Chapter 15.

3 - Randomized Complete Block Design

Assignment: Read section 15.3. Homework Problems 15.9 and 15.10.

Introduction:
• The Completely Randomized Design (CRD) is best to use when we have homogeneous
experimental units.

• If the experimental units are widely dispersed over time and/or space, we might expect
them to be heterogeneous.

• If we use a CRD on heterogeneous units, then

1. if we get a “bad” randomization we could get spurious results,

2. experimental error could be inflated.

• The solution is to use a Randomized Complete Block Design (RCBD).

• Blocks will be used to control for other sources of variation.

RCBD - defined

1. Experimental units are partitioned into b blocks each containing at least a number of
experimental units (where a is the number of treatments).

2. Within each block, treatments are randomly assigned to units belonging to that block.

3. Blocks should be chosen so that,

• units within blocks are as uniform as possible.

• differences between blocks should be as large as possible.

Case 1: RCBD w/ one (of each) treatment per block

Notation:
a = number of treatments
b = number of blocks
N = a × b, which equals the total number of observations
yij = response of treatment i for block j
yi. = treatment total bj=1 yij
P

ȳi. = yi. /b = treatment mean

y.j = ai=1 yij = block total
P

ȳ.j = y.j /a = block mean

y.. = ai=1 bj=1 yij = grand total
P P

ȳ.. = y.. /N = grand mean

1
Model:
yij = µ + βj + τi + ij , for i = 1, . . . , a, j = 1, . . . , b.
where µ is the overall mean, βj is the jth block effect, and τi is the ith treatment effect, and
ij is the experimental error.
Model Assumptions:

1. E[ij ] = 0

2. Var( ij ) = σ 2 , i.e. constant variance

3. ij ’s are independent.

4. ij ’s are normally distributed.

Pa Pb
5. i=1 τi = 0 and j=1 βj = 0

6. No treatment × block interaction. (i.e. the effect of the treatment does not depend on
the block.)

ANOVA

Source d.f. SS MS F

1 Pb SSblock
Blocks b−1 a j=1 y.j2 − y..2 /(ab) b−1

1 Pa SStreat M Streat
Treatment a−1 b i=1 yi.2 − y..2 /(ab) a−1 M SE

SSE
Error (a − 1)(b − 1) SStotal − SSblock − SStreat (a−1)(b−1)

Pa Pb
Total ab − 1 i=1 j=1 yij2 − y..2 /(ab)

To test the hypothesis

H0 : µ1 = µ2 = . . . = µa Vs. At least one of the means differs

we use
M Streat
F = ∼F df 1 = a−1, df 2 = (a−1)(b−1)
M SE
and reject H0 if F > Fα, df 1 = a−1, df 2 = (a−1)(b−1) .

Treatment Means:

2
µ̂i = ȳi.
var( ȳi. ) = σ 2 /b
var(ȳ
ˆ i. ) = M Serror /b
A (1 − α) 100 % confidence interval for µi is
q
ȳi. ± tα/2; d.f. = (a−1)(b−1) var(ȳ
ˆ i. )

Example: Redwing Flaxseed (Steel and Torrie, p. 136)

A biologist wants to determine the influence of time of inoculation with Septoria linicola (a
fungus) on the oil content of Redwing Flaxseed.
Treatments:
A - seedling
B - early bloom
C - mid bloom
D - late bloom
E - ripening
F - control
Response: Percent oil content
Also, of interest to the researcher is whether there are differences in oil content from
1. Inoculation Vs. None
2. Time of inoculation - Seedlings Vs Blooming/ripe
3. Time of inoculation - Blooming Vs Ripening
4. Time of inoculation - Within Blooming
Experimental Design: Randomized Complete Block Design

Block 1 Block 2 Block 3 Block 4

F D F D A D D E
36.4 36.8 37.3 36.6 36.0 37.0 36.4 37.1
E A C E E B F B
36.3 34.4 34.0 34.9 35.9 34.9 36.7 37.1
C B A B F C C A
34.4 33.3 35.9 31.9 37.7 34.5 33.1 34.1

