Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Chapter Six

Nested and Split - Plot Designs

Contents

6. Nested and Split Plot Designs

6.1 The two - stage nested design

6.2 The split - plot design

6.3 The split - split plot design

Nested and Split - Plot Designs

“Statistics are like a bikini. What they reveal is suggestive, but what

they conceal is vital”. A. LEVENSTEIN

Under this title we introduce two important types of designs, the nested design and split-plot
design. Both of these designs are reasonably widespread application in the industrial use of
designing experiments.

One consistent feature of the designs discussed in the course so far is that the size of the
experimental units is the same for all the treatment factors. However, it is not always desirable or
possible to use the same size experimental units for all the treatment factors.

Moreover, in some experiments the levels of one factor (e.g. factor B) are similar but not
identical for different levels of another factor (e.g. factor A). This arrangement is call a nested or
“hierarchical” design. Therefore, this chapter, unlike the pervious chapters, deals with the
analysis of split plot and nested designs.

[email protected]
160
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

6.1 The two - stage nested designs

In certain multifactor experiments, the levels of one factor, say factor B, are similar but not
identical for different levels of another factor, say A. Such an arrangement is called a nested or
hierarchical, design, with the levels of factor B nested under the levels of factor A.

For the one who understand two-factor factorial design may ask the difference between
two - stages nested design and factorial design. The answer is quite straight forward. In case of
factorial design, each level of factor B is the same across different levels of A. But in case of
nested design, each level of factor B is not the same for each level of factor A.

For example consider a company that purchases its row material from three different suppliers.
The company wishes to determine if the purity of row materials is the same from each supplier.
There are four batches of row material available from each supplier, and three determinations of
purity are to be taken from each batch. The following picture indicating this concept:-

This is a two - stage nested design, with batch nested under suppliers.
Nested factors can be random or fixed, though they are usually random and often arise from
some kind of sub sampling. Accordingly, the hypothesis of interest will vary.
The layout of the data is then:-

[email protected]
161
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Statistical Analysis

The linear statistical model for the two stage nested design is:

That is there are a levels of factor A, b levels of factor B nested under each level of A, and n
replicates. The subscript j(i) indicates that the jth level of B is nested under the ith level of factor
A. it is convenient to think of the replicates as being nested within the combination of levels of A
and B; thus the subscript (ij)k is used for the error terms. This is a balanced nested design
because there are an equal number of replicates. Because every level of factor B does not appears
with every level of factor A, there can be no interaction between A and B. we may write the total
corrected sum of squares as

∑∑∑ ̅ ∑∑∑ ̅ ̅ (̅ ̅ ) ̅

Expanding the right hand equation it yields

∑∑∑ ̅ ∑ ̅ ̅ ∑∑ ̅ ̅ ∑∑∑ ̅

Because the three cross product terms are zero. The above equation indicated that the total sum
of squares can be partitioned in to sum of squares due to A, a sum of squares due to factor B
under the levels of A, and a sum of squared due to error. Symbolically we may write:

With degrees of freedom the sum of squares is:






[email protected]
162
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

As usual, let us consider only the balanced nested design. This is called balanced because there
are an equal number of replications.
Note that: Since every level of factor B doesn’t appear with every level of factor A, there can be
no interaction between A and B. Therefore, the model of nested design never contains an
interaction term.

To handle the appropriate hypothesis, we have to know whether the factors are random or fixed.
If factors A and B are fixed, we assume that ∑ = 0 and ∑ =0( ). That
is, the A treatment effects sum to zero and the B treatment effects sum to zero within each level
of A. Alternatively, if A and B are random, we assume that
( ) . Mixed models with A fixed and B random are also
encountered.
Case 1: If A and B are fixed, the hypotheses of interest are:
:

This null hypothesis can be tested by

o :

This null hypothesis can be tested by

Case 2:At last, if both A and B are random factors, then:

This null hypothesis can be tested by

[email protected]
163
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

This null hypothesis can be tested by

Case 3:If A is fixed factor and B is random, we test

This null hypothesis can be tested by

This null hypothesis can be tested by

∑∑∑ ∑

∑∑ ∑ ∑∑∑ ∑∑

The Anova table seems like the following:

Source of variation Sum of squares DF Mean square

A a-1

B(A) a(b-1)

Error ab(n-1)

Total

NB: The F - ratio is computed based on the nature of factors.

