AUGMENTED EXPERIMENT DESIGNS WITH RECOVERY OF

INTERBLOCK AND INTERVARIETY INFORMATION

Walter T. Federer

Biometrics Unit
Cornell University
Ithaca, NY 14853

BU-1266-M November, 1994

 Abstract

The class of augmented experiment designs has been found useful for experiments involving
comparisons of standard check treatments with a set of new and untried treatments. Often
the new treatments have limited experimental material, or the experimenter may wish to
use few experimental units and screen a large number of new treatments. The promising
new treatments from this stage are then tested in replicated experiments or they can become
the checks for another experiment using an augmented experiment design. By combining
treatments from various stages of testing, efficiency of experimentation can be greatly in-
creased. With regard to the new treatments, usually included only once in an experiment,
every augmented design is incomplete for the new treatments. This means that interblock,
interrow, and/or intercolumn information is available to use in obtaining solutions for new
treatment effects. Also, since the new treatment effects can often be considered to be random
effects, their distributional properties can be used to increase the efficiency of experimenta-
tion. We demonstrate the statistical procedures for recovering this information in block and
row-column designed experiments.
1 Introduction
Various augmented experiment designs have been presented in the literature (Federer, 1955,
1961; Federer and Raghavarao, 1975; Federer, Nair and Raghavarao, 1975; Federer and
Wright, 1988). The purpose of this paper is to present a statistical analysis for these ex-
periment designs making use of the information obtained from the random blocking effects
and from the distributional effects of the augmented (or new) treatments in the experiment.
Since augmented designs are used to include treatments (varieties) for which there is little
information and often limited material, these treatment or varietal effects can be considered
to be distributed around some mean and with a common variance a;.Herein we consider
that each new treatment is included once in the experiment but this need not be the case
as the procedure is easily extendable to take additional replication into account. There are
c check or standard treatments which are used to obtain the experiment design prior to
adding the augmented treatments. The check treatment yields are used to obtain solutions
for blocking and check treatment effects. The former are used to adjust the new treatment
effects. From the mean square for the new treatments, an estimate of the variance component
a; is obtained and used for adjusting new treatment means for their distributional effects.
Adjustment for distrinutional properties of the random effects makes use of all information
from an experiment.

In Section 2, augmented block experiment designs are considered. In Section 3, aug-

mented row-column experiment designs such as those described by Federer and Raghavarao
(1975) and Federer, Nair and Raghavarao (1975) are the subject of statistical analyses. In
Section 4, we present statistical analyses for resolvable row-column designs such as those
presented by Federer and Wright (1988) and those which could be obtained from the row-
column designs of Russell et al. (1981), Nguyen and Williams (1993), and John and Whitaker
(1993) by a "variety cutting" procedure. Finally, some comments on the analyses are given
in Section 5.

2 Augmented Block Experiment Designs

Among the experiment designs in this class are the augmented randomized complete blocks
(ARCBD), augmented balanced incomplete block (ABIBD), and augmented partially bal-
anced incomplete blocks (APBIBD). With respect to the augmented or new treatments all
these designs are incomplete in that all the new treatments do not appear together in the
blocks. Thus, recovery of inter block information is needed for a more efficient analysis. First
consider an ARCBD with c checks and n new treatments for a total of v = c + n treatments
in r blocks. Let the c check treatments appear once in each of the blocks (note that the c
treatments could appear in the proportions n 1 : n2 : ... : nc in each of the r blocks and the
design would still be an orthogonal one (See Federer, 1991, ch. 7)). An analysis of variance,
ANOVA, for an ARCBD is given in Table 1. Since the n new treatments each occur once
in the experiment the observation can only contribute to the new treatment estimate and
nothing to block, overall mean or error estimation (Federer and Raghavarao, 1975). From

1
an analysis on check treatment results only, the expected value of the block mean square is
CJ; + cCJ~, and an estimate of CJ~ is (B -E) j c. This is the variance component to be used to
obtain adjusted new treatment means recovering interblock information. For our analysis,
we use the following linear model:

