Вы находитесь на странице: 1из 17

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/250117441

Identifying Mega-Environments and Targeting


Genotypes

Article in Crop Science January 1997


DOI: 10.2135/cropsci1997.0011183X003700020002x

CITATIONS READS

169 228

2 authors, including:

Richard W. Zobel
United States Department of Agriculture
85 PUBLICATIONS 3,312 CITATIONS

SEE PROFILE

All content following this page was uploaded by Richard W. Zobel on 05 February 2016.

The user has requested enhancement of the downloaded file.


Published March, 1997

Identifying Mega-Environments and Targeting Genotypes


Hugh G. Gauch, Jr.,* and Richard W. Zobel

ABSTRACT locations in the same mega-environment (but of course


To maximizeyield throughout a crops heterogeneous growing re- not outside the locations own mega-environment).
gion, despite differences in cultivar rankings from place to place due The term mega-environment appears to have been
to genotype-environmentinteractions, frequently it is necessary to coined by researchers at CIMMYT. "Mega-environments
subdivide a growingregion into several relatively homogeneous mega- are broad, not necessarily contiguousareas, usually inter-
environments and to breed and target adapted genotypes for each national and frequently transcontinental, defined by simi-
mega-environment. Theobjectives of this study are to identify relevant lar biotic and abiotic stresses, cropping system require-
criteria for evaluating mega-environmentanalyses and to apply the ments, consumer preferences, and, for convenience, by
Additive Main Effects and Muitiplicative Interaction (AMMI) model a volume of production of the relevant crop sufficient
to mega-environmentanalysis. The proposed analysis is illustrated to justify ... attention," for example, "tropical lowland,
using a Louisianacorn (Zea maysL.) trial. Statistical strategies for late-maturing, white dent" corn with relevant disease
identifying mega-environments should meet four criteria: flexibility
resistances, which occupies 3.8 million hectares across
in handlingyield trials with various designs, focus on that fraction of
the total variation that is relevant for identifying mega-environments, 18 countries (CIMMYT,1989, p. 58). Note that this
duality in giving integrated informationon both genotypesandenviron- definition encompasses environmental, genotypic, geo-
ments, and relevance for the primary objective of showing which graphical, and even economic aspects of mega-envi-
genotypes win where. The AMMI modelmeets these criteria effectively ronments. Mega-environments are used to allocate re-
whenthe usual biplots are supplemented with several new types of sources in a breeding or research program, to rationalize
graphs designed to address questions about mega-environments. Pre- germplasm and information exchanges between breeding
liminary results indicate that a small and workablenumberof mega- programs (allowing even small programs to progress by
environmentsoften suffices to exploit interactions andincrease yields. focusing on the most promising material), to increase
heritabilities within relatively well-defined and predict-
able environments, to increase the efficiency of testing
N A REGIONAL YIELD TRIAL with heterogeneous envi- and breeding programs, and to target genotypes to appro-
.ronments and numerous genotypes, different genotypes priate production areas (Brown et al., 1983; Peterson
are superior in different locations (or years or both). and Pfeiffer, 1989; Abdalla et al., 1996). Manyother
optimize growers yields, despite genotype-environment terms, however, have essentially the same meaning, such
interactions that cause no one genotype to win every- as agroclimatic or ecogeographic regions. Because the
where and always, the growing region must be subdivided need for mega-environments arises from genotype-envi-
into relatively homogeneousmega-environments and ap- ronment interactions and furthermore these interactions
propriate genotypes must be targeted for each of these involve both genotypes and environments, there are also
mega-environments. A mega-environment is defined as corresponding terms applied to genotypes for identifying
a portion (not necessarily contiguous) of a crop species adaptation traits, such as the maturity groups in corn
growing region with a fairly homogeneousenvironment and soybean [Glycine max (L.) Merr.] that are used
that causes similar genotypes to perform best. With no for delineating adaptation zones. Virtually the entire
literature on interaction, clustering environments, yield
subdivision, only broad adaptation can be exploited,
but with subdivision, narrow adaptations can also be stability and dependability, wide and narrow adaptation,
exploited. Subdivision of a crops growing region into and heritability is relevant for understanding and charac-
terizing mega-environments, whether or not the word
several mega-environments implies more work for plant
mega-environment appears in a given paper.
breeders and seed producers, but subdivision also implies
Subdivision of a crops growing region into several
higher heritabilities and faster progress for plant breed-
mega-environments could be avoided if genotypes could
ers, potentially stronger competitiveness for seed produc-
be found with yield superiority throughout the region,
ers, and higher yields for growers. For example, Annic- meaning that cultivars bred in favorable environments
chiarico (1992) compared alfalfa (Medicago sativa L.) wouldalso perform best in different or unfavorable envi-
yields in three mega-environments in Italy with yields for ronments. But this expectation assumes that "the same
the overall regional averages. The spread and statistical genetic system controls yield" in diverse environments,
significance of yield differences Wasgreater in the indi- contrary to extensive evidence from genetics, physiol-
vidual mega-environmentsthan in the overall averages, ogy, and yield trials (Ceccarelli and Grando, 1993; Cec-
and the highest yielders in each mega-environmentwere carelli, 1989; Simmonds,1991). Consequently, for many
different from the overall winner. Also, results from a crops, acceptable yields have necessitated targeting nar-
given location had good predictive reliability for other rowly adapted genotypes to several different, well-

H.G. Gauch, Jr., Soil, Crop and Atmospheric Sciences, Cornell Univ., Abbreviations:AMMI, Additive MainEffects and Multiplicative Interac-
1021 Bradfield Hall, Ithaca, NY14853-1901; and R.W. Zobel, USDA- tion; AOV,analysis of variance; CIMMYT, Centro Internacional de
ARS-NAA, 1017 Bradfield Hall, Ithaca, NY14853-1901. Received 19 Mejoramientode Maizy Trigo; df, degrees of freedom; IPCA,interaction
April 1996. *Corresponding author (hggl@comell.edu). PCAaxis; PCA,principal componentsanalysis; S/N, signal-to-noise; SS,
sum of squares; G, genotype; E, environment; L, location; Y, year; R,
Published in Crop Sci. 37:311-326 (1997). replication.

311
312 CROPSCIENCE, VOL. 37, MARCH-APRIL
1997

defined mega-environments.Indeed, one can hardly ex- is especially clear whendevelopmentof more widely
pect a single cultivar of a crop to flourish the world adaptedcultivars revises andsimplifies a previousmega-
over, under all environmentsand management practices environment classification. Indeed,if differencesin envi-
(Rosielle and Hamblin,1981; Ceccarelli, 1989). A culti- ronmentalfactors fromplace to place (and year to year)
var planted outside its mega-environmentfrequently affected only environmentmaineffects but not genotype-
suffers yield reductions. Furthermore,evenif the breed- environment interactions, then these environmental
ing goal is wide adaptation (rather than mega-envi- differences wouldbe of interest to crop physiologists
ronmentdirected breeding as in Abdallaet alo, 1996), and agricultural economistsbut not plant breeders. Mega-
the best strategy could be to identify several mega- environmentdifferences exert their importance - for
environmentsand place a test location in each to select plant breeders - by changing genotype rankings from
for wide adaptation. one mega-environmentto another. Accordingly, this
For several reasons, identifying mega-environments studys title presents "identifying mega-environments"
has attracted greater attention recently. Interest has and "targeting genotypes"as deeply and intrinsically
grownin providing adapted material for marginal envi- interrelated tasks.
ronments,particularly becausemillions of people live The mainstatistical strategy for groupinglocations
in such places and yet they often lack the economic into mega-environments has been various classification
meansneededto purchaselarge quantities of grain from and ordination methods(or both, whichis termedpattern
moreproductiveagricultural areas. Thesemarginalenvi- analysis). Amongthese numerousordination methods,
ronmentsare stressed in a variety of waysthat typically the AMMI model has provided biplot graphs useful for
generate large genotype-environmentinteractions, so delineating mega-environments(Annicchiarico, 1992;
genotypesthat win in highly favorable environmentsmay Annicchiaricoand Perenzin, 1994; Annicchiaricoet al.,
rank poorly there (Ceccarelli, 1989; Simmonds,1991; 1995). Despite this success, however,the usual biplot
Nachitet al., 1992;Zavala-Garcia et al., 1992;Ceccarelli graphsreveal only a portion of the informationavailable
and Grando, 1993; Annicchiarico and Perenzin, 1994; from AMMI that bears on mega-environments.In partic-
Falconer, 1989, p. 313-335). On the other hand, within ular, these graphs fail to makeevident whichgenotype
the favorable environmentsof major production areas, has the highest yield in various domainsof the AMMI
conventional managementprocedures have produced in graphs. Accordingly, this paper presents several new
the past a rather coherent environmenteasily targeted extensions of AMMI that enhance its informativeness.
by plant breeders. In the future, however,it is likely The merits of this new methodologyare (i) focus
that greater concernabout long-term soil conservation that fractionof the total yield variationthat is relevantfor
and about loweringpesticide and fertilizer usages will identifying mega-environments and targeting genotypes,
stimulate a greater diversity of management practices, (ii) duality in giving integratedinformationon both geno-
hence creating a variety of mega-environmentswithin types and environments,and (iii) relevancefor the pri-
major crop regions previously managedin a muchmore mary objective of showingwhich genotypes win where.
uniform manner(Smith and Zobel, 1991). Finally and This analysis is also flexible in handlingyield trials with
moregenerally, most plant breeders feel that they are various designs. Preliminaryresults indicate that a small
exploiting rather than ignoring the potential for yield and workable number of mega-environments often
increases that resides in genotype-environment interac- suffices to exploit interactions and increase yields.
tions (Cooperet al., 1993). Nevertheless,despite breed-
ers strong interests in interactions andspecific adapta-
tions, one can question whetherthey have routinely had FOUR CRITERIA
adequatestatistical tools to exploit interactions aggres- Rational assessments of mega-environmentanalyses
sively and to grasp clearly the inevitable implications are possibleonly after specifyingthe criteria for evalua-
for identifying mega-environments. tion. Fourkeycriteria are emphasized here: (i) selective
The primary proposal adopted here for interlinking focuson the variation in yield that is relevant for identi-
mega-environmentsand genotypes is to define a mega- fying mega-environments, (ii) relevance to agronomists
environmentas a portion of a crops overall growing and breeders questions, especially "Whatwinswhere?",
region in whicha particular genotype wins. A comple- (iii) dual analysis of both genotypesand environments,
mentaryproposalis to define mega-environments in terms and(iv) flexibility with handlingvariousdata structures.
of major environmentalfactors, such as irrigated low- First, mega-environment analyses should focus on rel-
lands and nonirrigated uplands. These genotype-based evantvariationin the data andignoreirrelevant variation.
and environment-basedproposals are interdependent and Considera regional variety trial fromthe perspective of
complementary(Peterson and Pfeiffer, 1989; DeLacyet a breeder makingselections or an agronomist offering
al., 1994). The genotype-basedconcept is relevant be- recommendations.Analysis of variance partitions the
cause planting the winning genotype in each mega- treatment variation into three basic sources: genotypes,
environmentoptimizes yield. Andthe environment-based environments, and genotype-environmentinteraction.
concept is helpful for assigning individual farms to an Here a treatment is defined as a genotypeand environ-
appropriate mega-environment (and hence genotype rec- mentcombination.The genotypeand interaction sources
ommendation), especially if the mainenvironmentalfac- affect genotype rankings within each environmentand
tor has an obviousand simple geographicaldistribution. hence are relevant for identifying mega-environments
This interlinking of mega-environmentsand genotypes and targeting genotypes. But the environment main
GAUCH& ZOBEL: IDENTIFYING MEGA-ENVIRONMENTS 313

