Chapter 30

Latin Square and Related Designs

We look at latin square and related designs.

30.1 Basic Elements

Exercise 30.1 (Basic Elements)
1. Different arrangements of same data.
Nine patients are subjected to three different drugs (A, B, C). Two blocks, each
with three levels, are used: age (1: below 25, 2: 25 to 35, 3: above 35) and
health (1: poor, 2: fair, 3: good). The following two arrangements of drug
responses for age/health/drugs,
health, i: 1 2 3
age, j: 1 2 3 1 2 3 1 2 3
drugs, k = 1 (A): 69 92 44
k = 2 (B): 80 47 65
k = 3 (C): 40 91 63

health ↓ age→ j=1 j=2 j=3

i=1 69 (A) 80 (B) 40 (C)
i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)

are (choose one) the same / different data sets. This is a latin square design
and is an example of an incomplete block design. Health and age are two blocks
for the treatment, drug.
2. How is a latin square created?
True / False Drug treatment A not only appears in all three columns (age
block) and in all three rows (health block) but also only appears once in every
row and column. This is also true of the other two treatments (B, C).

261
262 Chapter 30. Latin Square and Related Designs (ATTENDANCE 12)

3. Other latin squares.

Which are latin squares? Choose none, one or more.

(a) latin square candidate 1

health ↓ age→ j=1 j=2 j=3
i=1 69 (A) 80 (B) 40 (C)
i=2 91 (B) 92 (C) 47 (A)
i=3 65 (C) 63 (A) 44 (B)
(b) latin square candidate 2
health ↓ age→ j=1 j=2 j=3
i=1 69 (A) 80 (C) 40 (B)
i=2 91 (B) 92 (A) 47 (C)
i=3 65 (C) 63 (B) 44 (A)
(c) latin square candidate 3
health ↓ age→ j=1 j=2 j=3
i=1 69 (B) 80 (A) 40 (C)
i=2 91 (C) 92 (B) 47 (B)
i=3 65 (A) 63 (C) 44 (A)

In fact, there are twelve (12) latin squares when r = 3. There are 161,280 latin
squares when r = 5.

4. Latin squares counteract/reduce order effect

Nine patients are subjected to three different drugs (A, B, C). Two blocks, each
with three levels, are used: time (1: morning 25, 2: afternoon, 3: evening) and
day (1, 2, 3).

day ↓ time→ j=1 j=2 j=3

i=1 69 (A) 80 (B) 40 (C)
i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)

Since the three drugs are given in a different order during each of the three
days, this should (choose one) decrease / increase any confounding effect due
to the time at which any of the drugs are given during the day.

5. Latin square, with replications.

Eighteen (18) patients are subjected to three different drugs (A, B, C). Two
blocks, each with three levels, are used: age (1: below 25, 2: 25 to 35, 3: above
35) and health (1: poor, 2: fair, 3: good). Two possible equivalent arrangements
of this data are given below.
Section 1. Basic Elements (ATTENDANCE 12) 263

health ↓ replication, ↓ age→ j=1 j=2 j=3

i=1 l=1 69 (A) 80 (B) 40 (C)
l=2 68 (A) 89 (B) 62 (C)
i=2 l=1 91 (C) 92 (A) 47 (B)
l=2 80 (C) 95 (A) 68 (B)
i=3 l=1 65 (B) 63 (C) 44 (A)
l=2 40 (B) 43 (C) 49 (A)

replication, ↓ health ↓ age→ j=1 j=2 j=3

i=1 69 (A) 80 (B) 40 (C)
l=1 i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)
i=1 68 (A) 89 (B) 62 (C)
l=2 i=2 80 (C) 95 (A) 68 (B)
i=3 40 (B) 43 (C) 49 (A)

There are (choose one) one / two / three replications of this latin square
design. Increased replication decreases error variance.
6. Additional latin squares.
Eighteen (18) patients are subjected to three different drugs (A, B, C). Two
blocks, each with three levels, are used: age (1: below 25, 2: 25 to 35, 3: above
35) and health (1: poor, 2: fair, 3: good). Two possible equivalent arrangements
of this data are given below.
health ↓ latin square, ↓ age→ j=1 j=2 j=3
i=1 l=1 69 (A) 80 (B) 40 (C)
l=2 68 (A) 89 (C) 62 (B)
i=2 l=1 91 (C) 92 (A) 47 (B)
l=2 80 (B) 95 (A) 68 (C)
i=3 l=1 65 (B) 63 (C) 44 (A)
l=2 40 (C) 43 (B) 49 (A)

latin square, ↓ health ↓ age→ j=1 j=2 j=3

i=1 69 (A) 80 (B) 40 (C)
l=1 i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)
i=1 68 (A) 89 (C) 62 (B)
l=2 i=2 80 (B) 95 (A) 68 (C)
i=3 40 (C) 43 (B) 49 (A)