Contrasts:
Treatments
Contrast A B C D E F
1. Control Vs Inoculants 1 1 1 1 1 -5
2. Seed Vs Bloom/ripe 4 -1 -1 -1 -1 0
3. Bloom Vs Ripening 0 1 1 1 -3 0
4. Within Bloom 0 1 -1 0 0 0
5. “ ” 0 1 1 -2 0 0

3
SAS Results:

ANOVA

Source d.f. SS MS F
Blocks 3 3.14 1.05
Treatment 5 31.65 6.33 4.82**
Ctrl Vs Inoc 1 10.74 10.74 8.17*
Sdlng Vs B&R 1 .08 .08 0.06
Blm Vs Rpn 1 3.31 3.31 2.52
w/i Blm 2 17.52 8.76 6.66**
Error 15 19.72 1.31
Total 23 54.51

Conclusions: Combine the results of the hypothesis testing with observing the treatment
means. Doing so we conclude that
1. Fungus reduces oil content of redwing flaxseed
2. Inoculation during early to midbloom reduces oil content more than by inoculation during
late bloom.
Case 2: RCBD w/ replicated treatments
In some cases, we may have an interest in interaction between the treatments and blocks. In
the example below, we have four blocks. There are six experimental units within each block.
We have three treatments. We will randomly assign two of each treatment to each block.

Block 1 Block 2 Block 3 Block 4

C C B C A C C B
B A C A B B A B
A B A B A C C A

Notation:
a = number of treatments
b = number of blocks
r = number of replicates of each treatment in each block
yijk = response for replicate k of treatment i in block j

Model:
yijk = µ + βj + τi + γij + ijk
where µ is the overall mean, βj is the jth block effect, τi is the ith treatment effect, γij is the
interaction effect from the ith treatment and jth block, and ijk is the experimental error.
We have the usual assumptions of normality of the error terms with homogeneous variance.
For model estimation, we have zero sum constraints on βj , τi , and γij .

4
 ANOVA

Source d.f. SS MS F

SSblock
Blocks b−1 SSblock b−1

SStreat M Streat
Treatment a−1 SStreat a−1 M SE

SSB×T M SB×T
B×T (a − 1)(b − 1) SSB×T (a−1)(b−1) M SE

SSE
Error ab(r − 1) SSerror ab(r−1)

Total abr − 1

To test the hypothesis

H0 : µ1 = µ2 = . . . = µa Vs. At least one of the means differs

we use
M Streat
F = ∼F df 1 = a−1, df 2 = ab(r−1)
M SE
and reject H0 if F > Fα, df 1 = a−1, df 2 = ab(r−1) .

To test the hypothesis

H0 : No Treat by Block Interaction Vs. Ha : Treatment by Block Interaction

we use
M SB×T
F = ∼F df 1 = (a−1)(b−1), df 2 = ab(r−1)
M SE
and reject H0 if F > Fα, df 1 = (a−1)(b−1), df 2 = ab(r−1) .

Treatment Means:
µ̂i = ȳi..
var( ȳi.. ) = σ 2 /(br)
var(ȳ
ˆ i.. ) = M Serror /(br)
A (1 − α) 100 % confidence interval for µi is
q
ȳi.. ± tα/2; d.f. = ab(r−1) var(ȳ
ˆ i.. )

5
Example: Wheat
An agronomist wants to compare the yields of three varieties of wheat. A field is divided
into 4 blocks, each containing 6 plots. Each variety is randomly applied to two plots in each
block.
Experimental Design: Randomized Complete Block Design with Replication of Treatments
within Blocks.

Block 1 Block 2 Block 3 Block 4

A A A B C B B C
27 21 17 16 15 26 18 11
C B A C A C A B
32 31 13 9 21 21 19 16
B C C B A B A C
33 36 5 18 17 26 17 15

Questions of interest to the researcher.

