The Analysis of Square Lattice Designs

Using R and SAS

Patchanok Srisuradetchai

Department of Mathematical Sciences

Montana State University

April 17, 2012

A writing project submitted in partial fulfillment

of the requirements for the degree

Master of Science in Statistics

 APPROVAL

of a writing project submitted by

Patchanok Srisuradetchai

This writing project has been read by the writing project advisor and has been found
to be satisfactory regarding content, English usage, format, citations, bibliographic
style, and consistency, and is ready for submission to the Statistics Faculty.

Date John J. Borkowski

Writing Project Advisor

Date Mark C. Greenwood

Writing Project Coordinator
 1

Contents
1 INTRODUCTION 3

2 SQUARE LATTICE DESIGNS 5

2.1 A Balanced Square Lattice Design . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Lattice Design Construction . . . . . . . . . . . . . . . . . . . 5
2.1.2 Randomization . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.3 Statistical Analysis . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Partially Balanced Square Lattice . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Construction and Randomization . . . . . . . . . . . . . . . . 10
2.2.2 Statistical Analysis for Partially Balanced Lattices . . . . . . 11

3 NUMERICAL EXAMPLES 14
3.1 Example of a balanced square lattice design . . . . . . . . . . . . . . 14
3.2 Example of a partially balanced square lattice design . . . . . . . . . 17

4 R AND SAS PROGRAMS 22

4.1 R and SAS Programs for a balanced lattice design . . . . . . . . . . . 22
4.1.1 R Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.2 SAS Program . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 R and SAS Programs for a partially balanced lattice design . . . . . . 27
4.2.1 R Program . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.2 SAS Program . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5 R PROGRAMS FOR SQUARE LATTICE DESIGNS 32

5.1 A Balanced lattice design . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 A Partially balanced lattice design . . . . . . . . . . . . . . . . . . . 33

A R Code for lattice.design 34

B References 40

List of Tables
1 3 × 3 balanced lattice (t = 9, k = 3, r = 4, λ = 1) . . . . . . . . . . . . 6
2 Two orthogonal latin squares of order 3. . . . . . . . . . . . . . . . . 6
3 Analysis of Variance for the balanced square lattice design . . . . . . 8
4 Analysis of Variance for the partially balanced square lattice design . 12
5 Gain in weight for 2 pigs . . . . . . . . . . . . . . . . . . . . . . . . . 14
 2

6 Treatment Totals and Adjustment Factors . . . . . . . . . . . . . . . 15

7 Analysis of Variance for the balanced square lattice design example . 16
8 The adjusted treatment totals . . . . . . . . . . . . . . . . . . . . . . 16
9 Data of yields in bushels per acre of 25 varieties of soybeans . . . . . 18
10 Treatment totals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
11 Value of Cl for each block . . . . . . . . . . . . . . . . . . . . . . . . 19
12 Analysis of variance table for example of yield data . . . . . . . . . . 19
13 Adjusted treatment totals . . . . . . . . . . . . . . . . . . . . . . . . 20
 3

1 INTRODUCTION
For experiments using randomized block designs, all treatment combinations may not
be run in each block because of shortages of experimental units or facilities. Yates
(1936a) formally introduced balanced incomplete block (BIB) designs in which every
treatment is not present in every block, but the number of pairs of each treatment
occurring together is the same. A BIB design can reduce the number of experimental
units used in experiment as we already knew. Although BIB designs are efficient
designs, these designs are still not appropriate for experiments with a large number
of treatments such as animal breeding experiments. Moreover, the minimum number
of blocks required for a BIB design may be too large to be practical. In the same
year, Yates (1936b) proposed a new method of arranging agricultural variety trials
involving a large number of crop varieties. These types of arrangements were named
a quasi-factorial or lattice designs. His paper contained numerical examples based on
the results of a uniformity trial on orange trees. A special feature of lattice designs is
that the number of treatments, t, is related to the block size, k, in one of three forms:
t = k 2 , t = k 3 , or t = k(k + 1). Even though the number of possible treatments
is limited, a lattice design may be an ideal design for field experiments with a large
number of treatments. Although lattice designs have been frequently used, there is
limited software that performs an appropriate statistical analysis for a lattice design.
The SAS package is one of the most commonly-used packages for the design and
analysis of experiments. The use of SAS for analyzing a lattice design was first
discussed in a colloquium on horticulture (Stroup and Paparozzi, 1989). At that time,
PROC MATRIX, PROC IML, and PROC LATTICE, which can analyze experiments
using lattice designs, were only found on main-frame computers. Fernandez G. C.J.
(1990) presented a PC-SAS program for the analyses of data obtained from a lattice
design. His program consisted of two parts. In the first part, PROC GLM was used
to calculate unadjusted block sum of squares (SS), adjusted block SS, unadjusted
treatment SS, and intra-block error. In the second part, PROC MEANS and the
MERGE option were used to calculate many statistics such as adjusted block values
and treatment totals. Now PROC LATTICE is available in PC-SAS and will analyze
data from balanced square lattices, partially balanced square lattices, and some
rectangular lattices.
Besides SAS software, R, a free software environment for statistical computing and
graphics is capable of analyzing experimental design data. Groemping (2011) provides
a summary of R packages related to the design and analysis of experimental data. The
 4

R package “agricolae” has the capability to analyze data obtained from lattice designs.
Although it was not specifically created for lattice designs, there is one function
which can be applied to use with lattice designs. Felipe de Mendiburu (2010) wrote
the“agricola” package as a master’s degree project, and it offers extensive functionality
on experimental design especially for plant breeding and agricultural experiments.
The strength of this package is that it allows for the randomization of treatments
in lattice designs, factorial designs, randomized complete block designs, latin square
designs, balanced incomplete block designs, alpha designs, cyclic, augmented block,
and split and strip plot designs. Moreover, this package can also perform an analysis
of variance for many designs. The R “agricolae” package is the newest tool for the
analysis of lattice design data. More details are in a document by Felipe de Mendiburu
(2010).
This paper provides a review of the analysis of two types of square lattice designs:
balanced and partially balanced lattices, as well as numerical examples and the R
and SAS computer code needed to analyze the experimental data. In the last section,
the R function are written to give the complete analysis of variance which both SAS
and R do not provide.
 5

2 SQUARE LATTICE DESIGNS

The feature of balanced square lattices is that the number of treatments, t, is equal to
the square of the number of units per block, k or t = k 2 . The numbers of replication
of partially balanced square lattices are similar to balanced square lattices, but only
some replications are selected. If the number of replications is less than required for
a balanced design, the analysis follows the same procedure as for a balanced design,
but some formulas are changed.

2.1 A Balanced Square Lattice Design

A balanced square lattice design is similar to a balanced incomplete block design
with k 2 treatments arranged in k(k + 1) blocks with k runs per block and r = k + 1
replications. So, each replication has k blocks and contains every treatment. In this
design, every pair of treatments occurs together once in the same incomplete block.
This property holds for all plans having an odd number of treatments which is also
a perfect square (e.g. 9, 25, 49, 81, 121, and 169 treatments). Let λ be an integer
number indicating how many times each treatment occurs together in same block,
and the relationship among the number of treatments t, block size k, and number of
replications r. Numerically, it is defined as λ = r(k − 1)/(t − 1) (Montgomery, 2005).
In balanced square lattice designs, r = k + 1 or t = k 2 , implying λ = 1 . Because
the designs are balanced, all treatment differences have the same estimated variance
or the same precision. Hinkelmann and Kempthorne (2005) referred to the number
of replications, r, as the number of different systems of confounding. Designs with
r = 2 are called simple (or double) lattices; designs with r = 3 are called triple
lattices; designs with r = 4 are called quadruple lattices, and balanced square lattices
require r = k + 1.

2.1.1 Lattice Design Construction

Assume there are t treatments labeled as 1, 2, . . . , k 2 with treatment numbers arranged

in a k × k square. For example, in a 3 × 3 square forming Replication I (see Table 1),
there are nine treatment numbers arranged in a specific order such that each row of
the square array is considered as a block containing three treatments. To construct
Replication II, each column of the array for Replication I is taken to form the three
blocks in Replication II (see Table 1). Replications III and IV are based on two
orthogonal latin squares. Two latin squares are orthogonal if, when superimposed,
 6

all ordered pairs are distinct (See Table 2). From the standard array, the treatment
numbers that fall on the same letter in a latin square are taken to form a block. For
example, from latin square 1 (See Table 2), treatment numbers 1, 6 and 8 fall on
the letter A, so treatments 1, 6 and 8 are in the same block in Replication III (See
Table 1). Analogous to Replication III, Replication IV is constructed from the second
latin square. Another method of constructing lattice square designs can be found in
Federer and Wright (1988) who proposed a simple method for constructing lattice
square designs when the number of treatments is greater than 3.

