected for WGAIM
and random WGAIM (RWGAIM) analyses.

Population Method C1.4 C1.8 C2.4 C2.8 C3.6 C4.4 C5.1 Total
100 WGAIM 0.350 0.465 0.675 0.425 0.700 0.590 0.695 3.900
RWGAIM 0.355 0.455 0.685 0.450 0.715 0.615 0.710 3.985
200 WGAIM 0.940 0.915 0.650 0.595 0.760 0.825 0.800 5.485
RWGAIM 0.945 0.905 0.685 0.615 0.770 0.820 0.810 5.550
400 WGAIM 0.955 0.935 0.900 0.900 0.985 1.000 0.995 6.670
RWGAIM 0.955 0.940 0.915 0.915 0.985 1.000 0.995 6.705

The rates of detection of extraneous QTL are presented in Table 2. The mean numbers
of linked extraneous QTL are presented for each chromosome, together with an overall
total. The mean number of unlinked extraneous QTL are presented as a total for chro-
Table 5: Two way tables for the QTL in repulsion (C1.4 and C1.8) with proportions of
the 200 simulations for each population size for the combinations of non-detected D̄ and
detected D QTL. C1.4 is on the left and C1.8 on the top of each 2×2 table.

Population size
100 200 400
Type Method D̄ D D̄ D D̄ D
Coupling WGAIM D̄ 0.445 0.205 0.025 0.035 0.040 0.005
D 0.090 0.260 0.060 0.880 0.025 0.930
RWGAIM D̄ 0.440 0.205 0.030 0.030 0.040 0.005
D 0.105 0.250 0.070 0.875 0.020 0.935
Repulsion WGAIM D̄ 0.085 0.240 0.055 0.295 0.025 0.075
D 0.490 0.185 0.350 0.300 0.075 0.825
RWGAIM D̄ 0.075 0.240 0.055 0.260 0.020 0.065
D 0.475 0.210 0.330 0.355 0.065 0.850

mosomes 6 to 9. The rates for linked extraneous QTL show that for population size 100,
where correct detection is difficult, WGAIM has higher rates than CIM7. For l = 200,
WGAIM has higher rates, except for chromosome 2 where the two QTL are in coupling.
Here CIM7 has much higher rates. The total rates are very similar for both methods. For
a population size l = 400, the rates for both methods are very good and are very similar.
The detection of unlinked extraneous QTL is always higher for WGAIM, but decreases
as the population size increases. Thus the increased power of detecting true QTL that
WGAIM affords, is accompanied by an increase (albeit small) in detection of false QTL.
Table 3 gives the rates for correct identification of QTL in repulsion and coupling
presented as two-way tables. For the two QTL in repulsion, CIM7 is very poor for
population sizes of 100 and 200. There is considerable improvement for a populations
size of 400. In contrast, WGAIM gradually improves detection of both QTL, and has
much better rates of detection than CIM across the board, with a marked improvement
in performance with a population size of 200. The patterns are similar for two QTL in
coupling. Again CIM7 performs poorly for population sizes of 100 and 200, improving
considerably for a size of 400. WGAIM gradually improves detection of both QTL, and
has much better rates of detection than CIM7 across the board. This study highlights
the difficulty in detecting QTL in coupling for small and moderate populations sizes, even
with a more powerful method, namely WGAIM.

3.3 Random WGAIM vs Fixed WGAIM

In this section the original fixed WGAIM method is compared to the newer random
WGAIM (RWGAIM) method. For each population size 200 simulations were carried out.
Table 4 presents the proportion of simulations in which each QTL was successfully
found using standard WGAIM and RWGAIM. The results are very similar with RWGAIM
Table 6: Proportion of the 200 simulations in which false QTL were detected for the 4
methods of analysis. Both linked (selected QTL are on chromosomes with QTL) and
unlinked (selected QTL are on chromosomes without QTL) are presented.

Population size
100 200 400
Method Linked Unlinked Linked Unlinked Linked Unlinked
WGAIM 0.770 0.275 0.445 0.125 0.115 0.030
RWGAIM 0.675 0.255 0.390 0.105 0.105 0.025

finding marginally more QTL for all population sizes. The effect of using random QTL
effects is minimal on the good performance of the general WGAIM approach.
Two-way tables for the proportions of QTL detected for the two QTL in coupling
and repulsion are given in Table 5. Results for WGAIM and RWGAIM are very similar.
Although the detection rates compared to repulsion were lower, RWGAIM shows some
improvement over WGAIM in detecting both QTL for all population sizes.

Table 7: Mean estimated size of QTL effects with empirical standard deviations for each
method across the 200 simulations for each population size.

Population Size
Method Interval 100 200 400
WGAIM C1.4 0.503 (0.127) 0.430 (0.131) 0.391 (0.073)
C1.8 -0.478 (0.110) -0.439 (0.144) -0.377 (0.109)
C2.4 0.576 (0.146) 0.471 (0.144) 0.381 (0.085)
C2.8 0.585 (0.188) 0.466 (0.112) 0.390 (0.083)
C3.6 0.501 (0.144) 0.382 (0.110) 0.395 (0.051)
C4.4 0.480 (0.122) 0.384 (0.108) 0.369 (0.047)
C5.1 0.461 (0.112) 0.377 (0.070) 0.412 (0.058)
RWGAIM C1.4 0.459 (0.112) 0.403 (0.078) 0.374 (0.068)
C1.8 -0.428 (0.108) -0.436 (0.085) -0.377 (0.065)
C2.4 0.537 (0.136) 0.431 (0.125) 0.369 (0.074)
C2.8 0.548 (0.137) 0.439 (0.105) 0.380 (0.078)
C3.6 0.470 (0.093) 0.374 (0.072) 0.388 (0.050)
C4.4 0.441 (0.093) 0.370 (0.070) 0.362 (0.046)
C5.1 0.432 (0.096) 0.356 (0.062) 0.405 (0.056)

Table 6 gives the proportion of false positives for the 2 methods. False positives can
be linked (on the same chromosome as true QTL, C1-C5) or unlinked (on chromosomes
where no QTL exist, C6-C9). The rates of false positives are the number found across
all 200 simulations divided by 200. RWGAIM shows reduced rates of false positives over
WGAIM. Population size again has a major impact.
One aspect of WGAIM that was not examined by Verbyla et al. (2007) was bias in the
estimated QTL sizes. This was examined in the current simulation study and the results
are given in Table 7. WGAIM shows higher bias (all true QTL sizes are ±0.378) than
RWGAIM. Thus RWGAIM reduces the bias particularly for the smaller population sizes,
though bias still persists. The bias at population size 200 is smaller for RWGAIM while
both method exhibit little bias at population size 400. Lastly, the standard deviations
are generally smaller for RWGAIM.
4 The R package wgaim: A casual walk through
A typical QTL analysis with wgaim can be viewed as series of steps with the appropriate
functions

1. Fit a base asreml() model

Fit a base asreml() (see the ASReml-R package) model as in (4) but without the added
marker/interval genetic information term Z g M E a. The asreml() call allows very com-
plex structures for the variance matrices G(ϕ) and R(φ) through its random and rcov
arguments. This makes it an ideal modelling tool for capturing non-genetic experimental
variation, such as design components and/or extraneous environmental variation.
For a comprehensive overview of the ASReml-R package, including thorough ex-
amples of its flexibility, users should, in the first instance, consult the documentation
that is included with the package. Note: On any operating system that has
ASReml-R installed, the documentation can be found using the simple com-
mand asreml.man() in R.