[email protected]
164
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

When there is unbalanced replication, then:

∑∑∑ ∑
∑ ∑

∑∑ ∑ ∑∑∑ ∑∑

Example:

Consider a company that buys raw material in batches from three different

suppliers. The purity of this raw material varies considerably, which causes

problems in manufacturing the finished product. We wish to determine if the

variability in purity is attributable to differences between the suppliers. Four

batches of raw material are selected at random from each supplier, and three

determinations of purity are made on each batch. This is, of course, a two –

stage nested design. The data are as follows:

Test the signicance of suppliers and batches at 5% level of significance? and

make anova table.

Case 3 is appropriate for this example because, suppliers (A) are fixed and
Batches (B) are random.

[email protected]
165
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Solution: a = 3 , b = 4 , n = 3 . abn = 3*4*3 = 36

= = = = 7.53 = = 0.97

= = = = 7.77 = = 2.94

= = = = 2.64

[email protected]
166
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Since, suppliers are fixed and batches are random, then the hypothesis of

interest is

Step 1: State the null hypothesis and alternative hypothesis.

i, for supplier

: vs :
ii, for batches

: vs :

Step 2: Identify the level of significance, .

Step 3: Identify the test statistic, F – test.

Step 4: Determine test statistic.

= = 2.94

Step 5: Identify the critical value

4.26
2.30
Step 6: Make decision:

o Reject if > , for supplier

0.97< 4.26, so Don’t reject

o Reject if > , for batch

2.94 > 2.30, so reject

Step 7: Conclusion:

At 5% level of significance we have enough evidence to conclude that

there is no significant effect on purity due to suppliers.

At 5% level of significance we have enough evidence to conclude that the

purity of batches of raw material from the same supplier does differ

significantly.

[email protected]
167
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

The ANOVA table becomes:

The Anova table seems like the following:

Source of variation Sum of squares DF Mean square

Suppliers(A) a-1

Batches(B(A)) a(b-1)

Error ab(n-1)

Total

Source of variation Sum of squares DF Mean square

Suppliers(A) 15.06 2 7.53 0.97

Batches(B(A)) 69.92 9 7.77 2.94
Error 63.33 24 2.64
Total 148.31

[email protected]
168
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

6.2 The split plot design

In some multifactor factorial experiments we may be unable to completely randomize the order
of the runs. This often results in a generalization of the factorial design called a split-plot design.

Split-plot designs are needed when the levels of some treatment factors are more difficult to
change during the experiment than those of others. It has three main parts: Whole plot, sub (split)
plot parts and Block (plot) and have a nested blocking structure: split plots are nested within
whole plots, whole plots are nested within blocks. It were originally used in agricultural
experiments and represent a randomized complete block design, with one more factors applied to
experimental units within each block.

An important strategy in agricultural and laboratory experiments of studying two different

experimental factors is to employ the so called split - plot design. For reasons of convenience or
necessity, the levels of one factor (the so called sub - plot treatments) are applied to the ultimate,
single experimental units and the levels of the other factor (the so called main-plot treatments)
are applied to set of ultimate units.

Notice that the two factors can essentially be applied at different time. As a result, a split-plot
design can be viewed as two experiments combined or superimposed on each other. One
experiment has the whole-plot factor applied to large experimental units (or it is a factor whose
levels are hard to change) and the other experiment has subplot factor applied to the smaller
experimental units (or it is a factor whose levels are easy to change).

In other words, the split-plot design is specifically suited for a two factor experiment that has
more treatments that can be accommodated by a complete block design. In this design, one of the
factors is assigned to the main-plot (whole-plot). The assigned factor is called the main-plot
factor. The main-plot is divided into subplots (split-plots) to which the second factor, called the
subplot factor, is assigned. Thus, each main - plot becomes a block for the subplot treatments
(i.e., the levels of the subplot factor).

[email protected]
169
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

With a split - plot design, the precision for the measurement of the effect of the main-plot factor
is sacrificed to improve that of the subplot factors. In this design, both the procedures for
randomization and that for analysis of variance are accomplished in two stages - one of the
main-plot level and another on the subplot level. At each level, the procedure of randomized
complete block design, as described in chapter 4, are applicable.

The typical layout of split plot design for r replicate, a factor and b block.

Block 1 Block 2 … Block r

Factor A 1 2 … a 1 2 … a 1 2 … a
1
2
Factor B

Split-plot design can be considered as two superimposed blocked designs, assume

two factors:

A: whole-plot factor (a level); B: sub-plot factor (b level), r replicates

– RCBD A: number of treatment = a, number of block = r.

– RCBD B: number of treatment = b, number of block = ra.

