Linear, bilinear, and linear-bilinear fixed and mixed models

for analyzing genotype
environment interaction in plant

breeding and agronomy
Jose Crossa1, Mateo Vargas1,2, and Arun Kumar Joshi3
1
Biometrics and Statistics Unit, Crop Research Informatics Laboratory, International Maize and Wheat
Improvement Center (CIMMYT), Apdo. Postal 6-641, 06600 México DF, México; 2Universidad Autónoma de
Chapingo, Chapingo, México; and 3Global Wheat Breeding Program, CIMMYT, South Asia Office, Kathmandu,
Nepal. Received 7 January 2010, accepted 11 March 2010.
Crossa, J., Vargas, M. and Joshi, A. K. 2010. Linear, bilinear, and linear-bilinear fixed and mixed models for analyzing
genotype
environment interaction in plant breeding and agronomy. Can. J. Plant Sci. 90: 561
574. The purpose of this
manuscript is to review various statistical models for analyzing genotype  environment interaction (GE). The objective is
to present parsimonious approaches other than the standard analysis of variance of the two-way effect model. Some fixed
effects linear-bilinear models such as the sites regression model (SREG) are discussed, and a mixed effects counterpart such
as the factorial analytic (FA) model is explained. The role of these linear-bilinear models for assessing crossover interaction
(COI) is explained. One class of linear models, namely factorial regression (FR) models, and one class of bilinear models,
namely partial least squares (PLS) regression, allows incorporating external environmental and genotypic covariables
directly into the model. Examples illustrating the use of various statistical models for analyzing GE in the context of plant
breeding and agronomy are given.

Key words: Least squares, singular value decomposition, environmental and genotypic covariables

Crossa, J., Vargas, M. et Joshi, A. K. 2010. Modèles linéaires, bilinéaires et linéaires-bilinéaires fixes et mixtes pour l’analyse
des interactions du génotype et de l’environnement en phytogénétique et en agronomie. Can. J. Plant Sci. 90: 561
574. Cet
article passe en revue divers modèles statistiques servant à analyser les interactions entre le génotype et l’environnement
(GE). L’objectif consiste à présenter des approches parcimonieuses différentes de l’analyse usuelle de la variance avec le
modèle à deux axes. On y présente des modèles linéaires-bilinéaires aux mêmes effets fixes, comme le modèle de régression
des sites (SREG), et propose une contrepartie à effets mixtes, comme le modèle à analyse factorielle. Les auteurs décrivent
comment les modèles linéaires-bilinéaires servent à évaluer l’interaction des effets croisés. Un type de modèle linéaire, ceux
à régression factorielle, et un type de modèle bilinéaire, ceux à régression partielle des moindres carrés, permettent
d’intégrer directement des covariables environnementales et génotypiques. Suivent des exemples illustrant l’utilisation de
divers modèles statistiques permettant d’analyser les GE dans le contexte de l’amélioration génétique des plantes et de
l’agronomie.

Mots clés: Moindres carrés, décomposition en valeurs singulières, covariables environnementales et génotypiques

The presence of genotype environment interactions
(GE) in plant breeding and agronomy experiments is
expressed either as inconsistent responses of some
genotypes relative to others due to genotypic rank
change or as changes in the absolute differences between
genotypes without rank change. Several models are
commonly used for describing the mean response
of genotypes over environments and for studying
and interpreting GE in agricultural experiments: linear
models, bilinear models, and linear-bilinear models. One
class of linear models, namely factorial regression (FR)
models, and one class of bilinear models, namely partial
least squares (PLS) regression, allow incorporating
Presented at the Statistics Symposium held during the
Annual Meeting of the Canadian Society of Agronomy,
7 August 2009, Guelph, Ontario.
561
external environmental and genotypic covariables di-
rectly into the model.
The basic two-way fixed effects linear model for GE
analyses considers that the empirical mean response, ȳij ;
of the ith genotype (i1, 2, . . ., I) in the jth environment
(j1, 2, . . ., J) with n replications in each of the IJ
cells is expressed as
ȳij mti dj (td)ij  ōij (1)
where m is the grand mean over all genotypes and
environments, ti is the additive effect of the ith
genotype, dj is the additive effect of the jth environment,
(td)ij is the non-additivity interaction (GE) of the ith
genotype in the jth environment, and ōij is the average
error assumed to be NID (0, s2/n) (where s2 is the
within-environment error variance, assumed to be con-
stant). For a complete random model, it is assumed that
562 CANADIAN JOURNAL OF PLANT SCIENCE
ti, dj, and (td)ij are normally and independently dis-
tributed, with variances s2t ; s2d ; and s2td ; respectively.
Yates and Cochran (1938) introduced the model in
which the GE term is linearly related to the environ-
mental main effect.
The purpose of this paper is to present parsimonious
approaches
other than the model in Eq. 1
to GE
analysis. Examples illustrating the use of various
statistical models for analyzing GE in the context of
agronomy and plant breeding experiments are given.
STATISTICAL MODELS FOR GENOTYPE