Table 1. AMMI-1analysis of variance table for the Lousiana corn


effects are wholly irrelevant for genotype rankings and yield trial. The grand mean is "6.69324 Mgha-~.~
hence for present objectives. Because of the central goal
of makinggood selections, "the breeder is particularly Source df SS MS Probability
interested in the rank orders of the genotypes in the Total 989 5336.17 5.396
different environments" (Hiihn et al., 1993). Fox and Treatments 255 3842.13 15.067 0.0000000
Genotypes 15 238.93 15.929 0.0000000
Rosielle (1982), Gauch (1992, p. 220-230), Cooper Environments 15 2865.17 191.011 0.0000000
DeLacy(1994), and Annicchiarico et al. (1995) strongly G x E 225 738.03 3.280 0.1)000019
emphasize that breeding decisions and mega-environment IPCA 1 29 229.81 7.925 0.0000000
Residual 196 508.22 2.593 0.0139627
choices dependon genotype and interaction effects exclu- Error 734 1494.04 2.035
sively (Yau, 1991). Often these relevant effects account
AMMI = Additive M__ainEffects and Multiplieative I_nteraction; df =
for just 10 to 40%of the treatment variation (because degrees of freedom; SS = sum of squares; MS= mean square; IPCA=
environmentmain effects account for the remaining varia- interaction principal components
analysis axis.
tion), so a statistical analysis modelingall of the variation
is mostly capturing irrelevant features of the data, which al. (1993) that this exampleis typical. Again,the relevant
greatly hinders effective mega-environment identification. variation for identifying mega-environmentsis often only
For example, Kang (1993) kindly supplied us the data ~ 10 to 40%of a yield trials treatment variation.
from a Louisiana corn yield trial. This experiment had a Incidentally, because the interaction has most of the
randomizedcomplete-block design with four replications treatment df and hence noise but has only a fraction of
(except that 32 treatments had only three replicates and the treatment SS, interaction is muchnoisier than the
1 had only two replicates). Because of the missing data, data (yield averages across replications). The data
for convenience, the analysis is approximated here as a of 3842.13 with 255 df contain 255 2.035 = 518.93
completely randomized design [but Piepho (1994) ex- noise and 3323.20 pattern, so its S/N ratio is 3323.20/
plains the calculations for randomized complete blocks 518.93 = 6.40. But the interaction S/N ratio is only
without missing data]. Table 1 gives the analysis of 280.15/457.88 = 0.61, which is ten times smaller. The
variance table for the AMMI-1 model. The SS for treat- general lesson to be learned here is that the interaction
ments is 3842.13. However, the goal of a mega- gets buried by noise more quickly than genotype and
environment analysis should not be to account for this environment main effects. Consequently, the breeding
entire treatment SS. Rather, the relevant portion of this strategy of sampling more locations with fewer replica-
SS, ignoring the environmentmain effect, is only 238.93 tions to characterize interactions better can be self-
+ 738.03 = 976.96, or 25.4%. defeating if there is not enoughaccuracy to makeinterac-
Furthermore, errors from uncontrolled variation tions rise abovethe noise. Yield trials with elite genotypes
within experimental fields are also irrelevant for present having relatively small interactions that are conducted
concerns and unfortunately cause inaccuracy or noise in as unreplicated tests are especially prone to this problem.
yield estimates. Because interaction has many more df Mega-environment analysis should capture genotype
than the genotype and environment main effects, most and (real) interaction effects but should also distinguish
of this noise appears in the interaction. (The grand mean, these two effects for reasons that are agricultural rather
genotype deviations, and environment deviations also than statistical. Genotypeeffects control wide adapta-
contain some noise, but these problems are neglected tions, whereasinteraction effects control narrow adapta-
here because this other noise is small comparedwith the tions, and these two kinds of adaptations offer breeders
interaction noise and because there is no way to removea different problems and opportunities, especially in re-
sizable portion of this other noise.) Noisegives yield-trial gards to designing testing systems and making good
results a spurious complexity, thereby multiplying the selections. Also, genotype and interaction effects often
number of mega-environments needlessly. Conse- relate to different environmental and genetic systems.
quently, reducing noise helps when identifying mega- For example, genotype main effects may involve water
environments. and nutrient uptake, whereas interaction effects may
For this corn trial, the error mean square is 2.035 involve photoperiod response. For these reasons, it is
and the interaction has 225 df, so the interaction contains preferable for a mega-environmentanalysis not to con-
approximately 225 2.035 = 457.88 noise SS and found the genotype and interaction SS but rather to
738.03 - 457.88 = 280.15 real structure SS or 62.0% treat them separately. Althoughgenotype and interaction
noise and 38.0% pattern (Gauch, 1992, p. 147). Dis- effects should be distinguished, they should also be inte-
counting for this interaction noise, the relevant variation grated in regards to their impact on yields because yield
is only 238.93 + 280.15 = 519.08 or 13.5% of the rankings are affected by these two effects jointly. So,
treatment SS. So, an ideal mega-environment analysis genotype and interaction effects should be distinguished,
would focus on this 13.5%of relevant variation in the not confounded, and should be integrated, not divorced.
data, and ignore the other 86.5 %pertaining to irrelevant Second, mega-environment analyses should provide
environment effects and interaction noise. By contrast, relevant answers to researchers primary questions. The
capturing less than this 13.5 %underfits real and relevant obvious question is: What wins where? An answer to
patterns in the data, and capturing morethan this 13.5 % this question provides a basis for grouping locations with
overfits noise and captures irrelevant features in the data. identical (or at least similar) winners into a mega-
One gathers from extensive results in Talbot (1984), environment and for targeting suitable genotypes for
DeLacyet al. (1990), Braun et al. (1992), and HiJhn each mega-environment. Mega-environment analyses
314 CROPSCIENCE, VOL. 37, MARCH-APRIL
1997