2. Read in genetic data using read.cross()

Read in genetic data using read.cross() (see the qtl package). This function allows the
reading in of genetic information in a number of formats including files generated from
commonly used genetic software programs such as Mapmaker and QTL Cartographer. At
the current printing of this document read.cross() accepts data in the following formats
(from the help for read.cross()),

• comma-delimited (“csv”)
• rotated comma-delimited (“csvr”)
• comma-delimited with separate files for genotype and phenotype data (“csvs”)
• rotated comma-delimited with separate files for genotype and phenotype data (“csvsr”)
• Mapmaker (“mm”)
• Map Manager QTX (“qtx”)
• Gary Churchill’s format (“gary”)
• Karl Broman’s format (“karl”).

For the exact requirements of all available file types and their nomenclature users should
consult the qtl documentation. The read.cross() function can also process more ad-
vanced genetic crosses. However, in wgaim the QTL analysis is restricted to populations
with two genotypes. Thus users should be aware that the class of the returned object from
read.cross() needs to have the structure c("bc","cross"). The "bc" is an abbreviated
form for “back-cross”. It is this class structure that is checked in the proceeding steps.
 read.cross() will also estimate map distances if they are not given in the genetic
file(s) before importation. It uses the Lander & Green (1987) hidden Markov model for
its estimation. This is an EM algorithm and therefore suffers from linear convergence.
On some occasions the algorithm may not converge or slowly reach convergence over an
extended length of time. In these instances users may need to patient.

3. Convert genetic "cross" object to an "interval" object

This can be done using the wgaim function

cross2int(fullgeno, missgeno = "MartinezCurnow", rem.mark = TRUE, id =

"id", subset = NULL)

The function contains a number of arguments that provide some linkage map manipulation
before calculation of the interval information for each chromosome. They are detailed as
follows,

1 Sub-setting: The map can be subsetted by giving the subset argument a character
string vector of chromosome names.
2 Co-locating markers: If rem.mark = TRUE then co-locating markers across the
genome are removed from the marker set. The correlated markers and how they are
connected is returned as part of the final object.
3 Missing values: If missgeno = "MartinezCurnow", missing values within a chro-
mosome are imputed using the rules of Martinez & Curnow (1992). If missgeno =
"Broman" the they are calculated using the default values of argmax.geno() in the
qtl package

Note: This step is crucial in the process of QTL analysis using wgaim. The
imputation of the missing markers ensures the genetic data being passed into
wgaim.asreml() in the proceeding step is a complete (i.e. no missing values)
genetic data set across all linkage groups.
After the linkage map manipulation, for each chromosome, the imputed marker data
matrix is returned as an element of the object. Along with this, several interval calcula-
tions are returned such as distances between markers, recombination fractions and most
importantly, the interval data matrix.
The final id argument is required to determine the unique rows of the genotypic data
and is passed to the imputed marker data and the interval data matrix. The final genetic
data object returned also retains the c("bc","cross") class for backward compatibil-
ity with other functions in the qtl package as well as inherits the class "interval" for
functionality within the wgaim package.
4. Perform QTL analysis with wgaim()
wgaim(baseModel, phenoData, intervalObj, merge.by = NULL, gen.type = "interval",
method = "random", TypeI = 0.05, attempts = 5, trace = TRUE,
verboseLev = 0, ...)

The baseModel argument must be an asreml.object and therefore have "asreml" as

its class attribute. Thus a call to wgaim() is actually a call to wgaim.asreml(). This
stipulation ensures that an asreml() call has been used to form the base model in step 2
before attempting QTL analysis. An error trapping function, wgaim.default() is called
if the class of the base model is not "asreml". The second argument phenoData is a data
object of phenotypic data usually used in the analysis of the base model in step 2. The
intervalObj contains the imputed genetic marker and interval data obtained from a call
to cross2int() in the previous step. Thus intervalObj must be of class "interval".
The gen.type allows the user to specify "interval" or "marker" depending on the
desired analysis. If gen.type = "marker" then the imputed marker matrix for each
linkage group in intervalObj is combined into a whole genome matrix before being
merged with phenoData. If gen.type = "interval" then the interval matrix for each
linkage groups is combined and used instead.
The character string merge.by is then used to identify the appropriate column of
phenoData and the newly formed whole genome genetic object with which to merge
the two data sets. This merging differs depending on the whether the problem is high
dimensional (r×m−c) or not. Note: Unmatched elements of merge.by are handled
differently depending on whether they are from the intervalObj or phenoData.
If elements of merge.by exist in phenoData and are unmatched with elements
in intervalObj then they are kept to ensure completeness of the phenotypic
data. If elements of merge.by exist in intervalObj and not in phenoData they are
dropped as there will be no phenotypic information available for that genetic
line.
TypeI argument allows users to change the significance level for the testing of QTL
effects variance component γa . As asreml() calls output components of the fit to the
screen there is an option to trace this to a file if desired. The level of reporting can be
changed using verboseLev.

The method argument in wgaim()

This argument is probably one of the most important arguments in the wgaim() call. In
the current version two choices are available.
If method = "fixed" the wgaim.asreml() code uses the legacy algorithm that was
available in pre-1.0 versions of wgaim. This version is the original algorithm discussed
in Section 2 and in more detail in Verbyla et al. (2007). Users need to be aware that
this algorithm uses the chromosome outlier statistic to hone its selection of a QTL at
each iteration. It is now known (as discussed in Section 2.2.1) that using the chromosome
statistic is flawed when there are small linkage groups. The problem dissipates as the
linkage group becomes larger. This method also places the selected QTL in the fixed
effects of the linear mixed model as the algorithm proceeds (hence the "fixed" argument).
The final set of QTL is then tested using an appropriate wald test as discussed in section
2.3.1
If method = "random" the wgaim.asreml code uses the updated algorithm and is
available with versions 1.0+ of wgaim. In this version the selected QTL are placed in the
random part of the linear mixed model as the algorithm proceeds as discussed in Verbyla
& Taylor (2011) and Section 2.2. This is known to reduce bias of the selected QTL effects.
This algorithm also only uses an interval outlier statistic to guide the selection of each
QTL and therefore is much more appropriate if the genetic map being used contains small
linkage groups.
Note: Both of these methods allow high dimensional genetic components
to be added to the wgaim call through intervalObj.

5. Summarise QTL with various method functions

summary(object, intervalObj, LOD = TRUE, ...)
print(x, intervalObj, ...)
tr(object, iter = 1:length(object$QTL$effects), diag.out = TRUE, ...)
link.map(object, intervalObj, chr, max.dist, marker.names = "markers",
list.col = list(q.col = "light blue", m.col = "red", t.col =
"light blue"), list.cex = list(t.cex = 0.6, m.cex = 0.6),
trait.labels = NULL, tick = FALSE, ...)

Various function can be used to summarize and diagnostically check the QTL obtained
from a wgaim() analysis. The summary() function assesses the significance of the QTL
effects (fixed or random) and retrieves genetic marker information. For an "interval"
analysis, chromosome, interval and distance of the left and right flanking markers are
outputted. For a "marker" analysis, chromosome, distance of the nearest marker are dis-
played. tr() displays diagnostic information of the forward selection process underlying
a wgaim() analysis. It shows a summary of the Residual Maximum Log-Likelihood ratio
tests of significance for the parameter γa at each iteration. There is also a triangular
p-value matrix that shows the significance of the QTL at each iteration.
Selected QTL can also be placed on a linkage map using link.map(). This function
neatly plots the linkage map and places "interval" or "marker" QTL at their appropriate
position. The function has some added flexibility for colouring of QTL regions as well as
colour and size of printed text for all components of the map.
5 Examples
5.1 A quick example: The Zinc Data
The zinc data is available in wgaim as a usable date set to display the functionality of
the package. The data consists of 200 observations of zinc concentration and shoot length
for a DH population of wheat. There are two replicates of 90 double haploid lines from a
crossing of the wheat varieties Cascades and Rac875-2 and ten each of the parents in an
id variable. The data also includes a Type variable to distinguish the parents from the
DH lines. The experiment also contained two blocks in a variable called Block.
A suitable base model for shoot length is explored by considering (4) without the
random regression effects, Z g M E a, attributed to genetic markers/intervals. As we are
interested in the genetic variance associated with the DH lines, Type is modelled as a
fixed effect (τ ) to ensure the removal of the genetic effect associated with the parents.
The Block is a random effect of non-interest (ue ) and id as a set of polygenic random
effects p.