ENVIRONMENT INTERACTION
Fixed Effect Linear-bilinear Models
Williams (1952) considered the model ȳij mti
lai gj  ōij where l is the largest singular value of ZZ?
and Z?Z (for Z ȳij  ȳi: ) and ai and gj are the
corresponding eigenvectors. Gollob (1968) and Mandel
(1969, 1971) extended Williams’ (1952) work by con-
sidering the bilinear GE term as (td)ij  atk1 lk aik gjk :
Thus, the general formulation of the linear-bilinear
model is
Xt
ȳij mti dj  l a g  ōij (2)
k1 k ik jk
where the constant lk is the singular value of the kth
multiplicative component that is ordered l1 ]l2 ]. . .]
lt ; the aik are elements of the kth left singular vector of
the interaction and represent the genotypic sensitivity to
hypothetical environmental factors represented by the
kth right singular vector with elements gjk. The aik
and gjk satisfy the constraints ai aik aik0  aj gjk gjk0  0 for
k "k? and ai a2ik aj g2jk  1: Gabriel (1978) described the
least squares fit of Eq. 2 and explained how the residual
matrix of the GE term
Z  ȳij  ȳi:  ȳ:j  ȳ::
is subjected to a singular value decomposition (SVD)
after adjusting for the additive (linear) terms. Gauch
(1988) called Eq. 2 the Additive Main Effect and
Multiplicative Interaction (AMMI) model.
Other classes of linear-bilinear models described by
Cornelius et al. (1996) are the Genotypes Regression
Model (GREG)
Xt
ȳij mi  l a g  ōij ;
k1 k ik jk
the Sites (environments) Regression Model (SREG)
Xt
ȳij mj  l a g  ōij ;
k1 k ik jk
the Completely Multiplicative Model (COMM)
Xt
ȳij m l a g  ōij ;
k1 k ik jk
and the Shifted Multiplicative Model (SHMM)
Xt
ȳij b lk aik gjk  ōij
k1
In matrix notation, these linear-bilinear models can be
expressed as Y am k1 bk XK ALG?E (Cornelius and
Seyedsadr 1997), where Y [ȳij ]; Xk [Xkij ]; E [ōij ];
L diag(lk, k1, 2, . . . , t), l1 ]l2 ]. . .]lt ; A 
(a1, . . . , at), G (g1,. . ., gt) and A?A G?G It.
The xkij are known constants and bk ; lk ; aik ; and gjk
are parameters to be estimated.
Fixed Effects Sites Regression Model
Note that in the SREG model, the bilinear term models
the main effects of genotypes (G) plus GE interaction
(usually called a GGE biplot), and the composition of
the two-way I J matrix to be subjected to singular
value decomposition is different than the one used in the
AMMI model. Furthermore, SREG with two compo-
nents (SREG2) can be perceived as consisting of a set of
multiple regression equations (one for each environ-
ment), each regression equation consisting of an envir-
onmental mean or environmental effect as intercept plus
two terms for regression on two genotypic regressor
variables, ai1 and ai2 (either observed or latent), with gj1
and gj2 as the regression coefficients.
When the correlation between the site means and the
first principal components of sites in the SREG2 model
is high (say ]0.80), then the SREG model has the
property that the first principal component of SREG2
accounts for non-crossover interaction (non-COI) and
the second principal component of SREG2 is due to
COI variability; this should hold for FA(2) as well
(Burgueño et al. 2008). This is an important property
that allows using the biplot for discriminating group
of sites and genotypes with and without crossover
interactions.
Recently, the merits and demerits of AMMI vs. GGE
biplots for genotype and environment identification
(Yan et al. 2007; Gauch et al. 2008) have been examined
and discussed. Yang et al. (2009) pointed out the
advantages and disadvantages of these fixed effects
linear-bilinear models and discussed relevant issues
concerning the use of biplot analysis as a descriptive
statistical tool. One of the main issues pointed out by
Yang et al. (2009) was related to the fact that genotypes
or environments, or both, may be considered realiza-
tions of random variables and therefore must be treated
as random effects. Another relevant issue pointed out by
Yang et al. (2009) is whether the biplot can detect
crossover interaction.
Mixed Effect Linear-bilinear Models
A mixed-model analogue of AMMI or SREG has been
developed using the factor analytic (FA) model for
approximating the variance-covariance GE structure
(Piepho 1998; Smith et al. 2002, 2005; Piepho and
Mohring 2005). Research conducted by Crossa et al.
(2006) and Burgueño et al. (2008) described how to model
 CROSSA ET AL. * ANALYZING GENOTYPE ENVIRONMENT INTERACTION
variance-covariance GE and GGE using the FA model
and how to incorporate the additive (relationship A)
matrix and the additive additive covariance matrix into
the FA model based on pedigree information. Burgueño
et al. (2008) also described the equivalence between
SREG2 and FA(2) for finding subsets of genotypes and
environments without COI.
Factor Analytic and Sites Regression Models for
Assessing Crossover Genotype
Environment
Interaction
In the FA model, the random effect of the ith genotype
in the jth environment (gij) is expressed as a linear
function of latent variables xik with coefficients djk for
k 1, 2, . . . t, plus a residual, hij ; i.e., gij mj 
t line at an interaction score of zero. Hence, this graph
ak1 xik djk hij ; so that the ijth cell mean can be written
as yij gijoij. With only the first two latent factors
being retained, gij is approximated by gij :mj xi1 dj1 
xi2 dj2 hij : Therefore, there is a clear connection be-
tween the SREG2 and FA(2) models, as described by
Burgueño et al. (2008). A similar connection between
the AMMI2 and FA(2) models was established by Smith
et al. (2002).
Under principal component rotation, the directions
and projections of the vectors of FA(2) and SREG2 in
the biplot are the same. Therefore, the property of the
SREG by which the first principal component of
SREG2 accounts for non-crossover interaction (non-
COI) and the second principal component of SREG2 is
due to COI variability should hold for FA(2) as well.
It should be pointed out that the absolute values of
genotypic and environmental scores under the FA(2)
and SREG2 models may not necessarily be the same; the
estimates of the random effects in the FA(2) model are
BLUPs (Best Linear Unbiased Predictions) (Henderson
1984), whereas the estimates in fixed effects SREG2
model are least squares estimates, that is, Best Linear
Unbiased Estimates (BLUE). Furthermore, the
standard errors of the estimable functions of fixed
effects under SREG differ from those of predictable
functions of a mixture of fixed and random effects under
FA, and FA models are more flexible in handling
unbalanced data (the SREG model does not handle
missing data).
Additional Options for Assessing COI and
Non-COI
The GREG linear-bilinear model defined above is a
reparameterization of the stability analysis models of the
Finlay and Wilkinson (1963) linear regressions of yields
on environment means and the Eberhart and Russell
(1966) model, with the first multiplicative term, l1 ai1 gj1 :
perceived as the genotype regressions, with coefficients
ai1 on environmental indices gj1 (the scale parameter l1
can be absorbed into ai1 or gj1 or partially into each), and
the deviation modeled as multiplicative components
provided that t 1.
563
The simple Finlay and Wilkinson shows COI and
non-COI because for any pair of genotypes, if their lines
cross, that is COI; if their lines do not cross but have
different slopes, that is non-COI; and if their lines are
nearly parallel, that is negligible COI. In the special
case that GE has a rather simple structure, so that these
regressions capture nearly as much GE as does the
AMMI1 or SREG2 models, this could be the preferred
model because of its conceptual simplicity.
In the most common case of a complex GE, yields for
each genotype can be graphed as a linear function of the
environment first interaction component scores using
the AMMI1 model, as in Fig. 3 of Gauch et al. (2008).
This shows COI and non-COI. Furthermore, the
genotype means are also shown by the height of each
may be considered comparable to the SREG2 counter-
part. In case of very complex GE, the AMMI2 model
can be used to show COI, as in Fig. 4 of Gauch et al.
(2008), which shows COI not for all genotypes, but for
those of greatest interest, namely those genotypes that
are the winners. Its counterpart, generally recovering
a similar amount of GE, would be the SREG3 model,
but no graphical representation of that model has yet
been published. However, the example used by Gauch
et al. (2008) is rather small in size as compared with
those usually present in international plant breeding
trials.
Given the Finlay-Wilkinson (and joint regression),
AMMI1, and AMMI2 counterparts of SREG2 for
graphical display of COI and non-COI, various circum-
stances for a given dataset might indicate a preference
for the Finlay-Wilkinson or AMMI1 or AMMI2 model.
However, and as described and explained by Yang et al.
(2009), the fixed effects AMMI as well as SREG models
have the drawback of not incorporating the natural
uncertainty present when estimating the GE parameters
from the data; on the other hand the mixed AMMI and
SREG versions consider this uncertainty when estimat-
ing the Best Linear Unbiased Prediction (BLUP).
Finally, relaying only in the display of genotypes and
sites offer by the biplot from the fixed effects AMMI
and SREG models does not seem to be a powerful
scientific methodology for predicting the performance of
genotypes in other sites in future years.?
There are several statistical as well as biological
reasons to prefer SREG over AMMI for assessing
COI and non-COI under the common situation of a
complex GE; (1) SREG is a more parsimonious model
than AMMI, (2) SREG incorporates the main effect of
genotypes directly into the statistical analysis; this is of
importance for breeders’ objectives that require includ-
ing the main performance of genotypes in the model, (3)
the mixed SREG model can be fitted much more easily
than the mixed AMMI model, and (4) the mixed SREG
model, as proved by Burgueño et al (2008), is useful for
delineating mega-environments using a formal statistical
approach based on the factor analytic model rather
564 CANADIAN JOURNAL OF PLANT SCIENCE
than a mere visualization of genotypes and sites given by
the biplot.
INCORPORATING EXTERNAL COVARIABLES
FOR EXPLAINING GENOTYPE
