BASIC EXPERIMENTAL DESIGNS

LATIN SQUARE DESIGN

Dr. Yousaf Hayat

Professor
Department of Mathematics, Statistics and Computer Science

05/31/24 Dr. Yousaf Hayat 1

 Latin Square Design
 RCBD is less efficient if there are two sources of variation A B C D
in the experimental units
 To control two sources of variation, Latin Square Design B C D A
is preferred. C D A B
 Two way blocking is used for controlling the sources of
D A B C
variation in the experimental units
 Blocks are always made orthogonal of variation i.e. Row blocks and
Column blocks
 Latin Square Design needs more restriction than RCBD i.e. number of
rows, columns and treatments must be equal, and total number of plots will be the
square of the number of treatments
 Each treatment will appears once and only once in each row and column
05/31/24 Dr. Yousaf Hayat 2
Sources of Variation in the Experimental Units
 Two sources of variation A B C D
 Slope and fertility
B C D A
 Furrow irrigation and shading
Orthogonal blocks are made i.e. C D A B
perpendicular to the sources of D A B C
variation.

 This design is preferred to be used if the

number of treatments are FOUR but fewer
than 10.

05/31/24 Dr. Yousaf Hayat 3

Statistical Model for Latin Square Design

 Yij(k) =  + i + j +k + ij

 = mean effect
 βi = ith row effect

 j = jth column effect

 k = kth treatment effect

 ij = random error

 Each treatment will occur once in each row and once in each column
 Number of treatments, rows and columns must be equal (say P).
05/31/24 Dr. Yousaf Hayat 4
Randomization and Layout
Randomization procedure take the following steps: Standard 4x4
5x5
1. Take a standard 4x4 Latin square plan A B C D
2. Randomize the rows B C D A
3. Randomize the columns B C D A C D A B
4. Randomize the treatments A B C D D A B C
D A B C
Row: 1 2 3 4
Random #: 220 100 821 440 C D A B
Rank #: 2 1 4 3
A D B C
Column: 1 2 3 4
Random # 678 445 230 425 D C A B
Rank # 4 3 1 2 C B D A
B A C D
Treatment: A B C D C B A D
Random# 340 150 558 319
B D C A
Rank# 3(C) 1( A ) 4(D) 2(B)
D A B C
05/31/24 Dr. Yousaf Hayat A C D B 5
4 x 4 Latin Square Design

No. of Cows/Group = 4, Period/Lactation = 4, No. of feeds =4

Period/Time
Group
I II III IV
Group-1 A B C D
Group-2 B C D A
Group-3 C D A B
Group-4 D A B C
05/31/24 Dr. Yousaf Hayat 6
 Statistical Analysis
 In a two way Table (actual plan of Latin Square) calculate the total of
each row and each column
 Calculate the total of each treatment in the experiment
 Calculate the grand total (GT), and after that compute the Correction
Factor (CF) by dividing the square of GT by the total number of
observations
 Compute the sum of squares for
 Total, Rows, Columns, Treatments and Error
 Construct ANOVA Table wherein FT will tests the differences among
treatment means, and FR and FC will test if row and column blocks
are effective in controlling the respective sources of variation
05/31/24 Dr. Yousaf Hayat 7
 Statistical Analysis
Consider that there are P-treatments in the experiment, so that we have PxP Latin
Square Design. The following steps will be taken to compute various sum of
squares:
CF = (GT)2/P2
ANOVA Table
 SSTotl = ijk2 – CF SOV df SS MS F-ratio
Row (P-1) SSR MSR FR
 SSR = 1/P Yi.2 - CF
Column (P-1) SSC MSC FC
 SSC = 1/P Y.j - CF
2
Treatment (P-1) SST MST FT
 SST = 1/P Y.(k)2 - CF Error (P-1) (P-2) SSE MSE -------
Total (P2 – 1) SSTotal -------- -------
 SSE = SSTotal – SSR –SSC – SST -

05/31/24 Dr. Yousaf Hayat 8

Numerical Example
To determine the effect of four different sources of seed inoculums, A, B,
C, and D, and a control, E, on the dry matter yield of irrigated alfalfa.
The plots were furrow irrigated and there was a line of trees that might
form a shading gradient. Analyze the data and draw conclusions at 5%
level of probability.
A (40) B (38) D (34) C (36) E (28) 176
D (32) E (26) B (33) A (35) C (29) 155
C (37) D (36) A (43) E (33) B (41) 190
E (24) A (33) C (31) B (31) D (32) 151
B (41) C (33) E (31) D (35) A (39) 179
05/31/24 Dr. Yousaf Hayat 9
174 166 172 170 169 851
DATA FORMAT