that answer questions other than "Whatwins where?" possibilities. Naturally, moreextensive data allow more
should be carefully checkedto determinewhether their extensive results, but smaller experimentsalso merit
answersbear effectively on the objective of subdividing effective analysis. For a trial with GLY structure, re-
a crops growingregion to increase yields. Otherwise, peated across both locations and years, mega-envi-
laborious calculations mayoptimizea mathematicalprop- ronmentanalysis should discriminate betweenpredict-
erty of dubiousor tangential relevanceto the agricultural able and nonpredictableinteractions, largely associated
task at hand. with locations andyears, respectively.
Interactionsdiffer in their relevancefor plant breeders. Whathappenswhenthese four criteria are neglected?
Breedershave emphasizedcrossover interactions because (i) Clusteranalysisof the yield data withoutfirst removing
noncrossover interactions do not change rankings or environmentmaineffects causesweakresults that mostly
selections (Crossa, 1990). But this insight should reflect irrelevant features of the data. Likewise,mistaking
supplementedwith a second emphasis, that crossovers spuriousnoise for real interactions causesirrelevant and
amongwinninggenotypes are moreimportant than cross- unrepeatable results and multiplies mega-environments
overs amonglosers. For a trial with manygenotypes, needlessly. (ii) Not groupinglocations by their winning
only a tiny fraction of the crossovers involve winners, genotypecausesindividual clusters with a mixof winners
so equal emphasison all crossovers diffuses attention or separate clusters with the samewinner or both prob-
from the mainaction and multiplies mega-environments lems, so these clusters are not suitable for identifying
needlessly. Furthermore, crossover interactions have mega-environments or for targeting genotypes.(iii) Not
three sources-spurious noise, unpredictable environ- analyzing both genotypes and environmentsmeansthat
mental causes, and predictable environmental causes- a satisfactory understanding of genotype-environment
of whichonly the last can be exploited for identifying interactions and their implications for mega-envi-
mega-environments. In summary, mega-environment ronments is just not possible. (iv) Nothandlinga diversity
analysis should emphasizecrossover interactions among of generousor sparce datasets limits the applicablity of
winninggenotypescaused by reasonably predictable en- an analysis. Furthermore, combinations of several of
vironmentalfactors. these problemsat once producevery confusing and mis-
Third, mega-environment analyses should provide an leading conclusions. For example,clustering yield data
integrated understandingof the genotypesand environ- with large environment maineffects anda sizable interac-
ments. The underlying concept of genotype-environment tion that is actually just noise could divide a region
interaction, and the derived concept of mega-envi- into several clusters or mega-environments, and yet this
ronmentsare fundamentallydual, involving both geno- subdivisionreflects only irrelevant environment effects.
types and environments.Hence,the objectives of identi- Unfortunately,were such a poor analysis taken seriously
fying mega-environmentsand targeting genotypes are becauseits problemswere not recognized, it could make
inherently and deeply interrelated and must be pursued a sensible analysis look bad becauseit reaches different
jointly, as this studystitle suggests.Werethe roster of conclusions.Clearly, it is essential for mega-environment
genotypeschangedin a regional trial, it is possible, analyses to be evaluated accordingto relevant criteria.
and even probable, that appropriate mega-environments
wouldalso change. Likewise,were the definition or the
numberof mega-environmentschanged, the number of THE ADDITIVE MAIN EFFECTS
recommended genotypes and the particular recommenda- AND MULTIPLICATIVE INTERACTION
tion for each mega-environmentwould probably also MODEL FAMILY
change. Accordingly,an ideal analysis or clustering of Considera yield trial with a two-wayfactorial design
mega-environmentsdoes not analyze just genotypes or of G genotypesand E environments,with R replications,
just environmentsseparately but rather both genotypes denoting yield observations by Yger- The AMMI model
and environmentsin a single integrated model.It would combines AOVwith additive parameters and PCAwith
be strange for a clustering analysis applied to mega- multiplicative parametersinto a single analysis. Having
environmentsto analyze only environmentsor only geno- first removedthe grand meanI~ from the data, AOV
types since genotype-environmentinteraction involves partitions the varianceinto three sources: genotypedevia-
both. tions from the grand meantg, environmentdeviations
Fourth and finally, mega-environment analyses needto from the grand mean I~e, and genotype-environment
be flexible regardingthe structure of availableyield-trial interactions 0g~. ThenPCAfurther partitions the interac-
data. The data mayhave a two-wayGEstructure or a tions 0g~into NaxesEn~.n3,gn~5~ with ~.n the singularvalue
three-way GLYstructure. The trial maybe replicated, for axis n, 3g~ the gth value in the genotypeeigenvector
resulting in a GERor GLYR structure, or it maybe for axis n, and 6~n the eth value in the environment
unreplicated. Experimentsmayhave completeor incom- eigenvector,plus residuals pg~if Nis less than the full
plete factorial designs, and there maybe somemissing model.If the experimentis replicated, there is also an
data. So, the least generoussetup is a GEdesign with error term, eg~r, further partitioned in accordancewith
somemissing data and no replication, whereasthe most the particular experimental design. The AMMI model
generous setup is a GLYR experiment with a complete equation is
factorial design,replication, andnomissingdata. Statisti-
Yger = l-i, q- (tg + ~e + ,~_~nLn~gnSend- Pge d- I~ger
cal ahalyses for identifying mega-environments should
be flexible, accepting data across this wholerange of AMMI
is not a single modelbut rather a modelfamily,
GAUCH& ZOBEL: IDENTIFYING MEGA-ENVIRONMENTS 315

with that model having N axes denoted by AMMI-N. broad implications for multivariate models in general.
(In the same way, the more familiar polynomial models This potential for gaining accuracy was first exploited
comprise a modelfamily including constant, linear, quad- for AMMI by Gauch (1988). Subsequently, agronomists
ratic, cubic, and higher-order models.) The most simple and breeders have used AMMIextensively for both of
and most complex extremes are AMMI-0with no axes, these purposes: understanding interactions and gaining
equivalent to AOV,and the full model AMMI-Fwith accuracy (Gauch and Zobel, 1996a). Both are relevant
F = min(G - 1, E - 1) axes, equivalent to means across in the present context of identifying mega-environments.
replications. Hence, AMMIanalysis provides models
intermediate to the extremes of the additive model and
ADDITIVE MAIN EFFECTS
the raw data.
AND MULTIPLICATIVE INTERACTION
With replication, data splitting can be used to diagnose
MEGA-ENVIRONMENT ANALYSIS
the most predictively accurate member of the AMMI
family for a given dataset (Gauch, 1988, 1992, p. 134- Mega-environmentanalysis is illustrated here using
153; Gauch and Zobel, 1988). As will be emphasized yield data from a Louisiana corn trial analyzed previously
momentarily, effective use of AMMI requires careful by Kang (1993). Parameters of the AMMI-1model for
model diagnosis. The Expectation-Maximization imple- these data are given in Table 2. This yield trial has 16
mentation of AMMIcan handle and impute missing genotypes tested in 16 environments comprised of four
data (Gauch and Zobel, 1990; Gauch and Furnas, 1991; locations for four years, with four replications (or occa-
Gauch, 1992, p. 157-162). A convenient scaling for the sionally only two or three, but to avoid excessively
multiplicative parameters, for a given IPCAaxis, is difficult computations, an approximate solution is ob-
~,0.5~g and ~,5~5obecause the product of these interaction tained here by treating these data as if they were bal-
scores gives the interactions expected value directly, anced). The genotype codes are transparent abbreviations
without needing further multiplication by the singular of the full names in Kang(1993), the location codes are
value. For balanced data, AOVfollowed by PCA of AXfor Alexandria, BR for Baton Rouge, BC for Bossier
the interaction provides the least-squares solution. The City, and SJ for St. Joseph, and the years are 1985 to
AMMI solution is unique up to simultaneous sign changes 1988. For example, using this table and the above model
of a given genotype and environment eigenvector pair equation, the AMMI-1model estimates the yield for
(because the product ~,3g~e remains unchanged), Genotype 3165 in Environment BR87as 6.693 + 0.472
different software (or different orderings of the genotypes
and environmentsin the dataset) can give reversed signs.
This possibility emphasizes that contrast and trend from Table 2. AMMI-1model for the Louisiana corn yield trial with
16 genotypes and 16 environments. Deviations are in units of
positive to negative scores is meaningful but not the Mgha -1, IPCA scores are in units of (Mg ha-~) -s, and the
happenstancethat a particular score is positive or negative grand meanis 6.693 Mgha- 1. Entries are listed in rank order
in the output from a given software package. A worked of the IPCAscores.~"
example of AMMI calculations is given by Gauch (1992, Entity Deviation IPCA 1
p. 85-88). The AMMIanalysis was done by MATMO-
DEL l, which includes a model validation procedure for 3165 0.472 1.295
3147 0.665 0.783
diagnosing the most predictively accurate memberof the 1860 0.619 0.745
AMMIfamily for a given dataset (Gauch and Furnas, CK21 0.182 0.554
8990 - 1.046 0.371
1991). 4765 - 0.105 0.322
The AOV and PCA portions of the AMMImodel 8951 - 0.519 0.231
3389 - 0.255 0.183
were invented by Fisher (1918) and Pearson (1901). 0.003
8172 0.840
Fisher and Mackenzie (1923) applied both AOVand 3320 - 0.259 - 0.021
PCAseparately to the same potato (Solanum tuberosum 4673 0.280 - 0.093
4733 - 0.202 - 0.475
L.) yield trial. But the two components of AMMI were 1802 - 0.332 - 0.793
not combinedinto a single analysis until 1952 by Williams 4522 - 0.431 - 0.985
(1952) and Pike and Silverberg (1952). Further develop- 8150 - 0.050 - 0.987
1827 0.144 - 1.135
ments are reviewed by Gauch (1992, p. 8-12), but the BC88 - 1.184 1.338
incisive paper by Kempton(1984) first brought AMMI AX87 0.625 1.181
AX88 - 2.422 0.723
to the widespread attention of agronomists and breeders BC86 - 3.436 0.593
for the purpose of understanding complex genotype- BC85 - 1.608 0.297
environment interactions. Coincidentally, at about the SJ87 1.884 0.189
BC87 - 1.142 0.136
same time that AMMI was invented, Stein (1955) first SJ85 0.164 0.122
demonstrated that estimators from a multivariate model SJ86 0.541 0.008
could be more accurate and efficient than the usual aver- AX85 2.177 - 0.186
BR87 2.130 - 0.620
age across replications, and this intriguing result had AX86 2.478 - 0.622
BR88 - 0.242 - 0.639
SJ88 - 1.127 - 0.699
~ Mentionof a trademark or proprietary product does not constitute a BR85 0.327 - 0.873
BR86 0.834 - 0.947
guarantee or warranty of the product by the USDA or Cornell University
and does not imply its approval to the exclusion of other products that AMMI = Additive M._ainEffects and M._ultiplicative I_nteraction; IPCA=
mayalso be suitable. interaction principal componentsanalysis axis.
316 CROPSCIENCE, VOL. 37, MARCH-APRIL
1997