YiJ = (f.l + Pi + TJ + Eij )niJ (1)

where f.l is a general mean effect, Pi is the ith block effect, TJ is the yth treatment effect,
i = 1, ... , r, j = 1, ... , v, niJ is one if the yth treatment occurs in ith block and zero otherwise,
and Eij is a random error effect distributed with mean zero and variance CJ;. The resulting
normal equations are:

[
Krxr
NB~xr
NBrxv
Rvxv
l[ /3rxl ]
Tvxl
= [ YBrxl
YT vxl
] (2)

where K is a diagonal matrix with block sizes on the diagonal, R is a diagonal matrix with
replication numbers on the diagonal and is

R _ [ ric 0cxn ] (3)

- 0' In '

where Ix is an identity matrix of side x, 0 is a matrix whose elements are all zeros, and NB
is the incidence matrix of blocks and treatments with a one appearing for the treatments
which occur in block i and zero for those not appearing. The check treatments all have ones
for every block. Therefore, a reasonable restriction is for the sum of the check treatment
effects to be zero and likewise for the sum of the block effects. With these restrictions a
solution of the normal equations is possible. Let

NB = [J rxc NN rxn] (4)

and
J0 = [Jrxc 0rxn], (5)
where J is a matrix whose elements are all ones. Then, the usual solutions for a randomized
complete block design are

---y.--y
/3 z - z. .. andf·--y.--y
J- ·J .. ' (6)

where the observations in the above equations are for check yields only. The new treatment
solutions are:
f1 = YiJ- fh , (7)
where YiJ is the yield of new treatment j in block i and fh is as in (6).
The various sums of squares in Table 1 are:
r n
(n - 1) N = "L...J "
L...JfJ· (Y;..
tJ - -y.z. ) n tJ
.. ' (8)
i=lj=l

where YiJ is the yield of new treatment j in block i, and the other sums of squares are from
standard procedures.

2
Source of Variation Degrees of Freedom Sum of Squares Mean Square

Total T*
Correction for mean 1 M = Y.~/T
Blocks (ignoring treatments) r-1 block tot. squared/no.-M
Treatments (eliminating blocks) v-1 see text
Among check treatments c-1 see below c
Check vs. new 1 see text
Among new treatments n-1 see text N
Residual (r- 1)(c- 1) subtraction E

ANOVA on check yields only (standard analysis)

Source of Variation Degrees of Freedom Sum of Squares Mean Square

r c
Total rc L:L:Yi;
i=lj=l

Correction for mean 1 Y.~ jrc

n
Block (r- 1) '"""~-
L...,; c
y2
rc B
i=l

Vc

Check (c- 1) "Y~jr-!::

L...,; .J rc c
j=l

Check x block (r- 1)(c- 1) subtraction E

*T = rc +n.
Table 1: ANOVA for an ARCBD

3
 To recover interblock and intervariety information formula (2.2) is changed as follows
before applying the restrictions and obtaining the solutions

K + ff;Ir/ ff~
[ ------- (9)
NB'

where ff; is the estimate for the error variance component O";, ff~ is the estimated block
variance component, and ff; is the estimate of the variance component O";. Here the new
treatment effects are random effects distributed with mean T. and variance O";. The check
variety effects are fixed effects. Substituting NB - J0 for NB in (9) results in solutions for
the Tj effects. The treatment effects adjusted for interblock and intervarietal information
are:

x [( :~ ) - NB' ( K + ~ r l
I, YB

Var x [ ( :~ ) - NB' ( K + ~ r lI, YB (10)

Variances of differences among treatment effects are obtained from the following variance-
covariance matrix.
ff; x Var (11)
An approximate average variance of a difference between two adjusted new treatment effects
may be obtained as
(12)
where II · II denotes the determinant of a matrix and Var is determined for new treatments
only. The value of (12) is less than or equal to the correct average variance of a difference but
may be useful for a quick scrutiny of the adjusted effects (or means) in certain situations.
The statistical analysis outlined above is directly applicable to an analysis for ABIBD,
APBIBD, and other augmented incomplete block designs. The K and NB matrices will
need to be adjusted to take into account the experiment design for the check treatments.
When the check treatments are in an incomplete block experiment design, their effects will
need to be adjusted for recovery of interblock information. The remainder of the analysis
proceeds as described above.

