Latin Square Design

If there are two directional sources of variation in experimental area, it is necessary to remove the
sources of variations in two way corresponding to two mutually perpendicular direction being designed
by rows and another is column. If there are p treatments, experimental area will be divided with p rows
and p columns resulting p2 experimental units.

The treatment are then arranged at random to experimental units such that each treatment must appear
once and only once in each row and each column. Such an arrangement is called Latin Square
Design(LSD).

Since this design is a square arrangement where the treatments are denoted by Latin letters, so the
design is called Latin Square Design

Incomplete 3-way layout

The LSD is an incomplete 3-way layout. In LSD there are three factors (says) row, column and
treatment, each factor is subdivided into p groups , so there are p3 combinations and so for a complete
layout there need p3 experimental units. But in LSD p3 combinations are arranged in p2 experimental
units. That’s the reason behind the incomplete 3-way lay out.

Advantage
i) With two way grouping LSD controls more variations than CRD and RBD

ii) LSD is more efficient than CRD and RBD

iii) Statistical analysis is simple even though with missing observation

iv) LSD is an incomplete 3-way layout. Its advantages over the complete 3-way layout is that, instead
of p3 experimental units only p2 units are needed

v) Three factors can be investigated simultaneously with fewer trials than complicated design

Disadvantage
i) The fundamental assumption that there is no interaction between factors may not be always true in
general.

ii) LSD is not suitable for large number of treatments

iii) LSD can not be used if the number of treatments becomes less than three
Examples of LSD

The Statistical model for a LSD

𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + 𝜏𝑘 + 𝜖𝑖𝑗𝑘 i = 1, 2, …, p; j = 1, 2, …, p and k = 1,2,…,p
Where 𝑦𝑖𝑗𝑘 is the ith row and jth column for kth treatment, 𝜇 is the overall mean, 𝛼𝑖 is the ith row
effect, 𝛽𝑗 is the jth column effect, 𝜏𝑘 is the kth treatment effect, and 𝜖𝑖𝑗𝑘 is the random error. Note that
this is an effects model. The model is completely additive: that is, there is no interaction between rows,
columns and treatments. Because there is only one observation in each cell, only two of the three
subscripts i, j, and k are needed to denote a particular observation.
if i=2 and j = 3, we automatically find k = 4 (formulation D) and if i =1 and k=3(formulation C), we
find j = 3. This is a sequence of each treatment appearing exactly once in each row and column.
Analysis of Variance for LSD
SV df SS MSS CalF
Row p-1 𝑆𝑆𝑅 𝑀𝑆𝑅
Column p-1 𝑆𝑆𝐶 𝑀𝑆𝐶
Treatment p-1 𝑆𝑆𝑇𝑟 𝑀𝑆𝑇𝑟 𝑀𝑆𝑇𝑟
𝑀𝑆𝐸
Error (𝑝 − 2)(𝑝 − 1) 𝑆𝑆𝐸 𝑀𝑆𝐸
Total 𝑝2 − 1 𝑆𝑆𝑇

Partition of Total Sum of Square

Total Sum of Square =
𝑝

∑ (𝑦𝑖𝑗𝑘 − 𝑦̅… )2
𝑖,𝑗,𝑘=1
𝑝

= ∑ [(𝑦̅𝑖.. − 𝑦̅… ) + (𝑦̅.𝑗. − 𝑦̅… ) + (𝑦̅..𝑘 − 𝑦̅… )

𝑖,𝑗,𝑘=1
2
+ (𝑦𝑖𝑗𝑘 − 𝑦̅𝑖.. − 𝑦̅.𝑗. − 𝑦̅..𝑘 + 2 𝑦̅… )]
𝑝 𝑝 𝑝
2
= 𝑝 ∑(𝑦𝑖.. − 𝑦̅… )2 + 𝑝 ∑(𝑦.𝑗. − 𝑦̅… ) + 𝑝 ∑(𝑦..𝑘 − 𝑦̅… )2
𝑖=1 𝑗=1 𝑖=1
𝑝
2
+ ∑ (𝑦𝑖𝑗𝑘 − 𝑦̅𝑖.. − 𝑦̅.𝑗. − 𝑦̅..𝑘 + 2𝑦̅… )
𝑖,𝑗,𝑘=1
+ product terms (vanish)
= SSRow + SSColumn + SSTreatment + SSError
Expectation of Sum of Squares
The mathematical model
𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + 𝜏𝑘 + 𝜖𝑖𝑗𝑘
 𝑦̅𝑖.. = 𝜇 + 𝛼𝑖 + 𝜖̅𝑖..
𝑦̅.𝑗. = 𝜇 + 𝛽𝑗 + 𝜖̅.𝑗.