Table 1: 3 × 3 balanced lattice (t = 9, k = 3, r = 4, λ = 1)

Block Replication I Block Replication II

1 1 2 3 4 1 4 7
2 4 5 6 5 2 5 8
3 7 8 9 6 3 6 9

Block Replication III Block Replication IV

7 1 6 8 10 1 5 9
8 2 4 9 11 2 6 7
9 3 5 7 12 3 4 8

Table 2: Two orthogonal latin squares of order 3.

Latin Square 1 Latin Square 2 Superimposed

A B C A B C
(1) (2) (3) (1) (2) (3) (A,A) (B,B) (C,C)
B C A C A B
(4) (5) (6) (4) (5) (6) (B,C) (C,A) (A,B)
C A B B C A
(7) (8) (9) (7) (8) (9) (C,B) A,C) (B,A)
 7

2.1.2 Randomization

Randomization is always important when designing an experiment. The

randomization procedure gives an equal chance for each assignment of treatments to
the experimental units. Randomization consists of the following steps: (1) randomly
allotting the treatments to the treatment numbers (labels), (2) randomizing the
replications, and (3) randomizing the blocks separately and independently within
each replication.

2.1.3 Statistical Analysis

Although the following procedure can be applied to partially balanced designs, there
are some differences which will be discussed later. We will adopt the following
notation. Let
t denote the total number of treatments,
k denote the number of units per block or block size,
s denote the number of blocks per replication which is equal to k,
r denote the number of replications (for balanced designs, r = k + 1).
Let yij(l) denotes the response value of the jth treatment in the lth block within ith
replication, i = 1, 2, . . . , k + 1, j = 1, 2, . . . , k 2 , l = 1, 2, . . . , rk. The model is

yij(l) = µ + πi + βi(l) + τj + εij(l) .

where µ, πi , βi(l) , and τj represent the effect of the mean, the replicate, the incomplete
block, and the treatment, respectively. εij(l) is the intra-block residual, assumed to
be normally and independently distributed with mean 0 and variance σe2 .
Various ANOVA sums of squares are now presented:

1. Total Sum of Squares:

X
2
SST ot = yij(l) − CF (1)
, where CF = ( yij(l) )2 /(rk 2 ),
P P
yij(l) is the grand total.

2. Unadjusted treatment sum of squares:

Tj2
P
SST rtU = − CF (2)
r
where Tj is the sum of observations for treatment j.

3. Replication sum of squares:

Ri2
P
SSR = − CF (3)
k2
 8

where Ri is the sum of observations in replication i.

4. For computing the adjusted block sum of squares, SSBAdj , several quantities
are required to be computed first. Let
Bj denote the sum of block totals for the blocks with treatment j, j = 1, 2, . . . , t,
Tj denote the total of the jth treatment total from all replications,
Wj denote the weight for the jth treatment which is used for adjustment for
block,
Wj = kTj − (k + 1)Bj + G (4)
P P
where G = yij(l) , or the grand total. Note that Wj = 0. The sum of
squares for blocks within replication, adjusted for treatment effects, SSBAdj , is
defined as
Wj2
P
SSBAdj = . (5)
k 3 (k + 1)
5. Intra-block error sum of squares:

SSE = SST − SSR − SST rtU − SSBAdj . (6)

Using the various sums of squares above, the analysis of variance (ANOVA)
table is given in Table 3.

Table 3: Analysis of Variance for the balanced square lattice design

Source of variation Degree of freedom Sum of squares Mean squares

Replication r−1 SSR M SR
2
Treatment(Unadjusted) k −1 SST rtU M ST rtU
Block within replication(Adj) k2 − 1 SSBAdj M SBAdj = Eb
Intra-block Error (k 2 − 1)(k − 1) SSE M SE = Ee
2
Total rk − 1 SST

It is important to note that the mean square of the unadjusted treatment cannot be
used for testing against the mean square of intra-block error because the mean square
of unadjusted treatment still contains block effects. The adjusted treatment sum of
the squares is defined as
P 0 2
(Tj + µWj )2
P
(Tj )
SST rtAdj = − CF = − CF (7)
r r
 9

0
where Tj is the adjusted treatment total defined as Tj + µWj , and µ is the adjustment
factor for error and treatment means defined as
M SBAdj − M SE Eb − Ee
µ= 2
= (8)
k M SBAdj k 2 Eb

If dividing the SST rtAdj by its degree of freedom, k 2 − 1, we will get the M ST rtAdj .
So, the F-statistics is the M ST rtAdj divided by the the intra-block error mean square.
For t-test, the mean square of intra-block error, Ee , needs to be adjusted to take into
account the sampling error. The new mean square is called the “effective error mean
0
squar”, or Ee , which is defined as
0
Ee = Ee (1 + kµ) = Ee [1 + (r − 1)µ]. (9)

The F-statistic is
M ST rtAdj
F = (10)
Ee
The degrees of freedom of the numerator and denominator for the F statistic are
k 2 − 1 and (k − 1)(k 2 − 1), respectively.
0
The variance of the adjusted treatment mean is Ee /r , and the error variance of the
difference between two adjusted treatment means in the same block is
0
2Ee [1 + (r − 1)µ] 2Ee
= (11)
r r
The square root of error variance of the difference between two treatment means is
the standard error used for t-tests between pairs of treatments.
If, however, Eb is less than Ee , then we conclude the blocks have no effect,
and the data will be analyzed as if it were a randomized block design but using
the replications as the blocks. One numerical value that summarizes the potential
advantage or disadvantage of one specific experimental design relative to another is
called a “relative efficiency”. The relative efficiency is expressed as an unbiased
estimator of the error variance that would have been present if the experiment
had been conducted as a randomized block divided by effective error mean square,
0
Ee . The sum of the SSBAdj and SSE, SSBAdj + SSE, is called the “randomized
complete block design (RCBD) error sum of squares” which is an estimate of the
block sum of squares in a randomized block design. The degrees of freedom for
RCBD sum of squares is the sum of the degrees of freedom of SSBAdj and SSE , or
(k 2 − 1) + (k 2 − 1)(k − 1) = rk(k − 1).
 10

2.2 Partially Balanced Square Lattice

Partially balanced square lattices are similar to balanced square lattices, but are more
flexible with respect to the number of replications. Partially balanced designs do not
use all replications of the basic plan. For example, for 3 × 3 lattice designs, we need 4
replications for a balanced design, but if we have only two or three replications, these
are partially balanced designs. The designs using only the first two replications from
the basic plan are called simple lattices, and designs using the first three replications
from the basic plan are called triple lattices.

2.2.1 Construction and Randomization

After construction of a balanced lattice design, the number of replications is selected.

The function named “design.lattice” in the R package “agricolae” can automatically
create partially balanced lattice designs, but only simple and triple lattice designs.
For example, the code for creating a 4 × 4 triple lattice is:

library(agricolae)
design.lattice(4,type="triple")

and the R results are:

$square1
[,1] [,2] [,3] [,4]
[1,] 6 3 8 12
[2,] 11 13 10 15
[3,] 14 7 2 5
[4,] 9 1 4 16
$square2
[,1] [,2] [,3] [,4]
[1,] 15 5 12 16
[2,] 13 7 3 1
[3,] 10 2 8 4
[4,] 11 14 6 9
$square3
[,1] [,2] [,3] [,4]
[1,] 11 7 8 16
[2,] 13 5 6 4
[3,] 10 14 12 1
[4,] 15 2 3 9
 11

2.2.2 Statistical Analysis for Partially Balanced Lattices

Computations for the analysis of variance for a partially balanced lattice design is
more complicated than it is for a balanced design. Actually, the partially balanced
square lattices could be considered as partially balanced incomplete block designs
(PBIB). This was shown by Bose and Nair (1939). Nair (1952) used the method of
analysis of a PBIB design for the simple square lattice design. However, this paper
actually presents the method from Cochran and Cox (1950) who simplified the original
method by Yates (1936b). In the analysis, let
t denote the total number of treatments,
k denote the number of units per block (i.e., the block size),
s denote the number of blocks per replication (which equals k), and
r denote the number of replications.
The analysis procedure involves the following steps:

1. The total sum of squares (SST ot), the unadjusted treatment sum of squares
(SST rtU ), and the replication sum of squares (SSR) are computed in the same
way as those in balanced designs.

2. Find (1) the block total, Bl , l = 1, 2, . . . , rk, (2) the replication total, Ri ,
i = 1, 2, . . . , r, (3) the treatment total, Tj , j = 1, 2, . . . , k 2 , and (4) the grand
total G.

3. For each block,

P P
(1) calculate Cl = Tj(l) − rBl , l = 1, 2, . . . , rk, where Tj(l) is the sum of
the treatment totals which are in block l.
0 P
(2) calculate the total of C s, RCi = Cl(i) , in the ith replication (e.g. i = 1, 2
for a simple lattice).

