You are on page 1of 2

Analyses for multi-site experiments using augmented designs

Walter T. Federer
Cornell University, Departments of Biometrics and Statistical Sciences
434 Warren Hall, Cornell University
Ithaca, New York, USA
wtf1@cornell.edu
Mathew Reynolds
CIMMYT, Department of Plant Breeding
Mexico D. F., Mexico
Jose Crossa
CIMMYT, Biometrics and Statistics Unit
Mexico D. F., Mexico
jcrossa@alphac.cimmyt.mx
1. Introduction
The class of augmented experiment designs (AEDs) contains two kinds of treatments, check or
standard and new. The latter are usually considered to be random and the former as fixed. The new
treatments are usually unreplicated while the checks are replicated to obtain an estimate of the error
and to obtain estimates of blocking effects. An experiment design is selected for the check treatments
and then the blocks, rows, and/or columns are enlarged to accommodate the new treatments and this
forms the augmented design. AEDs were introduced by Federer (1956) (Also see Federer et al., 1975
and Federer, 1995, 1998) as an alternative to the systematically arranged check and new treatments.
AEDs have several advantages over the systematic arrangement. They are useful for screening new
treatments such as genotypes, insecticides, herbicides, drugs, etc.
An appropriate response model needs to be selected for the results from each AED. A mixed
effects model with two kinds of treatments needs to be formulated. For multi-site trials, this may mean
a different experiment design and a different model for each site, thus complicating the combination of
results. The errors at each site may, and usually do, vary, further complicating combining results.
Blocking effects and site effects may be considered to be random effects.
2. Data Sets
The data are yields of 120 new wheat genotypes grown at three sites. The experiment design at
two of the sites was a 15 row by 12 column layout of the 120 new entries included once and two
checks which were included 30 times each. The response models used at these two sites for the checks
were:
Y
hij
= +R1+R2+R4+R8+R10+C1+C2+C3+C4+C6+C8+R1*C1+R1*C2+R1*C3+%
hij
and
Y
hij
= +R2+C1+C4+C10+R1*C1+R2*C2+R2*C3+R3*C4+R4*C2+R4*C3+R4*C4+%
hi
j
where Y
hij
is the yield of the j
th
check in the h
th
row and i
th
column, is a general mean effect, Rh and
Ci are polynomial regressions of row and column effects, respectively, and %
hij
is a random error
effect distributed with mean zero and variance 5
2
. A mixed model effect model for the Rh and Ci and
for new genotypes was used to obtain the new treatment means. Considerable reduction in the residual
mean squares were obtained for these models over those obtained from a model with only row,
column, and entry effects. The designs were also unconnected for these three effects considered
simultaneously.
The experiment design at the third site was an incomplete block design for v = bk = 120
treatments in b = 15 incomplete blocks of size k = 8 in each of two replicates. The response model
selected was
Yield = rep + entry + block within rep + linear gradients within block and rep + random error
A reduction in the error mean square of 21% over the textbook incomplete block design analysis was
obtained by taking the linear gradients within each each incomplete block into account.
3. Methods of Combining Over Sites
Several procedures are available for combining results over sites. Some of these are:
(i) a fixed effects analysis for the selected model
(ii) a mixed model analysis for blocking and new treatments random
(iii) a fixed effects analysis on site means for entries
(iv) a mixed effects model with site and new as random on fixed effect site means
(v) a site by entry analysis with REML solutions for site means
(vi) a site by entry analysis of site REML means divided by standard errors
The method selected will be to obtain the best estimate of the entry means for each site and to divide
these means by their standard errors. Then a two-way site by entry analysis will be performed on these
means. Since the variance will be one, an estimate of the site by entry interaction variance component
may be obtained. REML solutions for entry means are obtained for sites and entries random. The
selected procedure will be compared with each of the other methods to determine the effect on the
rankings of entry means over sites.
References
Federer, W. T. (1956). Augmented (or hoonuiaku) designs. Hawaiian Planters Record LV(2):191-208.
Federer, W. T. (1993). Statistical Design and Analysis for Intercropping Experiments. Volume I. Two
Crops. Springer-Verlag, New York, Heidelberg, Berlin, Chapter 10.
Federer, W. T. (1998). Recovery of interblock, intergradient, and intervariety information in
incomplete block and lattice rectangle designed experiments. Biometrics 54:471-481.
Federer, W. T., Nair, R. C., and Raghavarao, D. (1975). Some augmented row-column designs.
Biometrics 31:361-373.
Resume
Dessians augmentes taient developpes par slection de traitements nouveux en une manire
capable. Ensembles donnes des trois emplacements taient employes illustrer le analyse statistique
par emplacement chaque et par rsultats combiner sur emplacements. Mthodes de modeles mles
taient employes.