ENVIRONMENT
INTERACTION
Factorial regression (FR) and partial least squares
(PLS) analysis (e.g., van Eeuwijk et al. 1996; Vargas
et al. 1999) are useful for studying the effects of both
genetic and environmental covariables and for explain-
ing the causes of GE. The structural equation model
(SEM) using endogenous and exogenous variables is a
useful alternative for overcoming some of the limitations
of the FR and PLS approaches and for developing
functional relationships and predictability with expla-
natory covariables.
Factorial Regression Model
The GE is modeled directly using regression on envir-
onmental (and/or genotypic) variables. A useful linear
model for incorporating external environmental (or
genotypic) variables is the factorial regression (FR)
model (Denis 1988; van Eeuwijk et al. 1996). The FR
models are ordinary linear models that approximate the
GE effects in Eq. 1 by the products of one or more (1)
genotypic covariables (observed) environmental po-
tentialities (estimated), (2) genotypic sensitivities
(estimated) environmental covariables (observed).
For k 1, . . . , G genotypic covariables (centered) re-
presented by xi1 ; . . . ; xiG ; Eq. 1 becomes ȳij mti 
dj aG g1 xig jjg  ōij ; G 5I 1, where jjg represents an
environmental factor (regression coefficient) with re-
spect to the genotypic covariable, xig : Constraints on the
parameters are ai ti  aj dj aj jjg  0: In matrix nota-
tion the expectation is
E(Y) m1I 1?J t1?J 1I d?XJ? (3)
where Y [ȳij ] is a IJ matrix; 1I and 1J are I 1 and
J 1 vectors with all elements equal to 1, respectively;
t [tI] is the I1 vector of main effects of genotypes;
d [dj] is the J1 vector of main effects of environ-
ments; X [xig ] is the IG matrix of known genotypic
covariables; J [jjg ] is the J G matrix of unknown
environmental factors (regression coefficients).
For h1, . . . , H environmental covariables (cen-
tered) represented by zj1 ; . . . ; zjH ; Eq. 1 is ȳij mti 
dj aHh1 wih zjh  ōij ; H5J 1, where wih represents a
genotypic sensitivity (regression coefficient) with respect
to the environmental covariable, zjh : Constraints on
the parameters are ai ti aj dj ai zih 0: In matrix
notation, the expectation is
E(Y) m1I 1?J t1?J 1I d?zZ? (4)
where Z [zjh ] is the J H matrix of known environ-
mental covariables; z[wih ] is the IH matrix of
unknown differential genotypic sensitivities.
The FR model including genotypic and environmen-
tal covariables simultaneously is
XG XH
ȳij mti dj  x j 
g1 ig jg
w z
h1 ih jh
XH XG
 h1 g1
xig ngh zjh  ōij
where nkh is a constant that scales the cross-product of
the genotypic covariables, xk ; with the environmental
covariables, zh ; and can be derived from the two
previous FR models by imposing the restriction jjg 
ngh zjh or wih xih ngh ; each cross-product represents one
degree of freedom in the GE subspace. In matrix
notation, the expectation is
E(Y)m1I 1?J t1?I 1J d?XnZXJ?zZ?
where the constraint XJ?zZ?0 (where 0 is a matrix
H G of zeros) is required. The model should be fitted
for all possible combinations of genotypic covariables
and environmental covariables.
When environmental (or genotypic) covariables show
high collinearity, interpretation of the least squares
regression coefficients is complicated because they are
estimated very imprecisely. Consequently, a stepwise
procedure for choosing the covariables to include could
be useful for model construction. Noise on the response
variable also complicates interpretation of the FR
parameters. Furthermore, least squares estimation of
the parameters in the FR models is not unique when the
number of covariables is larger than the number of
observations, so an alternative estimation method is
needed. Partial least squares (PLS) regression, which
overcomes some of these problems, can be used as an
alternative estimation method.
Partial Least Squares Regression
Multivariate partial least squares (PLS) regression
models (Aastveit and Martens 1986; Helland 1988) are
a special class of bilinear models. When genotypic
responses over environments (Y) are modeled using
environmental covariables, the JH matrix Z of H
(h 1, 2, . . . , H) environmental covariables can be
written in bilinear form as
Z t1 p?1 t2 p?2 . . .tM p?M EM  TP?E (5)
where the matrix T contains the tJ J 1 vectors called
latent environmental covariables or Z-scores (indexed
by environments), and the matrix P contains the p1 . . .
pH H 1 vectors called Z-loadings (indexed by environ-
mental variables) and E has the residuals. Similarly, the
response variable matrix Y in bilinear form is
Y t1 q?1 t2 q?2 . . .tM q?M FM  TQ?F (6)
where the matrix T is as in Eq. 5 and the matrix Q
contains the q1 . . . qI I1 vectors called Y-loadings
 CROSSA ET AL. * ANALYZING GENOTYPE ENVIRONMENT INTERACTION
(indexed by genotypes) and F has the residuals. The
relationship between Y and Z is transmitted through
the latent variable T. The PLS algorithm performs
separate (but simultaneous) principal component
analysis of Z and of Y that allows reducing the number
of variables in each system to a smaller number of
hopefully more interpretable latent variables.
Helland (1988) showed that a reduced number of PLS
latent variables gives a low rank representation of the
least squares estimates of the FR with environmental
covariables because the expectation of Y? is
E(Y?)QT? Q(ZW)?(QW?)Z? zZ?
SH
h1 wih Zjh (7)
as in Eq. 4, where T, Q, and Z are defined as before and
the vector W is H1 and contains the Z-loadings (or
weights) of the environmental covariables; z contains the
PLS approximation to the regression coefficients of the
responses in Y to the environmental covariables in Z.
The matrices T (with J coordinates for environments),
Q (with I coordinates for genotypes), and W (with H
coordinates for environmental covariables) can be
represented in the PLS biplot such that projecting the
jth environment (row) of T on the ith genotype (row) of
Q [Y?(TQ?)?] approximates the GE; projecting the hth
environmental covariable (row) of W on the ith geno-
type (row) of Q (QW?z) approximates the regression
coefficient of the ith genotype on the hth environmental
covariable (Vargas et al. 1999; van Eeuwijk et al. 2000).
When genotypic covariables are used to model environ-
mental responses over genotypes, the latent genotypic
covariables are T XW, where vector W is G 1 and
contains the weights of the genotypic covariables. The
expectation of Y is
E(Y) TQ?XWQ? XJ? SG
g1 Xig jjg (8)
as in Eq. 3 (Vargas et al. 1999; van Eeuwijk et al. 2000)
where J contains the PLS approximation to the regres-
sion coefficients of the responses in Y to the genotypic
covariables in X. The matrices T (with I coordinates for
genotypes), Q (with J coordinates for environments),
and W (with G coordinates for genotypic covariables)
can be represented in a PLS biplot such that projection
of the ith genotype (row) of T onto the jth environment
(row) of Q (Y TQ?) approximates the GE; projection
of the gth genotypic covariable (row) of W onto the jth
environment (row) of Q (WQ?J) approximates the
regression coefficient of the jth environment on the gth
genotypic covariable.
Structural Equation Model (SEM)
The SEM approach is similar to multiple regression,
because it analyzes a system of equations in which each
equation describes a causal relationship among variables
considered in the system. The SEM is useful because
it can integrate and model intermediate traits, such
565
as yield components in cereals and their interrelation-
ships with other variables, as well as with final grain
yield. The SEM allows a researcher to test hypotheses
on cause-effect relationships between variables in a
complex system where the initial definition of SEM
comprises a path diagram that outlines the various levels
of observed (or latent) independent or dependent vari-
ables, as well as the directions of causal relationships
among variables.
In an agricultural context, the SEM was first proposed
by Dhungana (2004) to study GE of grain yield in wheat
and its components, and to account for the importance
of intermediate traits associated with yield components.
The authors explained yield GE with cross-products of
genotypic and environmental covariates as exogenous
(independent) variables and observed yield component
GE as endogenous (dependent/independent) variables.
Dhungana (2004) concluded that SEM on observed
variables was an effective way of describing yield GE
in wheat, given that the interrelationships and role of
yield component GE can be incorporated simultaneously
in a single model. Diagrams representing the structural
models known as path diagrams are useful for visualizing
complex models and variable relationships.
EXAMPLES OF THE USE OF FACTORIAL
REGRESSION, PARTIAL LEAST SQUARES AND
STRUCTURAL EQUATION MODELS IN PLANT
BREEDING AND AGRONOMY EXPERIMENTS
Treatment
Environment Interaction Analysis
in Agronomy
A description of the treatment environment (T E)
interaction of 24 agronomic treatments (1
24) [tillage,
summer crop, manure, and nitrogen (N)] evaluated over
10 yr (1988
1997) was provided by Vargas et al. (2001).
Results of the final FR were compared with those of a
partial least squares (PLS) to achieve extra insight into
both the T E and the final FR model.
The FR was applied to year tillage, year summer
crop, year manure, year N, year summer crop 
N, and year manure N. Results for the stepwise
multiple factorial regression model of the interaction
between 27 environmental covariables and tillage
showed that evaporation in December (EVD) tillage
sum of squares accounted for 68% of the whole year 
tillage interaction. For year summer crop, evaporation
in April (EVA) accounted for 36% of the year summer
crop interaction. For year manure, covariables pre-
cipitation in December (PRD) and sun hours in
February (SHF) contributed 56% of the year manure
sum of squares. Year nitrogen (N) interaction deter-
mined the major part of year treatment interaction
sum of squares.
The PLS biplot shown in Fig. 1 separated the nine
highest yielding treatments (T9, T19, T21, T17, T11,
T12, T10, T23, and T18) from the nine lowest yielding
treatments (T1, T2, T3, T4, T5, T6, T7, T8, and T16).
566 CANADIAN JOURNAL OF PLANT SCIENCE