R> data("zinc", package = "wgaim")

R> sh.fm <- asreml(shoot ~ Type, random = ~Block + id, data = zinc)

A simple summary of the variance parameters in the model can be achieved with

R> summary(sh.fm)$varcomp

gamma component std.error z.ratio constraint

Block!Block.var 0.1902258 0.03721561 0.05539823 0.6717835 Positive
id!id.var 9.1030840 1.78091965 0.28195227 6.3163871 Positive
R!variance 1.0000000 0.19563915 0.02674723 7.3143694 Positive

The summary reveals there is only a small difference between Blocks. However, the genetic
variance is more than nine times the residual variance of the model making the response
an ideal candidate for QTL analysis.
wgaim is prepackaged with a genetic map associated with the zinc data. The data has
already been read in using read.cross() and can be loaded using

R> data("raccas", package = "wgaim")

Alternatively to illustrate the use of read.cross() in conjunction with this package the
same data is available from the extdata directory of the package library as a CSV file.
A subset of the data from the CSV file is given in Table 8. This reveals that the CSV file
is in the rotated CSV format (see read.cross() from the qtl package). The genotypes
are set as AA or AB and missing values are "-". Thus a call to read.cross() is
 V1 V2 V3 V4 V5 V6 V7 V8 V9 V10
1 id DH01 DH02 DH03 DH04 DH05 DH06 DH07
2 wmc469 1A1 0.00 AB AA AB - AB AB AB
3 wPt.5914 1A1 7.26 AB AA AB AB AA AB AB
4 wPt.0751 1A1 8.88 - AA AB AB AA AB AB
5 P42.M49.235 1A1 20.11 AA AA AB AB AA AB AB
6 bcd304C 1A2 0.00 AA AB AB AA AA AA AA
7 P31.M55.148 1A2 31.52 AA - AB AA AA AA AA
8 barc213 1A2 42.60 AA AB AB AA AA AA AA
9 P34.M48.83 1A2 43.84 AA AB AB AA AA AA AA
10 gwm99 1A2 54.30 AA AA AA AA AA AA AA
Table 8: A subset of the genetic data from the comma delimited file raccas.csv.

R> wgpath <- system.file("extdata", package = "wgaim")

R> raccas <- read.cross("csvr", file="raccas.csv", genotypes=c("AA","AB"),
+ dir = wgpath, na.strings = c("-", "NA"))
R> class(raccas)

[1] "bc" "cross"

The returned object has the required class structure.

Looking inside the "cross" object you will see the following

R> names(raccas$geno)

[1] "1A1" "1A2" "1B" "1D1" "1D2" "2A1" "2A2" "2A3" "2B" "2D1" "2D2" "3A1"
[13] "3A2" "3A3" "3A4" "3D1" "3D2" "4A1" "4A2" "4A3" "4B" "4D1" "4D2" "5A1"
[25] "5A2" "5A3" "5B" "5D1" "5D2" "6A1" "6A2" "6B1" "6B2" "6D1" "6D2" "6D3"
[37] "7A1" "7A2" "7A3" "7B" "7D1" "7D2"

The genetic marker information is a named list format with the appropriate name for
each linkage group. Looking deeper into the genetic object we see

R> names(raccas$geno$"1B")

[1] "data" "map"

For each linkage group, "data" contains to the actual marker data matrix, converted into
R/qtl format (AA = 1, AB = 2, missing values = NA). Marker names are placed as the
column names and the genotype lines (DH001, DH002, ...) are the row names.

R> raccas$geno$"1B"$data[1:8,1:8]
 gwm18 gwm11 P40.M54.358 bcd338 P37.M48.217 wPt.7030 wPt.3457 gwm413
DH01 1 2 2 2 2 2 2 2
DH02 1 1 1 1 1 1 1 1
DH03 1 2 2 2 2 2 2 2
DH04 2 1 1 1 1 NA 1 1
DH05 1 1 1 1 1 1 1 1
DH06 1 1 1 1 1 NA 1 1
DH07 1 1 1 1 1 NA 1 1
DH08 1 1 2 2 2 2 2 2

The "map" element contains the map distances that were estimated using the Lander-
Green hidden markov algorithm (Lander & Green, 1987) during the read.cross() pro-
cess.

R> raccas$geno$"1B"$map

gwm18 gwm11 P40.M54.358 bcd338 P37.M48.217 wPt.7030

0.00000 26.48862 54.73602 57.35473 65.13091 74.86437
wPt.3457 gwm413 P34.M58.126 barc8 wPt.2838 Clone117500
76.34429 83.66295 92.58413 93.91230 96.80020 98.29318
wmc406 gwm374 P37.M48.136 P34.M50.124 P40.M56.75 P39.M47.204
100.09763 110.71761 112.82703 115.99925 121.75534 167.12669
bcd808A P31.M60.286 gwm403 P39.M50.180 wPt.1770 wPt.6802
174.27171 175.41956 194.79081 238.51156 244.04986 245.39910
P44.M54.74 P39.M50.114 wPt.4290 ksuI27B cdo393B gwm140
268.40321 275.31194 281.08452 297.39544 303.74771 308.84706

The first marker of the linkage group is always set to zero.

We can now convert the "cross" object to an "interval" object using

R> raccas <- cross2int(raccas, missgeno = "Mart", id = "id", rem.mark = TRUE)

R> summary(raccas)
R> class(raccas)

[1] "bc" "cross" "interval"

After removing coincident markers, the summary shows there is total of 458 markers
across 40 linkage groups. It also reveals that just over 5% of markers were missing and
imputed using the rules of Martinez & Curnow (1992). The classes of raccas and their
ordering is retained and it now also inherits the class "interval" for use with functions
in wgaim.
For a specific linkage group in the "interval" object, there are now additional com-
ponents

R> names(raccas$geno$"1A1")
[1] "data" "map" "dist" "theta" "imputed.data"
[6] "intval"

Thus for each linkage group, "data" and "map" are as before. "dist" contains the interval
distances and "theta" are the recombination fractions between adjacent markers based
on "dist". "imputed.data" contains the marker data with all missing values imputed

R> raccas$geno$"1A1"$imputed.data[1:8,1:8]

gwm18 gwm11 P40.M54.358 bcd338 P37.M48.217 wPt.7030 wPt.3457 gwm413

DH01 1 -1 -1 -1 -1 -1.0000000 -1 -1
DH02 1 1 1 1 1 1.0000000 1 1
DH03 1 -1 -1 -1 -1 -1.0000000 -1 -1
DH04 -1 1 1 1 1 0.9971324 1 1
DH05 1 1 1 1 1 1.0000000 1 1
DH06 1 1 1 1 1 0.9971324 1 1
DH07 1 1 1 1 1 0.9971324 1 1
DH08 1 1 -1 -1 -1 -1.0000000 -1 -1

and "intval" contains the interval data based on the midpoint pseudo-interval calculation
of Verbyla et al. (2007)

