DESIGN _2

RANDOMIZED COMPLETE BLOCK DESIGN

(RCBD)
(with one factor)
Or
TWO-WAY ANOVA

Recap CRD
CRD all experimental units are

considered (ALMOST)
homogeneous.
randomization helps BALANCE or
AVERAGE the EXTRANEOUS
VARIATION you cannot see (not
under direct control of
experimenter)
No sources of variation other
than treatment effects are known.

V1

V3

V2

V1

V2

V2

V1

V3

V1

V3

V3

V2

EXPERIMENTAL UNITS
although there might exist minor
differences, can safely assume to
be almost homogeneous

Heterogeneity of experimental units

What happens if they are not

homogeneous?
Now, there is a known source
of additional variation
Have to take this effect into

consideration
BLOCKING can be used to
control the factors you can see
(more on blocking later)
dry

moist
SOIL

wet

Heterogeneity examples

You want to study the effects of

different vitamins (treatments) on

growth rate of mice. The weight after a
certain period after being fed with the
vitamin is taken.
Mice (experimental units) of the same
age may be of different weights (not
homogeneous) it is known or can be
anticipated that heavier animals may
exhibit different growth rate than
lighter animals

 You want to study the effects of different diets

(treatments) on lipid level in blood.

A group of volunteers are of different ages (not
homogeneous) it is known or anticipated that older
volunteers are more likely to have higher lipid level in
blood than younger volunteers.

You want to study the effect of different incentive pay

schemes (treatments) on employee productivity.

A group of employees (experimental units) with
differing manual dexterity (heterogeneous) are used
for the study. It is known or anticipated that
productivity is correlated with manual dexterity.

 So what to do if the experimental units are not

homogeneous? There are additional source of variation

(known beforehand or anticipated) like in our examples,
weights of mice

age of volunteers
manual dexterity

which may influence the outcome of the experiment

apart from the treatment effects you are investigating

Solution_1: Use of concomitant variable(s)

(Analysis of Covariance, ANCOVA)
Basic idea
Controlling error variance can be accomplished by means of
covariates or by experimental design.
The use of covariance to control error is to remove by
regression certain recognized effects that cannot be or have
not been controlled effectively by experimental design.
ANCOVA is a combination of regression and ANOVA
We discuss briefly on this topic.

Refer to our earlier example,

You want to study the effect of different vitamins
(treatments) on growth of mice. The weight after a certain
period fed with the vitamin is taken.
Mice (experimental units) of the same age may be of
different weights (not homogeneous) it is known or can be
anticipated that heavier animals may exhibit different
growth than lighter animals
Now if initial weight is correlated with the growth, then a

portion of the experimental error for growth can be the

result of differences in initial weight
By covariance analysis, this portion may be computed and
eliminated from experimental error

How covariance analysis reduces error variability

Final
weight

Vit A

Vit B

Vit C

Scatter around treatment means

large indicating large error

variance

3
2

mouse

Covariance analysis to reduce

Final
weight

error
Scatter around treatment
regression lines is much less
reflects smaller error variability
Initial weight

Suppose the experiment is carried out using CRD

Yij i X ij X ij
dependent
variable
final weight

treatment
different
vitamins

error
regression Y on X
or the covariate
Xij = initial weight

Solution_2 Random Complete Block

Design, RCBD
Block

Experimental units are sorted

into homogeneous groups or
blocks
Random
Treatments then assigned at
random WITHIN the blocks
(restricted randomization)
Complete
Each block contains ALL
treatments
Blocks are replications

Block1
10-15 g

Block 2
16-20 g

Block 3
21-25 g

Age
20-30

31-40

41-50

Objectives of Blocking the experimental units

Obj_1

To make the experimental units as

homogeneous as possible within blocks with
regard to the dependent variable

Obj_2

To account for the known (or anticipated)

source of variation (eg. weight, age, manual
dexterity, in our earlier examples)

Obj_3

To remove this source of variation from the

experimental error (reduce the experimental
error and make the experiment more
powerful)

Criteria for blocking

Characteristics associated with the experimental units
Persons: sex, age, income, intelligent, education, job

experience, attitude
Plants: maturity stage, disease level
Animals: weight, age, family or litter, breed
Characteristics associated with the experimental setting
Observer, batch of materials, shades, shelves, locality