+ 2.130 + [1.295 (-0.620)] = 8.492 Mg -l , other words, specific and wide adaptations are equally
which approximates the actual yield of 8.554. important. Genotypesnear the top of Fig. 1 are mostly
The AMMI-1model was diagnosed for this particular long-season varieties, those in the middle medium-season
dataset by validation calculations that identified AMMI-1 varieties, and those near the bottom short-season varie-
as the most predictively accurate memberof the AMMI ties, and correspondingly the environments near the top
family. This diagnosis is confirmed by a much simpler have longer growing seasons, so the interaction appears
calculation: estimation of the interaction noise. Recall to be caused by different maturity requirements (M.S.
that the interaction SS is 738.03, which includes 457.88 Kang, 1996, personal communication). Incidentally, al-
noise and 280.15 real interaction. The AMMI analysis though the environment main effects are irrelevant for
first three IPCAcapture 229.81, 133.99, and 94.87. delineating mega-environments, they may be of interest
Hence, the AMMI model that comes closest to the goal for other reasons (such as looking for correlations or
of capturing 280.15 of the interaction is AMMI-1. Recall relationships between environment main effects and envi-
that the target for mega-environment analysis of this ronment IPCA1 scores). At any rate, they can also be
Louisiana corn trial is 13.5% of the treatment SS not shown in this biplot without any confounding or extra
100%. The AMMI-1model captures 12.2%, which is effort because one kind of marker is for environments
close to the target. This modelfocuses on relevant geno- and one of the axes is for main effects.
type and real interaction effects while ignoring irrelevant In this graph, distances in the direction of the abscissa
environment effects and muchinteraction noise. showmain effect differences, whereas the ordinate shows
The AMMI parameters in Table 2 can also be presented interaction differences. For example, consider the three
in a biplot graph, so namedbecause it has two kinds of environments SJ87, BC87, and BR87. SJ87 and BC87
markers, one for genotypes and one for environments differ markedly in main effects (1.884 vs. -1.142) but
(Gabriel, 1971, 1981; Bradu and Gabriel, 1978). Two have similar interaction scores (0.189 and 0.136),
different biplot axis systems are common:the abscissa whereasSJ87 and BR87differ in interaction scores (0.189
and ordinate showing main effects and IPCA1 (Zobel vs. -0.620) but have similar main effects (1.884 and
et al., 1988; Gauch, 1992, p. 92) or else IPCA1 and 2.130), and BC87and BR87differ in both since they
IPCA 2 (Kempton, 1984; Gauch, 1992, p. 94). Figure lie in diametrically opposite quadrats. Interactions are
1 showsthis first kind of biplot. estimated by multiplying a genotype and environment
Figure 1 is extremely informative, showing both main IPCA 1 score. For example, the AMMI-1model esti-
mates the interaction for Genotype 3165 in Environment
and interaction effects for both genotypes and environ-
BR87 as 1.295 (-0.620) = -0.803 Mg -t, which
ments. This biplot captures the genotype SS of 238.93
approximates the observed interaction effect of -0.742.
and the environment SS of 2865.17, and IPCA1 captures
Data in Fig. 2 show the AMMI-1nominal yields for
229.81 of the interaction SS of 738.03, so the biplot each of the genotypes as a function of the environment
reveals 86.8% of the treatment SS. Note that nearly as
much variation in yield is caused by the interaction
pattern captured in IPCA 1 as by the genotype main
effect, so interaction is important for this dataset. In

BC88 0 3165
AX87
1.0
03147
AX88 O 1860
BC86
0.5 O CK21
O 8990 4765
BC85 SJ87
8951O,-, $J85
BC87 338~ s..J8~8172
0
3320 AX85
-0.8 -0.4 0 0.4 0.8 1.2
-0.5 4733 0
BR88~ AX86 05
EnvironmentIPCA1 Score (Mg/ha)
SJ88~18020 BR87
BR85 Fig. 2. Additive__MainEffects and Multiplicativelnteraction (AMMI)-I
-I.0 4522 O0 8150 BR86 nominal yields for a Louisiana corn trial as a function of the
O~ 1827 = ~ environmentinteraction principal componentsanalysis axis (IPCA)
4 5 6 7 8 9 I score. Thescores for the 16 environmentsare indicated by triangles
along the abscism. The four genotypes that win over someregion of
Mean(Mg/ha) this score are shownwith thick lines andare identified individually,
Fig. 1. Additive Main Effects and Multiplicative Interaction (AMMI)-I whereas the remaining 12 genotypes that never win are shownwith
biplnt for a Louisiana corn trial, showingboth mainandinteraction thin lines. Hypothetical Genotype A or any line above it would be
effects for both genotypes and environments. Opencircles denote a universal winner. Hypothetical GenotypeB is the universal winner
genotypes; closed circles denote environments(IPCA= interaction with a genotype IPCAI score of 0 andthe smallest possible genotype
principal componentsanalysis axis). mean.
GAUCH& ZOBEL: IDENTIFYING MEGA-ENVIRONMENTS 317

IPCA1 score, where the nominal yield is defined as relatively well or relatively poorly." By providing an
the yield from the AMMImodel equation without the effective summaryor overview of yield trial results,
environment deviation IBe. Hence, nominal yields repre- standard biplots can provide helpful insight for delineat-
sent an average environment (with IBe = 0), and they ing mega-environments. However, such graphs fail to
can be adjusted to give actual AMMI yield estimates for make evident what wins where, which limits their value
an individual environment by adding that environments for analyzing mega-environments. Accordingly, several
deviation. However, this adjustment merely shifts all new kinds of AMMIgraphs are developed here that do
yields within a given environment up or downequally, so address mega-environment issues directly. The key is
the genotype rankings and overall picture are unchanged. combiningfeatures of Fig. 1 and 2 and decreasing atten-
Again, environment deviations are irrelevant for present tion on losing genotypes while focusing attention on
concerns because they do not affect genotype rankings winning genotypes.
within each environment. Genotypes affect mega-envi- Figure 2 showed that the winning genotype in each
ronments through their main and interaction effects, but environment, e, is determinable for the AMMI-1model
environments affect mega-environments through their from a single parameter, the environment IPCA1 score
interaction effects only. In Fig. 2, the four genotypes (~,0.56e) shownon the abscissa, and that in this particular
that win in portions of the IPCA1 space are identified case, there are four winners. That information is depicted
individually, whereas the other 12 genotypes that never differently in Fig. 3, where this environment score is
win are represented by thin lines. Scores for the 16 nowon the ordinate instead of the abscissa.
environments are shown by the 16 triangles along the Figure 3 shows the four mega-environmentsidentified
abscissa (which are the same scores listed previously in by AMMI.Note for example that in Fig. 2, the lines
Table 2 and shown in the ordinate of Fig. 1). Note for for winning Genotypes 3147 and 3165 cross at an envi-
example, that Genotype 3165 has the highest AMMI-1 ronment score of 0.376. Accordingly, there is a line
estimated yield for four environments, namely those at across Fig. 3 at this score of 0.376, and Genotype3165
environment IPCA1 scores of 0.593, 0.723, 1.181, and wins in the four environments that happen to have scores
1.338 (which can be identified more specifically in Fig. above this value. Likewise, from 0.376 to 0.224, Geno-
1 as BC86, AX88, AX87, and BC88). type 3147 has 1 win; from 0.244 to -0.612, Genotype
The mathematical relationship between Fig. 1 and 2 8172 has 5 wins; and below -0.612, Genotype 1827
should be understood clearly. Correspondingto the above has 6 wins. On the right side of Fig. 3, nominal yields
AMMI model equation, Table 2 gives the AMMI-1model are indicated, copied from the previous figures ordinate.
parameters for this Louisiana corn trial, with two values For example, from Fig. 2, at an environment IPCA 1
for each genotype (and each environment). For example, score of 1.0, the winning Genotype 3165 has a nominal
Genotype 3165 has a mean of 6.693 + 0.472 = 7.166 yield of 8.461 Mgha-~, as is also indicated in Fig. 3.
Mgha -1 5.
and an IPCA 1 score of 1.295 (Mg ha-I) Nominal yields form a response surface in Fig. 3
Figure 1 uses these two values directly in its abscissa
and ordinate to place its marker for Genotype3165. But
Fig. 2 uses this same information differently, depicting 1.4
[] BC88 8.899
each genotypeby a straight line (instead of a point). For 1.2 3165 Wins 8.8
O AX87
example, Genotype 3165 is depicted by a line with the 8.6
1.o 8.4
equation y = 7.166 + 1.295x, where y is the nominal
yield and x is the environment IPCAscore, so this line O AX88 8.2
uses these same two parameters. The left-most x of 0.8 ] BC86 8.o
-0.947 implies a yield of 5.940 Mgha-l, whereas the
right-most score of 1.338 implies 8.899, and these two 7.653
3147 WinsOBC85 7.534
0.2
coordinate pairs give two points that define the line for 8172 Wins [3 BC87 SJ85 SJ87
Genotype 3165, as shown in Fig. 2. More generally, 13.0 SJ86
each point in Fig. 1 corresponds to a line in Fig. 2. -13.2 AX85 O
Movinga point left or right in Fig. 1 equates to moving
a line downor up in Fig. 2, and moving a point up or -0.4
downin Fig. 1 equates to tilting a lines slope upward 7.531
7.6
or downwardin Fig. 2. -13.8 1827 Wins SJ88 BR88 BR87 (3
BR85 AX8,~ 7.8
Whena biplot of the AMMI-1model, as in Fig. 1, 7.912
-1.13
fails to capture all of the real structure in the interaction, 4 5 6 7 8 9
another kind of biplot using IPCA1 and IPCA2 as its Environment
Mean(Mg/ha)
two axes can showmore of the interaction, as in Kempton
Fig. 3. ~dditive MainEffects and Multiplicative Interaction (AMMI)
(1984) and Gauch (1992, p. 94, 222). But Westcott mega-environmentanalysis for a Louisiana corn trial. There are
(1987) remarks that Kemptonsbiplot shows interactions, four genotypes that win in various regions of the AMMI-1 models
not actual performance meaningyield, in line with Kemp- parameter space, such as Genotype 8172 that wins in the five
tons ownremarkthat his biplot "indicates whichvarieties environments with an interaction principal componentsanalysis
axis (IPCA) score between 0.224 and - 0. 612. The fo ur Louisiana
have similar and which have very different patterns of locations are shownwith different symbolsto facilitate comprehend-
response over the environments"but "it does not identify ing the results for each location across the four years. Onthe right
the particular environments in which a variety performs is the nominalyield, which correspondsto the values in Fig. 2.
318 CROPSCIENCE, VOL. 37, MARCH-APRIL
1997

