Lecture 8

Latin Square Design and

Incomplete Block Design
Latin Square (LS) design
 Latin square (LS) design
• It is a kind of complete block designs.
• A class of experimental designs that
allow for two sources of blocking.
• Can be constructed for any number of
treatments, but there is a cost. If there
are t treatments, then t2 experimental
units will be required.
 Latin square design
• If you can block on two (perpendicular)
sources of variation (rows x columns) you
can reduce experimental error when
compared to the RBD
• More restrictive than the RBD
• The total number of plots is the square of
the number of treatments
• Each treatment appears once and only
once in each row and column A B C D
B C D A
C D A B
D A B C
 Facts about the LS Design
• With the Latin Square design you are
able to control variation in two directions.
• Treatments are arranged in rows and
columns
• Each row contains every treatment.
• Each column contains every treatment.
• The most common sizes of LS are 5x5
to 8x8
 Advantages
• You can control variation in two
directions.
• Hopefully you increase efficiency
as compared to the RBD.
 Disadvantages
• The number of treatments must equal the
number of replicates.
• The experimental error is likely to increase
with the size of the square.
• Small squares have very few degrees of
freedom for experimental error.
• You can’t evaluate interactions between:
• Rows and columns
• Rows and treatments
• Columns and treatments.
 Examples of Uses of the
Latin Square Design
• 1. Field trials in which the experimental
error has two fertility gradients running
perpendicular each other or has a
unidirectional fertility gradient but also has
residual effects from previous trials.
 Examples of Uses of the
Latin Square Design
• 2. Animal science feed trials.
• 3. Insecticide field trial where the insect
migration has a predictable direction that
is perpendicular to the dominant fertility
gradient of the experimental field.
 Examples of Uses of the
Latin Square Design
• 4. Greenhouse trials in which the
experimental pots are arranged in a
straight line perpendicular to the glass
walls, such that the difference among rows
of pots and distance from the glass wall
are expected to be the major sources of
variability. ‘Row’
1 2 3 4

A B C D B C D A C D A B D A B C

1 2 3 4
‘Column’
How to randomize the design
• If all standard Latin squares of size t x t
are available, randomization is
accomplished with the following steps:
– Step 1. Randomly select one of the standard
squares
– Step 2. Randomly order all but the first row
– Step 3. Randomly order all columns
– Step 4. Randomly assign treatments to the
letters
 Example
• All possible randomizations can be generated
without including the first row in Step 2 if a
standard square is randomly selected.
• If all standard squares are not available for
selection, then it is recommended in Step 2 that
all rows be included in the randomization.
• Not all possible Latin squares are generated in
this way but a number of possibilities is increased
considerably. Suppose the standard square
selected at Step 1 for the 4 x 4 Latin square
experiment with automobile tires is A B C D
B C D A
C D A B
D A B C
 Step 2
• Obtain a random permutation of numbers
to order the last three rows:
Permutation Original row
3 2
1 3
2 4
• The placement of the rows for the
standard square with row 1 in its original
position is Original row
1 A B C D
3 C D A B
4 D A B C
2 B C D A
 Step 3
• Obtain a random permutation of number to order
the four columns from Step 2.
Permutation Original row
1 1
4 2
3 3
2 4
• The placement of the columns for the standard
square is Original column
1 4 3 2
A D C B
C B A D
D C B A
B A D C
 Step 4
• Obtain a random permutation to assign
treatments to the letters. This assignment is
not necessary if the standard square has
been selected at random from all possible
standard squares. The method of
assignment is shown here for illustration.
Suppose the treatment as W, X, Y and Z.
Permutation 4 (D), 2 (B), 3 (C), 1 (A).
• The treatment labels W, X, Y, Z replace the
Latin square letters in the order D, B, C and
A in the randomized arrangement.
 Step 4
Tire position
Auto 1 2 3 4
1 Z W Y X
2 Y X Z W
3 W Y X Z
4 X Z W Y
Analysis for LS
 Linear Model
• Linear Model: yij =  + ρi + j +k + ij
–= mean effect
– ρi = ith row effect (i=1,…, r)
– j = jth column effect (j=1, …, c)
– k = kth treatment effect (k=1,…, t)
– ij = random error
• Each treatment occurs once in each row
and once in each column
–r=c=t
– N = t2
 Analysis
• Set up a two-way table and compute the row and
column means and deviations
• Compute a table of treatment means and
deviations
• Set up an ANOVA table divided into sources of
variation
– Rows
– Columns
– Treatments
– Error
• Significance tests
– FT tests difference among treatment means
– FR and FC test if row and column groupings are
effective
 Size of the Square on Error
Degrees of Freedom
Source D.F. 2x2 3x3 4x4 5x5 8x8
Total t2-1 3 8 15 24 63
Rows t-1 1 2 3 4 7
Columns t-1 1 2 3 4 7
Treatments t-1 1 2 3 4 7
Error (t-1)(t-2) 0 2 6 12 42