1.0 March are related to precipitation in December,

January, and March.
T18 The most highly productive treatments are associated
mTJ
0.6 mTUJ T12
T23 T11 with high N levels (100 and 200 kg ha 1) and no
T24
PRM
T20 T22 T19 precipitation. The explanation maybe that precipitation
mTUA
1992
T14 PRJ
T15 1997 T17 is associated with leaching of N (especially if the texture
1990 T9
T10 mTA T21
of the soil is coarse). In addition, higher precipitation is
0.2 mTM
also associated with clouds, which reduce radiation.
Factor 2

PRFmTUD13 EVD
1994
1993
mTUM 1989 MTD 1988 While radiation is the major yield-limiting factor when
SHF
T16 mTD
PRD EVF
N and water are non-limiting, high radiation may also
–0.2 T2 MTM EVJ
1991 SHD be associated with higher temperatures and excessive
T6
EVM evaporative demand. Accelerated development rate may
SHJ
T1
MTJ be especially prejudicial to yield during spike growth
1995 1996
–0.6 mTUF
T5
EVA
(February) and, to a lesser extent, during grain filling
MTA
T7 T4
T8 (March
April). Excessive evaporative demand may
mTF T3
reduce the ability of the plant to cool itself directly by
–1.0 MTF not permitting sufficient evapotranspiration or indir-
ectly by reducing soil moisture.
–1.0 –0.6 –0.2 0.2 0.6 1.0
Factor 1

Fig. 1. Biplot of the first and second PLS factors representing

the Z-scores of 10 yr (1988
1997) and the Y-loadings of the 24
practice treatments (1
24) enriched with the Z-loadings of 27
environmental variables [extracted from Vargas et al. (2001)].
EV total monthly evaporation; PRtotal monthly precipi-
tation; SHsun hours per day; mTmean minimum
temperature sheltered; MT mean maximum temperature
sheltered; mTUmean minimum temperature unsheltered;
D December; J January; FFebruary; MMarch;
A April.
The nine lowest yielding treatments had a positive
interaction with year 1995, which had high mTUF
(mean minimum temperature unsheltered in February),
mTF (mean minimum temperature in February), and
MTA (mean maximum temperature in April) but
negative interaction with year 1988 (opposite quadrant).
The PLS biplot contains roughly five clusters of
correlated environmental covariables. The order of
inclusion of these covariables in the FR using the
stepwise procedure for each factor effect corresponds
to selecting covariables for the different cluster groups
depicted in Fig. 1.
In general, SH, EV, and MT are grouped in the right
quadrants of the biplot, whereas PR, mT, and mTU are
grouped in the left quadrant of the biplot. It is expected
that with more sun hours, there will be higher maximum
temperatures and more evaporation, and that with more
precipitation, there will be fewer sun hours and thus
lower temperature. This is clear for the lower right
cluster of variables comprising MT, EV, and SH. The
group of environmental variables located in the right
upper quadrant indicates that minimum temperature in
April with maximum temperature and evaporation in
December had a similar effect on the T E for the
treatments located in that quadrant. The two groups
of variables in the left upper quadrant indicate that
minimum temperatures in December, January, and
Genotype
Environment Interaction for Zinc
and Iron Concentration of Wheat Grain
Zinc and iron are important micronutrients for human
health; there is widespread deficiency of these micro-
nutrients in many regions of the world including South
Asia. Breeding aimed at enriching wheat grains with
more zinc and iron are in progress in India and
Pakistan, and at CIMMYT. Greater knowledge of GE
of these nutrients in the grain is expected to increase our
understanding of the magnitude of this interaction and
to help identify more stable genotypes for this trait. Elite
wheat lines from CIMMYT were evaluated in multi-
environment trials in the Eastern Gangetic Plains of
India during three years to study GE interactions for
agronomic and nutrient traits. Soil and meteorological
data from each of the locations were also used (Joshi
et al. 2010).
Joshi et al. (2010) showed the results of factorial
regression with contributions by significant environ-
mental and soil covariables to explain GE variability
(Table 1). For iron concentration in the grain, four
covariables (maximum temperature before flowering,
rainfall after flowering, zinc in 30
60 soil depth, and RH
after flowering) were significant, accounting for 59.46%
of GE variation. For grain Zn concentration, five
covariables explained 82.41% of GE variation. These
covariables were, in descending order of their contribu-
tion, minimum temperature after flowering, Zn in 30
60
cm soil depth, rainfall after flowering, minimum
temperature before flowering, and Zn in 0
30 soil
depth. Zn content in 30
60 cm soil depth was also a
significant determinant for grain Fe concentration.
These results suggest that the GE was substantial for
grain Fe and Zn.
For Fe and Zn concentrations in wheat grain,
genotypic responses varied widely across environments,
as indicated by vectors that radiated in all directions
in the PLS biplot depicting genotypic variables,
 CROSSA ET AL. * ANALYZING GENOTYPE ENVIRONMENT INTERACTION 567

Table 1. Proportion of GE variation accounted for by factorial regression the fact that most locations in 2005 and 2006 were
analysis for each significant covariable for grain iron and zinc not correlated with respect to GE and often placed
concentration opposite to each other. The biplot clearly differentiated
Variable SS variable % variation explained the 2005 and 2006 sown environments at all locations,
showing the complexity of evaluating grain Fe and Zn
Fe concentration in the grain concentrations.
TMXBFz 1045.64 22.21
RAFz 644.96 13.69
Maximum temperature before flowering (TMXBF)
Zn_60z 644.89 13.69 appeared to play an important role for both Fe and Zn.
RHAFz 463.93 9.85 This was perhaps because temperature before flowering
% contribution of 4 variables 59.46 contributes to proper development of the embryo, where
Zn concentration in the grain a large portion of micronutrients reside. PLS analysis
TMNAFz 1369.06 29.11
Zn_60 751.67 15.98
of genotypic variables and GE for grain Fe concentra-
RAFz 653.90 13.90 tion placed variables such as Fe_30 and Zn_30 in the
TMNBFz 604.58 12.85 right uppermost quadrant, while variables Fe_60 and
Zn_30z 495.65 10.54 Zn_60 were in the left uppermost quadrant (Fig. 2).
% contribution of 5 variables 82.41
PLS analysis of environmental variables and GE
z
TMXBFmaximum temperature before flowering; TMNBFmini- for grain Zn concentration placed variables such as
mum temperature before flowering; TMNAFminimum temperature minimum temperature before and after flowering
after flowering; RAFrainfall after flowering; Zn_30zinc concen- (TMNBF, TMNAF) in the right uppermost quadrant
tration in 0
30 cm soil depth; Zn_60zinc concentration in (Fig. 3), indicating their similar role. On the other hand,
30
60 cm soil depth; RHAFrelative humidity after flowering.
maximum temperatures before and after flowering
(TMXBF, TMXAF) were in the opposite quadrants,
environments, and genotypes (Figs. 2 and 3 for Fe and indicating their opposite role in grain Zn concentration.
Zn concentration in wheat grain). Variations in the Variables Zn_30 and Zn_60 were also in the opposite
pattern of response within the location are evident from quadrants, indicating that their contribution to grain Zn

1.0
Zn_60
0.8 Fe_60
12 1 5
0.6
TMNBF
Ghu_6 BHU_6
0.4 15 17 Zn_30
TMNAF
Bhd_5
Factor 2 (16.17%)

4 Fe_30
7 8
0.2 Bha_5 Bhu_6Pid_6 3
11 16 6 RHBFRHAF
0.0 TMXBF
19
–0.2
RBF
14 BHU_5
Ghu_513
–0.4
18 2
TMXAF
–0.6
Mau_5 Bhu_5
20 910
–0.8
RAF
–1.0

–1.0 –0.8 –0.6 –0.4 –0.2 0.0 0.2 0.4 0.6 0.8 1.0
Factor 1 (18.52%)

Fig. 2. Partial least squares (PLS) biplot of the number of locations and years with environmental and soil covariables on the
performance of iron grain concentration in 20 wheat lines in 10 environments of the eastern Gangetic plains of south Asia. The more
significant variables in FR are given in italics. Bhu_5 Bhurkura in 2005; Ghu_5Ghurahoopur in 2005; Bhd_5 Bhadawal in
2005; Bha_5 Bhagwanpur in 2005; Mau_5 Mauparasi in 2005; BHU_5Banaras Hindu University in 2005; Bhu_6Bhurkura
in 2006; Mau_6 Mauparasi in 2006; BHU_6Banaras Hindu University in 2006; Pid_6Pidkhir in 2006. TMXBF maximum
temperature before flowering; TMXAF maximum temperature after flowering; TMNBF minimum temperature before flower-
ing; TMNAF minimum temperature after flowering; RHBFrelative humidity before flowering; RHAF relative humidity after
flowering; RBF rainfall before flowering; RAF rainfall after flowering; Zn_30 zinc concentration in 0
30 cm soil depth;
Zn_60 zinc concentration in 30
60 cm soil depth; Fe_30 iron concentration in 0
30 cm soil depth; Fe_60iron concentration in
30
60 cm soil depth.
568 CANADIAN JOURNAL OF PLANT SCIENCE