Row Column Treatment Data

1 1 1 40
5 3 5 31
2 1 4 32
1 4 3 36
3 1 3 37
4 1 5 24 2 4 1 35
5 1 2 41 3 4 5 33
1 2 2 38
4 4 2 31
2 2 5 26
5 4 4 35
3 2 4 36
4 2 1 33 1 5 5 28
5 2 3 33 2 5 3 29
1 3 4 34
3 5 2 41
2 3 2 33
4 5 4 32
3 3 1 43
4
05/31/24 3 3 31 Dr. Yousaf Hayat5 5 1 39 10
ANOVA of Dry Matter Yield
Latin Square ANOVA Table

Source DF SS MS F P
Row 4 220.560 55.1400 20.89 0.000
Column 4 7.360 1.8400 0.697 0.6085
Treatment 4 279.360 69.8400 26.45 0.0000
Error 12 31.680 2.6400
Total 24 538.960

Grand Mean = 34.040 CV = 4.77%

LSD All-Pairwise Comparisons Test of data for Treatment

Treatment Mean Homogeneous Groups
1 38.000 a
2 36.800 a
3 33.200 b
4 33.800 b
5 28.400 c
05/31/24 Dr. Yousaf Hayat 11
Means and Standard Errors
45
a a
Treatment Mean SE 40
b b
35
c

Average Yield (Kg/plot)

30
A 38.00 a 1.79
25
20
B 36.80 a 2.06 15

10

C 33.20 b 1.50 5

0
A B C D E
D 33.80 b 0.80 Treatments

E 28.40 c 1.63
05/31/24 Dr. Yousaf Hayat 12
Report of Statistical Analysis
 F-test for treatments is highly significant (P< 0.01) indicating that there
exist significant differences among the yield of treatments.
 The F-test for Row blocking is significant which indicates that Row
blocking is effective in controlling the experimental error. On the other
hand, F-test for Column grouping is non-significant suggesting that it is
not effective in reducing the experimental error.
 Treatment A produced higher yield though not significantly different from
B. Similarly, there was no significant difference between the yield of
treatment C and D. However, the yield of treatment E was significantly
lower from rest of the treatments.
05/31/24 Dr. Yousaf Hayat 13
Relative Efficiency
 To compare with an RBD using columns as blocks
RE = MSR + (P-1) x MSE
P x MSE
 To compare with an RBD using rows as blocks
RE = MSC + (P-1) x MSE
P x MSE
 To compare with a CRD
RE = MSR + MSC + (P-1) x MSE
(P+1) x MSE
If RE > 1, then Latin square design will be more efficient; otherwise
RCBD with column or row as block will be efficient; and in case of
comparing it with CRD, if RE < 1 then CRD will be efficient.

05/31/24 Dr. Yousaf Hayat 14

Estimating Missing Value
 To estimate one missing value in Latin Square Design, the following formula will be
used:
^
Yij(k) = P x (Ri + Cj + Tk)-2G
[(P-1)(P-2)]
Where:
Ri = sum of observations in the ith row without missing value
Cj = sum of observations in the jth column without missing value
Tk = sum of observations in the kth treatment without missing value
G = grand total without missing value
P = number of treatments
05/31/24 Dr. Yousaf Hayat 15
ADVANTAGES AND LIMITATIONS
 This design is used to control two sources of variation in the experimental
units.
 This design requires that the number of rows, columns and treatments are
equal, which is the greatest disadvantage.
 For two treatments i.e. 2x2 LS design is not exist, because the degrees of
freedom for error term is zero and hence MSE is not exist (undefined).
 Generally (4x4) to (9x9) LS Design are used.
 In the presence of more than two sources of variations this design is
inefficient.
 The analysis becomes complicated if there exist missing observation(s) in
the data.
 Wrong assignment of treatments to the experimental units can cause great
difficulty in the analysis, and the design is no more LS design.
05/31/24 Dr. Yousaf Hayat 16
 DISCUSSION

05/31/24 Dr. Yousaf Hayat 17