One way to increase the Error df for small

squares is to use more than one square in
the experiment (i.e. repeated squares).
 ANOVA
Source of Degree Sum of Mean Expected F
variation of squares square MS (EMS)
freedom
Total t2-1  y
i j
ij  y 
2

Rows t   yi .  y 
2
t-1 MSR MSR/MSε
i

Columns t   y. j  y 
2
t-1 MSC MSC/MSε
j

t   yk  y 
2
Treatments t-1
i
MST  2  t T2 MST/MSε
Error (t-1)(t-2) SSε MSε
 ANOVA
• Where
SSTotal    yij  y    yij - t 2 y 2
2 2

i j i j

SS R  t   yi .  y   t  yi2.  t 2 y 2
2

i i

SSC  t   y. j  y   t  y.2j  t 2 y 2
2

j j

SSTreatment  t  yk2  t 2 y 2
i

SS  SSTotal  SS R  SSC  SSTreatment

Standard errors for treatment means
• Standard Error estimate for a
treatment mean s  MS t
yk 

• Standard Error estimate for a

distance between two treatment
means
s yk  ym  2 MS t
 Did both blocking factors
increase precision?
• The efficiency of the Latin Square design
with two blocking criteria is determined
relative to the randomized complete block
design with only one blocking criterion.
• Relative efficiency measures can be
computed separately for the row and
column blocking criteria of the Latin
square.
Relative efficiency of experiment designs
• Relative efficiency measures the effectiveness of
blocking in experiment designs to reduce experimental
error variance
• The variance of a treatment mean is a measure of the
precision of the estimated treatment mean in an
 y  r 
experiment, i.e., 2 1 2

• Say, error variances are 1 and 2 in two designs, and

replications are r1 and r2
y 
2
1
1
r1 y 
2
2
2
r2

• The two variances will be same only if r2 = 2 r1. We say

Design 1 is more efficient than Design 2 with respect to
the number of replications required to have the same
precision for an estimate of the treatment mean
 When error variance has to be
estimated from the data
• Fisher (1960) proposed the concept of
Information (I), i.e.,
( f  1) 1
I
( f  3) s 2

• where s2 is the estimated experimental

error variance with f degrees of freedom
 The relative efficiency
• The relative efficiency of two experiment designs is
defined as the ratio of information in the two
designs, i.e., ( f1  1) 1 ( f 2  1) 1
I1  I2 
( f1  3) s1
2
( f 2  3) s22
I1 ( f1  1)( f 2  3) s22
RE  
I 2 ( f1  3)( f 2  1) s12
• When RE=1, the designs require the same number
of replications to have the same variance of
treatment mean, i.e.  y
2

• Say RE=1.5, Design 2 requires 1.5 times as many

replications as Design 1 to have the same
variance of a treatment mean
 Relative Efficiency of LS design
• Relative Efficiency of column blocking
MSC  (t - 1) MS
2
srcb  (The estimated mean square for error in RCB)
t
MSC  (t - 1) MS
REC 
tMS
• Relative Efficiency of row blocking
MS R  (t - 1) MS
s 2
rcb  (The estimated mean square for error in RCB)
t
MS R  (t - 1) MS
RER 
tMS
 Relative Efficiency
• To compare with a completely
randomized design