1.0 TMNAF

8
0.8 Bhu_6 TMNBF
RHAF 10 19 7
0.6
2 5 Pidk_6 20 18
RHBF TMXBF
0.4 RFAF RFBF

Factor 2 (16.30%)
6
Ghu_5 Zn_30
0.2 9
BHU_5
0.0
12 17 11
Bhd_5 Bhu_5 TMXAF
Fe_30
–0.2 13
Bha_5 Ghu_6
–0.4
Mau_5 16
–0.6 3 BHU_6
1
Fe_60 Zn_60
–0.8 4 1415
–1.0

–1.0 –0.8 –0.6 –0.4 –0.2 0.0 0.2 0.4 0.6 0.8 1.0
Factor 1 (21.53%)

Fig. 3. Partial least squares (PLS) biplot of the number of locations and years with environmental and soil covariables on the
performance of zinc concentration in the grain of 20 wheat lines in 10 environments of the eastern Gangetic plains of south Asia.
The more significant variables in FR are given in italics. Bhu_5 Bhurkura in 2005; Ghur_5Ghurahoopur in 2005; Bhd_5 
Bhadawal in 2005; Bha_5 Bhagwanpur in 2005; Mau_5 Mauparasi in 2005; BHU_5 Banaras Hindu University in 2005;
Bhu_6 Bhurkura in 2006; Mau_6Mauparasi in 2006; BHU_6Banaras Hindu University in 2006; Pidk_6 Pidkhir in 2006.
TMXBF maximum temperature before flowering; TMXAF maximum temperature after flowering; TMNBFminimum
temperature before flowering; TMNAF minimum temperature after flowering; RHBF relative humidity before flowering;
RHAF relative humidity after flowering; RBFrainfall before flowering; RAF rainfall after flowering (RAF); Zn_30 zinc
concentration in 0
30 cm soil depth; Zn_60 zinc concentration in 30
60 cm soil depth; Fe_30 iron concentration in 0
30 cm soil
depth; Fe_60 iron concentration in 30
60 cm soil depth.

concentration is quite different. Variable Zn_60 anthesis); (5) foliar, 2.0 kg ha1 at two stages; (6) foliar,
appeared to play a greater role in grain Zn concentra- 4.0 kg ha1 at two stages; (7) soil (25 kg ha1)2 foliar
tion. Genotypes 2, 5, and 6 were on the same quadrant (2.0 kg ha 1); (8) soil (50 kg ha 1)2 foliar (2.0 kg ha 1);
as Zn_30, indicating they can have greater Zn in the (9) soil (0 kg)2 foliar @ 0.2%; (10) soil (25 kg)2 foliar
grain when the micronutrient is higher in top soil. On @ 0.2%; (11) soil (50 kg)2 foliar @ 0.2%, and (12) local
the other hand, genotypes 1, 11, 13, 14, 15, and 16 were farmers’ practice for Zn, i.e., ZnSO4 @ 5 kg ha 1.
in the same quadrant as Zn_60, indicating their capacity Factorial regression analysis showed that for grain
to take up Zn from deeper soil. Zn, two covariables were significant and contributed
around 91% of the variation (Table 2). They were
Zinc
Environment Interaction for Zinc and
Iron Concentration in Wheat Grain
Current genotypes have relatively low zinc concentra-
Table 2. Percentage of variation accounted for by each significant
tion in the grain. A low-cost agronomic intervention covariable for grain yield and grain iron and zinc concentration
using Zn fertilizers could be a complementary approach
to enrich wheat grain. A multi-environment trial includ- Variable SS variable % variation explained
ing 12 zinc application treatments was organized in late-
Fe concentration in the grain
sown environments in four sites of the Eastern Gangetic Zn_30z 1660.29 60.85
Plains of India using the most widely sown wheat RHBFz 861.96 31.59
cultivar in that region. Agronomic performance and % contribution of 2 variables 92.44
Zn and Fe concentrations in the grain were recorded. Zn concentration in the grain
Factorial regression (FR) and partial least squares Fe_30z 2025.53 73.16
(PLS) analysis was performed on the Zn treatment Zn_30z 511.36 18.47
% contribution of 2 variables 91.63
environment (TE).
For Zn, 12 treatments of ZnSO4 were used in each z
RHBFrelative humidity before flowering; Zn_30zinc concentra-
location: (1) control, (2) soil, 25 kg ha1; (3) soil, 50 kg tion in 0
30 cm soil depth; Fe_30 iron concentration in 0
30 cm soil
ha1; (4) foliar, 1.0 kg ha1 at two stages (flag leaf and depth.
 CROSSA ET AL. * ANALYZING GENOTYPE ENVIRONMENT INTERACTION 569

Fe and Zn in the soil (0
30 cm depth). For Fe
concentration in the grain, two covariables (Zn in 0

30 cm soil depth and relative humidity before flowering)
were significant and accounted for around 92% of the
variation.
For Zn concentration in wheat grain, location
responses varied widely, as indicated by vectors that
radiated in all directions in the PLS biplot depicting
treatment variables, environments, and treatments
(Fig. 4). However, for Fe, some locations, such as
Banaras Hindu University (BHU) and Ghurahoopur
(Ghur), had similar genotypic responses (Fig. 5). The
TE interaction was significant for Zn concentration in
the grain. Zinc concentrations increased significantly
when the micronutrient was applied as a foliar spray.
Soil application alone was found not to show an
enhanced effect. The highest Zn concentration was
recorded when soil and foliar combinations were applied
together. Results suggest that foliar applications can be
utilized to increase Zn concentration in wheat grain. The
combination of soil application and foliar spray can
be used to improve grain Zn concentration and increase
grain yield, which could lead to better human nutrition
in South Asia.
Causes of Genotype
Environment Interaction
and its Effects on Grain Yield, Biomass, Yield
Components, and Other Traits in Wheat Trials
Using a Structural Equation Model
Vargas et al. (2007) showed how the structural equation
model (SEM) methodology may be applied to observed
yield GE, yield component GE, and other intermediate
traits using environmental covariates, for studying the
causes of GE and its effects on grain yield, biomass,
yield components, and other interrelated traits acting at
different development stages in wheat trials.
The given structural equation model explained 0.96 of
total variability of yield GE (Table 3). The variables that
contributed most to explaining yield GE were GEs of
yield components grains per square meter (GM2GE),
1000-kernel weight (TKWGE), grains per spike (GSPGE),
and spikes per square meter (SM2GE), with total effects of
1.09, 0.64, 0.56, and 0.54, respectively. The GEs
of GM2GE, TKWGE, GSPGE, and SPMGE explained
0.90, 0.43, 0.44, and 0.42, respectively, of total variability

1.0 RHAF 11

0.8 TMXAF
TMXBF 5
Bhad
0.6 8

0.4 6
Factor 2 (21.40%)

RHBF TMNBF
0.2 Zn_30
9 Fe_30
0.0 Pidk 12 7
RFAF
1 BHU
–0.2 TMNAF
2
–0.4 10
4
–0.6 Ghur Fe_60

–0.8 3

–1.0 RFBF Zn_60

–1.0 –0.8 –0.6 –0.4 –0.2 0.0 0.2 0.4 0.6 0.8 1.0
Factor 1 (63.75%)