R> raccas$geno$"1A1"$intval[1:6,1:6]

gwm11 P40.M54.358 bcd338 P37.M48.217 wPt.7030 wPt.3457

DH01 0.0000000 -0.9742251 -0.9997715 -0.9979892 -0.9968539 -0.9999270
DH02 0.9772501 0.9742251 0.9997715 0.9979892 0.9968539 0.9999270
DH03 0.0000000 -0.9742251 -0.9997715 -0.9979892 -0.9968539 -0.9999270
DH04 0.0000000 0.9742251 0.9997715 0.9979892 0.9954246 0.9984933
DH05 0.9772501 0.9742251 0.9997715 0.9979892 0.9968539 0.9999270
DH06 0.9772501 0.9742251 0.9997715 0.9979892 0.9954246 0.9984933

The genetic object is now ready to be used in wgaim QTL analysis. For this analysis
we will use the calculated genetic intervals or "intval" components of each linkage group

R> zn.qtlI <- wgaim(sh.fm, phenoData = zinc, intervalObj = raccas,

+ merge.by = "id", na.method.X = "include", gen.type = "interval",
+ method = "fixed")
R> summary(zn.qtlI, raccasM, LOD = FALSE)

Chromosome Left Marker dist(cM) Right Marker dist(cM) Size z.ratio Pr(z)
1 3D2 gdm8 31.51 gdm136 32.64 0.436 4.13 0
2 4B Rht1mut 54.8 gwm6 70.12 0.54 4.86 0
3 4D1 barc098 0 P42.M49.70 1.13 0.422 3.63 3e-04
4 4D2 wPt.2573 0 Rht2W.type 23.48 -0.383 -2.82 0.0048
The analysis reveals four significant QTL in four linkage groups. Verbyla et al. (2007)
recommends the use of p-values, rather than the commonly used LOD scores, as the
overall test of significance for each of the QTL. The argument LOD = TRUE can be given
to summary.wgaim() if LOD scores are necessary.
As the gen.type used in this analysis was "interval", the summary displays the left
and right marker information for the interval. We can do a similar analysis using the
markers by changing the argument for gen.type.

R> zn.qtlM <- wgaim(sh.fm, phenoData = zinc, intervalObj = raccas,

+ merge.by = "id", na.method.X = "include", gen.type = "markers",
+ method = "fixed")
R> summary(zn.qtlM, raccasM)

Chromosome Marker dist(cM) Size z.ratio Pr(z) LOD

1 3D2 gdm8 31.51 0.450 4.928 0.000 5.273
2 4B Rht1mut 54.8 0.454 5.027 0.000 5.487
3 4D1 barc098 0 -1.194 -2.728 0.006 1.616
4 4D1 P42.M49.70 1.13 1.554 3.618 0.000 2.842
5 4D2 Rht2W.type 23.48 -0.470 -4.188 0.000 3.808
6 7D1 P42.M54.113 0 -0.332 -3.571 0.000 2.770

Now the summary information only contains the markers that were selected during the
analysis. The argument LOD = TRUE argument is the default to summary.wgaim().

5.2 Example 2: Sunco-Tasman data

This example stresses the importance of modelling extraneous variation to a ensure a more
appropriate QTL analysis. The Sunco-Tasman data is available in the data directory of
wgaim and contains the results of a field trial conducted in the year 2000 with 175 double
haploid lines from a crossing of wheat varieties Sunco and Tasman. The original field
trial was arranged in a 31 rows by 12 columns with two replicates of each line. A milling
experiment was then performed which replicated 23% of the field samples producing 456
samples milled over 38 mill days with 12 samples per day. The focus is on the trait milling
yield.

R> stpheno <- asreml.read.table(paste(wgpath,"\\stpheno.csv", sep =""),

+ header = TRUE, sep = ",")
R> names(stpheno)

[1] "X" "Expt" "Type" "id" "Range" "Row"

[7] "Rep" "Millday" "Milldate" "Millord" "myield" "lord"
[13] "lrow"

Smith et al. (2006) provides a phenotypic analysis of the data. They give a base model
of the form
R> st.fmF <- asreml(myield ~ Type + lord + lrow, random = ~ id + Rep +
+ Range:Row + Millday, rcov = ~ Millday:ar1(Millord),
+ data = stpheno, na.method.X = "include")
R> summary(st.fmF)$varcomp

gamma component std.error z.ratio constraint

id 7.0925458 1.92573995 0.23965934 8.0353220 Positive
Rep 0.2843737 0.07721201 0.15604795 0.4947967 Positive
Range:Row 1.4973306 0.40654927 0.06206771 6.5500926 Positive
Millday 1.7795039 0.48316385 0.15646257 3.0880476 Positive
R!variance 1.0000000 0.27151604 0.08035809 3.3788264 Positive
R!Millord.cor 0.7109431 0.71094307 0.12682697 5.6056142 Unconstrained

The model realistically accounts for extraneous plot variation occurring in the field as
well as variation due to the design components of the milling experiment. The lord and
lrow components of the fixed model are mean centred covariates of Millord and Row that
capture the natural linear trends that occur in the samples across milling order on any
given day and across rows in the field. The summary reveals a large genetic variance
component. For comparison a NULL model (no extraneous effects) is also fitted.

R> st.fmN <- asreml(myield ~ 1, random = ~ id, data = stpheno,

+ na.method.X = "include")

The genetic map consists of 287 unique markers across 21 chromosomes and can be read
in and converted using

R> stmap <- read.cross("csv", file="stgenomap.csv", genotypes=c("A","B"),

+ dir = wgpath, na.strings = c("-", "NA"))
R> stmap <- cross2int(stmap, missgeno="Bro", id = "id")
R> names(stmap$geno)

[1] "1A" "1B" "1D" "2A" "2B" "2D" "3A" "3B" "3D" "4A" "4B" "4D" "5A" "5B"
[15] "5D" "6A" "6B" "6D" "7A" "7B" "7D"

It is possible to view the genetic map using link.map(). The function allows sub-setting
according to distance (cM) and/or chromosome. Figure 1 shows the genetic map resulting
from the first of these two commands

R> link.map(stmap, cex = 0.5, marker.names = "dist")

R> link.map(stmap, cex = 0.5, marker.names = "markers")

For larger maps a more aesthetic plot is reached by adjusting the character expansion
(cex) parameter and increasing the plotting window width manually. You can also subset
the map by distance and chromosome if you wish

R> link.map(stmap, cex = 0.5, marker.names = "markers",

+ chr = names(nmar(stmap)[1:12]), max.dist = 200)
 Genetic Map

0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
8.57 9.01 1.71 3.39 0.85 8.13 2.37 2.14 10.32 11.98 1.84 8.62 2.26 4.62 3.56
10.69 17.26 5.12 4.54 5.22 5.92 11.58 4.52 7.73 5.07 13.67 16.51
19.6 11.39 15.81 18.87 15.39
23.63 11.02 5.72 6.94 7.87 11.06 11.93 4.96 8.24 7.21 26.11
25.23 20.22 18.69 29.98 30.92 15.93
13.71 36.63 7.02 33.35 13.37 26.79 19.39 13.17 19.96 11.78 12.26 8.92 34.85 29.48
29.48 19.26 20.17 31.51 14.16 40.14 9.45 19.95
43 16.59 21.34 42.1 22.18 32.41
37.45 29.67 51.16 21.7 44.64 52.05 50.08 15.85 50.35 10.51 43.51 34.99
22.32