3 Augmented Row-Column Experiment Designs

A number of augmented row-column designs (ARRCD) have been presented by Federer and
Raghavarao (1975) and Federer, Nair, and Raghavarao (1975). One particular row-column

4
design for few checks and many new treatments is to have the checks repeated more than
once in each row and or column. Jose Crossa, CIMMYT, personal correspondence, is using
a row-column design of 15 columns and 11 rows with 2 checks which appear either once or
twice in each column and twice in each row with 42 check plots and 123 new treatments.
The design will be used over several sites.
To recover interrow and intercolumn information on both check treatments and new
treatments, as well as intervariety information on new treatments, we proceed as follows for
an ANOVA as presented in Table 2. The linear model used here is

Yghi = + Pu + /h + Ti + Eghi) nghi

(J-L (13)
where J-L is a general mean effect, Pu is the effect of the lh row
distributed with mean zero
and variance CJ; , 1 is the effect of the hth column distributed with mean zero and variance
CJ; , Ti is the effect of the ith treatment where the check treatments are fixed effects and the
new treatments are random effects distribution with mean T. and variance CJ;, and nghi is
one if treatment i occurs in row g and column h and zero otherwise.
The normal equations for check plot yields only are:

[
Ka
RC
NR'
+ Y RCa;2b1
Kb + ~
NC'
u"'
NRaxc
NCbxc
Vc
l[ l [
Paxl
/bxl
Tcxl
=
YRaxl
YCbxl
YTcxl
l (14)

where Ka is a diagonal matrix with the number of check plots occurring in each row, Kb is
a comparable diagonal matrix for columns, Vc is a diagonal matrix of replication numbers
for the check treatments and is rIc for equal replication of checks, RC is the a x b row-
column design matrix of nghi values, NR is the row-check treatment design matrix, NC is
the column-check treatment design matrix, YR is a column vector of row totals, YC is a
column vector of column totals, and YT is a column vector of check treatment totals. When
there are an equal number of check plot yields for each row, column, and check treatment,
fl ... is a solution for J-L and YR, YC, and YT are totals of Yghi - fl... check plot yields.
Otherwise fl ... will not be a solution for P, and P, will need to be evaluated. In this case,
Yghi- P, values will be used to obtain totals for YR, YC, and YT. The various sums of
squares in Table 2 are computed as follows.

Columns (ignoring treatments and eliminating rows):

[[Kb- RC' K;;- 1 RC + Jbxb/kr 1 [YC- RC' K;;- 1 YR]J'

x [YC - RC' K;; 1 YR] (15)

( k an appropriate scalar, J a matrix of ones).

Treatment (eliminating rows and columns)

Let

5
Source of Variation Degrees of Freedom Sum of Squares Mean Square

a b c
Total T 2: 2: 2: nghiyg~i
g=l h=l i=l

Correction for mean 1

a
Row (ignor. col. and treat.)* a-1 2:Yi,./b9 - Y,~/T
g=l

Column
(ignor. treat. and elim. rows) b-1 see text (15)

Treatment
(elim. rows and col.) c-1 see text (17) v
Remainder T-a-b-c+2 subtraction E

Row
(elim. col. and treat.)* a-1 see text (19) R

Column
(elim. rows and treat.) b-1 see text (21) c
Augmented treatment
(elim. rows and col.) n-1 see text (23) N

*b 9 =number of check treatment yields in row g.

Table 2: ANOVA for check treatment yields in an ARCD

6
 x [YT- (NR' NC1M [ ~~ ]] (16)

where M = [
K
RC' RC- J
Kb
l- 1
.