There are (choose one) one / two / three latin squares in this design. Each
additional latin square acts as another blocking variable and so reduces the
error.
264 Chapter 30. Latin Square and Related Designs (ATTENDANCE 12)

7. Latin square crossover design.

Six1 (6) patients are subjected to three different drugs (A, B, C). Two blocks,
each with three levels, are used: dose (1: below 25 mm, 2: 25 to 35 mm, 3:
above 35 mm) and frequency (1: once/day, 2: twice/day, 3: thrice/day).

frequency ↓ patient, ↓ dose→ j=1 j=2 j=3

i=1 m=1 69 (A) 80 (B) 40 (C)
m=2 68 (A) 89 (B) 62 (C)
i=2 m=1 91 (C) 92 (A) 47 (B)
m=2 80 (C) 95 (A) 68 (B)
i=3 m=1 65 (B) 63 (C) 44 (A)
m=2 40 (B) 43 (C) 49 (A)

The six (not two!) patients are each measured repeatedly, (three times) for the
three drugs (at three different dosages) which appear in a latin square design.
In other words, this design combines a repeated measures design with a latin
square design and so is called a latin square crossover design2 . In this case, the
subject is nested inside (choose one) frequency / dose.

30.2 Latin Square Model

The latin square model is stated in this section.

Exercise 30.2 (Latin Square Model)

Nine patients are subjected to three different drugs (A, B, C). Two blocks, each with
three levels, are used: age (1: below 25, 2: 25 to 35, 3: above 35) and health (1: poor,
2: fair, 3: good).

health ↓ age→ j=1 j=2 j=3

i=1 69 (A) 80 (B) 40 (C)
i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)

Using the following latin square model, match appropriately,

Yijk = µ··· + ρi + κj + τk + εijk ,

1
Even though the notation in the tables seem to suggest there is only two patients, there is, in
fact, intended to be six different patients in the study. Furthermore, it is not sensible to set up a
second arrangement of data, like in previous questions because there six patients, not two.
2
If two different latin squares had been used, the design would have been called a latin square
double crossover design.
Section 3. Analysis of Latin Square Experiments (ATTENDANCE 12) 265

model example
(a) Yijk (a) error
(b) µ··· (b) health (block, fixed) effect
(c) ρi (c) drug (treatment, fixed) effect
(d) κj (d) age (block, fixed) effect
(e) τk (e) (grand) average of all responses
(f ) εijk (f ) individual response

model (a) (b) (c) (d) (e) (f)

example

30.3 Analysis of Latin Square Experiments

SAS program: att12-30-3-drugs-latin-inference

Exercise 30.3 (Analysis of Latin Square Experiments: Drug Responses)

Nine patients are subjected to three different drugs (A, B, C). Two blocks, each with
three levels, are used: age (1: below 25, 2: 25 to 35, 3: above 35) and health (1: poor,
2: fair, 3: good).
health ↓ age→ j=1 j=2 j=3
i=1 69 (A) 80 (B) 40 (C)
i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)
1. Residual Plot Versus Predicted.
The residual plot, from the SAS output, indicates
(choose one) constant / nonconstant variance.
2. Normal Probability Plot of Residuals.
The normal probability plot, also obtained from the SAS output, indicates
(choose one) normal / non–normal residuals.
3. ANOVA Table
The ANOVA table is given by,

Source df SS MS E{M P S}
2 ρ2i
Row blocking variable r−1 SSROW MSROW σ + rP r−1
κ2i
Column blocking variable r−1 SSCOL MSCOL σ2 + r Pr−1
τ2
Treatment r−1 SSTR MSTR σ 2 + r r−1i
Error (r − 1)(r − 2) SSRem MSRem σ2
Total r2 − 1 SSTO
266 Chapter 30. Latin Square and Related Designs (ATTENDANCE 12)

True / False where, in this case,

Source df SS MS
Health (rows) 2 592.67 296.33
Age (columns) 2 2194.67 1097.33
Drug (treatments) 2 32.67 16.33
Error 2 216 108
Total 8 3036

4. Test drugs at α = 0.05

H0 : τi = 0 versus
Ha : at least one τi 6= 0, i = 1, 2, 3.
MSTR
since p–value P (F > MSRem = 16.33
108
= 0.151; 2, 8) = 0.86 > α = 0.05
accept null; that is, drug effect (choose one) is / is not significant

5. 95% pairwise confidence intervals for drug effect, bonferroni.