𝑦̅..𝑘 = 𝜇 + 𝜏𝑘 + 𝜖̅..𝑘
𝑦̅… = 𝜇 + 𝜖̅…
SSTreatment =𝑝 ∑𝑝𝑘=1(𝑦..𝑘 − 𝑦̅… )2 = 𝑝 ∑𝑝𝑘=1(𝜇 + 𝜏𝑘 + 𝜖̅..𝑘 − 𝜇 − 𝜖̅… )2 =
𝑝 𝑝 𝑝 𝑝
𝑝 ∑𝑘=1[𝜏𝑘 + (𝜖̅..𝑘 − 𝜖̅… )]2 = 𝑝 ∑𝑘=1 𝜏𝑘2 + 𝑝 ∑𝑘=1(𝜖̅..𝑘 − 𝜖̅… )2 + 2𝑝 ∑𝑘=1 𝜏𝑘 (𝜖̅..𝑘 − 𝜖̅… )
𝑝 𝑝 𝑝

𝐸 [SSTreatment ] = 𝐸 [𝑝 ∑ 𝜏𝑘2 + 𝑝 ∑ (𝜖̅..𝑘 − 𝜖̅… )2 + 2𝑝 ∑ 𝜏𝑘 (𝜖̅..𝑘 − 𝜖̅… )]

𝑘=1 𝑘=1 𝑘=1

𝑝 𝑝
𝜎2
= 𝑝 ∑ 𝜏𝑘 + 𝑝(𝑝 − 1) + 2𝑝. 0 = 𝑝 ∑ 𝜏𝑘2 + (𝑝 − 1)𝜎 2
2
𝑝
𝑘=1 𝑘=1

Now
𝑝
𝑝
𝐸 [MSTreatment ] = 𝜎 2 + ∑ 𝜏𝑘2
(𝑝 − 1)
𝑘=1

Similarly
𝑝
𝑝
𝐸 [MSRow ] = 𝜎 2 + ∑ 𝛼𝑖2
(𝑝 − 1)
𝑖=1
𝑝
𝑝
𝐸 [MSColumn ] = 𝜎 2 + ∑ 𝛽𝑗2
(𝑝 − 1)
𝑗=1
 𝑝
2
SSError = ∑ (𝑦𝑖𝑗𝑘 − 𝑦̅𝑖.. − 𝑦̅.𝑗. − 𝑦̅..𝑘 + 2𝑦̅… )
𝑖,𝑗,𝑘=1
𝑝
2
= ∑ (𝜖𝑖𝑗𝑘 − 𝜖̅𝑖.. − 𝜖̅.𝑗. − 𝜖̅..𝑘 + 2𝜖̅... )
𝑖,𝑗,𝑘=1

𝑝
𝜎2 𝜎2 𝜎2 2
𝜎2 𝜎2 𝜎2 𝜎2
𝐸 [SSError ] = ∑ [𝜎 + + + +4 2 −2 −2 −2
𝑝 𝑝 𝑝 𝑝 𝑝 𝑝 𝑝
𝑖,𝑗,𝑘=1
2
𝜎 𝜎2 𝜎2 𝜎2 𝜎2 𝜎2 𝜎2
+ 4 2 + 2 2 + 2 2 − 4 2 − 4 2 + 2 2 − 4 2]
𝑝 𝑝 𝑝 𝑝 𝑝 𝑝 𝑝

= 𝜎 2 (𝑝2 − 3𝑝 + 2) = 𝜎 2 (𝑝 − 2)(𝑝 − 1)

𝐸 [MSError ] = 𝜎 2
Analysis of Variance for LSD
SV df SS MSS E(MSS) CalF
𝑝 𝑀𝑆𝑅
Row p-1 𝑆𝑆𝑅 𝑀𝑆𝑅 𝑝
𝜎2 + ∑ 𝛼𝑖2 𝑀𝑆𝐸
(𝑝 − 1)
𝑖=1
𝑝 𝑀𝑆𝐶
Column p-1 𝑆𝑆𝐶 𝑀𝑆𝐶 𝑝
2
𝜎 + ∑ 𝛽𝑗2 𝑀𝑆𝐸
(𝑝 − 1)
𝑗=1
𝑝 𝑀𝑆𝑇𝑟
Treatment p-1 𝑆𝑆𝑇𝑟 𝑀𝑆𝑇𝑟 𝑝
2
𝜎 + ∑ 𝜏𝑘2 𝑀𝑆𝐸
(𝑝 − 1)
𝑘=1
2
Error (𝑝 − 2)(𝑝 − 1) 𝑆𝑆𝐸 𝑀𝑆𝐸 𝜎
Total 𝑝2 − 1 𝑆𝑆𝑇

Parameter Estimation of LSD

𝑦𝑖𝑗𝑘 = 𝜇 + 𝛼𝑖 + 𝛽𝑗 + 𝜏𝑘 + 𝜖𝑖𝑗𝑘 i = 1, 2, …, p; j = 1, 2, …, p and k = 1,2,…,p
Where 𝑦𝑖𝑗𝑘 is the ith row and jth column for kth treatment, 𝜇 is the overall mean, 𝛼𝑖 is the ith row
effect, 𝛽𝑗 is the jth column effect, 𝜏𝑘 is the kth treatment effect, and 𝜖𝑖𝑗𝑘 is the random error. Note that
this is an effects model. The model is completely additive: that is, there is no interaction between rows,
columns and treatments. Because there is only one observation in each cell, only two of the three
subscripts i, j, and k are needed to denote a particular observation.
Assumptions
(1) The population, where the observations have been taken is normal and the observations are
independent
(2) Treatment and block effects are defined as deviations from the overall mean, so ∑𝑝𝑖=1 𝛼𝑖 = 0,
∑𝑝𝑗=1 𝛽𝑗 = 0 and ∑𝑝𝑘=1 𝜏𝑘 = 0.
(3) 𝜖𝑖𝑗𝑘 ~NID(0, 𝜎 2 )