4. Calculate the adjusted block sum of squares, SSBAdj , given by

Prk 2
Pallrep 2
l=1 Cl i=1 (RCi )
SSBAdj = − . (12)
rk(r − 1) rk 2 (r − 1)

5. Calculate the intra-block error sum of squares given by

SSE = SST − SSR − SST rtU − SSBAdj (13)

From these sums of squares, the analysis of variance table is shown in Table 4.
From Table 4, the mean square of the unadjusted treatment, however, cannot be
 12

Table 4: Analysis of Variance for the partially balanced square lattice design

Source of variation Degree of freedom Sum of squares Mean squares

Replication r−1 SSR M SR
Treatment(Unadjusted) k2 − 1 SST rtU M ST rtU
Block within replication(Adj) r(k − 1) SSBAdj = Ba M SBAdj = Eb
Intra-block Error (rk − k − 1)(k − 1) SSE M SE = Ee
Total rk 2 − 1 SST

used for testing against the mean square of intra-block error in an F test. Instead,
the unadjusted treatment mean square can be used for testing the data as if the
experiment were in randomized block. The denominator for the F-test is the pooled
mean square for adjusted blocks and intra-block error. For lattice designs, if we want a
test for the treatment effects, the treatment totals must be adjusted. The adjustment
factor, similar to the weight in a balanced lattice, is
M SBAdj − M SE Eb − Ee
µ= = . (14)
k(r − 1)M SBAdj k(r − 1)Eb
0
Then, the adjusted treatment total, Tj , is defined as
r
0
X
Tj = Tj + µ Ci (in Block l containing Trt j). (15)
i=1

0
Then, for a simple lattice, Tj = Tj + µ[C1∗ + C2∗ ], where C1∗ and C2∗ are the values
of C’s in the particular blocks containing the treatment jth in the replication 1,
and 2, respectively. The procedure to obtain the adjusted treatment sum of squares
(Cochran and Cox, 1950) in a simple lattice is the following:

1. Calculate the unadjusted sum of squares for blocks within replication,

P 2 P 2
Bl Bl
SSB(W ithin Rep)U nadj = ( − CF1 ) + ( − CF2 ) = Bu (16)
| k {z } | k {z }
replication 1 replication 2

where CFi is the correction factor in ith replication.

2. Let Ba be the adjusted sum of squares for blocks within replication, or SSBAdj
(see Table 4). The adjusted treatment sum of squares, SST rtAdj is calculated
 13

by
r
SST rtAdj = SST rtU nadj − k(r − 1)µ[ Bu − Ba ] (17)
(r − 1)(1 + kµ)
Like the adjusted mean square error in balanced designs, the “effective error
0
mean square”, or Ee , is defined as

0 rkµ
Ee = Ee [1 + ]. (18)
k+1
The effective error mean square is used for estimating the gain in accuracy over
randomized blocks. However, the F-statistic for testing the treatment effect
uses the intra block mean square as the denominator.
M ST rtAdj
F = (19)
Ee

with the degrees of freedom (k 2 − 1) and (rk − k − 1)(k + 1).

The formula for the error variance of the difference between two treatment means
in the same block is the same as those in for a balanced lattice design which is
2Ee [1+(r −1)µ]/r. Because of the partial balance, the error variance of the difference
between two treatment means in the same block is slightly smaller than the error
variance of the difference between two treatment means in different blocks which is
2Ee
[1 + rµ]. (20)
r
Except for small designs, we can use the average variance given by
 
2Ee rµk
1+ . (21)
r k+1

To estimate the gain in accuracy over the randomized block designs, the effective
error mean square is used to compare with the error mean square in randomized
block analysis.
 14

3 NUMERICAL EXAMPLES
3.1 Example of a balanced square lattice design
In the following example, the original data appeared in a paper by Comstock et al
(1948). They studied the effects of nine feeding treatments on the growth rates of
pigs. The experimental unit in this study was a pair of pigs which were fed in the
same pen. The 3 × 3 balanced lattice design is used, so six pigs were required for each
block. From past experience, the litter has an effect on the variance in the growth
rate; hence, two sets of three uniform littermates were assigned to each block. Within
a block, each treatment was applied to only one member of each set. The response
variable is the sum of gains in weight for 2 pigs. The data is given in Table 5. In this
example, we have r = k + 1 = 4 replications, block size of k = 3, k 2 =32 treatments,
and 32 − 1 = 8 blocks.

Table 5: Gain in weight for 2 pigs

Block Replication I Total Block Replication II Total

1 (1) (2) (3) 4 (3) (4) (8)
2.20 1.84 2.18 6.22 1.71 1.57 1.13 4.41
2 (4) (5) (6) 5 (2) (6) (7)
2.05 0.85 1.86 4.76 1.76 2.16 1.80 5.72
3 (7) (8) (9) 6 (1) (5) (9)
0.73 1.60 1.76 4.09 1.81 1.16 1.10 4.08
Block Replication III Total Block Replication IV Total
7 (1) (4) (7) 10 (3) (5) (7)
1.19 1.20 1.15 3.54 2.04 0.93 1.78 4.75
8 (2) (5) (8) 11 (2) (4) (9)
2.26 1.07 1.45 4.78 1.50 1.60 1.42 4.52
9 (3) (6) (9) 12 (1) (6) (8)
2.12 2.03 1.63 5.78 1.77 1.57 1.43 4.77

The total sum of squares, the unadjusted treatment sum of squares, the unadjusted
block sum of squares, and the replication sum of squares are computed below:
from equation (1), SST = (2.202 + 1.842 + . . . + 1.432 ) − CF = 97.55 − 91.58 = 5.97;
 15

from equation (2),

Tj2
P
6.972 + 7.362 + . . . + 5.922
SST rtU = − CF = − 91.58
r 4
379.24
= − 91.58 = 3.23;
4
from (3),
15.072 + 14.212 + 14.102 + 14.042
SSR = − 91.58 = 91.66 − 91.58 = 0.08
32
Then, to find the adjusted sum of squares for blocks, (1) the sum of block totals for
block with the jth treatment, Bj , (2) the total of treatment j for all replications, Tj ,
and (3) the weights for treatment j, Wj , are required. For example, for treatment 7,
T7 = 0.73 + 1.80 + 1.15 + 1.78 = 5.46,
B7 = 4.09 + 5.72 + 3.54 + 4.75 = 18.10,
and from equation (4), W7 = 3(5.46) − 4(18.10) + G = 1.40, where G is the grand
total. The values of Bj , Tj , and Wj are summarized in Table 6. From Table 6, we

Table 6: Treatment Totals and Adjustment Factors

Trt j Tj Bj Wj = kTj − (k + 1)Bj + G

1 6.97 18.61 3.89
2 7.36 21.24 -5.46
3 8.05 21.16 -3.07
4 6.42 17.23 7.76
5 4.01 18.37 -4.03
6 7.62 21.03 -3.84
7 5.46 18.10 1.40
8 5.61 18.05 2.05
9 5.92 18.47 1.30

can find SSBAdj from equation (5):

P 2
Wj 3.892 + (−5.46)2 + . . . + 1.302 153.43
SSBAdj = 3 = 3
= = 1.42;
k (k + 1) 3 (3 + 1) 108
the intra-block error sum of squares,
SSE = SST − SSR − SST rtU − SSBAdj = 5.97 − 0.08 − 3.23 − 1.42 = 1.24. The
complete ANOVA table is summarized in Table 7.
 16

Table 7: Analysis of Variance for the balanced square lattice design example

Source of variation Degree of freedom Sum of squares Mean squares

Replication 4−1=3 0.08 0.026
Treatment(Unadjusted) 32 − 1 = 8 3.23 0.404
2
Block within replication(Adj) 3 −1=8 1.42 0.178 = Eb
Intra-block Error (32 − 1)(3 − 1) = 16 1.24 0.077 = Ee
Total 4(32 ) − 1 = 35 5.97

To test for a treatment effect, the adjustment factor for treatment means is calculated
as follows:
1.42
Eb = M SBAdj = = 0.178
32 − 1

1.24
Ee = M SE = = 0.078.
(32 − 1)(3 − 1)
From equation (8), we can calculate the adjustment factor,
Eb − Ee 0.178 − 0.078
µ= 2
= = 0.062.
k Eb 32 (0.178)
0
So, using the adjustment factor to calculate the adjusted treatment total, Tj =
Tj + µWj , j = 1, 2, . . . , 9; all adjusted treatment totals are summarized in Table
8.