Calculate all the 95% pairwise confidence intervals for the drug treatment using
the Tukey procedure.
From SAS,
Y ··1 = 68.3, Y ··2 = 64.0, Y ··3 = 64.7
L̂1 = Y ··1 − Y ··2 = 4.3,
L̂2 = Y ··1 − Y ··3 = 3.6,
L̂3 = Y ··2q− Y ··3 = −0.7,
¢ q
s{L̂i } = MSRem 1r + 1r = 108 13 + 13 ≈ 8.49
¡ ¡ ¢

T = √12 q(1 − α; r, (r − 1)(r − 2)) = √12 q(0.95; 3, 2) = 8.33

√ ≈ 5.89
2
and so the CIs are
4.3 ± 8.49(5.89) =
(choose one) (−45.71, 54.31) / (−15.12, 32.78) / (−21.83, 19.83)
3.6 ± 8.49(5.89) =
(choose one) (−20.83, 18.83) / (−46.41, 53.61) / (−21.83, 19.83)
−0.7 ± 8.49(5.89) =
(choose one) (−20.83, 18.83) / (−15.12, 32.78) / (−50.71, 49.31)
In other words, none of the average responses to the three drugs are different
from one another.

30.4 Planning Latin Square Experiments

We look at the efficiency and power of tests in latin square designs.

Exercise 30.4 (Planning Latin Square Experiments: Drug Responses)

Nine patients are subjected to three different drugs (A, B, C). Two blocks, each with
Section 4. Planning Latin Square Experiments (ATTENDANCE 12) 267

three levels, are used: age (1: below 25, 2: 25 to 35, 3: above 35) and health (1: poor,
2: fair, 3: good).

health ↓ age→ j=1 j=2 j=3

i=1 69 (A) 80 (B) 40 (C)
i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)

where, recall,
Source df SS MS
Health (rows) 2 592.67 296.33
Age (columns) 2 2194.67 1097.33
Drug (treatments) 2 32.67 16.33
Error 2 216 108
Total 8 3036

1. Efficiency measure, latin square compared to CRD

Since
MSROW + MSCOL + (r − 1)MSRem
Ê1 =
(r + 1)MSRem
296.33 + 1097.33 + (3 − 1)(108)
=
(3 + 1)(108)
=

(circle one) 0.00 / 3.73 / 0.87,

which indicates the Latin square design reduces the error variance 3.73–fold, as
compared to the comparable completely randomized design (CRD).

2. Efficiency measure, latin square (column) compared to RBD

Since
MSCOL + (r − 1)MSRem
Ê2 =
rMSRem
1097.33 + (3 − 1)(108)
=
3(108)
=

(circle one) 0.00 / 0.34 / 4.05.

which indicates the Latin square (column alone) design reduces the error vari-
ance 4.05–fold, as compared to the comparable randomized block design (RBD).
268 Chapter 30. Latin Square and Related Designs (ATTENDANCE 12)

3. Efficiency Measure, latin square (row) compared to RBD

Since
MSROW + (r − 1)MSRem
Ê3 =
rMSRem
296.33 + (3 − 1)(108)
=
3(108)
=

(circle one) 1.58 / 0.34 / 0.87.

which indicates the Latin square (row alone) design reduces the error variance
1.58–fold, as compared to the comparable randomized block design (RBD).

4. Power of F Test.
Assume σ = 100. What is the power of the test for treatment effects
if τ1 = 4, τ2 = 0 and τ3 = −4? Since
1 X 2
q
φ = τk
σ
1 p 2
= (4) + (0)2 + (−4)2
100
= 0.06

and ν1 = r − 1 = 3 − 1 = 3
ν2 = (r − 1)(r − 2) = (2)(1) = 2
and so, using Table B.11 (p 1356), the power 1 − β is
(circle one) 0.00 / 0.11 / 0.87.
In other words, the power (where zero (0) is “poor” and one (1) is “excellent”)
of the test associated with the latin square in this case is
(choose one) poor / good.