Fig. 4. Partial least squares (PLS) biplot of the number of locations and treatments with environmental and soil covariables on the
performance of grain zinc concentration in wheat cultivar HUW 234 in four locations in the eastern Gangetic plains of south Asia.
Pidk Pidkhir; Ghur Ghurahoopur; Bhad Bhadawal; BHUBanaras Hindu University. The 12 treatments are: (1) control, (2)
soil, 25 kg ha 1; (3) soil, 50 kg ha 1; (4) foliar, 1.0 kg ha1 at two stages (flag leaf and anthesis); (5) foliar, 2.0 kg ha1 at two
stages; (6) foliar, 4.0 kg ha1 at two stages; (7) soil (25 kg ha 1)2 foliar (2.0 kg ha 1); (8) soil (50 kg ha1 ZnSO4)2 foliar (2.0
kg ha1); (9) soil (0 kg)2 foliar @ 0.2%; (10) soil (25 kg)2 foliar @ 0.2%; (11) soil (50 kg)2 foliar @ 0.2% (12) local farmers’
practice for Zn, i.e., 5 kg ZnSO4 ha 1. Environmental and soil variables are: TMXBF maximum temperature before flowering;
TMXAF maximum temperature after flowering; TMNBFminimum temperature before flowering; TMNAF minimum
temperature after flowering; RHBFrelative humidity before flowering; RHAF relative humidity after flowering; RBFrainfall
before flowering (RBF); RAF rainfall after flowering. Zn_30 zinc concentration in 0
30 cm soil depth; Zn_60 zinc
concentration in 30
60 cm soil depth; Fe_30 iron concentration in 0
30 cm soil depth; Fe_60 iron concentration in 30
60 cm
soil depth.
570 CANADIAN JOURNAL OF PLANT SCIENCE

1.0 11 RFAF

0.8

0.6 TMXBF
7
TMNBF
BHU10 Zn_60
0.4 3

Factor 2 (33.17%)
Pidk
0.2 2 8
RHBF
12
0.0
4
–0.2
6
Fe_60 RHAF
–0.4 Bhad
Ghur TMNAF
TMXAF RFBF
–0.6 Fe_30

–0.8 5
9
–1.0 1 Zn_30

–1.0 –0.8 –0.6 –0.4 –0.2 0.0 0.2 0.4 0.6 0.8 1.0
Factor 1 (59.11%)

Fig. 5. Partial least squares (PLS) biplot of the number of locations and treatments with environmental and soil covariables on the
performance of grain iron concentration in wheat cultivar HUW 234 in four locations in the Eastern Gangetic Plains of South Asia.
Pidk Pidkhir; Ghur Ghurahoopur; Bhad Bhadawal; BHU Banaras Hindu University. The treatments are: (1) control, (2)
soil-25 kg ha1; (3) soil, 50 kg ha1; (4) foliar, 1.0 kg ha 1 at two stages (flag leaf and anthesis); (5) foliar 2.0 kg ha1 at two stages;
(6) foliar, 4.0 kg ha 1 at two stages; (7) soil (25 kg ha1)2 foliar (2.0 kg ha 1); (8) soil (50 kg ZnSO4 ha1)2 foliar (2.0 kg ha1);
(9) soil (0 kg)2 foliar @ 0.2%; (10) soil (25 kg)2 foliar @ 0.2%; (11) soil (50 kg)2 foliar @ 0.2%; (12) local farmers’ practice for
Zn, i.e., 5 kg ZnSO4 ha1. The environmental and soil covariables are: TMXBF maximum temperature before flowering;
TMXAF maximum temperature after flowering; TMNBFminimum temperature before flowering; TMNAF minimum
temperature after flowering; RHBFrelative humidity before flowering; RHAF relative humidity after flowering; RBFrainfall
before flowering; RAF rainfall after flowering. Zn_30 zinc concentration in 0
30 cm soil depth; Zn_60 zinc concentration in
30
60 cm soil depth; Fe_30iron concentration in 0
30 cm soil depth; Fe_60 iron concentration in 30
60 cm soil depth.

(Table 3). Yield component SM2GE had a very small R2 components GM2GE and TKWGE had the largest posi-
value (0.04), but a significant indirect effect on grain yield tive direct association with yield GE (1.09 and 0.64,
GE (0.54). The model indicated that GEs of yield respectively) and no indirect effects (0.0), while GSPGE

Table 3. Direct, indirect, and total effects of yield component GEs and adjusted cross-product covariates on grain yield GE (R2 0.96) [extracted from
Vargas et al. (2007)]

Variable Direct effect Indirect effect Total effect R2

Grains per square meter (GM2GE) 1.09 0.00 1.09 0.90

1000-kernel weight (TKWGE) 0.64 0.00 0.64 0.43
Grains per spike (GSPGE) 0.05 0.61 0.56 0.44
Spikes per square meter (SM2GE) 0.00 0.54 0.54 0.04
Spike mass (SPMGE) 0.00 0.05 0.05 0.42
Relative duration of spike growth (RSGGE) 0.00 0.09 0.09
Crop growth rate during spike growth (dBMbGE) 0.00 0.07 0.07
Biomass at anthesis (BMAGE) 0.00 0.03 0.03
Biomass at the vegetative stage (BMVGE) 0.00 0.11 0.11
MXT4GM2zGE 0.00 0.39 0.39
MXT4GSPGE 0.00 0.23 0.23
RAD2SM2GE 0.00 0.09 0.09
MNT4TKWGE 0.00 0.59 0.59
RAD2TKWGE 0.00 0.40 0.40
MXT3BMAGE 0.00 0.10 0.10
MNT1BMAGE 0.00 0.01 0.01

z
MXTmean daily maximum temperature; MNTmean daily minimum temperature; RADsolar radiation. The suffixes 1, 2, 3, and 4 denote the
first, second, third, and fourth growth developmental stages, respectively.
 CROSSA ET AL. * ANALYZING GENOTYPE ENVIRONMENT INTERACTION 571

and SM2GE GEs had the greatest indirect effects on
yield GE (0.61 and 0.54, respectively) and a low negative
direct effect (GSPGE 0.05) or no direct effect at all
(SM2GE 0.0) on yield GE (Table 3).
The SEM using GE effects is a powerful method that
gives a more complete overview of the external and
internal variables acting and interacting in the GE
of various traits than do the PLS or FR methods. In
the current analysis, climatic variables were related
mostly to final main yield components GM2GE, GSPGE,
and TKWGE. Only more intermediate endogenous
spike mass (SPMGE) and SM2GE were affected by
minimum temperature in the first developmental stage
(MNT1), along with solar radiation in the second
developmental stage (RAD2). SEM analysis of GE of
variables showed that weather conditions during the
spike primordia and early grain-filling stages influence
GE of other traits; this result is consistent with PLS
results obtained by Reynolds et al. (2004). Furthermore,
the result of SEM analysis of GE effects shows the
influence of biomass at anthesis (BMAGE) on SPMGE
and of MNT1 on SPM.
Interpreting Genotype
Environment Interaction
of Fruit Yield in a Tomato Multi-Environment
Trial
This example describes results from a multi-environment
trial comprising 15 tomato genotypes [seven hybrids (H)
and eight open-pollinated (OP)] evaluated in 18 locations
of Latin America and the Caribbean; the results were
analyzed using FR and PLS (Ortiz et al. 2007). Several
environmental covariables were included in the FR and
PLS analyses for studying and interpreting GE. Envir-
onmental factors, such as days to harvest, soil pH, mean
temperature, potassium available in soil, and phospho-
rus fertilizer, accounted for a sizeable portion of GE for
marketable fruit yield, whereas trimming, irrigation, soil
organic matter, and nitrogen and phosphorus fertilizers
were important environmental covariables for explaining
GE of average fruit weight.
The factorial regression model with a stepwise regres-
sion procedure for variable selection was used to
determine the most informative subset of environmental
covariables affecting marketable fruit yield. The subset
of independent environmental covariables that ex-
plained 62% of total GE included days to harvest
(DHA gen), soil pH (PH gen), mean temperature
(MET gen), potassium (K gen), extra phosphorus
(EX_P gen), and minimum temperature (MNT gen)
(Table 4). Days to harvest (DHA) and soil pH (PH)
jointly explained 34% of total GE variability with only
28 degrees of freedom (from a total of 238 degrees of
freedom). The ability to use nitrogen fertilizer (EX_N)
explained a small portion of GE variability. Severe N
stress can reduce tomato fruit yield by 60 to 70%
(Scholberg et al. 2000). The remaining environmental
covariables were significant, but did not explain much of
the GE variability for marketable fruit yield.
The first two PLS factors with all 15 tomato
genotypes evaluated in 18 environments along with 16