Table 8: The adjusted treatment totals

Adjusted treatment totals

0 0 0 0 0 0 0 0 0
T1 T2 T3 T4 T5 T6 T7 T8 T9
7.21 7.01 7.85 6.90 3.76 7.38 5.55 5.74 6.00

Thus, the adjusted treatment sum of squares (see equation (7)) is

P 0 2
(Tj ) 7.212 + 7.012 + . . . + 6.002
SST rtAdj = − CF = − 91.58
r 4
379.02
= − 91.58 = 3.16,
4
 17

0
and the effective error mean squares, equation (9), is Ee = Ee (1 + kµ) = Ee [1 +
(r − 1)µ] = 0.078[1 + 3(0.062)] = 0.092. The adjusted treatment mean square which
equals to SST rtAdj /(k 2 − 1) = 3.16/8 = 0.395. Therefore, the F-ratio equals to
0.395/0.092 = 4.30 having 8 and 16 degrees of freedom. Normally, the intra-error
mean square is the denominator of F-statistic, but in this example the effective mean
square is used instead to take account of sampling errors. The standard error of the
p p
adjusted treatment mean is Ee0 /r = 0.092/4 = 0.151; and the standard error
of the difference between 2 adjusted means in the same block (see equation (11)) is
p 0 p
2Ee /r = (2 × 0.092)/4 = 0.215.
The randomized complete block design error sum of squares is SSBAdj + SSE =
1.42 + 1.24 = 2.66 with 24 degrees of freedom. So, the unbiased estimate of error
variance for a randomized block design is 2.66/24, or 0.1108, and the efficiency of the
experiment relative to a randomized complete block design is 0.1108/0.092, or 120%.

3.2 Example of a partially balanced square lattice design

This example is from Cochran and Cox (1967). The data are yields in bushels per acre
of 25 varieties of soybeans. The data are collected in two replications of 25 varieties
in five blocks. Each block contains 5 varieties. So, this is a simple lattice design. The
data is given in Table 9. In this example, we have r = 2 replications, block size of
k = 5, k 2 =52 = 25 treatments, and rk = 2(5) = 10 blocks.
From the data in Table 9, the total sum of squares, the unadjusted treatment sum of
squares, the unadjusted block sum of squares, and the replication sum of squares are
calculated from equation(1),
SST = (62 + 72 + . . . + 142 ) − CF = 10, 767 − 9, 275.22 = 1, 491.78
and from equation (2),
Tj2
P
302 + 282 + . . . + 332
SST rtU = − CF = − 9, 275.22
r 2

= 9, 834.50 − 9, 275.22 = 559.28

The jth treatment totals are summarized in the Table 10. For example, T1 = 6 + 24 =
30.
From equation (3),
2892 + 3922
SSR = − 9, 275.22 = 9, 487.4 − 9, 275.22 = 212.8
52
, where CF is the correction factor, 6812 /50 = 9, 275.22. To calculate the adjusted
block sum of squares (equation (12)) we need Cl and RCi . For example, for block 2
 18

Table 9: Data of yields in bushels per acre of 25 varieties of soybeans

Bk Replication I TotalBl Bk Replication II TotalBl

1 (1) (2) (3) (4) (5) 6 (1) (6) (11) (16) (21)
6 7 5 8 6 32 24 13 24 11 8 80
2 (6) (7) (8) (9) (10) 7 (2) (7) (12) (17) (22)
16 12 12 13 8 61 21 11 14 11 23 80
3 (11) (12) (13) (14) (15) 8 (3) (8) (13) (18) (23)
17 7 7 9 14 54 16 4 12 12 12 56
4 (16) (17) (18) (19) (20) 9 (4) (9) (14) (19) (24)
18 16 13 13 14 74 17 10 30 9 23 89
5 (21) (22) (23) (24) (25) 10 (5) (10) (15) (20) (25)
14 15 11 14 14 68 15 15 22 16 19 87
Total 289 Total 392

Table 10: Treatment totals

T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11 T12 T13

30 28 21 25 21 29 23 16 23 23 41 21 19
T14 T15 T16 T17 T18 T19 T20 T21 T22 T23 T24 T25
39 36 29 27 25 22 30 22 38 23 37 33

P
(which is in replication 1) C2 = Tj(2) − rB2 = (T6 + T7 + T8 + T9 + T10 ) −2(61) =
| {z }
see T able 10
(29 + 23 + 16 + 23 + 23) − 2(61) = −8. Other Cl ’s, l = 1, 2, . . . , 10 are computed in
the same way. The complete set of values of Cl ’s are in Table 11.

Then, calculate the total of Cl s in each replication:

for replication 1, RC1 = C1 + C2 + C3 + C4 + C5 = 61 + (−8) + 48 + (−15) + 17 = 103;
for replication 2, RC2 = C6 + C7 + C8 + C9 + C10 = (−9) + (−23) + (−8) + (−32) +
(−31) = −103.
 19

Table 11: Value of Cl for each block

C1 C2 C3 C4 C5 C6 C7 C8 C9 C10
61 -8 48 -15 17 -9 -23 -8 -32 -31

From the equation (12), the adjusted block sum of squares is calculated:
Prk 2 Pallrep
l=1 Cl (RCi )2
SSBAdj = − i=12
rk(r − 1) rk (r − 1)

612 + (−8)2 + . . . + (−31)2 1032 + (−103)2

= − = 501.84.
2(5)(2 − 1) 2(52 )(2 − 1)
The intra-block error sum of square (equation (13)) is

SSE = SST −SSR−SST rtU −SSBAdj = 1, 491.78−212.18−559.28−501.84 = 218.48.

The complete ANOVA table is summarized in Table 12.

Table 12: Analysis of variance table for example of yield data

Source of variation Degree of freedom Sum of squares Mean squares

Replication 2−1=1 212.18 212.18
Treatment(Unadjusted) 52 − 1 = 24 559.28 23.30
Block within replication(Adj) 2(5 − 1) = 8 501.84 = Ba 62.73 = Eb
Intra-block Error 16 (Substraction) 218.48 13.66 = Ee
2
Total 2(5 ) − 1 = 49 1, 491.78

To get the adjusted treatment total, the adjustment factor, equation (14), is
calculated:
Eb − Ee 62.73 − 13.66
µ= = = 0.1564.
k(r − 1)Eb 5(2 − 1)62.73
After getting the the adjustment factor, this quantity is used for calculating the
adjusted treatment total (see equation (15)). For example, the adjusted treatment
total 2, treatment 2 are in the blocks 1 and 7 in replications 1 and 2 , respectively.
So,
0
T2 = T2 + µ[ C1∗ + C2∗ ] = 28
|{z} +0.1564[61 + (−23)] = 33.945.
|{z} |{z} | {z }
in block 1 in block 7 f rom T able 10 see T able 11
 20

Another example for treatment 23 which are in the blocks 5 and 8 in replications 1
and 2 , respectively. So,
0
T23 = T23 + µ[ C1∗ + C2∗ ] = 23
|{z} +0.1564[17 + (−8)] = 24.41.
|{z} |{z} | {z }
in block 5 in block 8 f rom T able 10 see T able 11

0
The complete set of values of Tj ’s are in Table 13.

Table 13: Adjusted treatment totals

0 0 0 0 0
T1 T2 T3 T4 T5
38.14 33.95 29.29 29.54 25.70
0 0 0 0 0
T6 T7 T8 T9 T10
26.34 18.15 13.50 16.74 16.90
0 0 0 0 0
T11 T12 T13 T14 T15
47.10 24.91 25.26 41.50 38.66
0 0 0 0 0
T16 T17 T18 T19 T20
25.24 21.05 21.40 14.65 22.80
0 0 0 0 0
T21 T22 T23 T24 T25
23.25 37.06 24.41 34.65 30.81

To test the treatment effect, we need the adjusted treatment sum of squares, equation
(17). Equations (16) and (17) would be used to obtain SST rtAdj . First, calculate the
unadjusted sum of squares for blocks within replication:

Bl2
P P 2
Bl
Bu = ( − CF1 ) + ( − CF2 )
| k {z } | k {z }
replication 1 replication 2

322 + 612 + 542 + 742 + 682 2892 802 + 802 + 562 + 892 + 872 3922
=[ − ]+[ − ]
5 25 5 25

= 211.36 + 138.64 = 350.

 21

Next, calculate the adjusted treatment sum of squares (equation (17)) in the following:
r
SST rtAdj = SST rtU nAdj − k(r − 1)µ[ Bu − Ba ]
(r − 1)(1 + kµ)

2
= 559.28 − 5(2 − 1)(0.1564)[ (350) − 501.84]
(2 − 1)(1 + 5(0.1564))

= 559.28 − (−85.26) = 644.54.

