You are on page 1of 16

10/18/2016

The Latin Square Design


Text reference, Section 4.2, pg. 158
These designs are used to simultaneously control
(or eliminate) two sources of nuisance
variability
A significant assumption is that the three factors
(treatments, nuisance factors) do not interact
If this assumption is violated, the Latin square
design will not produce valid results
Latin squares are not used as much as the RCBD
in industrial experimentation
Chapter 4 Design & Analysis of Experiments 8E 2012 1
Montgomery

The Rocket Propellant Problem


A Latin Square Design

Response Variable
Burning Rate

This is a 5 5 Latin square design


Page 159 shows some other Latin squares
Table 4-13 (page 162) contains properties of Latin squares
Statistical analysis?
Chapter 4 Design & Analysis of Experiments 8E 2012 2
Montgomery

1
10/18/2016

Statistical Analysis of the


Latin Square Design
The statistical (effects) model is
i 1, 2,..., p

yijk i j k ijk j 1, 2,..., p
k 1, 2,..., p

The statistical analysis (ANOVA) is much like the


analysis for the RCBD.
See the ANOVA table, page 160 (Table 4.10)
The analysis for the rocket propellant example
follows
Chapter 4 Design & Analysis of Experiments 8E 2012 3
Montgomery

Chapter 4 Design & Analysis of Experiments 8E 2012 4


Montgomery

2
10/18/2016

Chapter 4 Design & Analysis of Experiments 8E 2012 5


Montgomery

Chapter 4 Design & Analysis of Experiments 8E 2012 6


Montgomery

3
10/18/2016

Chapter 4 Design & Analysis of Experiments 8E 2012 7


Montgomery

Chapter 4 8

4
10/18/2016

Chapter 4 9

Chapter 4 10

5
10/18/2016

Chapter 4 11

*Select one from a list of all


possible Latin Squares at random
Chapter 4 or follow this process 12

6
10/18/2016

Chapter 4 Design & Analysis of Experiments 8E 2012 13


Montgomery

Chapter 4 14

7
10/18/2016

Chapter 4 15

Graeco Latin Square Design

Chapter 4 Design & Analysis of Experiments 8E 2012 16


Montgomery

8
10/18/2016

Chapter 4 Design & Analysis of Experiments 8E 2012 17


Montgomery

Chapter 4 18

9
10/18/2016

Chapter 4 Design & Analysis of Experiments 8E 2012 19


Montgomery

Relative Efficiency of the RCBD


It is often helpful to estimate the relative efficiency of the RCBD compared to a
completely randomized design (CRD). One way to define this relative efficiency is

(df b 1)(df r 3) 2r
R
(df b 3)(df r 1) b2

where 2r and b2 are the experimental error variances of the completely randomized and
randomized block designs, respectively, and df r and df b are the corresponding error
degrees of freedom. This statistic may be viewed as the increase in replications that is
required if a CRD is used as compared to a RCBD if the two designs are to have the same
sensitivity. The ratio of degrees of freedom in R is an adjustment to reflect the different
number of error degrees of freedom in the two designs.
20

10
10/18/2016

To compute the relative efficiency, we must have estimates of 2r and b2 . We can use
the mean square for error MSE from the RCBD to estimate b2 , and it may be shown [see
Cochran and Cox (1957), pp. 112-114] that
(b 1) MS Blocks b(a 1) MS E
2r
ab 1
is an unbiased estimator of the error variance of a the CRD. To illustrate the procedure,
consider the data in Example 4.1. Since MSE = 7.33, we have
b2 7.33
and
(b 1) MS Blocks b(a 1) MS E
r2
ab 1
(5)38.45 6(3)7.33

4(6) 1
14.10
Chapter 4 21

(dfb 1)(df r 3) r2
R
(dfb 3)(df r 1) b2
(15 1)(20 3) 14.10

(15 3)(20 1) 7.33
1.87
This implies that we would have to use approximately twice times as many replicates
with a completely randomized design to obtain the same sensitivity as is obtained by
blocking on the metal coupons.

Chapter 4 22

11
10/18/2016

Clearly, blocking has paid off handsomely in this experiment. However, suppose that
blocking was not really necessary. In such cases, if experimenters choose to block, what
do they stand to lose? In general, the randomized complete block design has (a 1)(b - 1)
error degrees of freedom. If blocking was unnecessary and the experiment was run as a
completely randomized design with b replicates we would have had a(b - 1) degrees of
freedom for error. Thus, incorrectly blocking has cost a(b - 1) (a - 1)(b - 1) = b - 1
degrees of freedom for error, and the test on treatment means has been made less
sensitive needlessly. However, if block effects really are large, then the experimental
error may be so inflated that significant differences in treatment means could possibly
remain undetected. (Remember the incorrect analysis of Example 4-1.) As a general rule,
when the importance of block effects is in doubt, the experimenter should block and
gamble that the block means are different. If the experimenter is wrong, the slight loss in
error degrees of freedom will have little effect on the outcome as long as a moderate
number of degrees of freedom for error are available.
Chapter 4 23

Chapter 4 Design & Analysis of Experiments 8E 2012 24


Montgomery

12
10/18/2016

{(1,3,4), (1,2,3), (2,3,4), (1,2,4)}

25

Chapter 4 Design & Analysis of Experiments 8E 2012 26


Montgomery

13
10/18/2016

Chapter 4 Design & Analysis of Experiments 8E 2012 27


Montgomery

Chapter 4 Design & Analysis of Experiments 8E 2012 28


Montgomery

14
10/18/2016

Chapter 4 Design & Analysis of Experiments 8E 2012 29


Montgomery

Chapter 4 Design & Analysis of Experiments 8E 2012 30


Montgomery

15
10/18/2016

Chapter 4 Design & Analysis of Experiments 8E 2012 31


Montgomery

16

You might also like