Then, the sum of squares is

f' [YT - (NR' NC') [ :c, R~~ l-l [~~ l] ·

J (17)

Row (eliminating columns and treatments)

Let

Kb NCV-c J
x [YRT- (RC NR) [ NC' l-1 [ l]·
YYCT (18)

Then the sum of squares is

p' x Kb NCV-c J
[ YR- (RC NR) [ NC' l-1 [ l]·
YYCT (19)

Column (eliminating treatments and rows):

Let

x [vc- (RC' NC) [:a, N~~J r [~~ l]· (20)

Then, the sum of squares is

'y' x [ YC - (RC' N C) [ :~, N~~ J r[~~ l]· (21)

7
The intra-row-column solutions for new treatments are

(22)

The sum of squares among new treatments is

n n
2::: fi (Yghi - P, - Pg - ih) = L fi2 . (23)
i=l i=l

The expected values for R, C, and N in Table 2 may be evaluated using various computer
software. Then estimates of the variances o-~ and o-;, , a-;
may be obtained. These solutions
are then substituted in (14) and the matrices are expanded to include the new treatments.
With these changes the new treatment effects with recovery of interrow, intercolumn, and
intervariety information are obtained. The variances of difference may be obtained as de-
scribed in Section 2.

4 Resolvable Augmented Row-Column Experiment

Designs
Lattice square experiment designs are resolvable row-column experiment designs and with
"variety-cutting" may be used to construct augmented resolvable row-column designs, AR-
RCD (Federer and Wright, 1988). For v = k 2 there are k- 1 suitable arangements from a
balanced lattice square design to construct resolvable row-column designs and a computer
package has been prepared for constructing such designs (Nam-Ky Nguyen, 1994, personal
communication). These designs may be used to construct ARRCDs by using "variety cut-
ting" in the manner described by Federer and Wright, (1988). To illustrate, consider the
following arrangements from a balanced lattice square plan with v = 9

Replicate 1 Replicate 2 Replicate 3 Replicate 4

123 147 159 168
4 56 258 672 924
789 369 834 573

Then "cut-varieties" 7, 8, and 9 and fill these spots with new treatments a, b, c, d, e and f.
Replicates 1 and 2 are not suitable as all new treatments would appear in a row or column
and solutions for these rows or columns would not be possible with the usual restrictions
for an analysis. With 2k check plots in k- 1 replicates, k(k- 1)(k- 2) new treatments
each replicated once could be included in an ARRCD. Note that the 2k checks could include
duplicated checks in each replicate. An ANOVA for an ARRCD is given in Table 3, for the
following linear model

Yghi = (J.L + /3g + Pgh + /gi + Tj + Eghii) nghii· (24)

J.L is a general mean effect, /3g is a lh replicate effect, Pgh is the kth row effect in replicate g
and is distributed with mean zero and variance u;, /gh is the ith column effect in replicate

8
g and is distributed with mean zero and variance o-~, Tj is the lh treatment effect, Eghi, is
a random error effect distributed with mean zero and variance a;, and 'T/ghij is an indicator
variable which is one when treatment j occurs in row h and column i in replicate g and zero
otherwise.

l[ l [ l
The normal equations recovering interrow and intercolumn information are:

Ira (b + 8-';/8-~) RC NR Praxl YRraxl

[ RC' Irb (a+ 8-'; /8-~) NC (rbxl = YCrbxl (25)
NR' NC' V Tcxl YTcxl

where YR, YC, and YT are totals of Yghij- (fl + Pu) values to remove effect of general mean
Jl and replicate (complete block) effects when the usual restrictions that sums of effects are
zero are imposed. When Jl and f3u effects are orthogonal to the remaining effects fl+Pu = Yg ... ·
The various sums of squares are computed in the manner described in the previous
section taking into account the nature of the various matrices involved. The intra-row-
column solution for a new treatment effect is
~ v (3 ~ ~ ~
(26)
Tj = .rghii- Jl- g- Puh- (gi,

and is
fi = Yghii - Yg ... - Puh - 'Y9i, (27)
when Y... on the checks contains an equal number of each of the effects Puh, (gi, and Tj· That
is, when the restrictions on the solutions is that
a b c
L Puh = L 'Yui = L fg = 0, (28)
h=l i=l j=l

these effects disappear in the total Y:... from the check yields. A sum of squares for new
treatments is n n
L
j=l
fg (Yghii- fl- P9 - Puh- 'Yui) = L:fJ.
j=l
(29)

The expected value for (29) needs to be evaluated to obtain the coefficients of and a';. a-;
With estimates of a;,
o-~, o-~, and a-;,
new treatment effects recovering interrow, intercolumn,
and intervariety information are obtained from (10) with the appropriate interpretation of
the various matrices. Likewise, variances for differences of these adjusted effects may be
obtained as explained for equation (11).