0.08
0.36 RSGGE
MXT4*GM2GE

–0.37
MXT4*GSPGE GM2GE
dBMdGE 0.13 1.04
0.31
1.09
RAD2*SM2GE SM2GE 1.01
0.16 –0.02
–0.56

MNT1*BMAGE HIAGE
–0.37
YLDGE
–0.21 GSPGE
–0.05

BMVGE –0.14
–0.76
–0.43
0.64
0.92
MNT4*TKWGE
TKWGE
–0.63
RAD2*TKWGE

0.15
MXT3*BMAGE

Fig. 6. Path estimates of the structural equation model for endogenous variables associated with grains per square meter GE
(GM2GE), grains per spike GE (GSPGE), 1000-kernel weight GE (TKWGE), spikes per square meter GE (SM2GE), spike mass GE
(SPMGE), relative duration of spike growth GE (RSGGE), crop growth rate during spike growth GE (dBMbGE), biomass at anthesis
GE (BMAGE), biomass at the vegetative stage GE (BMVGE), and yield GE (YLDGE), and cross-products (variables 
environmental covariates) [extracted from Vargas et al. (2007)]. MXTmean daily maximum temperature; MNT mean daily
minimum temperature; RAD solar radiation. Suffixes 1, 2, 3, and 4 stand for the first, second, third, and fourth crop development
stages. Arrows represent the direction of the variables’ influence, and the numbers on the arrow lines represent the estimated
standardized coefficients.
572 CANADIAN JOURNAL OF PLANT SCIENCE

Table 4. Analysis of variance for the stepwise multiple factorial regression model with environmental covariables for marketable fruit yield. The terms in
the factorial regression model appear in the order of inclusion [extracted from Ortiz et al. (2007)]

Sourcez df Sum of squares Mean squares ProbF % of GE explained

Environment 17 701980 41293 B0.001

Genotype 14 31669 2262 B0.001

GE 238 160674 675 B0.001

DHAGen 14 36363 2597 B0.001 22.63
PHGen 14 17989 1285 B0.001 11.19
METGen 14 14854 1061 B0.001 9.24
MNTGen 14 7627 545 B0.001 4.74
OMGen 14 6054 432 B0.001 3.76
MXTGen 14 5713 408 B0.001 3.56
IRRGen 14 6796 485 B0.001 4.23
PRC Gen 14 5802 414 B0.001 3.61
TRMGen 14 4100 293 B0.001 2.55
DRIGen 14 5092 364 B0.001 3.16
EX_NGen 14 5459 390 B0.001 3.39
EX_PGen 14 8108 579 B0.001 5.05
P Gen 14 6306 450 B0.001 3.92
EX_KGen 14 4114 294 B0.001 2.56
DAYGen 14 7157 511 B0.001 4.45
K Gen 14 15013 1072 B0.001 9.34
Residual 14 4123 294
Total 269 894324 3324

z
GEgenotypeenvironment; MXTmaximum temperature in (8C); MNTminimum temperature in (8C); METmean temperature in (8C);
PRCrainfall (mm); DAYdegree day; PHsoil pH; OMorganic matter; P phosphorus; K potassium; EX_Nextra nitrogen; EX_P
extra phosphorus; EXT_Kextra potassium; TRMtrimming; DRI drivings; IRRirrigation; DHAdays to harvest.

environmental covariables are depicted in Fig. 7.
The environmental covariables that most explained
GE in the FR analyses (DHA, PH, MET, MNT,
EX_P, and K) tend to be located farther away from
the center of the PLS biplot, indicating that these
covariables caused large GE for marketable fruit yield,
as had already been detected by FR analysis.
The PLS biplot for marketable fruit yield shows
general patterns in GE with respect to environments,
genotypes, and environmental covariables. Environ-
ments located to the right of the PLS biplot (Es_Gu,
Co_Sa, SA_Sa, Co_Ho, VS_Ni, SA_CR, SC_DR,
Pa_Co, Be_Br, and Ce_TT) had relatively higher values
for environmental covariables located in the same
direction, MET, MNT, and DAY, whereas sites located
on the opposite side of the biplot (BV_Gu, Co_DR,
LM_Pe, Sa_Ch, Ch_Ch, Cu_Ch, Co:Ch, and Ca_Pa)
tended to have high PH and DAH values. Concerning
genotypes, the first PLS axis clearly separates the hybrid
tomato genotypes (5, 6, 7, 8, 9, 11, and 15) (located
towards the left) from the open-pollinated tomato
genotypes (on the right) (1, 2, 3, 4, 10, 12, and 13),
whereas the second PLS axis separates open-pollinated
tomato genotypes 1, 2, and 4 from genotypes 3, 10, 12,
and 13. These results indicate that, in terms of GE,
open-pollinated genotypes 1, 2, and 4 performed better
in environments Es_Gu, Ca_Sa, SA_Sa, SC_DR, and
Be_Br, whereas open-pollinated genotypes 4, 10, 12, and
13 performed better in environments Co_Ho, VS_Ni,
SA_CR, Pa_Co, and Ce_TT (i.e., they tended to have
positive GE in those sites) and thus are favored by the
relatively higher values of DAY, MNT, and MET
prevailing in those environments.
The PLS biplots show more specific GE between
genotypes and environments. The covariables MNT and
MET are in the same direction as environments Es_Gu
(Estanzuela), Co_Sa (Cogutepeque), SA_Sa (San
Andrés), VS_Ni (Valle del Sábaco), SC_DR (San
Cristobal), Pa_Co (Palmira), and Be_Br (Belem), in-
dicating that these locations had relatively high mini-
mum temperature [MNT] and mean temperature
[MET]; these conditions favored the marketable fruit
yield of AVRDC OP heat-tolerant lines CL 5915-223
(12) and CL 5915-93 (13), which are located in the same
direction. The reproductive processes in tomato are
sensitive to high temperatures (Abdul-Baki 1991), and
the number of pollen grains in heat tolerant genotypes
was higher than that of heat sensitive genotypes. It
appears that proline accumulates in tomato leaf tissue at
high temperatures, which leads to its depletion in the
reproductive tissue, thus seriously reducing either pollen
formation or its viability (Kuo et al. 1986).
The amount of potassium (K) in the soil of environ-
ments Cogutepeque (Co_Sa) and Belém (Be_Br) was
relatively high, which favored positive GE interaction of
OP cultivar Triuque (4) in both locations. The soil in
Comayagua (Co_Ho), San Antonio de Belen (SA_CR),
and Centeno (Ce_TT) had relatively high organic matter
(OM) content; these environments are in the same
direction in the biplot (Fig. 7), which favored the
positive GE of OP cultivars Licapal 21 (3) and Angela
Gigante (10) in these locations. Since the first two PLS
 CROSSA ET AL. * ANALYZING GENOTYPE ENVIRONMENT INTERACTION 573

1.0
1
0.8 DAY

P
0.6 PH LM_Pe
2SA_SaMNTBe_Br
95 PRC
0.4 11 Co_Sa

Factor 2 (16.17%)
Sa_Ch 8 Cu_Ch
7 SC_DR
K
0.2 6 Es_Gu
15
4
0.0 Ch_Ch EX_N
Co_DR 14 VS_Ni
Pa_Co
MET
EX_K
–0.2
13
DHA Co_Ch Co_Ho
–0.4 IRR 12
BV_Gu

–0.6 EX_P MXT SA_CR

Ce_TT
OM 3
–0.8 Ca_Pa
TRM DRI

–1.0 10

–1.0 –0.8 –0.6 –0.4 –0.2 0.0 0.2 0.4 0.6 0.8 1.0
Factor 1 (18.52%)

Fig. 7. Plot of the first two partial least squares regression factors (factor 1 and factor 2) for marketable fruit yield of tomato for 15
cultivars tested across 18 environments in Latin America and the Caribbean (extracted from Ortiz et al. 2006). MXTmaximum
temperature in (8C); MNT minimum temperature in (8C); MET, mean temperature in (8C); PRCrainfall (mm); DAY degree
day (base 10); PHsoil pH; OM organic matter (%); Pphosphorus (P2O5 ppm); K potassium (K2O meq 100 g 1); EX_N
extra nitrogen (kg ha1); EX_Pextra phosphorus (kg ha1); EXT_K extra potassium (kg ha1); TRM trimming;
DRI drivings; IRRirrigation; DHA days to harvest. Estanzuela, Guatemala (Es_Gu), Baja Verapaz, Guatemala, (BV_Gu),
Cogutepeque, El Salvador (Co_Sa), San Andrés, El Salvador (SA_Sa), Comayagua, Honduras (Co_Ho),Valle de Sabaco,
Nicaragua (VS_Ni), San Antonio de Belén, Costa Rica (SA_CR), San Cristóbal, Dominican Republic (SC_DR), Constanza,
Dominican Republic (Co_DR), Palmira, Colombia (Pa_Co), La Molina, Perú (LM_Pe), Santiago, Chile (Sa_Ch), Chillán, Chile
(Ch_Ch), Curacavı́, Chile (Cu_Ch), Colina, Chile (Co_Ch), Belém, Brazil (Be_Br), Caacupé, Paraguay (Ca_Pa), Centeno, Trinidad
Tobago (Ce_TT).