composedof a convex set of planes, with yields increas- of -0.612 and increasing yields in either direction away.
ing both above and below the lowest yield of 7.531 Mg from this value). Consequently, the thin isoyield lines
ha -] located at an IPCAscore of -0.612. The yield indicating equal yields form a herringbone pattern. For
response surface for Genotype8172 is nearly flat (as is example, environments BC88, SJ86, and BR85 all have
also evident in Fig. 2), whereas 1827 and 3165 are much nearly equal yields of ~ 8 Mgha- i, but they achieve this
steeper. The lowest yield for a winning genotype always yield in different ways: SJ86 has almost no interaction,
occurs for that genotype with the highest main effect whereas BC88and BR85have sizable and opposite inter-
and hence an automatic win at an environment IPCA1 actions, which is why these three environments have
score of zero, which is Genotype 8172 in this case. three different genotype winners and hence fall into three
Higher yields are achieved by positive interactions that different mega-environments.
more than compensate for comparatively small reduc- A sequence of mega-environment assignments with
tions in genotype main effects. Hence, Fig. 3 shows why one to many mega-environments can be determined on
specific adaptations (interaction effects) are important the basis of the above AMMI results, while calculating
well as broad adaptations (main effects) in the pursuit the implications for yield. Wereeach of the 16 environ-
of high yields. ments planted to its own winner according to the raw
Although Fig. 3 has the advantage, unlike standard data, there would need to be eight different genotypes
biplots, of revealing genotypes winning domains, it has and hence eight mega-environments, and this highest
the disadvantage of providing information on main and possible average yield would be 8.416 Mg ha-I. But
interaction effects in different scales and units, namely were each environment planted instead to its AMMI-1
yield and the square root of yield. Consequently, the winner indicated in the above figures, then only four
relative importance of main and interaction effects is genotypes or mega-environments are required, and the
not apparent. Accordingly, Fig. 4 presents the AMMI average yield based on raw means would be 7.585 Mg
mega-environmentresults in still another format. ha- or a 9.9% yield drop. Were the entire region
The abscissa of Fig. 4 shows the environment main declared a single mega-environment and planted to the
effect plus the genotype main effect for the particular overall winner 8172, the average yield based on raw
genotype that wins in each environment (as identified in means would be 7.533 Mgha-l, or a 10.5% yield drop.
the right of Fig. 4, as well as earlier in Fig. 3). The So, according to the raw yield data, one, four, and
ordinate shows the interaction effect, again for each eight mega-environments allow average yields of 7.533,
environment using its winning genotype. For example, 7.585, and 8.416 Mg ha -~. In particular, based on
Environment BC88 has winning Genotype 3165 and AMMI-F, switching from the eight AMMI-Fwinners
hence an AMMI-1estimated interaction that equals the to the four AMMI- 1 winners causes an estimated average
product of this environment and genotypes IPCA 1 yield loss from 8.416 to 7.585 Mg ha-l or 12.4% of
scores, namely 1.338 1.295 = 1.733 Mgha -]. Main the grand mean.
effects and interaction effects are shownon the abscissa However, were AMMI- 1 yield estimates trusted across
and ordinate to the same scale, to make evident their AMMI-Fraw data because the AMMI-1model is more
relative importance. The AMMI-1yield estimate for predictively accurate than its data, then switching to
each environment is simply the sum of its two values AMMI-1winners causes an estimated yield gain from
shown here. For example, for its winning Genotype 7.527 to 7.806 Mg ha -] or 4.2% of the grand mean.
3165, Environment BC88has main effects of 6.693 + AMMI-Fand AMMI-1pick the same winner in only 2
0.472 - 1.184 = 5.981 and an interaction effect of of the 16 environments. Also, according to AMMI-1
1.295 1.338 = 1.733, resulting in a yield of 5.981 estimates, were the one Genotype 8172 planted every-
+ 1.733 = 7.714 Mgha-]. This ordinate preserves the where, the average yield would drop substantially to
convex topology of the ordinate of Fig. 3 (which has 7.533 Mg ha -]. AMMI-0 and AMMI-1 pick the same
lowest nominal yield at an environment IPCA1 score winner in only five environments. So, according to
AMMI-1yield estimates, one, four, and eight mega-
environments allow average yields of 7.533, 7.806, and
\
\ \ rl BC88 \ T 7.527 Mgha-]. AMMI-0is worse than AMMI-1because
it underfits real interaction patterns in the data, whereas
AMMI-Fis worse than AMMI-1because it overfits
spurious noise in the data.
~ ~BC85[] BC87. J87 .-- Because AMMI-1estimates are more predictively ac-
curate than AMMI-F estimates, there are strong rational
/ and empirical reasons for giving priority to the AMMI-1
/ SJ88i/ I BR85 BR8710 assessment. But even without accepting that argument,
the AMMI-1 analysis with four mega-environments could
4 5 6 7 8 9 10 be regarded as a useful simplification if practical consid-
Geno~peand EnvironmentMeans(Mg~a) erations impose a limit of about four mega-environments
Fig. 4. Maineffectsand interaction effects for a Louisiana corn trial, (rather than allowing the rather numerous eight mega-
showing yields for each environmentwhenplanted to its genotype
winner. Isoyield lines forma herringbonepattern. Mainand interac- environments of the AMMI-Fresults).
tion effects are shownto the same scale to make their relative Figure 5 reviews the AMMI-1yield estimates and
importance evident. AMMI-Fraw data for average yields achievable with
GAUCH& ZOBEL: IDENTIFYING MEGA-ENVIRONMENTS 319

one mega-environmentusing the AMMI-0 winner 8172,


four mega-environments using the AMMI-1 winners,
and eight mega-environmentsusing the AMMI-F win-
ners. TheAMMI- 1 curve reflects a predictive perspective 8.2
in which the data are deemeda sample from a larger AM
populationof inference, with optimalyield at rather few
mega-environments. Theshape of this curve is a typical 8.0
Ockhamshill, with optimal performancefor a fairly
parsimoniousmodeland penalty to the left for underfit- 78 "~
ting real patterns and penaltyto the right for overfitting
spurious noise (Jefferys and Berger, 1992;Gauch,1993).
But the AMMI-F curve reflects a postdictive perspective
in whichthe data are deemedthe entire population of
inference, so increasingly complexmodelsautomatically 1 2 3 4 5 6 7 8
seem better. The AMMI-1 and AMMI-Fyields are identi- Number
of Mega-environments
cal for one mega-environmentplanted to the overall
Fig. 5. Averageyield for a Louisian corn trial using one, four, or
winner8172because discarding part of the interaction eight mega-environmentsand estimating yields by the Additive
by the AMMI-1model has no effect on an average _MainEffects and Multiplicative I_nteraction (AMMI)-Imodel(solid
computedacross all environments.Note that predictive line) or the AMMI-F averages across replications (dashed line).
assessmentbased on AMMI-1 is qualitatively different The AMMI-Imodel merits more serious consideration because it
is morepredictively accurate than AMMI-F.Unfortunately, the
from postdictive assessmentbased on AMMI-F raw data. AMMI-Fraw data are doubly confusing because they both overesti-
For the postdictive perspective, more mega-environ- mate the yield with eight mega-environmentsand underestimate
mentscontinueto increase yields until there are enough the yield with four mega-environments. Noise encourages a choice
genotypesfor every environmentto be assigned its win- of too many mega-environments, which both increases costs and
ner, which number eight in this case. But from the decreases yields.
predictive perspective, both too few and too manymega-
environments result in a yield loss, and four mega- replications, whichhave, for this Louisianacorn trial,
environmentsare ideal in this case. a standard error of 0.713 Mgha-l. Theseaverages are,
Figure 5 emphasizesthat different membersof the fromstatistical theory, unbiasedestimators of the true
AMMI family differ in their predictive accuracy and in means.However,after ranking the 16 genotypesyields
their value for delineating mega-environments. For this within each environment,the highest yields are no longer
particular yield trial, AMMI-1 happensto be best, but unbiasedbecausethey tend to havepositive errors, and
for another dataset, the optimal modelmightbe AMMI-2 likewise the lowest yields tend to have negative errors
or some other member.If the optimal AMMI model for (Gauch and Zobel, 1989; Gauch, 1992, p. 171-177).
a given dataset has not been diagnosed, then selection The AMMI-F value is larger than AMMI-1 by 0.889/
of somearbitrary memberof the AMMI family is likely 0.713 = 1.247 standard errors. Nowthe first order
to give a suboptimal analysis, which weakens AMMI statistic for 16 drawsfromthe standardnormaldistribu-
results and confuses mega-environment choices. For ex- tion is 1.766, so wereall 16 genotypesessentially equal,
ample,if the interaction is essentially all noise, thenthe the upwardbias in the winners wouldapproximate1.766
PCAportion of AMMI will find a large number of x 0.713 = 1.259 Mgha-~. However,the 16 genotypes
small eigenvalues,exactly as describedby Mandel(1969, are not all nearly equal, and if the top group of real
1995). Thenit will look like AMMI fails to capture the contenderswere morelike about six entries, that first
interaction effectively, since numerousaxes are required order statistic of 1.267 wouldvery nearly explain the
to get most of the interaction SS, such as the AMMI-5 observed discrepancy. So, the AMMI-Fexcessively large
modelto capturejust 60 %of the interaction. This verdict yield at eight mega-environments is easily attributed to
leads to a complex model with numerous mega- spurious noise, to positive samplingerrors accumulated
environments.But the truth of the matter, whichwould after rankingthe data. Butif the effects of experimental
be revealed by not neglecting modeldiagnosis, is the noise are not understoodproperly, researchers will be
exact opposite. Theinteraction is buried in the noise, tempted to identify too manymega-environments,which
so only maineffects are statistically significant. Conse- will actually reduce averageyields, while also imposing
quently, only one winner is detected, the main-effect unnecessary burdens on plant breeders and seed pro-
winner, whichleads to no mega-environment subdivision ducers.
whatsoever.If no distinction is madebetweenreal inter- But the most interesting feature of Fig. 5 is that
actions and spurious noise, mega-environment analysis AMMI-1beats AMMI-Fat the optimal four mega-
becomesa disaster. environments, whichcan be explained as follows. When
It mayseemsurprising that these two models, AMMI-1 AMMI-F yield data are used to calculate the average
and AMMI-F, differ by such a large amount, 0.889 Mg yield but each environmentis assigned its AMMI-1 win-
ha- ~, at eight mega-environments. Butthis discrepancyis ner rather than its ownAMMI-F winner, only 2 of the
easily explained, showingthat AMMI-Fis unrealistically 16 environments receive that genotypewhichis also their
high, rather than AMMI-1unduly low. The AMMI-F AMMI-F winner. Forced away from its ownwinners 14
yield estimates are simplythe averages across the four out of 16 times, the AMMI-F average yield with AMMI-
320 CROP
SCIENCE,
VOL.37, MARCH-APRIL
1997