0
The effective error mean square, Ee , is

0 rkµ 2(5)(0.1564)
Ee = Ee [1 + ] = 13.66[1 + ] = 17.2144.
k+1 5+1
Therefore, the F-ratio equals to M ST rtAdj /Ee = (644.54/24)/13.66 = 1.966
having 24 and 8 degrees of freedom. The standard error of the adjusted
p 0 p
treatment mean is Ee /r = 17.2144/2 = 2.933; and the standard error of the
p
difference between 2 adjusted means in the same block is 2Ee [1 + (r − 1)µ]/r =
p √
2(13.66)[1 + (2 − 1)(0.1564)]/2 = 15.796 = 3.975; the standard error of the
difference between two adjusted treatment means in different blocks (equation (20))
is r
2Ee 2(13.66)
(1 + rµ) = (1 + 2(0.1564)) = 4.235.
r 2
The average of variance, equation (21), is

2Ee rµk 2(13.66) 2(0.1564)(5)

[1 + ]= [1 + ] = 17.22.
r k+1 2 5+1
The randomized complete block design error sum of squares is SSBAdj + SSE =
501.84 + 218.48 = 720.32 with 24 degrees of freedom. While the mean square of
error variance for randomized block design is 720.32/24 = 30.01, the effective error
mean square in this design is 17.22. The efficiency of the experiment relative to a
randomized complete block design is 30.01/17.22, or 174.27%.
 22

4 R AND SAS PROGRAMS

4.1 R and SAS Programs for a balanced lattice design
4.1.1 R Program

In R Package “agricolae”, although there is no function that can directly analyze data
obtained from a lattice design, there are 2 useful functions for use with lattice designs.
The function “design.lattic” can randomize treatments into a k×k lattice, but only for
a partially balanced lattice design (simple and triple lattice designs). This function,
however, does not generate the analysis of variance. The function “PBIB.test” which
is aimed at the analysis of the partially balanced incomplete block design (PBIB) can
be applied to resoluble designs (lattice and alpha designs). For the previous example
of the study of the effects of nine feeding treatments on the growth rates of pigs by
Comstock et al (1948), the data in Table 5 must be structured in an appropriate form
for use in R. Variables referring to replications, blocks, treatments, response values
must be created. Sample R code is given below:

First, we have to load packages “agricolae” and MASS”. Package “MASS” is also
required because package “agricolae” needs a function “ginv” from the MASS package.

library(agricolae)
library(MASS)

Second, create column vectors for the replication, the block, treatment, and response
which are the argument of the function “PBIB.test”. In Table 5, we have 4
replications, 12 blocks, and 9 treatments.

# Creates a replication vector

rep<-rep(1:4,each=9)
# Creates a block vector
block<-rep(1:12,each=3)
# Creat a treatment vector
trt<-c(1,2,3,4,5,6,7,8,9,
3,4,8,2,6,7,1,5,9,
1,4,7,2,5,8,3,6,9,
3,5,7,2,4,9,1,6,8)
# Creat a response vector
gain.wt
<-c(2.20,1.84,2.18,2.05,0.85,1.86,0.73,1.60,1.76,
1.71,1.57,1.13,1.76,2.16,1.80,1.81,1.16,1.11,
 23

1.19,1.20,1.15,2.26,1.07,1.45,2.12,2.03,1.63,
2.04,0.93,1.78,1.50,1.60,1.42,1.77,1.57,1.43)
Now, specify the value of k , the block size. For this example, k = 3.
model<- PBIB.test(block,trt,rep,gain.wt,k=3)
After running the above program and display “data”, the following output is
generated:
ANALYSIS PBIB: gain.wt

Class level information

Blocks: 12
Trts : 9

Number of observations: 36
Analysis of Variance Table
Response: gain.wt
Df Sum Sq Mean Sq F value Pr(>F)
replication 3 0.0774 0.02580 0.3337 0.801132
trt.unadj 8 3.2261 0.40326 5.2168 0.002467 **
replication:block.adj 8 1.4206 0.17758 2.2972 0.074630 .
Residuals 16 1.2368 0.07730
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

coefficient of variation: 17.4 %

gain.wt Means: 1.595

Treatments

Parameters PBIB
treatmeans : 9
Block size : 3
Blocks/rep : 3
Replication: 4

Efficiency factor 0.75

The mean sums of squares corresponds to those in Table 7 except for the F tests. As
we saw in the previous calculations, to test the treatment effect, both the treatment
sum of squares and the error mean square have to be adjusted. The actual F value
is 4.30 instead of 5.21. The function “PBIB.test” also gives the adjusted treatment
means which are presented below (and correspond to adjusted treatment totals in
Table 8).
 24

model$means
<<< to see the objects: comparison and means >>>
trt means mean.adj N std.err
1 1.7425 1.8035178 4 0.1554838
2 1.8400 1.7543554 4 0.1554838
3 2.0125 1.9643445 4 0.1554838
4 1.6050 1.7267220 4 0.1554838
5 1.0025 0.9392861 4 0.1554838
6 1.9050 1.8447665 4 0.1554838
7 1.3650 1.3869601 4 0.1554838
8 1.4025 1.4346559 4 0.1554838
9 1.4800 1.5003916 4 0.1554838
p p
The standard error above does not exactly corresponds to the Ee0 /r = 0.092/4 =
0.151. We can ask “PBIB.test” to show the difference between 2 adjusted mean in
the same block. This follows from equation (11) as well as the previous example.
model$comparison
Comparison between treatments means
<<< to see the objects: comparison and means >>>

Difference stderr pvalue

1 - 2 0.04916241 0.2143020 0.821460
3 - 1 0.16082670 0.2143020 0.463868
1 - 4 0.07679588 0.2143020 0.724764
1 - 5 0.86423169 0.2143020 0.000964
6 - 1 0.04124861 0.2143020 0.849788
1 - 7 0.41655769 0.2143020 0.069722
1 - 8 0.36886191 0.2143020 0.104484
1 - 9 0.30312627 0.2143020 0.176386
3 - 2 0.20998911 0.2143020 0.341734
2 - 4 0.02763347 0.2143020 0.899008
2 - 5 0.81506928 0.2143020 0.001562
6 - 2 0.09041103 0.2143020 0.678722
2 - 7 0.36739528 0.2143020 0.105760
2 - 8 0.31969950 0.2143020 0.155200
2 - 9 0.25396386 0.2143020 0.253292
3 - 4 0.23762257 0.2143020 0.283896
3 - 5 1.02505839 0.2143020 0.000204
3 - 6 0.11957808 0.2143020 0.584576
3 - 7 0.57738439 0.2143020 0.015958
3 - 8 0.52968860 0.2143020 0.025060
3 - 9 0.46395297 0.2143020 0.045852
4 - 5 0.78743581 0.2143020 0.002050
6 - 4 0.11804449 0.2143020 0.589358
 25

4 - 7 0.33976181 0.2143020 0.132432

4 - 8 0.29206603 0.2143020 0.191798
4 - 9 0.22633039 0.2143020 0.306596
6 - 5 0.90548031 0.2143020 0.000644
7 - 5 0.44767400 0.2143020 0.053044
8 - 5 0.49536978 0.2143020 0.034454
9 - 5 0.56110542 0.2143020 0.018634
6 - 7 0.45780631 0.2143020 0.048454
6 - 8 0.41011052 0.2143020 0.073720
6 - 9 0.34437489 0.2143020 0.127616
8 - 7 0.04769578 0.2143020 0.826692
9 - 7 0.11343142 0.2143020 0.603860
9 - 8 0.06573564 0.2143020 0.762994
Done.

The P-values are based on a t statistic. For example, for comparing treatments 1 and
2, t = 0.049/0.214 = 0.229 with an associated P-value of 2P (t16 > 0.229) = 0.821.

4.1.2 SAS Program

The LATTICE procedure in SAS can analyze data from balanced square lattices,
partially balanced square lattices, and some rectangular lattices. The LATTICE
procedure determines the type of lattice design from the data set. It also checks
whether the data is valid and gives the message if it is not working. The data
that we will create must consist of variables named “Group”, “Block”, “Treatmnt”,
and “Rep”. The variable Group indicates which orthogonal replication in the basic
plan (balanced design) includes the experimental unit. The values of ”Group” are
1, 2, . . . , r where r is the number of replicates for a balanced design. The variable
“Rep” is needed when there are more than 1 repetition of the entire basic plan. The
values of “Rep” are 1, 2, . . . , p, where p is the number of replications of the entire basic
plan. Hence, the experiment has a total of r × p replications. Then, for a balanced
square lattice design, there will be a variable “Rep”. More details are in SAS/STAT
9.2 (2008). A sample program for the pig data example is given below.
First, create a dataset with variables Group, Treatmnt, Rep, and response Weight.
data pigs;
input Group Block Treatmnt Weight @@;
cards
1 1 1 2.20 1 1 2 1.84 1 1 3 2.18 1 2 4 2.05
1 2 5 0.85 1 2 6 1.86 1 3 7 0.73 1 3 8 1.60
1 3 9 1.76 2 1 1 1.19 2 1 4 1.20 2 1 7 1.15
2 2 2 2.26 2 2 5 1.07 2 2 8 1.45 2 3 3 2.12
 26

2 3 6 2.03 2 3 9 1.63 3 1 1 1.81 3 1 5 1.16

3 1 9 1.11 3 2 2 1.76 3 2 6 2.16 3 2 7 1.80
3 3 3 1.71 3 3 4 1.57 3 3 8 1.13 4 1 1 1.77
4 1 6 1.57 4 1 8 1.43 4 2 2 1.50 4 2 4 1.60
4 2 9 1.42 4 3 3 2.04 4 3 5 0.93 4 3 7 1.78
;

Second, use PROC LATTICE to generate the analysis of variance.

proc lattice data = pigs;
var Weight;
run;

The results are the following:

The Lattice Procedure
Analysis of Variance for Weight

Sum of Mean
Source DF Squares Square

Replications 3 0.07739 0.02580

Blocks within Replications (Adj.) 8 1.4206 0.1776
Component B 8 1.4206 0.1776
Treatments (Unadj.) 8 3.2261 0.4033
Intra Block Error 16 1.2368 0.07730
Randomized Complete Block Error 24 2.6574 0.1107
Total 35 5.9609 0.1703

Additional Statistics for Weight

Variance of Means in Same Block 0.04593

LSD at .01 Level 0.6259
LSD at .05 Level 0.4543
Efficiency Relative to RCBD 120.55

Adjusted Treatment
Means for Weight

Treatment Mean

1 1.8035
2 1.7544
3 1.9643
4 1.7267
5 0.9393
 27

6 1.8448
7 1.3870
8 1.4347
9 1.5004

The mean sums of squares are the same as those from function PBIB.test in R except
that PROC LATTICE did not perform the F tests. PROC LATTICE also gives the
randomized complete block error, the pooled sums of squares for blocks and intra-
block error. The relative efficiency of the lattice design was 120.5. However, both
SAS and R did not give the adjusted treatment sum of squares. Therefore, we cannot
obtain the test of treatment effect from both R and SAS.