For whole-plots, subdivisions to smaller sub-plots are ignored. For sub-plots,

whole-plots considered blocks.

The statistical linear model for split-plot design is

Where and represent the whole-plot and correspond respectively to blocks (or
replicates), main-plot (whole-plot) treatment (factor A), and whole-plot error (replicate (or
block) × A ); and , , and represent the subplots and correspond
respectively to the subplot treatments (factor B), the replicates (or blocks)×B and AB

[email protected]
170
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

interactions, and the subplot error (block × AB). Note that the whole plot error is the replicate (or
block) × A interaction and the subplot interaction is the three - factor interaction block × AB.

The sums of squares for these factors are computed as in the three-way analysis of variance
without interaction. What makes a split-plot design different form other designs with factorial
treatment structure? Here are three ways to think about what makes the split-plot design
different.

First, the split-plot has two sizes of units and two separate randomizations. Whole - plots act as
experimental units for one randomization and other randomization assigns levels of the split -
plot factor variety to split-plots. In this randomization, split-plots act as experimental units, and
whole - plots act as block for the split-plots. In general, there are two types of separate
randomizations, with two different kinds of units that can be identified before randomization
starts.

Second, a split-plot randomization can be done in one stage, assigning factor - level
combinations to split-plots in any whole - plot get the same level of the whole-plot factor and no
two split-plots in the same whole-plot get the same level of the split-plot factor. Thus, a split-plot
design has a restricted randomization.

Third, a split-plot is a factorial design in incomplete blocks with one main effect confounded
with blocks. The whole-plots are the incomplete blocks, and the whole block factor is
confounded with block

Note that: You can’t distinguish a split-plot design from some other design simply by looking at
a table of factor levels and responses. You must know how the randomization was done. We
have also been speaking as if the whole-plot randomization was done first; this is often true, but
is not required.

[email protected]
171
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Then, the sums of squares are

∑∑∑ ∑

∑ ∑

∑∑

∑∑

∑∑

, SS = sum of squares.
o main plot = whole plot, subplot = split plot

The ANOVA table is then

Source of Variation Sum square DF Mean

Square
Replication(Blocks) (r-1)
whole plot factor(A) (a-1) /
whole plot error( A) (r-1)(a-1)
Subplot factor (B) (b-1) /

Replicates (or blocks) (r-1)(b-1)

AB (a-1)(b-1) /

Subplot(split plot) error (a-1)(r-1)(b-1)

(replicates(or blocks) AB)
Total rab-1

[email protected]
172
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Example: As an example consider a paper manufacturer who is interested in three different

pulp preparation method and four different cooking temperatures for the pulp and who wishes to
study the effect of these two factors on the tensile strength of the paper. Each replicate of a
factorial experiment requires 12 observations, and the experimenter has decided to run three
replicates .However , the pilot plant is only capable of making 12 runs per day , so the
experimenter decides to run one replicate on each of the three days and to consider the days or
replicates as blocks.
On any day, he conducts the experiment as follows. A batch of pulp is produced by one of the
three methods under study .Then this batch is divided into four samples, and each sample is
cooked at one of the four temperatures. Then a second batch of pulp is made up using another of
the three methods .This second batch is also divided into four samples that are tested at the four
temperatures. The process is then repeated, using a batch of pulp produced by the third method.
Initially, we might consider this to be a factorial experiment with three levels of preparation
method (factor A) and four levels of temperature (factor B) in a randomized block.
The split plot design is consider. Each replicate or block in the split-plot design is divided into
three parts called whole plots, and the preparation methods are called the whole plot or main
treatments. Each whole plot is divided in to four parts called subplots (or split-plot), and one
temperature is assigned to each.

The data are shown below.

IS there significant effect on pulp preparation method (A), Temperature (B)

and their interaction at 5% level of significance? and perform anova table.

[email protected]
173
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Solution:

r = 3, a = 3, b = 4, rab = 3*3*4 = 36 ,

∑∑∑ ( )

∑∑

∑∑

[email protected]
174
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

∑∑

= = = = 38.78 = = 7.08

= = = = 64.20

= = = = 9.07 = = 41.94

= = = 3.45

= = = 12.53 = = 2.96

= = = 4.24

Step 1: State the null hypothesis and alternative hypothesis.

i, for factor A

: vs :
i, for factor B

: vs :
ii, for interaction(AB)

: vs :
Step 2: Identify the level of significance, .

Step 3: Identify the test statistic, F – test.

[email protected]
175
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Step 4: Determine test statistic.