ls four mega-environments rises only a slight amount of variation is 10.7 %. Accordingto the morepredictively
above AMMI-/~s average with only one mega-envi- accurate AMMI-1models yield estimates, planting the
ronment. But AMMI-1yield estimates used with its AMMI-1winner 8172 instead of the AMMI-Fwinner
own winners exploit sizable interactions to achieve a 4673 results in a yield increasefrom 7.513 to 8.075 Mg
considerable yield increase when changing from one to ha -1. Crossa et al. (1991) describe AMMI adjustments
four mega-environments. Indeed, the first IPCAaxis of for a bread wheat (Triticutn aestivum L.) trial.
AMMIcaptures a SS of 229.81, which is comparable This experiments 16 environments have a factorial
in magnitude to the genotype main effect of 238.93, so design with four locations and four years, so it is interest-
interactions are quite important for this dataset. Equally ing to compile the results for each location across the
importantly, the AMMI- 1 model leaves behind a residual years, using the AMMI-1 yield estimates. From Fig. 3,
SS of 508.22, which is the SS of the differences between location BRis always in the winning domainof Genotype
AMMI-1and AMMI-Fyield estimates, and this dis- 1827. Its average yield in BRis 8.473 Mgha-I. Location
carded residual contains mostly noise. So, the differences BRis the most predictable, so variety recommendation
between these two models have somewhat more than is easy. Location SJ gives three wins to 8172 and one
twice as muchvariation as do the genotype main effects. to 1827, with an average winners yield of 7.924 Mg
That is, the interaction noise is about twice as large ha-1. Here the best recommendation, especially in a
as the genotype main effects, and this noise is largely predictive context, is to always plant Genotype 8172,
disregarded by AMMI-1,but unfortunately, it is taken reducing the average yield only imperceptibly to 7.899
seriously by AMMI-F.Because these two models are Mgha-1. Location BC gives two wins to 3165 and one
so different, AMMI-ls average is low using AMMI-Fs each to 3147 and 8172, with an average of 6.147 Mg
eight winners, and AMMI-Fsaverage is low using ha-~. At this location, these three genotypes average
AMMI-lsfour winners. Anyway,the net effect of these 6.089, 5.979, and 5.693 Mgha-~, respectively. Because
factors is that at four mega-environments,AMMI-1 beats 3165 has the most wins and the highest average yield
-I.
AMMI-Fby a substantial 0.221 Mg ha at BC,the best strategy, especially in a predictive contex.t,
The lesson from Fig. 5 is that from the perspective wouldbe to always plant 3165. Finally, the least predict-
of identifying mega-environments,experimental noise is able and most problematic site is AX, with winners 3165
doubly confusing. Noise causes an overestimate of the (twice), 8172, and 1827 spanning the entire range
yield increase for an excessive number of mega- interaction patterns, and an average winners yield of
environments and an underestimate of the benefit for a 8.682 Mgha -~. Location AXhas large interactions in
small and optimal number of mega-environments. Conse- hindsight but unpredictable ones in foresight, so the best
quently, effective mega-environmentidentification re- predictive strategy is not to try to exploit interactions
quires aggressive reduction of experimental errors, and a but rather to plant the AMMI-0 main effect winner 8172,
multivariate model such as AMMI can help considerably. resulting in 8.248 Mgha -I. The general outcome is
Furthermore, although this lesson is illustrated here that predictable interactions, typically associated with
with a particular Louisiana corn trial, it seems relevant locations, make mega-environments more numerous and
for manyregional yield trials. This corn trials standard useful, whereas unpredictable interactions, typically as-
error is 0.713 and its grand mean is 6.693 Mgha-~, so sociated with years, make mega-environments less nu-
this experiment has a middling coefficient of variation merous and useful.
of 10.7%. Hence, the substantial difference in mega- In summary, the recommendations are 1827 for BR,
environment conclusions from choosing the AMMI-1or
AMMI-Fperspective is not due to some enormous or AMMI-F AMMI-1
atypical noise level. Furthermore, manyregional trials 8172
have more than 16 genotypes, making order statistics 8.0 ~ 3147
problems larger than those encountered here. So, the Yield 1860
(ig/ha) 3165
general pattern exhibited in Fig. 5 can usually be expected
whenevera trial has substantial interactions and typical 7.5 4673
CK21
noise. It should not seem too surprising that when at- 1827
(Mean)
tempting to identify rational mega-environments, it 81504765
makes a difference whether or not a statistical analysis 7.0 ~ 4733
33203389
is used to filter out someof the noise in the data. 1802
Figure 6 shows the difference between AMMI-F data //\ )< 4522
8951
and AMMI-1estimates for a typical environment, SJ86. 6.5
In this example, Genotype 4673 that ranks first in the
raw data ranks fifth in the AMMI-1model, whereas 8990
Genotype 8172 that ranks first in the AMMI-1model
ranks third in the raw data. Especially because Environ- 6.0
ment SJ86 has almost no interaction (Table 2 and Fig. Fig. 6. AdditiveMainEffects andMultiplicativel_nteraction
1), there is a strong case for declaring its winner to be (AMMI)-Fyield dataandAMMI-1 estimatesfor environmentSJ86
of a Louisianacorntrial. Genotype
4673ranksfirst in the raw
the main effect winner, Genotype 8172. Furthermore, databut ranksfifth in the morepredictivelyaccurateAMMI-I
that 8172ranks third in the raw data, not first, is easily model.AMMI-1 identifies Genotype
8172as the best genotypein
explained by noise, given that the experiments coefficient this environment.
GAUCH& ZOBEL: IDENTIFYING MEGA-ENVIRONMENTS 321

8172 for SJ and AX, and 3165 for BC, giving an overall slope equals the negative reciprocal of the largest negative
average of 7.677 Mgha-~. In hindsight, planting the environment IPCA1 score. From Table 2, this is -0.947
AMMI-1 winner every time results in a somewhat higher for Environment BR86, so the upward slope is -1/
average yield of 7.806 Mgha- 1. But hindsight is clearer -0.947 = 1.056. Similarly, the downwardslope equals
than foresight. The present recommendations capture the negative reciprocal of the largest positive environ-
98.3% of the potential yield, which seems to be about ment IPCA1 score. This is for Environment BC88, so
the limit of predictive capability, with muchof the penalty the downward slope is -1/l.338 = -0.747. The line
incurred at AXand none at BR. Incidentally, were only with an upward slope (of 1.056) represents the family
two mega-environments feasible, the recommendation of all lines going through the point at the left extreme
would be to plant 1827 in BR and 8172 elsewhere, of a genotypes line in Fig. 2, whereas the line with a
capturing 98.1%of the potential yield. downwardslope (of -0.747) represents the family
Figure 7 uses AMMIparameters to show which geno- all lines going through the point at the right extreme.
types win over others, as well as to locate the parameter Hypothetical GenotypeA is constructed as a line going
space that identifies potential new winners. Each of the through two points, the highest yields at the left and
16 actual genotypes are located in this graph by their right extremes of Fig. 2. The point on the left has an
two AMMI-1parameters, the genotype mean (kt + ~tg) abscissa of -0.947 and ordinate of 7.912, and the point
and the genotype IPCA1 score (~,53,g), as listed on the right has coordinates of 1.338 and 8.899. Solving
Table 2 and shown in Fig. 1. Also shown are the two the equation of a straight line given two points, calling
hypothetical genotypes in Fig. 2, with GenotypeA having the environment score x and the nominal yield y, the
a mean of 8.321 and score of 0.432 and Genotype B a equation for GenotypeA is y = 8.321 + 0.432x. Accord-
mean of 8.899 and score of 0. ingly, this genotype appears in Fig. 7 at exactly these
An X is drawn at the markers for the four winning coordinates, having a nominal yield of 8.321 and score
genotypes but, to avoid clutter, is not indicated for the of 0.432. Incidentally, a line from the winning genotype
other 12 genotypes that never win. The purpose of these with the largest negative IPCA1 score, namely Genotype
X is to show which genotypes win over which others, 1827, connecting to Genotype A defines the upward
as explained momentarily. Every X has exactly the same slope. Similarly, a line from Genotype 3165 with the
shape, that is, all upwardlines (/) are at the sameslope largest positive IPCA1 score connecting to Genotype
and all downwardlines (\) are at the same slope. Casual A defines the downwardslope. Because Fig. 7 has the
inspection might suggest that these upward and down- same axes as Fig. 1, every point in Fig. 7 corresponds
ward slopes are of equal magnitude (apart from opposite to a line in Fig. 2, as explained earlier for Fig. 1 and
signs), but actually the upwardslope is steeper than the 2. The points on the line through 3165, A, and B in
downwardslope, and in general, these two slopes are Fig. 7 correspondto the family of all lines that go through
not equal. the right-most top yield in Fig. 2 (at coordinates x
These two .slopes are calculated as follows. The upward 1.338 and y = 8.899). Likewise, points on the line
through 1827 and A in Fig. 7 correspond to the family
of all lines that go through the left-most top yield in
Fig. 2 (at coordinates x = - 0.947 and y = 7.912). These
two lines in Fig. 7 intersect at one point, hypothetical
Genotype A, which correspondingly is the unique line
in Fig. 2 that goes through both the left-most and right-
. C, K2~ / ~ /~. / Universal most top yields. GenotypeA is also uniquely the universal
809908~1/0034~ ~ ~Winners winner with the lowest genotype main effect.
To begin the interpretation of this graph, focus on
just one genotype, namely 8172 near the middle. The
~ / 4673/" / ~ ~ ~ X drawn at this genotypes marker divides the parameter
space into four sectors. There are three interpretive
principles. First, the sector to the left is the shadow
cast by 8172 over the other genotypes that it always
oe- -1 ~ential ~~ outperforms, which include nine genotypes in this case.
~) These nine complete wins by 8172 can be confirmed
i ~ ~ i ~ i i i ~ and understood by reference to Fig. 2, which Showsthat
6 7 8 9 the line for 8172 is always above nine other lines for
Geno~pe Mean (Mg~a) the losing genotypes, such as the unlabeled lowest line
with an upwardslope that is actually for Genotype8990.
Fi~. ~. ~dditi~e~alnEff~tsand~ultiplicati~e~nteraction (A~[-])
modelfor the 16 ~enoty~sof a Louisianacorn trial. Four of these Second, in Fig. 7, the sectors above and below 8172s
~eeoty~sare the ~nnersthat shadethe other 12 ~enot~s to the marker identify mixed wins where 8172 sometimes wins
left. Parameter
spaceto the right of the ja~edthick line identi~es and sometimes loses, which involve six genotypes. For
~teetial newwinne~that wouldshadeat least oneof the current example, 3147 is in the sector above 8172, and 1827 is
winners,andparameterspacete the H~ht of the s~ondthick line
in the sector below 8172. Accordingly, Fig. 2 shows
identifies ~tential ~iver~! winners that would out~rform all
current ~enoty~s. ~noty~s A and B are hy~tbetical universal that the lines for 3147and 1827 cross the line for 8172,
winners(IPCA = interaction principal com~nents analysis axis). so 8172 sometimes wins and sometimes loses against
322 CROP SCIENCE, VOL. 37, MARCH-APRIL 1997