To find out the least squares estimators of 𝜇, 𝛼𝑖 , 𝛽𝑗 and 𝜏𝑘 we form the sum of squares of the errors
𝑝 𝑝 2
𝐿 = ∑𝑖,𝑗,𝑘=1 𝜖𝑖𝑗𝑘 2 = ∑𝑖,𝑗,𝑘=1(𝑦𝑖𝑗𝑘 − 𝜇 − 𝛼𝑖 − 𝛽𝑗 − 𝜏𝑘 )

And choosing values of 𝜇, 𝛼𝑖 , 𝛽𝑗 and 𝜏𝑘 say 𝜇̂ , 𝛼̂𝑖 , 𝛽̂𝑗 and 𝜏̂ 𝑘 which minimized L
𝜕𝐿
| = 0 gives
𝜕𝜇 𝜇
̂ ,𝛼 ̂𝑗 ,𝜏̂𝑘
̂ 𝑖 ,𝛽
𝑝
−2 ∑ (𝑦𝑖𝑗𝑘 − 𝜇̂ − 𝛼̂𝑖 − 𝛽̂𝑗 − 𝜏̂ 𝑘 ) = 0 (𝑖)
𝑖,𝑗,𝑘=1
𝑝 𝑝 𝑝

⇒ 𝑦… = 𝑝 𝜇̂ + 𝑝 ∑ 𝛼̂𝑖 + 𝑝 ∑ 𝛽̂𝑗 + 𝑝 ∑ 𝜏̂ 𝑘
2

𝑖=1 𝑗=1 𝑘=1

Similarly
𝜕𝐿
| = 0 gives
𝜕𝛼 𝜇
̂ ,𝛼 ̂𝑗 ,𝜏̂𝑘
̂ 𝑖 ,𝛽
𝑝
−2 ∑ (𝑦𝑖𝑗𝑘 − 𝜇̂ − 𝛼̂𝑖 − 𝛽̂𝑗 − 𝜏̂ 𝑘 ) = 0 (𝑖𝑖)
𝑗,𝑘=1
𝑝 𝑝

⇒ 𝑦𝑖.. = 𝑝𝜇̂ + 𝑝𝛼̂𝑖 + ∑ 𝛽̂𝑗 + ∑ 𝜏̂ 𝑘

𝑗=1 𝑘=1
𝜕𝐿
| = 0 gives
𝜕𝛽 𝜇
̂ ,𝛼 ̂𝑗 ,𝜏̂𝑘
̂ 𝑖 ,𝛽
𝑝
−2 ∑ (𝑦𝑖𝑗𝑘 − 𝜇̂ − 𝛼̂𝑖 − 𝛽̂𝑗 − 𝜏̂ 𝑘 ) = 0 (𝑖𝑖𝑖)
𝑖,𝑘=1
𝑝 𝑝

⇒ 𝑦.𝑗. = 𝑝𝜇̂ + ∑ 𝛼̂𝑖 + 𝑝𝛽̂𝑗 + ∑ 𝜏̂ 𝑘

𝑖=1 𝑘=1
𝜕𝐿
| = 0 gives
𝜕𝜏 𝜇
̂ ,𝛼 ̂𝑗 ,𝜏̂𝑘
̂ 𝑖 ,𝛽
𝑝
−2 ∑ (𝑦𝑖𝑗𝑘 − 𝜇̂ − 𝛼̂𝑖 − 𝛽̂𝑗 − 𝜏̂𝑘 ) = 0 (𝑖𝑣)
𝑖,𝑗=1
𝑝 𝑝

⇒ 𝑦..𝑘 = 𝑝𝜇̂ + ∑ 𝛼̂𝑖 + ∑ 𝛽̂𝑗 + 𝑝𝜏̂ 𝑘

𝑖=1 𝑗=1

In the above system of equations from (i) to (iv) there are (3p+1) equations. Out of these equations
there are 3 dependent equations so we may impose 3 restrictions to estimate the parameters which are
as follows
𝑝 𝑝 𝑝

∑ 𝛼̂𝑖 = ∑ 𝛽̂𝑗 = ∑ 𝜏̂𝑘 = 0

𝑖=1 𝑗=1 𝑘=1

using these restrictions we have

𝑦…
𝜇̂ = = 𝑦̅… from eq (i)
𝑝2

Again, from eq(ii)

𝛼̂𝑖 = 𝑦̅𝑖.. − 𝑦̅…
Similarly
𝛽̂𝑗 = 𝑦̅.𝑗. − 𝑦̅…
and
𝜏̂ ..𝑘 = 𝑦̅..𝑘 − 𝑦̅…