50
58.5 40.92 54.83 29.95 54.99 50.62 18.1 56.72 12.07 30.63
54.76 51.04 23.71 53.39 50.04 35.69 36.63
66.49 65.32 42.84 63.64 35.14 61.22 63.07 19.21 59.9 21.58 37.19
68.93 59.59 70.29 63.56 56.08 24.29 51.23 38.46
73.07 58.29 60.75 70.1 40.29 69.53 19.76 25.68 40.21
73.14 73.51 76.48 65.43 33.09 76.49 40.3
74.37 74.07 67.62 61.28 73.18 70.69 20.29 42.12 49.68
79.55 80.59 70.08 35.52 52.6 86.55
74.9 76.93 72.04 61.78 91.25 77.77 20.82 49.18
80.71 81.14 41.42 63.81 93.99
81.06 78.67 96.78 95.92 99.56 101.21 99.07 90.58 21.88
100.71 84.7 52.37 77.36 98.23
83.21 22 106.56

100
107.77 83.19 110.32 105.55 95.11
85.79 101.22 102.08 24.11 107.74 87.22
119 102.58 119.05 111.36 117.88 91.06
90.88 113.65 27.95 108.33
122.52 103.71 41.03 112.12 97.52
116.93 133.7 132.12 135.39
106.03 98.42 139.05
108.15 123.91
111.07 150.21 133.68 152.24

150
145.56 161.48 154.66 158.71 137.5 161.39
157.12 157.55 167.19 138.01
173.71 175.27 160.21 149.07
162.47 179.17
178.31 160.97 183.5 150.5
161.92
165.05 201.28

200
172.53

Location (cM)
239.51 235.91
247.32 237.09
237.62
256.69

250
238.62
242.98

286.65
297.82
302.49

300
305.29
306.43
308.58
310.48
317.55

350
1A 1B 1D 2A 2B 2D 3A 3B 3D 4A 4B 4D 5A 5B 5D 6A 6B 6D 7A 7B 7D

Chromosome

Figure 1: The genetic map for the Sunco-Tasman data. Names of chromosomes are given at the bottom and genetic distances between
markers are placed alongside each of the chromosomes.
5.2.1 QTL analysis and diagnostics

A QTL analysis is now performed for the full model st.fmF and the null model st.fmN.
This time we pipe the non-essential output to a text file using a file name for the argument
trace.

st.qtlN <- wgaim(st.fmN, phenoData = stpheno, intervalObj = stmap,

+ merge.by = "id", gen.type = "interval", method = "fixed",
+ na.method.X = "include", trace = "nullmodel.txt", verboseLev = 1)
st.qtlF <- wgaim(st.fmF, phenoData = stpheno, intervalObj = stmap,
+ merge.by = "id", gen.type = "interval", method = "fixed",
+ na.method.X = "include", trace = "fullmodel.txt", verboseLev = 1)

The process of selecting QTL is determined from the outlier statistics calculations in
Section 2. These are saved for each QTL selection and can be viewed using the out.stat()
command. For the first two iterations of the process the chromosome and interval outliers
statistics given in Figure 2 are produced with

R> out.stat(st.qtlF, stmerge, int = FALSE, iter = 1:2, cex = 0.6,

+ ylim = c(0, 6.5))
R> out.stat(st.qtlF, stmerge, int = TRUE, iter = 1:2, cex = 0.6)

For each iteration the interval outlier statistics are plotted across the whole genome ac-
cording to their distance and separated by differentiable colours according to their chro-
mosome. There is also an additional argument that allows the user to subset the genetic
map to specific chromosomes which is only available when int = TRUE.

R> out.stat(st.qtlF, stmerge, int = TRUE, iter = 1:5, cex = 0.6,

+ chr = c("2B","6B","5A","7D"))

From a statistical standpoint the QTL selected across the genome cannot be expected
to be orthogonal. Thus the introduction of the next QTL in the forward selection process
will inevitably affect the significance of the previously selected QTL. A post diagnostic
evaluation of the QTL p-values in the forward selection process can be displayed using

R> tr(st.qtlF, iter = 1:10, digits = 3)

Incremental QTL P-value Matrix.

===============================
2B.5 7D.2 4D.1 4B.1 1B.13 6B.5 5A.13 1B.1 4A.2 3D.5
Iter.1 <0.001
Iter.2 <0.001 <0.001
Iter.3 <0.001 <0.001 0.001
Iter.4 <0.001 <0.001 0.001 <0.001
Iter.5 <0.001 <0.001 <0.001 <0.001 0.001
Iter.6 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001
 Iteration: 2
7D.2
6
5
4
3
2
1

Iteration: 1
6 2B.5

outlier statistic
4
3
2
1

1A 1B 1D 2A 2B 2D 3A 3B 3D 4A 4B 4D 5A 5B 5D 6A 6B 6D 7A 7B 7D

Chromosome

Iteration: 2
15

10
7D.2

0
Iteration: 1
2B.5
15

outlier statistic
10

0
.1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1.1 .1 .1 .1 .1
1A 1B 1D 2A 2B 2D 3A 3B 3D 4A 4B 4D 5A 5B 5D 6A6B 6D 7A 7B 7D

Interval

Figure 2: Chromosome and interval outlier statistics for the first two iterations of the wgaim fit for the full model.
Iter.7 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001
Iter.8 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.021
Iter.9 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.008 0.009
Iter.10 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 0.005 0.008 0.013

Outlier Detection Diagnostic.

=============================
L0 L1 Statistic Pvalue
Iter.1 -309.563 -250.669 117.787 <0.001
Iter.2 -279.483 -243.213 72.541 <0.001
Iter.3 -272.049 -240.762 62.574 <0.001
Iter.4 -269.746 -238.879 61.734 <0.001
Iter.5 -256.599 -235.993 41.211 <0.001
Iter.6 -251.322 -232.332 37.98 <0.001
Iter.7 -236.172 -225.886 20.572 <0.001
Iter.8 -225.186 -221.5 7.372 0.003
Iter.9 -225.029 -221.577 6.904 0.004
Iter.10 -223.37 -221.047 4.647 0.016
Iter.11 -223.008 -220.78 4.456 0.017
Iter.12 -221.138 -220.029 2.219 0.068

The first of these displays shows the p-values of the selected QTL for the first ten iterations
occurring in the WGAIM process. An example of the dynamic changes in significance can
be seen for the selected QTL interval 4B.1. The second display presents the likelihood
ratio tests, −2 log Λ, for the significance of the QTL variance parameter, γa , in (4), with
the inclusion of the last hypothesis test where the null model is retained.