4.2 R and SAS Programs for a partially balanced lattice

design
4.2.1 R Program

We use the same function as we did for a balanced lattice design. Column vectors
are created for the replication, the block, treatment, and response which are the
arguments of the function “PBIB.test”. The example in Table 9 has 2 replications,
10 blocks, and 25 treatments.

library(agricolae)
library(MASS)
rep<-rep(1:2,each=25)
block<-rep(1:10,each=5)
trt<-c(1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,
21,22,23,24,25,1,6,11,16,21,2,7,12,17,22,3,8,13,18,23,
4,9,14,19,24,5,10,15,20,25)
yield<-c(6,7,5,8,6,16,12,12,13,8,17,7,7,9,14,18,16,13,13,14,14,15,
11,14,14,24,13,24,11,8,21,11,14,11,23,16,4,12,12,12,17,10,
30,9,23,15,15,22,16,19)
PBIB.test(block,trt,rep,yield,k=5)

The results from R are the following:

ANALYSIS PBIB: yield

Class level information
Blocks: 10
Trts : 25
Number of observations: 50

Analysis of Variance Table

 28

Response: yield
Df Sum Sq Mean Sq F value Pr(>F)
replication 1 212.18 212.180 15.5386 0.001166 **
trt.unadj 24 559.28 23.303 1.7066 0.135789
replication:block.adj 8 501.84 62.730 4.5939 0.004629 **
Residuals 16 218.48 13.655
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

coefficient of variation: 27.1 %

yield Means: 13.62

Treatments

Parameters PBIB
treatmeans : 25
Block size : 5
Blocks/rep : 5
Replication: 2
Efficiency factor 0.75

All mean squares are the same as those in Table 12 except for the F statistics.
Again, the estimates of adjusted treatment means which are obtained by dividing
the adjusted treatment totals (in Table 13) by 2 (the number of replications) are
p
given below. The standard error below is not exactly the same as Ee0 /r = 2.933.
PBIB.test(block,trt,rep,yield,k=5)$means
Comparison between treatments means

<<< to see the objects: comparison and means >>>

trt means mean.adj N std.err
1 15.0 19.068070 2 2.993997
2 11.5 16.972820 2 2.993997
3 20.5 14.646302 2 2.993997
4 10.5 14.768731 2 2.993997
5 9.5 12.846963 2 2.993997
6 19.5 13.170054 2 2.993997
7 18.0 9.074805 2 2.993997
8 14.5 6.748286 2 2.993997
9 13.5 8.370716 2 2.993997
10 12.5 8.448948 2 2.993997
11 11.0 23.551052 2 2.993997
12 14.0 12.455803 2 2.993997
13 15.0 12.629284 2 2.993997
 29

14 11.0 20.751714 2 2.993997

15 19.0 19.329946 2 2.993997
16 11.5 12.622429 2 2.993997
17 18.5 10.527180 2 2.993997
18 16.5 10.700662 2 2.993997
19 10.5 7.323091 2 2.993997
20 12.5 11.401323 2 2.993997
21 10.5 11.625857 2 2.993997
22 14.5 18.530607 2 2.993997
23 11.5 12.204089 2 2.993997
24 8.0 17.326518 2 2.993997
25 11.5 15.404751 2 2.993997

Again, we can generate the estimates of differences in treatment effects from

PBIB.test. There are 2 standard errors (see below). For example, treatments 1
and 2 are in the same block (see Table 9), and formula (11) is applied. Treatments 1
and 7 are in different blocks, and formula (20) is applied.

PBIB.test(block,trt,rep,yield,k=5)$comparison[1:24,]

Difference stderr pvalue

1 - 2 2.095249 3.973854 0.605248
1 - 3 4.421768 3.973854 0.282270
1 - 4 4.299338 3.973854 0.295332
1 - 5 6.221106 3.973854 0.137026
1 - 6 5.898015 3.973854 0.157188
1 - 7 9.993265 4.234151 0.031300
1 - 8 12.319783 4.234151 0.010234
1 - 9 10.697354 4.234151 0.022446
1 - 10 10.6191216 4.234151 0.023298
11 - 1 4.4829826 3.973854 0.275902
1 - 12 6.6122669 4.234151 0.137930
1 - 13 6.4387853 4.234151 0.147856
14 - 1 1.6836442 4.234151 0.696154
15 - 1 0.2618763 4.234151 0.951450
1 - 16 6.4456400 3.973854 0.124338
1 - 17 8.5408895 4.234151 0.060780
1 - 18 8.3674079 4.234151 0.065638
1 - 19 11.7449785 4.234151 0.013552
1 - 20 7.6667464 4.234151 0.089006
1 - 21 7.4422127 3.973854 0.079486
1 - 22 0.5374621 4.234151 0.900574
1 - 23 6.8639806 4.234151 0.124534
1 - 24 1.7415511 4.234151 0.686302
 30

4.2.2 SAS Program

The following SAS code is similar to the code used for a balanced design:

data Soy;
do Group = 1 to 2;
do Block = 1 to 5;
do Plot = 1 to 5;
input Treatmnt Yield @@;output;
end;
end;
end;
cards;
1 6 2 7 3 5 4 8 5 6
6 16 7 12 8 12 9 13 10 8
11 17 12 7 13 7 14 9 15 14
16 18 17 16 18 13 19 13 20 14
21 14 22 15 23 11 24 14 25 14
1 24 6 13 11 24 16 11 21 8
2 21 7 11 12 14 17 11 22 23
3 16 8 4 13 12 18 12 23 12
4 17 9 10 14 30 19 9 24 23
5 15 10 15 15 22 20 16 25 19
;
proc lattice data=Soy;
run;

The following SAS output was generated by the code above:

The Lattice Procedure

Analysis of Variance for Yield

Sum of Mean
Source DF Squares Square

Replications 1 212.18 212.18

Blocks within Replications (Adj.) 8 501.84 62.7300
Component B 8 501.84 62.7300
Treatments (Unadj.) 24 559.28 23.3033
Intra Block Error 16 218.48 13.6550
Randomized Complete Block Error 24 720.32 30.0133
Total 49 1491.78 30.4445

Additional Statistics for Yield

 31

Variance of Means in Same Block 15.7915

Variance of Means in Different Bloc 17.9280
Average of Variance 17.2159
LSD at .01 Level 12.1189
LSD at .05 Level 8.7959
Efficiency Relative to RCBD 174.34

The mean sums of squares are the same as those in Table 12 as well as the results
from R. PROC LATTICE provides the variances of the mean differences in the same
block and in different blocks. Their square roots are equal to standard errors in R
output. Unlike the PBIB.test in R, PROC LATTICE also gives the average variance
corresponding to formula (21). The following results are the estimates of adjusted
treatment means.

Adjusted Treatment
Means for Yield

Treatment Mean

1 19.0681
2 16.9728
3 14.6463
4 14.7687
5 12.8470
6 13.1701
7 9.0748
8 6.7483
9 8.3707
10 8.4489
11 23.5511
12 12.4558
13 12.6293
14 20.7517
15 19.3299
16 12.6224
17 10.5272
18 10.7007
19 7.3231
20 11.4013
21 11.6259
22 18.5306
23 12.2041
24 17.3265
25 15.4048
 32

5 R PROGRAMS FOR SQUARE LATTICE

DESIGNS
The function PBIB.test in library agricolae did not give the right F-test because the
mean square of unadjusted treatment still contains block effects. PROC LATTICE
in SAS did not give the F-test for the treatment effect. Therefore, in this paper the R
function, named “lattice.design” is written for balanced and partially balanced square
lattice designs. The arguments for this function are exactly the same as those for the
function PBIB.test. The R code for the function “lattice.balance” is in appendix.