Geographic areas: population size, average income

Advantages and Disadvantages

Advantages
With effective grouping, provides more precise results

than CRD of comparable size

Can accommodate any number of treatments and
replications
Statistical analysis relatively simple
Disadvantages

When variation among experimental units within block

is large, giving large error term. Occurs when number of

treatments is large, may not be possible to have
sufficient uniform units for blocks.
More assumptions required than CRD ( no
Block*Treatment interaction)

Analysis -Theory
Model

Yij = + i + j + ij
i = i-th block effect, i = 1,, r
j = j-th treatment effect, j = 1,, t and ij is random error

Yij = + (i. ) + (.j ) + (Yij - i. .j + )

(Yij ) = (i. ) + (.j ) + (Yij - i. .j + )
Estimators

i i . Yi . Y ..

Y..

j . j Y . j Y ..
Y

ij

ij

Y .. Yi . Y .. Y . j Y .. Yij Yi . Y . j Y ..

SSTO

ij

SSB

ij

SSTR

ij

SSE

Computational formulas
Source
Blocks

df
r-1

SS

Yi .

Y ..2

bt

t
Treatment

t-1

Y .

2
j

Y ..2

bt

Error

(r-1)(t-1)

by difference

Total

rt-1

ij

ij

Y ..

rt

VENTILATION
SYSTEM

RCBD worked
example

To investigate the effect of

variety (C, W, U) on
growth seedlings (mm).
The experimental setting
is not homogeneous, so
RCBD is used

Results
Layout
Block

BI
BII

C
C

W
U

U
W

BIII
BIV
BV

U
W
W

W
U
C

C
C
U

Block

14

II

14

24

III

16

32

IV

13

18

37

12

14

17

43

28

49

73

150

Calculation
CF = 1502/15
SSTO = (12 + 22 + .+ 172) CF
= 398.0
SSB = (142 + 242 + 322 + 372 +432)/3 CF
= 171.3
SSTR = (282 + 492 + 732)/5 CF
= 202.8
SSE = 398.0 171.3 202.8
=23.9

ANOVA table
Source

df

SS

MS

Block
Treatment
Error
Total

4
2
8
14

171.3
202.8
23.9
398.0

42.8
101.4
2.99

14.3
33.9

Test for Treatment

Ho: 1 = 2 = 3 = 0, Ha: not all j equal zero
F* = MSTR/MSE = 101.4/2.99 = 33.9 > F(0.95; 2,8) = 4.46 so conclude
means for the treatments differ (or treatment effect is significant.
Test for blocks
Ho: 1 = 2 = . = 0, Ha: not all i equal zero
F* = MSB/MSE = 42.8/2.99 =14.3 > F(0.95; 4,8) = 3.84 so conclude theres
evidence suggesting block effect is significantly different

RCBD
Source
Block
Treatment
Error
Total

df
4
2
8
14

SS
171.3
202.8
23.9
398.0

MS
42.8
101.4
2.99

CRD
Treatment
Error
Total

2
12
14

202.8
195.2
398.0

101.4
16.3

F
14.3*
33.9*
Reduction in Error
Variance

6.22*

Blocking as alternative to covariance analysis

A choice exists between;
i) a CRD with concomitant variable as covariate to reduce
experimental errors
ii) a RCBD with blocks formed from concomitant variable
Usually the later is preferred because;
Computation with RCBD is simpler than
covariance analysis
RCBD is free of assumptions about the nature of relationship
between blocking variable and the dependent variable while
covariance analysis assumes a definite form of relationship.
If regression is linear, RCBD and covariance analysis are equally
efficient. But if not linear, then covariance analysis with linear
relationship used with CRD is less effective than RCBD

Relative efficiency
dkB (MSB) + (dkT + dkE) MSE
MSE (CRD) = -----------------------------dkB + dkT + dkE

= [4(42.8) + 10 (2.99)] /14 = 14.36

Relative Efficiency
= ( f1 + 1)(f2 + 3) MSE(CRD)
---------------------------------(f2 +1)(f1 + 3) MSE (RCBD)
= (8 + 1) (12 + 3) 14.36
----------------------------------- = 646.42 / 427.57 = 1.5
(12 + 1) (8 + 3) 2.99

Meaning if CRD had been used, need about 6.5 replications

instead of 5 to obtain the same standard error of treatment
mean as RCBD