5.2.2 Visualising your QTL results

Full summaries are available throught the usual summary.wgaim() command

R> summary(st.qtlF, stmap, LOD = FALSE)

Chromosome Left Marker dist(cM) Right Marker dist(cM) Size z.ratio Pr(z)
1 1B gwm550 0 P36.M67.1 9.01 0.182 2.905 0.004
2 1B gwm11 90.88 gwm140 239.51 -0.804 -5.920 0.000
3 2A wmc198 29.67 wmc170 40.92 -0.201 -3.117 0.002
4 2B wmc474USQ 54.76 wmc35a 59.59 0.828 13.283 0.000
5 3D TeloPAGG2 54.99 TeloPAGG1 61.22 -0.162 -2.670 0.008
6 4A germin 10.32 cdo795 11.39 0.162 2.700 0.007
7 4B barc193 0 csME1 11.98 -0.447 -7.104 0.000
8 4D Rht2.mut 0 csME2 1.84 0.318 5.353 0.000
9 5A PAACTelo2 95.11 P46.M37.4 102.08 -0.348 -5.697 0.000
10 6B cdo507 8.92 barc354 9.45 -0.455 -7.748 0.000
11 7D gwm437 86.55 wmc94 93.99 0.283 4.592 0.000
 Genetic Map with QTLs

0
0 0 0 0 0 0 Full 0 0 0 0
Full Full
9.01 3.39 2.14 Null 1.84 8.62 Full 3.56
Full 10.32 Null 11.98
10.69 5.12 5.92 11.58 Null 5.07
11.39 15.81 18.87
11.02 5.72 11.06 7.21
11.93 18.69
13.71 7.02 29.98 30.92 8.92
33.35 13.17 19.96
Full 19.26 9.45
16.59 21.34 31.51
29.67 43 10.51
44.64 22.32
40.92 51.16 50.08 12.07
51.04 52.05

50
42.84 Full 54.99 23.71 21.58
54.76 Full 56.08 53.39 50.62
58.29 Null 61.22 24.29 25.68
65.32 59.59 63.07
67.62 63.56 33.09 42.12
73.07 60.75 69.53
72.04 65.43 35.52 49.18
74.37 61.28 70.69
70.08 41.42
74.9 61.78 77.77 86.55
91.25 52.37 Full
81.06 90.58 93.99
96.78 95.92 Full Null
83.21 101.21 99.07 95.11 98.23
Null 102.08

100
85.79 105.55 106.56
90.88 107.74
119 108.33
122.52 112.12
133.7 135.39

150.21

150
154.66 158.71

Location (cM)
161.48 157.55
Full 167.19
160.21
175.27
160.97 179.17
161.92 183.5
165.05
172.53 201.28

200
239.51
247.32

250
256.69

1B 2A 2B 3D 4A 4B 4D 5A 6B 7D

Chromosome

Figure 3: Genetic map with QTL for the Full and Null models obtained from an analysis of the Sunco-Tasman data. Markers and intervals
for the QTL are highlighted and trait names are placed on the left hand side of the chromosomes.
The summary actually produces a data.frame of results that can be easily exported to
a spreadsheet program if desired. For multiple tables a simple table binding function
is provided which stacks the QTL tables making it instantly useful for exporting with
programs such as the the LaTeX table package xtable.

R> qtlTable(st.qtlF, st.qtlN, intervalObj = stmap, labels = c("Full",

+ "Null"), columns = 1:7)

The full and the NULL QTL models can be summarised visually using link.map().
In this case it calls the method link.map.wgaim() to plot the QTL on the genetic map.

R> link.map(st.qtlF, stmap, marker.names = "dist", cex = 0.6,

+ trait.labels = "Full")

Multiple models or traits can be handled through link.map.default(). For example,

Figure 3 is produced with

R> link.map.default(list(st.qtlF, st.qtlN), stmap, marker.names = "dist",

trait.labels = c("Full", "Null"))

The multiple QTL map reveals that an extra six QTL were detected in the full model
compared to the null model, highlighting the importance of modelling extraneous variation
appropriately in QTL analyses.
The QTL plotting procedures link.map.wgaim() and link.map.default are highly
customisable. Through an argument list.col it allows the user to specify the QTL
colour between markers, the colour of the flanking QTL marker names, the colour of the
trait names and the rest of the marker names. If no colours are chosen q.col and t.col
defaults to rainbow(n) where n is the number of traits. You can also change the size of
the marker and trait name text with the argument list.cex.
Some customized examples are given below for the Full and Null QTL models for the
Sunco-Tasman data and can be seen in Figure 4. These have been produced using the
following criteria: Changing the colour of the QTL regions and the names and setting the
background marker text grey.

R> link.map.default(list(st.qtlF, st.qtlN), stmap, marker.names = "dist",

+ trait.labels = c("Full", "Null"), list.col = list(q.col = c("turquoise",
+ "salmon"), m.col = "red", t.col = c("turquoise", "salmon")), col = "gray")

A monochromatic plot with increased sizes for the trait labels.

R> link.map.default(list(st.qtlF, st.qtlN), stmap, marker.names = "dist",

+ trait.labels = c("Full", "Null"), list.col = list(q.col = rep(gray(0.8), 2),
+ m.col = "black", t.col = "black"), list.cex = list(t.cex = 0.8), col = "gray")
 Genetic Map with QTL

0
Full 0 0 0 0 0 Full 0 Full 0 0 0 0
Full Full
9.01 3.39 2.14 10.32 Null 11.98 Null 1.84 8.62 3.56
Null
10.69 Full 5.12 33.35 5.92 11.39 15.81 18.87 11.58 5.07
11.02 5.72 43 11.06 11.93 18.69 29.98 30.92 7.21

50
13.71 7.02 Full Full 54.99 13.17 19.96 52.05 31.51 8.92
51.16
65.32 19.26 Null 61.22 16.59 21.34 53.39 50.08 9.45
54.76
73.07 29.67 59.59 63.56 44.64 22.32 50.62 10.51 Full 86.55
74.37 40.92 65.43 51.04 23.71 Full 63.07 12.07 93.99
60.75 Null
74.9 42.84 70.08 56.08 24.29 Null 69.53 21.58 98.23
61.28
81.06 58.29 61.78 101.21 91.25 33.09 70.69 25.68
135.39
83.21 67.62 95.92 105.55 99.07 35.52 77.77 42.12
85.79 72.04 119 161.48 150.21 41.42 90.58 49.18
Full 167.19
90.88 96.78 122.52 175.27 154.66 52.37 95.11 106.56
157.55 179.17 102.08 107.74

Location (cM)
160.21 133.7 108.33
160.97 158.71 112.12
161.92 183.5
239.51 165.05 201.28
247.32 172.53

250 200 150 100

256.69

1B 2A 2B 3D 4A 4B 4D 5A 6B 7D

Chromosome

Genetic Map with QTL

0
Full 0 0 0 0 0 Full 0 Full 0 0 0 0
9.01 3.39 2.14 Full 10.32 11.98 1.84 8.62 Full 3.56
Null Null
10.69 5.12 33.35 5.92 11.39 15.81 18.87 11.58 Null 5.07
11.02
Full 5.72 11.06 11.93 18.69 29.98 30.92 7.21
43

50
13.71 7.02 Full 51.16 Full 54.99 13.17 19.96 52.05 31.51 8.92
65.32 19.26 Null 54.76 61.22 16.59 21.34 53.39 50.08 9.45
73.07 29.67 59.59 63.56 44.64 22.32 50.62 10.51 Full 86.55
74.37 40.92 60.75 65.43 51.04 23.71 Full 63.07 12.07 93.99
74.9 42.84 70.08 56.08 24.29 69.53 21.58 Null 98.23
61.28 Null
81.06 58.29 61.78 101.21 91.25 33.09 70.69 25.68
135.39
83.21 67.62 95.92 105.55 99.07 35.52 77.77 42.12
85.79 72.04 119 161.48 150.21 41.42 90.58 49.18
Full 90.88 96.78 154.66 52.37 167.19 95.11 106.56
122.52 175.27
157.55 179.17 102.08 107.74

Location (cM)
160.21 133.7 108.33
160.97 158.71 112.12
161.92 183.5
239.51 165.05 201.28
247.32 172.53

250 200 150 100

256.69

1B 2A 2B 3D 4A 4B 4D 5A 6B 7D

Chromosome

Figure 4: Customized genetic maps with QTL for the full and Null models obtained from an analysis of the Sunco-Tasman data.
5.2.3 Marker analysis

Similar to the previous example a whole genome marker QTL analysis of the Sunco-
Tasman data can be performed by changing the gen.type argument.