5 Comments and Discussion

Computation of sums of squares and adjusted effects for the proposed analyses poses little
difficulty owing to the availability of computer software packages such as GAUSS. Some
other standard statistical packages such as SAS and GENSTAT can be used to recover
interblock and intervariety information. Some packages such as MATHEMATICA may be
used to obtain expected values for mean squares. Then, solutions for treatment effects

9
 Source of Variation Degrees of Freedom Sum of Squares Mean Square

r a b c
Total rc* 2:: L
g=l h=l
2:: LY9~ii nghii
i=l j=l

Correction for mean 1 Y_~.frc

r
Replicate = B r-1 L Yg~ .. jc - Y.~.jrc
g=l
Row in B
r a y2 r y2
(ignor col. and checks) r(a- 1)+ 2::2:: gh .. _l:~
g=l h=l bgh g=l c

Column in B
(ignor. treat., elim. rows) r(b - 1) see text

Check
(elim. rows and columns) c-1 see text T

Error rc- ra- rb subtraction E

+r-c+1 E

Row
(elim. col. and checks) r(a- 1) see text R

Column
(elim. rows and checks) r(b- 1) see text c
New treatment
(elim. B, rows, and col.) (n- 1) see text N

*for c check treatments included once in each replicate. Otherwise, the total here is the total
number of check treatment yields and c needs to be changed as required.

+ b9 h equals the number of check plot yields in row gh.

Table 3: ANOVA for an ARRCD with rc check plot yields.

10
recovering inter block, interrow, intercolumn, and intervariety information are possible. These
computations pose little difficulty using GAUSS and concatanating submatrices or multiples
thereof.
Using only intrablock or intrarow-column information leads to inefficient statistical anal-
yses for a set of data. Since other information such as interblock, interrow, intercolumn
and/or intervariety information is available in experimental data it should be used to obtain
solutions for treatment effects. As demonstrated by Federer and Speed (1987), the efficiency
of an experiment design can change considerably when using a measure of design efficiency
that incorporates interblock information as compared to an intrablock measure.

References

Federer, W.T. (1955). Augmented (or hoonuiaku) designs. Hawaiian Planter's Record 55,
191-208.

Federer, W.T. (1961). Augmented designs with one-way elimination of heterogeneity. Bio-
metrics 17, 447-473.

Federer, W.T. (1991). Statistics and Society - Data Collection and Interpretation. Marcel
Dekker, Inc., New York. Chapter 7.

Federer, W.T., Nair, R.C. and Raghavarao, D. (1975). Some augmented row-column designs.
Biometrics 31, 361-374.

Federer, W.T. and Raghavarao, D. (1975). On augmented designs. Biometrics 31, 29-35.

Federer, W.T. and Speed, T.P. (1987). Measures of block design efficiency recovering in-
terblock information. Transactions of the Fourth Army Conference on Applied Mathe-
matics ARO Report 87-1, 781-786.

Federer, W.T. and Wright, J. (1988). Construction of lattice square design. Biometrical
Journal 30 (1), 77-85.

John, J.A., and Whitaker, D. (1993). Construction of resolvable row-column designs using
simulated annealing. Australian J. of Statistics 35, 237-245.

Nguyen, N. and Williams, E.R. (1993). An algorithm for constructing optimal resolvable
row-colum designs. Australian J. of Statistics 35 (3), 363-370.

Russell, K.G., Eccleston, J.A., and Knudsen, G.J. (1981). Algorithms for construction of
(M, S)-optimal block and row-column designs. J. Statistics, Computation, and Simula-
tion12, 93-105.

11