1. Does variety effect yield? If so, explain the relationship.
2. Is there interaction between variety and block? If so, explain the relationship.
SAS Results:

ANOVA

Source d.f. SS MS F
Blocks 3 996 332
Treatment 2 112 56 8.00**
Interaction 6 216 36 5.14**
Error 12 84 7
Total 23 1408

Conclusions:
1. Yield is effected by variety. Examining the means one can see that variety B has the
highest yield with variety A and C about the same.
2. There is interaction present between variety and blocks. Plot the treatment by block
means to describe the interaction.
3. Note from the SAS output that blocks are significant. If blocks had not been significant
then this indicates that there is no need to use blocks in the design.

6
Case 3: RCBD w/ subsampling
Notation:
a = number of treatments
b = number of blocks
r = number of subsamples in each experimental unit
yijk = response for subsample k of treatment i in block j

Model:
yijk = µ + βj + τi + ijk
where µ is the overall mean, βj is the jth block effect, τi is the ith treatment effect, and ijk
is the experimental error.

ANOVA

Source d.f. SS MS F

SSblock
Blocks b−1 SSblock b−1

SStreat M Streat
Treatment a−1 SStreat a−1 M SE

SSerror
Error (a − 1)(b − 1) SSerror (a−1)(b−1)

SSsub
Subsamples ab(r − 1) SSsub ab(r−1)

Total abr − 1

To test the hypothesis

H0 : µ1 = µ2 = . . . = µa Vs. At least one of the means differs

we use
M Streat
F = ∼F df 1 = a−1, df 2 = (a−1)(b−1)
M SE
and reject H0 if F > Fα, df 1 = a−1, df 2 = (a−1)(b−1) .

Treatment Means:
µ̂i = ȳi..
var( ȳi.. ) = σ 2 /(br)
var(ȳ
ˆ i.. ) = M Serror /(br)
A (1 − α) 100 % confidence interval for µi is
q
ȳi.. ± tα/2; d.f. = (a−1)(b−1) var(ȳ
ˆ i.. )

7
Example: Wireworms (Snedecor and Cochran 1980)
An agronomist wishes to determine the efficacy of two fumigants C and S for controlling
wireworms. A field is divided into 5 blocks, each containing 3 plots. Each of the three
treatments fumigant C, fumigant S, and the control 0 were randomly assigned to one plot
within each block. In each experimental plot, the number of wireworms was counted in each
of four subsamples.
Experimental Design: Randomized Complete Block Design with Subsampling

Block 1 Block 2 Block 3 Block 4 Block 5

S O S S C
5, 5, 1, 2 7, 4, 4, 5 2, 9, 3, 7 6, 4, 8, 4 4, 9, 8, 6
O C O C S
12, 20, 8, 8 0, 9, 3, 3 9, 6, 7, 11 7, 3, 5, 12 2, 9, 7, 3
C S C O O
5, 4, 5, 2 6, 4, 5, 4 4, 4, 3, 9 12, 22, 17, 13 7, 8, 5, 9

The researcher is interested in the following questions:

1. Is there a difference in wireworm means among the treatments?
2. If there is a difference, is there a difference between the control and the fumicants?
3. Is there a difference between the fumicants C and S?
SAS results:
Full Analysis - See SAS program.

ANOVA

Source d.f. SS MS F
Blocks 4 151.2 38.0
Treatment 2 293.4 146.7 05.98*
Control Vs Other 1 291.4 291.4 11.88**
C Vs S 1 2.0 2.0 00.08
Error = Block*Treat 8 196.2 24.5
Subsamples (w/i plots) 45 409.8 9.1
Total 59 1050.6

8
Analysis of Plot Means - See SAS program.

ANOVA

Source d.f. SS MS F
Blocks 4 37.8 9.4
Treatment 2 73.4 36.7 05.98*
Control Vs Other 1 72.9 78.9 11.88**
C Vs S 1 .5 .5 00.08
Error 8 49.1 6.1
Total 14 160.2

Conclusions:
1.

2.

3.