MS R  MSC  (t - 1) MS
RE 
(t  1) MS
 An example
• Grain yield of a wheat variety for five different
seeding rates in a Latin square design [Treatment
label (A, B, C, D, or E) in parentheses following
yield value]
Row Column
1 2 3 4 5
1 59.45 (E) 47.28 (A) 54.44 (C) 50.14 (B) 59.45 (D)
2 55.16 (C) 60.89 (D) 56.59 (B) 60.17 (E) 48.71 (A)
3 44.41 (B) 53.72 (C) 55.87 (D) 47.99 (A) 59.45 (E)
4 42.26 (A) 50.14 (B) 55.87 (E) 58.74 (D) 55.87 (C)
5 60.89 (D) 59.45 (E) 49.43 (A) 59.45 (C) 57.31 (B)
Treatment A B C D E
Seed rate 30 80 130 180 230
 An example
Row / 1 2 3 4 5 Row yi .
Column means
1 59.45 (E) 47.28 (A) 54.44 (C) 50.14 (B) 59.45 (D) 54.15
2 55.16 (C) 60.89 (D) 56.59 (B) 60.17 (E) 48.71 (A) 56.30
3 44.41 (B) 53.72 (C) 55.87 (D) 47.99 (A) 59.45 (E) 52.29
4 42.26 (A) 50.14 (B) 55.87 (E) 58.74 (D) 55.87 (C) 52.58
5 60.89 (D) 59.45 (E) 49.43 (A) 59.45 (C) 57.31 (B) 57.31
Column 52.43 54.30 54.44 55.30 56.16 y.. 
means ( y. j ) 54.53
Treatment A B C D E
Seed rate 30 80 130 180 230
Mean ( yk ) 47.13 51.72 55.73 59.17 58.88
 ANOVA
Source of Degree of Sum of Mean F Pr>F
variation freedom squares square
Total 24 716.61
Rows 4 99.20 24.80 5.26 0.011*
Columns 4 38.48 9.62 2.04 0.153
Seed rate 4 522.30 130.57 27.67 0.000**
Error 12 56.63 4.72
Standard errors for treatment means
• The standard errors for a treatment mean
is
s yk  MS t  4.72 / 5  0.97

• The standard error estimate for a

difference between two treatment means
is
s yk  ym  2 MS t  2  4.72 5  1.37
 Relative Efficiency
• Estimated values

9.62  4  4.72
REC   1.21
5  4.72

24.80  4  4.72
RER   1.85
5  4.72
 Explanation for RER
• There is a 85% gain in efficiency over the
randomized complete block design in which
only the column criterion of the Latin square
design is used for blocking.
• Thus, the row blocks for soil gradients
across the field effectively reduced the
variance by 85%. The randomized block
design without the row blocks for soil
gradients would required 1.85*5=9.25 for
10 replications to have an estimated
variance of the treatment mean equal to
that from the Latin square design.
 Explanation for REC
• There is a 21% gain in efficiency over the
randomized complete block design in which
only the row criterion of the Latin square
design is used for blocking.
• Thus, the column blocks for soil gradients
across the field effectively reduced the
variance by 21%. The randomized block
design without the column blocks for soil
gradients would required 1.21*5=6
replications to have an estimated variance of
the treatment mean equal to that from the
Latin square design.
 Correction for estimating σ2
• The correction for estimation σ2 by s2 is
( f ls  1)( f rcb  3) 13  19
  0.97
( f ls  3)( f rcb  1) 15  17
• Where fls =12 and frcb =16 are the error
degrees of freedom for the Latin square and
randomized complete block design. The
correction reduces the RE from 1.85 to
0.97*1.85=1.79 for row blocking and from
1.21 to 0.97*1.21=1.17 for column blocking.
The correction has a small effect on the
efficiency estimates.
Incomplete Block Design
(IBD)
 An introduction
• It is sometimes necessary to block
experimental units into groups smaller
than a complete replication of all
treatments with a randomized complete
block or Latin square design.
• The incomplete block design is utilized to
decrease experimental error variance and
provide more precise comparisons among
treatments than is possible with a
complete block design.
 Incomplete block design
• If the block size, k, is less than the number
of treatments v (k < v) then all treatments
can not appear in each block. The design
is called an Incomplete Block Design.
• For example, consider the block design
with 6 treatments and 6 blocks of size two.