5.1 A Balanced lattice design

Using the same data in section 3.1 (gain in weight in pigs), the result generated by
the lattice.balance function is below.

> lattice.design(block,trt,rep,y,3)

S.E.Adjusted Treatment Mean : 0.1515

S.E.Diff of Adj Treatment Mean : 0.2143
Effective error mean square : 0.0919
Efficiency Relative to RCBD : 120.5494

Analysis of Variance Table: Balanced

Df Sum Sq Mean Sq F value Pr(>F)

Replication 3 0.0774 0.02580 0.3337
Block.Adj 8 1.4206 0.17758 2.2972 0.074630 .
Treatment.Adj 8 3.1717 0.39646 5.1289 0.002689 **
Intra-block Error 16 1.2368 0.07730
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1
 33

5.2 A Partially balanced lattice design

Using the same data in section 3.2 (yield data), the result generated by the
lattice.balance function is below.

> lattice.design(block,trt,rep,y,5)

S.E.Adjusted Treatment Mean : 2.9339

S.E.Diff of Adj Trt Mean : 3.9739
S.E.Diff of Adj Trt Mean (not the same bk): 4.2342
The average variance : 17.2159
Efficiency Relative to RCBD : 174.3353

Analysis of Variance Table: Partially Balanced

Df Sum Sq Mean Sq F value Pr(>F)

Replication 1 212.18 212.180 15.5386
Block.Adj 8 501.84 62.730 4.5939 0.004629 **
Treatment.Adj 24 644.63 26.859 1.9670 0.082442 .
Intra-block Error 16 218.48 13.655
---
Signif. codes: 0 ’***’ 0.001 ’**’ 0.01 ’*’ 0.05 ’.’ 0.1 ’ ’ 1

The class of the lattice.balance is the list. In this object, the sums of squares, their
degrees of freedom, and the adjusted treatment means can be extracted. The function
“names” gives all objects in the “lattice.balance”.
 34

A R Code for lattice.design

lattice.design<- function(block,trt,rep,y,bksize.sclar)
{
if(max(rep)==(bksize.sclar+1)){
############## FIND Adj SSB ####################
k<-bksize.sclar
Total.InBj<- as.matrix(tapply(y,list(block),sum))
Tj<- as.matrix(tapply(y,list(trt),sum))
NumOfBlock<- max(dim(Total.InBj))
NumOfTrt<- max(dim(Tj))
BTotal.long<-rep(Total.InBj,each=k)
dat<-as.data.frame(cbind(BTotal.long,trt,block))
Bj<-matrix(0,ncol=1,nrow=NumOfTrt)

for(i in 1:NumOfTrt){
SumVec<-rep(0,(NumOfBlock*k))
dat<-as.data.frame(cbind(BTotal.long,trt,block,SumVec))
Bj[i]<-sum(ifelse(dat$trt==i,dat$SumVec<- dat$BTotal.long,dat$SumVec<- 0))
}
##############################################
FindW<-as.data.frame(cbind(Tj,Bj))
FindW$W<- k*FindW$V1-(k+1)*FindW$V2+sum(y)
FindW$WSquare<-(FindW$W)^2
SSB.Adj<-sum(FindW$WSquare)/((k^3)*(k+1))
#############################################

MSB.Adj<-SSB.Adj/(max(trt)-1)
SSRep<- anova(lm(y~factor(rep)))$"Sum Sq"[1]
MSRep<- SSRep/(max(rep)-1)
SSTrt.UnAdj<- anova(lm(y~factor(trt)))$"Sum Sq"[1]
MSTrt.UnAdj<- SSTrt.UnAdj/(max(trt)-1)
SSTotal<- sum(anova(lm(y~factor(trt)))$"Sum Sq")
SSE<- SSTotal-SSRep-SSTrt.UnAdj-SSB.Adj
MSE<- SSE/((max(trt)-1)*(sqrt(max(trt))-1))

mu<- (MSB.Adj-MSE)/(max(trt)*MSB.Adj)

CF<- (sum(y))^2/length(rep)
Adj.Total<-FindW$V1+mu*FindW$W
Adj.Trt.Mean<- Adj.Total/(k+1)

SSTrt.Adj<- sum(Adj.Total^2)/max(rep)-CF
MSTrt.Adj <- SSTrt.Adj/(max(trt)-1)

RBD.MSE<-(SSB.Adj+SSE)/((max(trt)-1)+(max(trt)-1)*(sqrt(max(trt))-1))
EffectiveMSE2<-MSE*(1+sqrt(max(trt))*mu)
EffectiveSSE2<-EffectiveMSE2*(k^2-1)*(k-1)
Efficiency<- (RBD.MSE/EffectiveMSE2)*100
F.Trt <- MSTrt.Adj/EffectiveMSE2
pval<-pf(F.Trt, df1=max(trt)-1, df2=(max(trt)-1)*(sqrt(max(trt))-1),lower.tail = FALSE)
se.Adj.trtMean<- sqrt(EffectiveMSE2/(sqrt(max(trt))+1))
se.Diff.same<- sqrt(2*EffectiveMSE2/(sqrt(max(trt))+1))
 35

############# bar plot ####################

#mp<-barplot(Adj.Trt.Mean, names.arg=seq(1:max(trt)),xlab="Treatments",
#ylab="Adjusted Means",col=terrain.colors(max(trt)),
#main="Balanced Square Lattice Design")
#tot <- round(colMeans(t(Adj.Trt.Mean)),2)
#text(mp, tot + 0.05, format(tot), xpd = TRUE, col = "black",cex=1)
#legend("bottomright",legend=paste(c("F-statistic=","p-value=","Efficiency="),
# c(format(F ,digits=4,justify = "right", scientific = TRUE)
# ,format(pval,digits=4,justify = "right", scientific = TRUE)
# ,format(Efficiency*100,,digits=4,justify = "right",
# scientific = TRUE)
# )
# ),bg="white",box.col="grey")

#error.bar <- function(x, y, upper, lower=upper, length=0.1,...){

#if(length(x) != length(y) | length(y) !=length(lower) | length(lower) != length(upper))
#stop("vectors must be same length")
#arrows(x,y+upper, x, y-lower, angle=90, code=3, length=length, ...)
#}
#error.bar(mp,Adj.Trt.Mean,rep(se.Adj.trtMean,length(Adj.Trt.Mean))

############## Make a Table Output (ANOVA)###############

ssrep<-SSRep
repdf<- k

SSBAdj<-SSB.Adj
SSBAdjrep<- max(trt)-1

SSTrt.Adj<-SSTrt.Adj
SSTrt.Adjdf<- max(trt)-1

SSE.Effective<-EffectiveSSE2
Errordf<-(max(trt)-1)*(k-1)

Df <- c(repdf, SSBAdjrep, SSTrt.Adjdf,Errordf,Errordf)

ssq <- c(ssrep, SSBAdj,SSTrt.Adj,SSE, SSE.Effective)

frep<-MSRep/MSE
fb<-MSB.Adj/MSE
pval.B<-pf(fb, df1=max(trt)-1, df2=(max(trt)-1)*(sqrt(max(trt))-1),lower.tail = FALSE)

ftrt<-MSTrt.Adj/EffectiveMSE2
fsse<-1

Fval<-c(frep,fb,ftrt,NA,NA)
Means.Adj<-Adj.Trt.Mean

Pvalue<-c(NA,pval.B,pval,NA,NA)

anovadf <- data.frame(Df, ‘Sum Sq‘=ssq, ‘Mean Sq‘=ssq/Df,

 36

‘F value‘=Fval ,‘Pr(>F)‘ = Pvalue, check.names=FALSE)

rownames(anovadf) <- c("Replication","Block.Adj[a]","Treatment.Adj[b]","Error[a]",
"Effective.Error[b]")
class(anovadf) <- c("anova","data.frame")
print(list(Adj.means=Means.Adj))
cat(c(" \n"))
cat(c(" \n"))
cat(c("S.E.Adjusted Treatment Mean : "),round(se.Adj.trtMean,4),"\n")
cat(c("Efficiency Relative to RCBD : "),round(Efficiency,4),"\n")
cat(c("S.E.Diff of Adj Treatment Mean : "),round(se.Diff.same,4),"\n")
cat(c(" \n"))
cat(c("Analysis of Variance Table: Balanced\n"))
cat(c(" \n"))
return(anovadf)
}
else{
k<-bksize.sclar

tmp<-as.data.frame(cbind(rep,block,trt,y))
class(tmp$block)

SSTrt.UnAdj<-anova(lm(y~factor(trt),data=tmp))$"Sum Sq"[1]
SSRep<-anova(lm(y~factor(rep),data=tmp))$"Sum Sq"[1]
SSTot<-anova(lm(y~factor(rep),data=tmp))$"SumSq"[1]+anova(lm(y~factor(rep),
data=tmp))$"Sum Sq"[2]

Total.InBj<- as.matrix(tapply(y,list(block),sum))
Tj<- as.matrix(tapply(y,list(trt),sum))