these competitors. Third and finally, the sector to the mega-environments. Assume for sake of argument that
right of 8172s marker identifies genotypes that always breeders expect to increase main effects by = 1% per
win over 8172. In this case, there are no such actual year, which has been typical for many major crops
genotypes, but there are two hypothetical genotypes, A (Evans, 1980; Russell, 1991). Then 1% of 7.533
and B. Again, it is evident from Fig. 2 that Genotypes 0.07533, so the needed increase will take = 1.366/
A and B always win over 8172. 0.07533 or 18 yr.
This X marking four sectors for Genotype 8172 is But now comparethe idiotype represented by hypothet-
also shownin Fig. 7 for the three other winners, namely ical GenotypeA. This genotype is not stable but rather
3165, 3147, and 1827. Note that the other 12 gentoypes, has a substantial interaction. FromFig. 1, this genotypes
which never win, are always in the shadow of one or interaction is positive for environments like BC88and
more of the four winners. AX87 and negative for environments like BR86 and
Havingexplained the technicalities of Fig. 7, its most BR85, which was interpreted earlier as meaning a posi-
interesting features can nowbe explored. The thick jag- tive interaction for yield in relatively long-season envi-
ged line down the middle of this graph separates the ronments. Although Genotype A is less stable than B,
parameter space to the left that is dominatedby the cur- Genotype A has the merit of requiring considerably less
rent four winners from the parameter space to the right mean yield to become a universal winner, only 8.321
that identifies potential newwinners. Note that the other instead of 8.899 Mgha- t. Assumethat the average main
12 genotypes, which never win, are located to the left effect for parental genotypes whose progeny could have
of this thick line, meaning that they are always in the Genotype As target interaction score of ~-0.432 equals
shadowof one or more of the four winners. On the other the average for nearby Genotypes 8172 and 3147, namely
hand, any new genotype with parameters to the right of 7.446. Then gaining 1%of 7.446 surpasses the required
this boundary will shadow and outperform at least one 8.321 in 12 yr. So, it is plausible that the target idiotype
of the current winners. Furthermore, any new genotype A could be reached several years faster than B. The
to the right of the second thick line is a universal winner general lesson here is that breeders should think carefully
that will outperform all current genotypes. Note that the about what they really want: stability or superiority.
parameter space for universal winners is defined by In essence, to optimize yield despite genotype-envi-
extending the line from 3165 to A and the line from ronment interactions, breeders have two options: to pur-
1827 to A because these two current winners have the sue wide adaptation and thereby minimize mega-en-
largest positive and negative genotype IPCA1 scores vironment subdivisions or to exploit narrow adaptations
and A is that potential universal winner with the smallest in several different mega-environments (Brennan and
mean yield. Byth, 1979). A graph such as Fig. 7 could help breeders
This figure stimulates insight on possible breeding comparethe feasibilities of these two options for a given
strategies. It shows the various combinations of main breeding program.
and interaction effects that could produce new winners. By knowing the landscape of the AMMIparameter
It also shows just what sort of new genotype could space, with the locations of current winners and the
outperform two or more current winners, thereby reduc- regions for potential and even universal new winners,
ing the number of mega-environments needed to optimize desirable idiotypes can be defined. And by knowing
yield everywhere. For example, consider the point where roughly how fast breeders can change main effects or
two thin lines cross between the labels for 3165 and interaction effects, one can predict roughly howlong it
3147. A new genotype located near this point would would take to reach various places in the parameter
combine a main effect like 8172 with an interaction space representing potential new winner~. In addition to
pattern intermediate to 3165 and 3147, and by outper- providing insight about possible idiotypes, this graph
forming both 3165 and 3147 (as well as the additional also gives clear implications for mega-environments.For
seven genotypes already shadowedby these two current example, in comparing the value of various possible
winners), it would merge two old mega-environments idiotypes, special priority can be given to those that
into one new mega-environment. Andif another genotype would shadow several current winners instead of just
could shadow 8172 and 1827, then the original entire one, thereby reducing the number of mega-environments.
yield trial with four mega-environmentswould need only Fewer mega-environments simplify and economize a
two mega-environments. breeders testing system, concentrate breeding efforts
The tradeoff between breeding for main effects or for on fewer needed cultivar types, and streamline seed
interaction effects can be illustrated with hypothetical production and distribution. So, a graph like Fig. 7
Genotypes A and B. Genotype B is most stable, as is helpful both for understanding present relationships
indicated by its fiat line in Fig. 2 and correspondingly amongactual genotypes and future implications of poten-
its score of 0 in Fig. 7. The current main-effect or tial idiotypes. It can help breeders understand where
AMMI-0winner, Genotype 8172, is also very stable, they have come and where they might go.
with a genotype IPCA1 score of just 0.003. The differ- Another application of a graph like Fig. 7 could be to
ence in mean yields between B and 8172 is 8.899 - compare the market strengths of various seed companies.
7.533 = 1.366 Mg ha-~. So, if a new genotype could Envision a collection of genotypes not all from a single
be developed with the same stability as 8172 but a main breeding program but rather including cultivars sold by
effect greater by 1.366, it wouldwin everywhere, thereby various companies. By identifying the commercial source
eliminating the need to subdivide the region into several of each genotype in this graph, it would be clear which
GAUCH& ZOBEL: IDENTIFYING MEGA-ENVIRONMENTS 323