R> st.qtlFR <- wgaim(st.fmF, phenoData = stpheno, intervalObj = stmap,

+ merge.by = "id", gen.type = "marker", method = "random",
+ na.method.X = "include", trace = "fullmodel.txt", verboseLev = 1)

The wgaim model can be summarised in the usual way

R> summary(st.qtlFR, stmap, LOD = TRUE)

Chromosome Marker dist(cM) Size Prob LOD

1 1B cdo473 85.79 -0.263 1.000 4.070
2 1B ksuI27a 247.32 -0.233 1.000 3.295
3 2A wmc170 40.92 -0.198 0.999 2.346
4 2B wmc474USQ 54.76 0.363 0.998 1.749
5 2B wmc35a 59.59 0.408 0.999 2.207
6 4B csME1 11.98 -0.423 1.000 10.600
7 4D Rht2.mut 0 0.285 1.000 4.991
8 5A PAACTelo2 95.11 -0.332 1.000 6.608
9 6B barc354 9.45 -0.438 1.000 11.553
10 7D wmc94 93.99 0.280 1.000 4.561

As discussed previously, for the random case the p-values are replaced by probabilties
and for whole genome marker regression the summary contains only the closest marker
that influences the trait. It is interesting to note that adjacent markers are chosen on
2B whereas the interval analysis found one large QTL in the interval between these two
markers.
The random method only calculates interval outlier statistics and for this analysis
places the outlier statistics at the marker position. The outlier statistics for the first first
five iterations can be seen in Figure 5

R> out.stat(st.qtlFR, stmap, int = TRUE, iter = 1:5, cex = 0.6)

Fitting the Null model in a similar manner.

R> st.qtlNR <- wgaim(st.fmN, phenoData = stpheno, intervalObj = stmap,

+ merge.by = "id", gen.type = "marker", method = "random",
+ na.method.X = "include", trace = "nullmodel.txt", verboseLev = 1)

Similar to the interval analysis, the results from the Full model and the Null model can
be plotted on the linkage map and is given in Figure 6. The QTL are now highlighted
with plotting symbols that can be altered with the usual arguments, pch and cex.
 Iteration: 5
15

10 4D.1

0
Iteration: 4
15

10
4B.2

0
Iteration: 3
15

10 1B.12

outlier statistic
0
Iteration: 2
15
6B.6
10

0
Iteration: 1
15 2B.5

10

0
.1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1 .1
1A 1B 1D 2A 2B 2D 3A 3B 3D 4A 4B 4D 5A 5B 5D 6A6B 6D 7A 7B 7D

marker

Figure 5: Outlier statistic plots for the first five iterations of the full marker regression model using the random WGAIM method
R> link.map.default(list(st.qtlFR, st.qtlNR), stmap, marker.names = "dist",
+ trait.labels = c("Full", "Null"), list.col = list(q.col = c("red",
+ "light blue"), m.col = "red", t.col = c("red", "light blue")),
+ list.cex = list(t.cex = 0.9, m.cex = 0.7), col = "black", cex = 2, pch = 16)

5.3 A quick high dimensional example

This section illustrates the accelerated performance of the WGAIM approach when the
genetic component required for analsyis is high dimensional. A high dimensional genetic
marker data set can be created from an existing marker map using the simulation function,
sim.geno() in the R/qtl package. In this example we use the Sunco-Tasman linkage map.
Firstly, the Sunco-Tasman data is re-read in using read.cross()

R> stmap <- read.cross("csv", file="stgenomap.csv", genotypes=c("A","B"),

+ dir = wgpath, na.strings = c("-", "NA"))
sum(nmar(stmap))

The original data has 287 markers across 21 chromosomes. An extended map can be
created with

R> tempmap <- sim.geno(stmap, n.draws = 1, step = 3)

The function uses the same hidden markov model that estimated the original distances
to infill the chromosomes with new simulated markers at distances specified by the user
(step = 3). By default sim.geno() returns the extended map in a separate list element
of the genetic object. Some trickery is needed to convert the object into a useable form

R> newmap$geno <- lapply(tempmap$geno, function(el){

+ el$data <- el$draws
+ el$data <- matrix(el$data, nrow = nrow(el$data))
+ el$map <- attr(el$draws, "map")
+ el$draws <- NULL
+ el})
R> nmar(newmap)

1A 1B 1D 2A 2B 2D 3A 3B 3D 4A 4B 4D 5A 5B 5D 6A 6B 6D 7A 7B
44 101 95 45 52 81 69 54 71 76 30 67 85 30 52 9 53 23 136 66
7D
50

This produces a total of 1289 markers spanning th 21 chromosomes. This can now be
converted to an ”interval” object and analysed using wgaim.

R> newmap <- cross2int(newmap, rem.mark = FALSE)

R> system.time(st.qtlHD <- wgaim(st.fmF, phenoData = stpheno, intervalObj = newmap,
 Genetic Map with QTL

0
0 0 0 Null ● 0 Full ●● 0 0 0 0
9.01 3.39 Null 1.84 8.62 Full 3.56
10.69 5.12 Full ● 11.98
11.58
●● 5.07
5.72 15.81 18.87 Null 7.21
11.02
7.02 18.69 8.92
13.71 29.98 30.92
19.26 33.35 19.96 9.45
21.34 31.51
Full ● 29.67 43 10.51
40.92 22.32 12.07
51.16 50.08

50
Full 23.71 52.05
42.84 ●● 54.76 50.62 21.58
● 24.29 53.39
65.32 58.29 Null 59.59 63.07 25.68
67.62 33.09 42.12
73.07 Full 60.75 69.53
72.04 35.52 49.18
74.37 61.28 70.69
Null ● 41.42
● 74.9 61.78 77.77 86.55
Full 52.37
81.06 90.58 Full 93.99
96.78 95.92 Full ● ●●
83.21 95.11 98.23
Null

100
85.79
Null ● 102.08 106.56
90.88 107.74
119 108.33
122.52 112.12
133.7 135.39

Location (cM)
150
158.71
167.19

179.17
183.5

201.28

200
239.51
Full ● 247.32

250
256.69

1B 2A 2B 4B 4D 5A 6B 7D

Chromosome

Figure 6: Genetic map with QTL for the Full and Null models. Marker QTL are highlighted according to user specifications.
+ merge.by = "id", gen.type = "interval", method = "fixed",
+ na.method.X = "include", trace = TRUE, verboseLev = 0))

user system elapsed

137.83 1.62 140.24

Checking the timing for the QTL model st.qtlF from the previous section

R> system.time(st.qtlHD <- wgaim(st.fmF, phenoData = stpheno, intervalObj = stmap,

+ merge.by = "id", gen.type = "interval", method = "fixed",
+ na.method.X = "include", trace = TRUE, verboseLev = 0))

user system elapsed

132.16 2.32 134.80

Each of the analyses finds 11 QTL and the high dimensional analysis is completed with
only a negligible increase in CPU time. As outlined in Section 2.2.3, this is due to the
substantial reduction in the dimension of the genetic component. In this analysis, 1289
columns of marker scores are transformed to 175 columns (i.e. the number of lines of
the genetic marker set that have phenotypic information) which are then passed to the
asreml fit during each iteration of the wgaim analysis.
 Bibliography

Ball, R. (2010). BayesQTLBIC: Bayesian Multi-Locus QTL Analysis Based on the

BIC Criterion. R package version 1.0-1, URL http://www.CRAN.R-project.org/src/
contrib/Archive/bayesQTLBIC/.