30.5 Regression Approach to Latin Square De-

signs
SAS program: att12-30-5-drugs-latin-regression

We use the regression approach to deal with missing observations in latin square
designs.

Exercise 30.5 (Regression Approach to Latin Square Designs)

Eight patients (not nine–one observation is missing!) are subjected to three different
drugs (A, B, C). Two blocks, each with three levels, are used: age (1: below 25, 2:
25 to 35, 3: above 35) and health (1: poor, 2: fair, 3: good).
Section 5. Regression Approach to Latin Square Designs (ATTENDANCE 12) 269

health ↓ age→ j=1 j=2 j=3

i=1 missing (A) 80 (B) 40 (C)
i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)

1. Regression Approach, Full Model

Using the SAS output, the full model is
Yijk = µ··· + ρ1 Xijk1 + ρ2 Xijk2
+ κ1 Xijk3 + κ2 Xijk4
+ τ1 Xijk5 + τ2 Xijk6
+ εijk
where 
 1, if case from row block 1
Xijk1 = −1, if case from row block 3
0, otherwise,


and Xijk2 is defined similarly,


 1, if case from column block 1
Xijk3 = −1, if case from column block 3
0, otherwise,


and Xijk4 is defined similarly


 1, if case from treat 1
Xij5 = −1, if case from treat 3
0, otherwise,


and Xijk6 is defined similarly

and so an estimate of the full model is
Ŷ = 68.67 + 15.33X1 + 9.67X2
+ 3.33X3 + 8X4
+ 8.67X5 − 4.67X6
where SSE (F ) = (choose one) 0.00 / 0.11 / 54.0.
2. Regression Approach, Reduced Model
Using the SAS output, the reduced model (assuming there is no drug response
effect) is
Yijk = µ··· + ρ1 Xijk1 + ρ2 Xijk2
+ κ1 Xijk3 + κ2 Xijk4
+ εijk
270 Chapter 30. Latin Square and Related Designs (ATTENDANCE 12)

which has the estimate,

Ŷ = 66.5 + 11X1 + 11.83X2 − X3 + 10.17X4

where SSE (R) = (choose one) 54.0 / 74.56 / 223.7.

3. Regression Approach, Test of Drug Response
H0 : τi = 0 versus
Ha : at least one τi 6= 0, i = 1, 2, 3.
SSE (R) − SSE (F ) SSE (F ) 223.67 − 54 54
÷ = ÷ = 1.57
dfR − dfF dfF 3−1 1
and so p–value is P (F > 1.57; 2, 1) ≈ (choose one) 0.34 / 0.44 / 0.49.
since p–value = 0.49 > α = 0.05
accept null; that is, same average responses for different drugs

30.6 Additional Replications with Latin Square

Designs
SAS program: att12-30-6-drugs-latin-replications-inference

We look at latin square designs with replications.

Exercise 30.6 (Additional Replications with Latin Square Designs)

Eighteen (18) patients are subjected to three different drugs (A, B, C). Two blocks,
each with three levels, are used: age (1: below 25, 2: 25 to 35, 3: above 35) and
health (1: poor, 2: fair, 3: good).
replication, ↓ health ↓ age→ j=1 j=2 j=3
i=1 69 (A) 80 (B) 40 (C)
l=1 i=2 91 (C) 92 (A) 47 (B)
i=3 65 (B) 63 (C) 44 (A)
i=1 68 (A) 89 (B) 62 (C)
l=2 i=2 80 (C) 95 (A) 68 (B)
i=3 40 (B) 43 (C) 49 (A)

1. Residual Plot Versus Predicted.

The residual plot, from the SAS output, indicates
(choose one) constant / nonconstant variance.
2. Normal Probability Plot of Residuals.
The normal probability plot, also obtained from the SAS output, indicates
(choose one) normal / non–normal residuals.
Section 7. Replications in Repeated Measures Designs (ATTENDANCE 12) 271

3. ANOVA Table
The latin square with replications model is given by

Yijkm = µ··· + ρi + κj + τk + εijkm ,

where the ANOVA table is,

Source df SS MS
Row blocking variable r−1 SSROW MSROW
Column blocking variable r−1 SSCOL MSCOL
Treatment r−1 SSTR MSTR
Error n(r − 1)(r − 2) SSRem MSRem
Total nr 2 − 1 SSTO