factors do not explain all the GE for marketable fruit genotypes without crossover genotype environment interac-
yield, some distortion occurred, e.g., environment tion. Crop Sci. 48: 1291
1305.
Co_DR (Constanza) has relatively high OM content, Cornelius, P. L., Crossa, J. and Seyedsadr, M. 1996. Statistical
but is not in the same direction as OM in the PLS tests and estimators for multiplicative models for cultivar
biplot. trials. Pages 199
234 in M. S. Kang and H. G. Gauch, Jr., eds.
Genotype-by-environment interaction. CRC Press, Boca
Raton, FL.
Cornelius, P. L. and Seyedsadr, M. 1997. Estimation of general
linear-bilinear models for two-way tables. J. Stat. Comput.
ACKNOWLEDGEMENTS
Sim. 58: 287
322.
The authors acknowledge the time and effort contrib- Crossa, J., Burgueño, J., Cornelius, P. L., McLaren, G.,
uted by two anonymous reviewers, which substantially Trethowan, R. and Krishnamachari, A. 2006. Modeling gen-
improved the quality of the manuscript. The authors are otype environment interaction using additive genetic covar-
grateful to national programs that carried the trials and iances of relatives for predicting breeding values of wheat
originated the data analyzed and presented in this genotypes. Crop Sci. 46: 1722
1733.
manuscript. Denis, J.-B. 1988. Two-way analysis using covariates. Statistics
19: 123
132.
Abdul-Baki, A. A. 1991. Tolerance of tomato cultivars and Dhungana, P. 2004. Structural equation modeling of gen-
selected germplasm to heat stress. J. Am. Soc. Hortic. Sci. 116: otype environment interaction. Ph.D. Dissertation, Univer-
1113
1116. sity of Nebraska, Lincoln, NE.
Aastveit, H. and Martens, H. 1986. Anova interactions Eberhart, S. A. and Russell, W. A. 1966. Stability parameter for
interpreted by partial least squares regression. Biometrics 42: comparing varieties. Crop Sci. 6: 36
40.
829
844. Finlay, K. W. and Wilkinson, G. N. 1963. The analysis of
Burgueño, J., Crossa, J., Cornelius, P. L. and Yang, R. C. 2008. adaptation in a plant breeding programme. Austr. J. Agric.
Using factor analytic models for joining environments and Res. 14: 742
754.
574 CANADIAN JOURNAL OF PLANT SCIENCE
Gabriel, K. R. 1978. Least squares approximation of matrices
by additive and multiplicative models. J. R. Stat. Soc. Ser. B.
40: 186
196.
Gauch, H. G., Jr. 1988. Model selection and validation for
yield trials with interaction. Biometrics 44: 705
715.
Gauch, H. G., Piepho, H.-P. and Annicchiarico, P. 2008.
Statistical analysis of yield trials by AMMI and GGE: Further
considerations. Crop Sci. 48: 866
889.
Gollob, H. F. 1968. A statistical model which combines
features of factor analytic and analysis of variance. Psychome-
trika 33: 73
115.
Helland, I. S. 1988. On the structure of partial least squares.
Commun. Stat. Part B Sim. Comput. 17: 581
607.
Henderson, C. R. 1984. Applications of linear models in animal
breeding. University of Guelph. Guelph, ON.
Joshi, A. K., Crossa, J., Arun, B., Chand, R., Trethowan, R.,
Vargas, M. and Ortiz-Monasterio, I. 2010. Genotype envir-
onment interaction for zinc and iron concentration of wheat
grain in the eastern Gangetic plains of India. Field Crop Res.
116: 268
277.
Kuo, C. G., Chen, H. M. and Ma, L. H. 1986. Effect of high
temperature on proline content in tomato floral buds and
leaves. J. Am. Soc. Hortic. Sci. 111: 746
750.
Mandel, J. 1969. The partitioning of interaction in analysis of
variance. J. Res. Natl Bur. Stand. Ser. B 73: 309
328.
Mandel, J. 1971. A new analysis of variance model for non-
additive data. Technometrics 13: 1
18.
Ortiz, R., Crossa, J., Vargas, M. and Izquierdo, J. 2007.
Studying the effect of environmental variables on the gen-
otypeenvironment interaction of tomato. Euphytica 153:
119
134.
Piepho, H. P. 1998. Methods for comparing the yield stability
of cropping systems
a review. J. Agron. Crop Sci. 180:
193
213.
Piepho, H. P. and Mohring, J. 2005. Best linear unbiased
prediction of cultivar effects for subdivided target regions.
Crop Sci. 45: 1151
1159.
Reynolds, M. P., Trethowan, R., Crossa, J., Vargas, M. and
Sayre, K. D. 2004. Physiological factors associated with
genotype by environment interaction in wheat. Field Crop
Res. 85: 253
274.
Scholberg, J., McNeal, B. L., Boote, K. J., Jones, J. W.,
Locascio, S. J. and Olson, S. M. 2000. Nitrogen stress effects
on growth and nitrogen accumulation by field-grown tomato.
Agron. J. 92: 159
167.
Smith, A., Cullis, B. R. and Thompson, R. 2002. Exploring
variety-environment data using random effects AMMI models
with adjustment for spatial field trends: Part 1: Theory. In
M. S. Kang, ed. Quantitative genetics, genomics and plant
breeding. CABI Publishing, Oxford, UK.
Smith, A. B., Cullis, B. R. and Thompson, R. 2005. The analysis
of crop cultivar breeding and evaluation trials: An overview of
current mixed model approaches. J. Agric. Sci. 143: 1
14.
Vargas, M., Crossa, J., Reynolds, M. P., Dhungana, P. and
Eskridge, K. M. 2007. Structural equation modelling for
studying genotype per environment interactions of physiolo-
gical traits affecting yield in wheat. J. Agric. Sci. 145:
151
161.
Vargas, M., Crossa, J., van Eeuwijk, F. A., Ramirez, M. E. and
Sayre, K. 1999. Using partial least squares, factorial regression
and AMMI models for interpreting genotype environment
interaction. Crop Sci. 39: 955
967.
Vargas, M., Crossa, J., van Eeuwijk, F. A., Sayre, K. and
Reynolds, M. P. 2001. Interpreting treatment  environment
interaction in agronomy trials. Agron. J. 93: 949
960.
van Eeuwijk, F. A., Denis, J.-B. and Kang, M. S. 1996.
Incorporating additional information on genotypes and en-
vironments in models for two-way genotype by environment
tables. In S. Kang and H. G. Gauch, eds. Genotype-by-
environment interaction, CRC Press, Boca Raton, FL.
van Eeuwijk, F. A., Crossa, J., Vargas, M. and Ribaut, J. M.
2000. Variants of factorial regression for analysing QTL by
environment interaction. Proceedings of the 11th Meeting of
the EUCARPIA Section Biometrics in Plant Breeding. In A.
Gallais, C. Dillmann, and I. Goldringer, eds. Quantitative
genetics and breeding methods: the way ahead. Paris, France.
Williams, E. J. 1952. The interpretation of interactions in
factorial experiments. Biometrika 39: 65
81.
Yan, W., Kang, M. S., Ma, B., Woods, S. and Cornelius, P. L.
2007. GGE biplot vs. AMMI analysis of genotype-by-
environment data. Crop Sci. 47: 643
655.
Yang, R. C., Crossa, J., Cornelius, P. L. and Burgueño, J. 2009.
Biplot analysis of genotype environment interaction: Pro-
ceed with caution. Crop Sci. 49: 1564
1576.
Yates, F. and Cochran, W. G. 1938. The analysis of groups of
experiments. J. Agric. Sci. 28: 556
580.