1 2 1 4 5 4
2 3 3 5 6 6
 Balanced Incomplete Block
Design (BIBD)
• The BIBD is arranged such that all
treatments are equally replicated and each
treatment pair occurs in the same block an
equal number of times somewhere in the
design.
• The balance obtained from equal
occurrence of all treatment pairs in the
same block results in equal precision for all
comparisons between pairs of treatment
means.
 Balanced Incomplete Block
Design (BIBD)
• There are t distinct treatments
• There are b blocks
• Each block contains exactly k distinct treatments
• Each treatments occurs in exactly r different blocks
• Every pair of distinct treatments occurs together in
exactly λ blocks
• Can be expressed as (t, k, λ) or (t, b, r, k, λ)
• For last example, t= 6, b=6, r=2, k=2, λ =1
 Properties
• 1. tr=bk (=total number of trials = N)
• 2. r(k-1)= λ(t-1)
• Proof: For each treatment, for example A1,
it will appear in r blocks. Each of these r
blocks contains k-1 non-A1 treatments. So
the total number of non-A1 treatments in
these r blocks is r(k-1).
• In another way, the numbers that A1 with
other treatments (t-1 treatments) should be
the same. So λ=r(k-1)/(t-1)
 How to randomize IBD
• After the basic design has been
constructed with the treatment code
numbers, the steps in randomization follow:
• Step 1. Randomize the arrangement of the
blocks of treatment code number groups
• Step 2. Randomize the arrangement of the
treatment code numbers within each block
• Step 3. Randomize the assignment of
treatments to the treatment code numbers
in the plan
 An example
• t=4 treatment, b=4 blocks, k=3 experimental
units each.
• Prior to randomization the plan is

Block I 1 2 3
Block II 1 2 4
Block III 1 3 4
Block IV 2 3 4
 Step 1
• The treatment groups (1,2,3), (1,2,4),
(1,3,4) and (2,3,4) must be randomly
assigned to the runs.
• Random permutation 2, 4, 1, 3
Run Original block
1 1 2 4 2
2 2 3 4 4
3 1 2 3 1
4 1 3 4 3
 Step 2
• Assign random treatment code numbers to
the three growth chambers in each run.
Choose a random permutation of the
numbers 1 to 4 for each chamber and omit
the treatment number absent in the run.
Chamber
Run A B C Permutation
1 2 4 1 2431
2 3 4 2 3412
3 1 2 3 4123
4 1 4 3 1432
 Step 3
• Suppose treatments T1 to T4. A random
permutation 2, 4, 3, 1 gives a random
assign to the treatment code number with
replacements 2 to T1, 4 to T2, 3 to T3, and
1 to T4.

Run A B C
1 T1 T2 T4
2 T3 T2 T1
3 T4 T1 T3
4 T4 T2 T3
 The model
• An appropriate linear model for
observations from an incomplete block
design is
• yij = μ + τi + ρj + εij (i = 1, 2, . . . , t; j = 1,
2, . . . , b),
• Where τi is the fixed effect of the ith
treatment, ρj the effect of the jth block, and
eij the error associated with the
observation yij. t b

   
i 1
i
j1
j 0
Sum of squares partitions for BIBD
• The sum of squares partitions can be
derived by considering alternative full and
reduced models for the design. Solutions
to the normal equations are obtained for
the full model, yij = μ + τi + ρj + εij , with
estimates yˆ ij  ˆ  ˆi  ˆ j , to compute the
experimental error sum of squares for the
full model,
SS f    yij  yˆ ij     yij  ˆ  ˆi  ˆ j 
2 2

i j i j
Sum of squares partitions for BIBD
• Solutions to the normal equations are
obtained for the reduced model, yij = μ +
ρj + εij , with estimates yˆ ij  ˆ  ˆ j , to
compute the experimental error sum of
squares for the reduced model,