True / False which, in this case, is

Source df SS MS
Health (rows) 2 2422.33 1211.16
Age (columns) 2 2006.33 100316
Drug (treatments) 2 129.33 64.67
Error 11 1562.5 142.045
Total 17 6120.5

4. Test drugs at α = 0.05

H0 : τi = 0 versus
Ha : at least one τi 6= 0, i = 1, 2, 3.
MSTR 64.67
since p–value P (F > MSRem = 142.045 = 0.46; 2, 11) = 0.64 > α = 0.05
accept null; that is, drug effect (choose one) is / is not significant

30.7 Replications in Repeated Measures Designs

SAS program: att12-30-7-drugs-crossover-inference

We look at a crossover design which combines a repeated measures design with a latin
square design.

Exercise 30.7 (Replications in Repeated Measures Designs)

Six3 (6) patients are subjected to three different drugs (A, B, C). Two blocks, each
with three levels, are used: dose (1: below 25 mm, 2: 25 to 35 mm, 3: above 35 mm)
and frequency (1: once/day, 2: twice/day, 3: thrice/day).
3
Even though the notation in the tables seem to suggest there is only two patients, there is, in
fact, intended to be six different patients in the study.
272 Chapter 30. Latin Square and Related Designs (ATTENDANCE 12)

frequency ↓ patient, ↓ dose→ j=1 j=2 j=3

i=1 m=1 69 (A) 80 (B) 40 (C)
m=2 68 (A) 89 (B) 62 (C)
i=2 m=1 91 (C) 92 (A) 47 (B)
m=2 80 (C) 95 (A) 68 (B)
i=3 m=1 65 (B) 63 (C) 44 (A)
m=2 40 (B) 43 (C) 49 (A)
The six (not two!) patients are each measured repeatedly (three times), for the three
drugs (at three different dosages) which appear in a latin square design.
1. Residual Plot Versus Predicted.
The residual plot, from the SAS output, indicates
(choose one) constant / nonconstant variance.

2. Normal Probability Plot of Residuals.

The normal probability plot, also obtained from the SAS output, indicates
(choose one) normal / non–normal residuals.

3. ANOVA Table
The latin square crossover model is given by

Yijkm = µ··· + ρi + κj + τk + ηm(i) + εijkm ,

where the ANOVA table is,

Source df SS MS E{M S} P
ρ2
Patterns (P) r−1 SSP MSP σ + rση2 +P
2
nr r−1i
κ2i
Order positions (O) r−1 SSO MSO σ 2 + nr P
r−1
τ2
Treatment (TR) r−1 SSTR MSTR σ 2 + nr r−1k
Subjects (S, within patterns) r(n − 1) SSS MSS σ 2 + rση2
Error (r − 1)(nr − 2) SSRem MSRem σ2
Total nr 2 − 1 SSTO

True / False which, in this case, is

Source df SS MS
Frequency (patterns, P) 2 2422.33 1211.16
Dose (order, O) 2 2006.33 1003.16
Drug (treatments, TR) 2 129.33 64.67
Patient(frequency) (subjects, S) 3 444.83 148.277
Error 8 1117.67 139.708
Total 17 6120.5
Section 7. Replications in Repeated Measures Designs (ATTENDANCE 12) 273

4. Test various at α = 0.05

H0 : τi = 0 versus
Ha : at least one τi 6= 0, i = 1, 2, 3.
MSTR 64.67
since p–value P (F > MSRem = 139.708 = 0.46; 2, 8) = 0.65 > α = 0.05
accept null; drug effect (choose one) is / is not significant
(average drug response same for different drugs)

H0 : ρi = 0 versus
Ha : at least one ρi 6= 0, i = 1, 2, 3.
1211.16
since p–value P (F > MSP
MSS
= 148.277 = 8.17; 2, 3) = 0.06 > α = 0.05
accept null; frequency effect (choose one) is / is not significant
(average drug response same for different frequency of drug administration)

H0 : κj = 0 versus
Ha : at least one κj 6= 0, k = 1, 2, 3.
MSO 1003.16
since p–value P (F > MSRem = 139.708 = 7.18; 2, 8) = 0.02 < α = 0.05
reject null; dose effect (choose one) is / is not significant
(average drug response different for different drug doses)