Beavis, W. D. (1994). The power and deceit of QTL experiments: lessons from com-
parative QTL studies. In Proceedings of the Forty-Ninth Annual Corn and Sorghum
Industry Research Conference. American Seed Trade Association, Washington, DC,
250–266.

Beavis, W. D. (1998). QTL analyses: power, precision and accuracy. In A. H. Patterson,

ed., Molecular Dissection of Complex Traits. CRC Press, New York, 145–162.

Broman, K. W. & Sen, S. (2009). A Guide to QTL Mapping with R/qtl. Springer-
Verlag. ISBN: 978-0-387-92124-2.

Broman, K. W. & Speed, T. P. (2002). A model selection approach for the identifica-
tion of quantitative trait loci in experimental crosses. Journal of the Royal Statistical
Society, Series B 64 641–656.

Broman, K. W. & Wu, H. (2010). qtl: Tools for Analayzing QTL Experiments. R pack-
age version 1.15-15, URL http://www.CRAN.R-project.org/src/contrib/Archive/
qtl/.

Gilmour, A. R. (2007). Mixed Model Regression Mapping for QTL Detection in Ex-
perimental Crosses. Computational Statistics and Data Analysis 51 3749–3764.

Gilmour, A. R., Gogel, B. J., Cullis, B. R. & Thompson, R. (2009). ASReml

User Guide. Release 3.0.

Gilmour, A. R., Thompson, R. & Cullis, B. R. (1995). Average information REML:

An efficient algorithm for variance parameter estimation in linear mixed models. Bio-
metrics 51 1440–1450.

107
Gogel, B. J. (1997). Spatial analysis of multi-environment variety trials. Ph.D. thesis,
Department of Statistics, University of Adelaide.

Gogel, B. J., Welham, S. J., Verbyla, A. P. & Cullis, B. R. (2001). Outlier de-
tection in linear mixed effects; summary of research. report p106. Tech. rep., University
of Adelaide, Biometrics.

Golub, G. & van Loan, C. (1996). Matrix Computations. The Johns Hopkins Univer-
sity Press: London, 3rd ed.

Hayley, C. S. & Knott, S. A. (1992). A simple regression method for mapping

quantitative trait loci in line crosses using flanking markers. Heredity 69 315–324.

Huang, B. & George, A. (2009). Look before you leap: a new approach to mapping
qtl. TAG Theoretical and Applied Genetics 119 899–911. 10.1007/s00122-009-1098-y,
URL http://dx.doi.org/10.1007/s00122-009-1098-y.

Kang, H. M., Zaitlen, N. A., Wade, C. M., Kirby, A., Heckerman, D., Daly,
M. J. & Eskin, E. (2008). Efficient control of population structure in model organism
association mapping. Genetics 178 1709–1723. URL http://www.genetics.org/cgi/
content/abstract/178/3/1709.

Lander, E. & Green, P. (1987). Construction of multilocus genetic linkage maps in

humans. Proceedings of the National Academy of Science 84 2363–2367.

Martinez, O. & Curnow, R. N. (1992). Estimating the locations and sizes of the
effects of quantitative trait loci using flanking markers. Theoretical and Applied Genetics
85 480–488.

Melchinger, A. E., Utz, H. F. & Schon, C. C. (1998). Quantitative trait loci

(QTL) mapping using different testers and independent population samples in maize
reveals low power of QTL detection and large bias in estimates of QTL effects. Genetics
149 383–403.

Meuwissen, T. H. E., Hayes, B. J. & Goddard, M. E. (2001). Prediction of total

genetic value using genome-wide dense marker maps. Genetics 157 1819–1829.

Oakey, H., Verbyla, A. P., S, P. W., Cullis, B. R. & Kuchel, H. (2006). Joint
Modelling of Additive and Non-Additive Genetic Line Effects in Single Field Trials.
Theoretical and Applied Genetics 113 809–819.

Patterson, H. D. & Thompson, R. (1971). Recovery of interblock information when

block sizes are unequal. Biometrika 58 545–554.

Schwarz, G. (1978). Estimating the dimension of a model. Annals of Statistics 6

461–464.
Seaton, G., Haley, C. S., Knott, S. A., Kearsey, M. & Visscher, P. M. (2002).
QTL Express: mapping quantitative trait loci in simple and complex pedigrees. Bioin-
formatics 18 339–340. URL http://bioinformatics.oxfordjournals.org/content/
18/2/339.abstract.

Shriner, D. & Yi, N. (2009). Deviance Information Criterion (DIC) in Bayesian Mul-
tiple QTL Mapping. Computational Statistics and Data Analysis 53 1850–1860.

Smith, A., Cullis, B. R. & Thompson, R. (2001). Analysing variety by environment

data using multiplicative mixed models. Biometrics 57 1138–1147.

Smith, A., Cullis, B. R. & Thompson, R. (2005). The analysis of crop cultivar breed-
ing and vvaluation trials: An overview of current mixed model approaches. Journal of
Agricultural Science 143 449–462.

Smith, A. B., Lim, P. & Cullis, B. R. (2006). The design and analysis of multi-phase
plant breeding programs. Journal of Agricultural Science 144 393–409.

Stram, D. O. & Lee, J. W. (1994). Variance components testing in the longitudinal

mixed effects model. Biometrics 50 1171–1177.

Taylor, J. & Verbyla, A. (2011a). R package wgaim: QTL analysis in bi-parental

populations using linear mixed models. Journal of Statistical Software 40. URL http:
//www.jstatsoft.org/v40/i07.

Taylor, J. D. & Verbyla, A. P. (2011b). A variable selection method involving

complex linear mixed models. Annals of Applied Statistics 5 000–000.

Verbyla, A. P. (1990). A conditional derivation of residual maximum likelihood. Aus-

tralian Journal of Statistics 32. 227-230.

Verbyla, A. P., Cullis, B. . R. & Thompson, R. (2007). The analysis of QTL by

simultaneous use of the of the full linkage map. Theoretical and Applied Genetics 116
95–111.

Verbyla, A. P. & Cullis, B. R. (2010). Multivariate whole genome average interval

mapping: QTL analysis for multiple traits and/or multiple environments. Theoretical
and Applied Genetics Submitted.

Verbyla, A. P., Hackett, C. A., Newton, A. M., Taylor, W. B. T. & Cullis,

B. R. (2010). Multi-treatment QTL analysis using whole genome average interval
mapping. Theoretical and Applied Genetics Submitted.

Verbyla, A. P. & Taylor, J. D. (2011). High-dimensional whole genome average

interval mapping and a random effects formulation. Theoretical and Applied Genetics,
Submitted. .
Whittaker, J. C., Thompson, R. & Visscher, P. M. (1996). On the mapping of
QTL by regression of phenotype on marker-type. Heredity 77 22–32.

Xu, S. (2003). Estimating polygenic effects using markers of the entire genome. Genetics
164 789–801.

Yandell, B. S., Mehta, T., Banerjee, S., Shriner, D., Venkataraman, R.,
Moon, J. Y., Neely, W. W., Wu, H., Smith, R. & Yi, N. (2005). R/qtlbim:
QTL with Bayesian Interval Mapping in Experimental Crosses. Bioinformatics 23
641–643.

Zeng, Z.-B. (1994). Precision mapping of quantitative trait loci. Genetics 136 1457–
1468.

Zhang, M., Zhang, D. & Wells, M. (2008). Variable selection for large p small n
regression models with incomplete data: mapping QTL with epistases. BMC Bioinfor-
matics 9.

Zhou, Q. (2010). A Guide to QTL Mapping with R/qtl. Journal of Statistical Software,
Book Reviews 32 1–3. URL http://www.jstatsoft.org/v32/b05.