companieshave special strengths or weaknessesin vari- interaction parameters exhibit some measure of ro-
ous portions of the AMMI parameterspace, in regards to bustness across different locations, years, cultivars,
both current actual winnersand future potential winners. crosses, and even crops. So, we wouldcaution against
Another advantage of Fig. 7, which is evident from overinterpreting AMMI graphs whenmakingdecisions
the logic of the calculation usedto place the Xover each about cultivar recommendations or breeding strategies.
winners marker, is that such a graph can be adjusted For example,broad features deserve moreconsideration
or tailored for a different range of target environments. than fine details. Also, current results can be taken
For example, the present graph showsresults for this more seriously whenthe genotypes are experiencing
entire Louisianacorn trial, with the interaction extremes incremental improvementsrather than drastic changes
represented by EnvironmentsBR86and BC88,leading from newidiotypes, management procedures, or disease
to slopes of 1.056 and -0.747. If instead of the entire pressures. Nevertheless,wewouldalso insist that various
trial, interest focuses on yields across years for just kinds of AMMI graphs, as illustrated above, can often
LocationSJ, then the newinteraction extremesare less supplementthe understanding and insight that agrono-
diverse, SJ88 at -0.699 and SJ87 at 0.189, leading to mists and breeders have already developed over the
slopes of 1.431 and -5.291. Consequently, genotypes years throughextensivefield experimentation,statistical
cast larger shadowsby their broadersector to the left, analysis, and experience.
whichalso leads to fewer winnersor mega-environments.
This makessense..From Fig. 2, it is clear that this
smaller range in environmentscores implies only two DISCUSSION
winners, 1827and 8172. Moregenerally, rather homoge- Subdividing a growing region into several mega-
neous environments makegenotype main effects rela- environmentsbecomesa daunting prospect if the recom-
tively moreimportant than genotype-environment inter- mendednumberof mega-environmentsis large. By con-
action effects. Onthe other hand, if interest focuses on trast, subdivision of a crops growingregion is most
even morediverse environmentsbecausea yield trial or inviting whenmerely a few mcga-cnvironmentssuffice
growingregion is expanded,wider interaction extremes to capturesubstantialyield increases,as for this Louisiana
lead to the opposite response of increasing the sectors corn trial. Fortunately,results for several diverseyield
above and belowthe genotypemarkers. This also makes trials indicate that often remarkablyfew mcga-cnvi-
sense. FromFig. 2, it is clear that were this graph ronmentsare sufficient (Annicchiarico,1992; Annicchi-
extendedto the left and right, moreand moreof these arico and Perenzin, 1994;DcLacyct al., 1994;Annicchi-
nonparallel lines wouldeventually cross each other, so arico et al., 1995;Abdallaet al., 1996).
there is an increase of mixedwins. So, graphs like Fig. Twofactors largely explain this nice outcome.First,
7 can be adjusted to suggestresults for various subsets substantial genotype-year (and/or genotype-location-
or extensions of environmentsof special interest in the year) interactions relative to genotype-locationinterac-
past or the future. tions reducethe portion of the interaction that has pre-
Theabovebrief remarkssketch only a few of the sorts dictive value for future years. Whena locations points
of questions that can be framedand addressedby graphs are spread widely by interaction differences (like AX
like Fig. 7 and the earlier graphs. These are highly but not BRin Fig. 3), less advantage can be taken
informative graphs that can promote insight on past of specific adaptations, whichresults in fewer usable
results and future options. There are other useful kinds mega-environments. Second,the roster of winnersoften
of AMMI graphs not included here. Gauch(1992, p. 222) includes somegenotypes with very small domains, so
extends the AMMI-1winning domains to the AMMI-2 the adjacent major winneror winnerscan be substituted
model,in whichdomainsform irregular polygonsinstead with negligible yield loss to retain just a few big mega-
of horizontal bands. Gauch(1992, p. 218) depicts loca- environments.
tions across years as regions, rather than points, in an However, experience with AMMI mega-environment
AMMI biplot by delineating each locations points across analysis is limited, so it is possiblethat there will also
years. Atypical outcomeis that somelocations differ be somesurprising cases requiring morenumerousmega-
from year to year mostlyin maineffects, others mostly environmentsthan expected.To introduce a newstatisti-
in interactioneffects, others in both, andstill othersare cal methodology for exploiting interactions, this study
rather stable. Fig. 2, 3, 4, and7 couldreverse the roles has used a small, convenient,previouslypublisheddata-
of genotypes and environmentsto create new kinds of set. But future studies are planned analyzing several
graphs addressing someother questions related to mega- large yield trials.
environments. As agronomistsand breeders consider various propos-
It must be stressed, however, that each and every als for exploiting interactions, includingthe statistical
analysis of interaction, including AMMI, pertains di- methodologyintroduced here, the maincomparisonmust
rectly to that particular collection of genotypesandenvi- be with prevailing current practices, described without
ronmentsin the dataset. Introduction of substantially illusions. Simmonds (1991)gives a fair appraisal, saying
different genotypes or environmentsin the future, or that there is "general awareness"of the challengesposed
deletion of distinctive genotypesor environments already by genotype-environment interactions but that "little is
in the dataset, can changeresults substantially. It is a donein practice beyondtesting potential newcultivars
statistical maxim that extrapolationis risky, althoughit in diverse environmentsin the hope of revealing wide
is also an empirical result that both main effects and adaptation (i.e., stability) of performanceas well
324 CROP SCIENCE, VOL. 37, MARCH-APRIL 1997

good mean performance." More specifically, breeders cal or impossible, so model interpretation must turn to
regularly select in high-yielding environments and hope other options. Fortunately, regardless how many IPCA
for good performance in low-yielding environments also, axes an AMMI model requires for a given dataset, it is
"even when they knowthat the procedure is suboptimal." always possible and simple to list the mega-envi-
Likewise, breeders often make selection decisions on ronments, listing the environments won for each winning
yield averages across wide area tests, but this can con- genotype (according to yield estimates from the AMMI
found mega-environments together, thus obscuring and model chosen to be most predictively accurate). At-
diminishing yield differences. Furthermore, as Nielsen tempting to relate these mega-environments to evident
(1992) remarkedin his presidential address to the Ameri- geographical or agroecological differences offers the best
can Society of Agronomy,"Westill rely on statistical hope for a straightforward interpretation. Frequently,
methods developed more than one-half century ago," mega-environments delineated by AMMIdo have an
such as "analysis of variance and regression techniques." evident geographical or biological interpretation (Annic-
Yet the regression technique developed by Yates and chiarico, 1992; Annicchiarico and Perenzin, 1994; An-
Cochran (1938) and popularized by Finlay and Wilkinson nicchiarico et al., 1995). WhenAMMI-3or higher mod-
(1963) and Eberhart and Russell (1966) remains the els are required, a preliminary clustering may be
commonmethod for analyzing interactions, even though considered, hopefully resulting in just a few clusters
an empirical comparison by Yau (1995) shows that being adequate to allow AMMI-1or AMMI-2to suffice
AMMI routinely captures much more of the interaction for each data subset. It maybe suspected that high-order
(and never less). To meet future needs for crops that modelswill occur mostly for international scientific yield
are highly productive in a diversity of environments, trials containing numerous entries that are obviously
muchof current practices wouldprove inadequate. Inter- nonstarters for local farmers. Finally, when three or
actions must be exploited vigorously, especially when more IPCA axes are recommended, the particular test
genotype-location interactions are sizable relative to ge- procedure or criterion used to reach this conclusion
notype main effects and genotype-year interactions. should be checked to assure its appropriateness relative
Mega-environmentresults are useful only if they are to research objectives. For example, F-tests tend to over-
reasonably stable-only if somewhatdifferent genotypes estimate statistical significance relative to predictive tests
or locations or years would cause little change. More (Berger and Berry, 1988; Gauch, 1992, p. 129-153), and
research is needed to quantify the robustness of the hence F-tests tend to recommendrelatively high-order
present methodology, particularly the quantity of data models. Likewise, an automatic goal of capturing, say,
required to achieve stable results. One simple test would 70 to 80%of the interaction can be seriously misguided
be to compare results for numerous subsets of several in the absence of a quantitative estimate of the amount
large datasets. However,there are substantial indications of signal and noise in a given datasets interaction. Fortu-
in the literature of long-term association of locations nately, "experience has shownthat only very infrequently
(DeLacy et al., 1994), stable mega-environmentsacross there are sufficient grounds for including more than two
years and even across crops (Annicchiarico and Perenzin, axes" (van Eeuwijk, 1995).
1994; DeLacy et al., 1994; Abdalla et al., 1996) and Breeding in marginal or stressed mega-environments
across different subsets of genotypes (Annicchiarico et calls for exceptional experimental technique to control
al., 1995), and heritability of interaction traits (Pham errors. Low-productivity environments are prone to large
and Kang, 1988; Berke et al., 1992; Jalaluddin and errors, less differentiation between genotypes, and less
Harrison, 1993; Romagosaet al., 1993). Romagosaand repeatability across years (Braun et al., 1992). Hence,
Fox (1993) claim that multivariate analysis (including it is important to select an optimal numberof replications
AMMI)of new genotypes with known genotypes based (Gauch and Zobel, 1996b). Also, AMMIanalysis
on just 1 yr of wide testing provides reliable predictions regional data can help to control errors and gain accuracy
for future years. Such rapid assessment of a new geno- (Gauch, 1988; Gauch and Zobel, 1996a).
types mega-environment is important because commer- Because historically muchgreater breeding effort has
cial cultivars for manycrops have an average lifetime been expended for major production areas than for mar-
of just several years. Especially if years are deemeda ginal environments, to a considerable extent, any verdict
randomeffect and a dataset is sizable, a different set of on the merit of exploiting narrow adaptations is likely
years is likely to give similar results. However,if the to underestimate the potential. However,Ceccarelli and
genotypes change drastically, as by introduction of a Grando(1993) report a rare case in which similar breed-
new idiotype, a new assessment of mega-environments ing efforts were given to local landraces and elite lines.
is needed. For example, Braun et al. (1992) report that They found that selection of locally adapted barley
the introduction of semi-dwarf wheat increased GEinter- (Hordeumvulgare L.) landraces in a low-yielding mega-
action. In any case, occasional review and refinement environment produced superior entries beating the check
of mega-environmentsis desirable (DeLacyet al., 1994). variety 30 times as frequently as did selection of elite
Somedatasets may require a higher-order model than barley lines in a high-yielding mega-environment.
AMMI-1.Extension to AMMI-2 is relatively straightfor- Clearly, in other cases, if the genotype roster contains
ward. The winning domains, which are horizontal bands only genotypes bred in and for highly productive environ-
for AMMI-1as in Fig. 3, become irregular polygons ments, the results may not be indicative of either the
for AMMI-2(Gauch, 1992, p. 222). For AMMI-3and number of needed mega-environments or the yield boost
higher models, however, analogous graphs are impracti- from exploiting narrow adaptations that would result
GAUCH & ZOBEL: IDENTIFYING MEGA-ENVIRONMENTS 325

were the roster expanded to also include genotypes vigor-


ously bred for marginal environments. As a quick diag-
nostic, whenever the genotype-location interaction SS
(discounted for noise) equals or exceeds the genotype
SS and the genotype-year SS, which is a common case,
subdivision of the growing region into several mega-
environments usually cannot be ignored without a sub-
stantial yield loss. If a breeding program in the past
focused on a single highly productive and well-defined
environment but in the future will serve a wider diversity
of mega-environments, then to maintain research effi-
ciency and productivity, stronger statistical tools will be
needed to comprehend interactions, reduce noise, iden-
tify mega-environments, and target genotypes.

ACKNOWLEDGMENTS
We appreciate helpful suggestions on earlier drafts of this
paper from P. Annicchiarico, P.L. Cornelius, J. Crossa, B.E.
Eisenberg, R.A. Fischer, M.S. Kang, H. Pham, H.TP. Piepho,
M. Sorrells, A.F. Troyer, W. Van, K. Yourstone, D. Wallace,
and three anonymous reviewers. The senior author especially
appreciates some penetrating questions from R.A. Fischer
several years ago at CIMMYT, which precipitated the develop-
ment of graphs like Fig. 3. This research was supported by
the Rhizobotany Project of the USDA-ARS.
326 CROP SCIENCE, VOL. 37, MARCH-APRIL 1997

View publication stats