= = 7.08 for factor A

= = 41.94 for factor B

= = 2.96 for interaction (AB)

Step 5: Identify the critical value

for factor A
for factor B

for interaction (AB)

Step 6: Make decision:

o Reject if > , for factor A

7.08> 6.94, so reject

o Reject if > , for factor B

41.94 > 24.76, so reject

o Reject if > , for interaction

2.96 < 3.00, so Don’t reject

Step 7: Conclusion:

At 5% level of significance we have enough evidence to conclude that

there is significant effect on preparation method (there is sufficient

evidence to conclude there is a significance difference in preparation

methods).

At 5% level of significance we have enough evidence to conclude that the

temperature is statistically significant (we conclude that these four

temperature levels have different effects on the paper manufacturer.).

At 5% level of significance we have enough evidence to conclude that the

preparation method and temperature (interaction) is not significant (we

conclude that the interaction is negligible.).

[email protected]
176
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

The ANOVA table looks like the following

Source of Variation Sum square DF Mean

Square
Replication(Blocks) (r-1)
Preparation method(A) (a-1) /
whole plot error( A) (r-1)(a-1)
Temperature (B) (b-1) /

Replicates (or blocks) (r-1)(b-1)

AB (a-1)(b-1) /

Subplot(split plot) error (a-1)(r-1)(b-1)

(replicates(or blocks) AB)
Total rab-1

Source of Variation Sum square DF Mean Square

replication(Blocks) 77.55 2 38.78
Preparation method(A) 128.39 2 64.20 7.08
whole plot error( A) 36.28 4 9.07
Temperature (B) 434.08 3 144.69

Replicates (or blocks) 20.67 6 3.45

AB 75.17 6 12.53

Subplot(split plot) error 50.83 12 4.24

(replicates(or blocks) AB)
Total 822.97 35

[email protected]
177
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

6.3 The split - split plot design

The concept of split-plot designs can be extended to situations in which randomization
restrictions may occur at any number of levels within the experiment. If there are two levels of
randomization restriction, the layout is called a split-split-plot design. In other words, what we
have split once, we can split again or split the split-plot is called split – split plot design.

Consider an experiment with three factors. The levels of factor are assigned at random to
whole-plots each (total of whole-plots). Each whole-plot is split into split-plots. The levels
of factor are assigned at random to split-plots, using whole plots as block. So far, this is just
like a split-plot design. Now, each split-plot is divided into split-plots (sub-subplots), and the
levels of factor are randomly assigned to split-split-plots using split-plots as blocks.

The linear statistical model for the split – split plot design is

Where represent the whole – plot and correspond to replicates or blocks, main
plot treatment (factor A), and whole – plot error (replicate(or block) ),respectively ; and ,

represent the subplot and correspond to the subplot treatments

(factor B) , the replicate (or block ) and AB interactions , and the subplot error respectively;
are correspond to sub – sub plot.
and the remaining are interactions. The four-
factor interaction is called the sub - sub plot error.
The statistical analysis of a split-split plot design is like that of a single replicate of a four-factor
factorial. The numbers of degrees of freedom for each test are determined in the usual manner.

[email protected]
178
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

Example:

A researcher is studying the absorption times of a particular type of antibiotic

capsule. There are three technicians, three dosage strengths, and four capsule

wall thicknesses of interest to the researcher. Each replicate of a factorial

experiment would require 36 observations. The experimenter has decided on

four replicates, and it is necessary to run each replicate on a different day

.Note that the days can be considered as blocks. Within a replicate

(or a block) (day), the experiment is performed by assigning a unit of antibiotic

to a technician who conducts the experiment on the three dosage strengths and

the four wall thicknesses. Once particular dosage strength is formulated, all

four wall thicknesses are tested. Finally, the third dosage strength and the four

wall thicknesses are tested. Meanwhile, two other laboratory technicians also

follow this plan, each starting with a unit of antibiotic.

Note that there are two randomization restrictions within each replicate (or

block): technician and dosage strength. The whole plots correspond to the

technician .The order in which the technicians are assigned, the units of

antibiotic is randomly determined. The dosage strengths from three subplots.

Dosage strength may be randomly assigned to a subplot. Finally, within

particular dosage strength, the four capsule wall thicknesses are tested in

random order, forming four sub-subplots. The wall thicknesses are usually

called sub – sub treatments. Because there are two randomization restrictions

in the experiment (some authors say two “splits” in the design), the design is

called a split-split-plot design.

[email protected]
179
Lecture notes for Design and analysis of experiments (Stat 2043) Chapter - 6

 See the figure below:

[email protected]
180