SumTj.L<-matrix(0,ncol=1,nrow=max(block))
for(i in 1:max(block)){
SumTj.L[i,1] <- sum(Tj[tmp[block==i,][,3]])
}
r<-max(rep)
CL<-SumTj.L-r*Total.InBj

RCi<-matrix(0,ncol=1,nrow=r)
for(j in 1:r){
RCi[j,1]<-sum(CL[unique(tmp[rep==j,][,2])])
}

SSB.Adj<- (sum(CL^2)/(r*k*(r-1)))-(sum(RCi^2)/(r*(k^2)*(r-1)))
SSE<- SSTot-SSRep-SSTrt.UnAdj-SSB.Adj

Df.Rep<- (r-1)
Df.Trt<- k^2-1
Df.Block<- r*(k-1)
Df.Error<- (r*k-k-1)*(k-1)

MSTrt.UnAdj<-SSTrt.UnAdj/Df.Trt
MSRep<-SSRep/Df.Rep
MSB.Adj<-SSB.Adj/Df.Block
 37

MSE<-SSE/Df.Error

mu<-(MSB.Adj-MSE)/(k*(r-1)*MSB.Adj)

Adj.Total<-matrix(0,ncol=1,nrow=k^2)

for(i in 1:(k^2)){
Adj.Total[i,1]<- Tj[i]+ mu*sum(CL[tmp[trt==i,][,2]])
}
Adj.means<-Adj.Total/r

BuI<-matrix(0,ncol=1,nrow=r)
for(i in 1:r){
BuI[i,1]<-(sum(Total.InBj[unique(tmp[rep==i,][,2])]^2)/k)-
(sum(Total.InBj[unique(tmp[rep==i,][,2])])^2)/(k^2)
}
Bu<-sum(BuI)

SSTrt.Adj<- SSTrt.UnAdj-k*(r-1)*mu*(r/((r-1)*(1+k*mu))*Bu-SSB.Adj)

Effective.Error <- MSE*(1+((r*mu*k)/(k+1)))

SSE.Effective <- (Effective.Error)*Df.Error

F.Trt<- (SSTrt.Adj/Df.Trt)/Effective.Error
se.Adj.Mean<- sqrt(Effective.Error/r)
se.Adj.Mean.Same<- sqrt((2*MSE*(1+(r-1)*mu))/r)
se.Adj.Mean.DiffB<- sqrt((2*MSE*(1+r*mu))/r)
averg.var<- 2*MSE*(1+(r*mu*k)/(k+1))/r

RBD.SSE<-SSB.Adj+SSE
RBD.MSE<- RBD.SSE/(Df.Block+Df.Error)

Efficiency<- (RBD.MSE/Effective.Error)*100

######################################
# MSTrt.UnAdj<-SSTrt.UnAdj/Df.Trt #
# MSRep<-SSRep/Df.Rep #
# MSB.Adj<-SSB.Adj/Df.Block #
# MSE<-SSE/Df.Error #
######################################

Df <- c(Df.Rep, Df.Block, Df.Trt,Df.Error,Df.Error)

ssq <- c(SSRep, SSB.Adj,SSTrt.Adj,SSE, SSE.Effective)

frep <- MSRep/MSE

fb<- MSB.Adj/MSE
ftrt<- F.Trt

Fval<-c(frep,fb,ftrt,NA,NA)
pval.B<-pf(fb, df1= Df.Block, df2= Df.Error,lower.tail = FALSE)
 38

pval.trt<-pf(ftrt, df1= Df.Trt, df2= Df.Error,lower.tail = FALSE)

Pvalue<-c(NA,pval.B,pval.trt,NA,NA)

anovadf <- data.frame(Df, ‘Sum Sq‘=ssq, ‘Mean Sq‘=ssq/Df,

‘F value‘=Fval ,‘Pr(>F)‘ = Pvalue, check.names=FALSE)
rownames(anovadf) <- c("Replication","Block.Adj[a]","Treatment.Adj[b]","Error[a]"
,"Effective.Error[b]")class(anovadf) <- c("anova","data.frame")

print(list(Adj.means=Adj.means))
cat(c(" \n"))
cat(c("S.E.Adjusted Treatment Mean : "),round(se.Adj.Mean,4),"\n")
cat(c("S.E.Diff of Adj Trt Mean : "),round(se.Adj.Mean.Same,4),"\n")
cat(c("S.E.Diff of Adj Trt Mean (not the same bk): "),round(se.Adj.Mean.DiffB,4),"\n")
cat(c("The average variance : "),round(averg.var,4),"\n")
cat(c("Efficiency Relative to RCBD : "),round(Efficiency,4),"\n")
cat(c(" \n"))
cat(c("Analysis of Variance Table: Partially Balanced\n"))
cat(c(" \n"))
return(anovadf)
}
}

#############################################################################
####################### BALANCED LATTICE DESIGN #############################
#############################################################################
# Creates a replication vector
rep<-rep(1:4,each=9)
# Creates a block vector
block<-rep(1:12,each=3)
# Creat a treatment vector
trt<-c(1,2,3,4,5,6,7,8,9,
3,4,8,2,6,7,1,5,9,
1,4,7,2,5,8,3,6,9,
3,5,7,2,4,9,1,6,8)

# Creat a response vector

y<-c(2.20,1.84,2.18,2.05,0.85,1.86,0.73,1.60,1.76,
1.71,1.57,1.13,1.76,2.16,1.80,1.81,1.16,1.11,
1.19,1.20,1.15,2.26,1.07,1.45,2.12,2.03,1.63,
2.04,0.93,1.78,1.50,1.60,1.42,1.77,1.57,1.43)

lattice.design(block,trt,rep,y,3)

#############################################################################
############# PARTIALLY BALANCED LATTICE DESIGN #############################
#############################################################################

rep<-rep(1:2,each=25)
block<-rep(1:10,each=5)
trt<-c(1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,
21,22,23,24,25,1,6,11,16,21,2,7,12,17,22,3,8,13,18,23,
 39

4,9,14,19,24,5,10,15,20,25)
y<-c(6,7,5,8,6,16,12,12,13,8,17,7,7,9,14,18,16,13,13,14,14,15,
11,14,14,24,13,24,11,8,21,11,14,11,23,16,4,12,12,12,17,10,
30,9,23,15,15,22,16,19)
lattice.balance(block,trt,rep,y,5)
 40

B References
References
[1] Bose, R. C., and Nair, K.R.(1939), “Partially Balanced Incomplete Block
Designs,” Sankhya, 4, 337-372.

[2] Cochran, W.G., and Cox, G.M. (1957), Experimental Designs(2nd ed.), New York:
John Wiley & Sons Inc.

[3] Comstock, R.E., Peterson, W.J., and Stewart, H.A. (1948), “An Application of
the Balanced Lattice Design in a Feeding Trial with Swine,” Journal of Animal
Science, 7, 320-331, URL http://jas.fass.org/content/7/3/320.

[4] Federer, W.T., and Wright, J. (1988), “Construction of Lattice Square Designs,”
Biometrical Journal, 30, 77-85.

[5] Felipe de Mendiburu.(2010), Statistical Procedures for Agricultural Research,

URL http://cran.r-project.org/web/packages/agricolae/agricolae.pdf.

[6] Fernandez, and George, C.J.(1990), “Analysis of Lattice Designs Using PC-SAS,”
HortScience, 2, 1450.

[7] Groemping U. (2010), “CRAN Task View: Design of Experiments (DoE) and
Analysis of Experimental Data,” HortScience, URL
http://cran.r-project.org/web/views/ExperimentalDesign.html.

[8] Hinkelmann, and Kempthorne (2005), Design and Analysis of Experiments -

Volume 2 Advanced Experiment Design, New Jersey:Wiley, 649.

[9] Monygomery, D. C. (2005), Design and Analysis of Experiments(6th ed.), New

York: Wiley, 146.

[10] Nair, K.R.(1952), “Analysis of Partially Balanced Incomplete Block Designs

Illustrated on the Simple Square and Rectangular Lattices,” Biometrics, Vol.8,
2, 122-155.

[11] SAS Institute Inc, SAS/STAT 9.2

c User’s Guide, Cary, NC:SAS Institute Inc.

[12] Sharma, V.K., Design and Analysis of Agricultural Experiments: Lattice

Designs.
 41

URL http://www.iasri.res.in/iasriwebsite/DESIGNOFEXPAPPLICATION
Electronic-Book/Module%202/4LATTICE%20Designs.pdf.

[13] Stroup, W. W., and Paparozzi, E.T. (1989), “Summary, Comments, and
Question and Answer Session,” HortScience, 24, 58-61.

[14] Yates, F. (1936a), “Incomplete Randomized Blocks,” Ann. Eugenics, 7, 121-140.

[15] Yates, F. (1936b), “A New Method of Arranging Variety Trials Involving a Large
Number of Varieties”, The Journal of Agricultural Science, 26, 424-455.