CSS 312: Design and Analysis of Experiments Second Semester 2010/11 Academic Year
Lecture 8: Multiway Factorial Analysis of Variance
Part II: The Latin square experimental design
Introduction
Recall the Randomized Complete Block Design from Lecture 5. RCBD is a two-factor analysis
of variance in which there is one fixed-effects factor, one random-effects factor, and no
replication. The levels of the random-effects factor are termed blocks, and it is assumed that no
variation exists within blocks except that due to differences among levels of the fixed-effects
factor.
There are times when two sources of variability may be accounted for by blocking, and such a
situation may be described by a Latin square experimental design, shown diagrammatically
below”
B A C
C B A
A C B
Here, there are not only three horizontal blocks (rows), but also three vertical blocks (columns)
of the data. There are three levels of the fixed-effects factor (designated A, B and C), and they are
assigned to each row and column, except that each level is represented exactly once in each row
and each column. In the Latin square design, the number of vertical blocks, the number of
horizontal blocks, and the number of levels in the fixed effects factor must be equal.
The design shown above is referred to as a 3 x 3 Latin square and is the smallest possible Latin
square experimental design. Twelve different configurations are possible for a 3 x 3 Latin square
analysis; there being 576 possible 4 x 4 squares, 161,250 possible 5 x 5 squares, etc.
The only hypothesis generally of interest in a Latin square analysis is the one concerning
equality among the levels of the fixed-effects factor. We test this hypothesis by considering the
experimental design to be a three-factor ANOVA with one fixed and two random factors, with
no replication.
1
Lecture 8 Wilson Wesley Lazaro Jere
CSS 312: Design and Analysis of Experiments Second Semester 2010/11 Academic Year
Example. Hypothetical experiment and dataset
A. Hypothesis
H0: A = B = C = D
There is no need to set up null hypotheses of the rows and columns since these are
random effects.
B. Setting
= 0.05
C. Data
i. Column and Row Totals
Columns
1 2 3 4 Row Totals
Rows 1 A10 B10 C9 D7 36
2 B9 C9 D9 A9 36
3 C10 D10 A11 B9 40
4 D8 A12 B10 C8 38
Column Totals 37 41 39 33 150
ii. Treatment Totals and Treatment Means
Treatment A Treatment B Treatment C Treatment D
10 9 10 8
12 10 9 10
11 10 9 9
9 9 8 7
Treatment 42 38 36 34
Totals
Treatment 10.5 9.5 9.0 8.5
Means
2
Lecture 8 Wilson Wesley Lazaro Jere
CSS 312: Design and Analysis of Experiments Second Semester 2010/11 Academic Year
D. Correction Factor
Correction factor = = 1406.25
E. Sums of Squares (SS)
Total SS = 1428 – 1406.25 = 21.75
Treatment SS = = 1415 – 1406.25 = 8.75
Column SS = = 1415 – 1406.25 = 8.75
Row SS = = 1409 – 1406.25 = 2.75
Error SS = Total SS – Treatment SS – Column SS – Row SS
=21.75 – 8.75 – 8.75 – 2.75 = 1.5
F. Degrees of freedom
i. Total degrees of freedom = 16 – 1 = 15
ii. Treatment degrees of freedom =4–1 =3
iii. Column degrees of freedom =4–1 =3
iv. Row degrees of freedom =4–1 =3
v. Error degrees of freedom = 15 – 3 – 3 - 3 =6
3
Lecture 8 Wilson Wesley Lazaro Jere
CSS 312: Design and Analysis of Experiments Second Semester 2010/11 Academic Year
G. Mean Squares (MS)
i. Mean squares for Treatment = = 2.9167
ii. Mean squares for Columns = = 2.9167
iii. Mean squares for Rows = = 0.9167
iv. Mean squares for Error = = 0.25
H. Fcalculated
i. F for Treatment = = 11.67
ii. There is no need to calculate the F value for columns and rows since these are not going
to be tested.
I. ANOVA Table
Source Degrees of Sums of Square Mean Square Fcalculated
Freedom
Totals 15 21.75
Treatment 3 8.75 2.9167 11.67
Rows 3 8.75 2.9167
Columns 3 2.75 0.9167
Errors 6 1.5 0.25
Fcritical = F(0.05)(1),3,6 = 4.76
J. Decision
The calculated F for treatments is greater than the critical F. We reject Ho. We conclude that
there were significant differences in the mean response of the plants to the four hormonal
treatments.
4
Lecture 8 Wilson Wesley Lazaro Jere