SS r    yij  yˆ ij     yij  ˆ   j 

ˆ
2 2

i j i j
Sum of squares partitions for BIBD
• The difference SSε -SSε , is the reduction
r f
in sum of squares as a result of including
τi in the full model. It is the sum of
squares due to treatments after block
effects have been considered in the
model. It is referred to as SS Treatment
(adjusted), implying that block effects are
also considered when estimating the
treatment effects in the full model.
Sum of squares partitions for BIBD
• For BIBD, the SS for treatments adjusted,
SSε -SSε , can be computed directly as
r f t
k  Qi2
SST (adjusted)  i 1
t
• With t-1 degrees of freedom. The quantity Qi
is an adjusted treatment total computed as
b
1
Qi  yi .  Bi , Bi   nij y. j
k j 1

• Where Bi is the sum of all block totals that

include the ith treatment and nij=1 if treatment i
appears in block j and nij=0 otherwise.
Sum of squares partitions for BIBD
• This correction to the treatment total has
the net effect of removing the block effects
from the treatment total.
• The SS for blocks is derived from the
reduced model with treatments ignored in
the model as SSB(unadjusted)=SST -SSε .
r
Treatment effects are not considered when
estimating the block effects, and SS is
called an unadjusted SS.
• SST=SSB(unadjusted)+SST(adjusted)+SSε
f
 ANOVA
Source of Degree of Sum of Mean F
variation freedom squares square
Total N-1 
 ij ..
y  y 2

i j

Blocks b-1 k   y. j  y..  MSB(unadj.) MSB(unadj.)

2

j /MSε
t
Treatments t-1 k  Qi2 MST(adj.) MST(adj.)
i 1
t /MSε
Error N-t-b-1 By MSε
subtraction
 Pressure (psi)
Example Run 250 325 400 475 550 y.j
1 16 18 - 32 - 66
• Percent 2 19 - - 46 45 110
conversion 3 - 26 39 - 61 126
4 - - 21 35 55 111
of methyl
5 - 19 - 47 48 114
glucoside 6 20 - 33 31 - 84
by 7 13 13 34 - - 60
acetylene 8 21 - 30 - 52 103

under high 9 24 10 - - 50 84
10 - 24 31 37 - 92
pressure yi. 113 110 188 228 311 950
in a BIBD Bi 507 542 576 577 648
Qi -56.0 -70.7 -4.0 35.7 95.0
Example: B1= y.1+y.2+y.6+y.7+y.8+y.9
Q1=y1.-B1/3=113-507/3=-56.0
 ANOVA
Source of Degree of Sum of Mean F Pr>F
variation freedom squares square
Total 29 5576.67
Blocks 9 1394.67 154.96 5.02 0.0025**
Treatments 4 3688.58 922.14 29.90 0.000**
Error 16 493.42 30.84
 Treatment means
• The least squares estimate for a treatment
mean μi is ˆi  ˆ  ˆi
• Where ˆ  y ,ˆ 
kQi
.. i
t

• For example, Q1=-56.00,

ˆ  y..  950 / 30  31.67
• So that
kQ1 3  ( 56.00)
ˆ1    11.20, ˆ1  31.67  11.20  20.47
t 35
 Least square estimates of
treatment means
Pressure (psi) Mean ̂ i
250 20.47
325 17.53
400 30.87
475 38.80
550 50.67
Standard errors of treatment means
• The standard error for a treatment mean
estimate is
MS  kr (t  1)  30.84  3  6  4 
sˆi  1   1    2.44
rt  t  65  35 

• A 95% confidence interval estimate of a

treatment mean is
2.44 ˆi  t0.025,16 ( sˆi ), wheret0.025,16  2.120

• The standard error of the estimated

difference between two treatment means is
2kMS 2  3  30.84
sˆi  ˆ j    3.51
t 35