The basic concepts of experimental design,

and completely randomized design

This section discusses the basic concepts of experimental design, and
completely randomized design:
. The entire topic is divided into the following subdivisions.
1. Terminologies
2. Principles of Experimentation
3. Completely randomised design (CRD)
4. Least square estimators of the parameters
5. Statistical Analysis
6. ANOVA Table
7. Advantages and disadvantages of CRD
8. Example
9. Conclusion
1. Terminologies
Experiment:
An experiment is a device to obtain answers to some scientific query. It Is a
process or study that results in the collection of data. The results of experiments
are not known in advance .In comparative experiment we compare the effects of
two or more factors on some population characteristics, e.g. comparison of
different varieties of crops, different fertilizers in a agricultural experiment
,different medicines in a medical expt.. etc.

Experimental design: Experimental design is the process of planning a study

to meet the specified objectives. Planning an experiment properly is very
important in order to ensure that the right type of data and a sufficient sample
size and power are available to answer the research questions of interest as
clearly and efficiently as possible.
 Experiment units: The smallest subdivision of the experimental material to
which the treatments are applied and on which the variable under study is
measured is called an experimental units. Thus in an agricultural field
experiment, the plot of land on which a treatment is applied is an experimental
unit. In a feeding experiment of animals, an animal is an experimental unit.

Treatment: Various objects of comparison in a comparative experiment are

called treatments. For eg. In an agricultural experiment, different fertilizers or
different varieties of crop are treatments.

Block- Experimental units are collected together to form a relatively

homogeneous group. This group is called a Block. A block is also a replicate.

Experimental error: Even when same treatment is applied to different

experimental units, the result will vary. A part of this variation is systematic and
can be ascribe to different known sources. The other unexplained part of
variation is called experimental error and includes all extraneous variations due
to inherent variability in the experimental units, errors of measurements and lack
of representativeness of the sample to the population of interest.
2. Principles of Experimentation: The experimental design involves the three
basic principles viz., randomisation, replication and local control
Randomization. The first principle of an experimental design is randomization,
which is a random process of assigning treatments to the experimental units.
The random process implies that every possible allotment of treatments has the
same probability. The purpose of randomization is to remove bias and other
sources of extraneous variation, which are not controllable. Hence the
treatments must be assigned at random to the experimental units
Replication. The second principle of an experimental design is replication;
which is a repetition of the basic experiment. In other words, it is a complete run
of all the treatments to be tested in the experiment. In all experiments, some
 variation is introduced because of the fact that the experimental units such as
individuals or plots of land in agricultural experiments cannot be physically
identical. This type of variation can be removed by using a number of trials. We
therefore perform the experiment more than once, i.e., we repeat the basic
experiment. An individual repetition is called a replicate. The number, the shape
and the size of replicates depend upon the nature of the experimental material.

Local Control. The term local control, referring to the amount of balancing,
blocking and grouping of the experimental units in to number of homogeneous
sub plots. Balancing means that the treatments should be assigned to the
experimental units in such a way that the result is a balanced arrangement of the
treatments. Blocking means that like experimental units should be collected
together to form a relatively homogeneous group.. The main purpose of the
principle of local control is to increase the efficiency of an experimental design
by decreasing the experimental error. It has been observed that all extraneous
sources of variation are not removed by randomization and replication. In other
words, we need to choose a design in such a manner that all extraneous
sources of variation are brought under control. For this purpose, Local Control
is implemented..

3. Completely randomised designs (CRD)

This is the simplest type of design, based on principles of randomisation and

replication. A total of N experimental units are available for use in the
experiment. These experimental units are as homogeneous as possible; that is
no source of variation can be recognized among them under any grouping or
arrangement. Suppose that we have p treatments whose effect on the response
has to be investigated. The experimental plan is subdivide the N plots randomly
in to p parts. Such that jth part consists of nj plots. ( ∑nj= N ). The first treatment is
allocated to the first set of n1 plots, the second treatment to the second set of n2
plots,, etc. The arrangement of N plots in to groups n1, n2,…, np plots is done in
 a completely random manner. Hence the design is called completely
randomised design.

The CRD utilises the principles of randomisations and replication in the

following way:
Randomisation: ni experimental units are selected at random and ith treatment
is allocated to these experimental units
Replication: Since each treatment appears ni times.

The data can be tabulated as follows

Treatment Observations Totals Averages
1 y11 y12 … y1n1 y1. y1.
2 y21 y22 … y2n2 y2. y2.
. . . … . . .
.. . . … . . .
. . . … . . .
p yp1 yp2 … ypnp yp. y p.

Statistical Analysis is similar to that of one-way classified data,

Represent Yij as the jth observation taken from treatment i. There will be in
general ni observations under the ith treatment.
Model for the data.: We define the model
Yij = μ + αi+ €ij, i=1,2…p, j=1,2,…ni (4)
Where µ general effect
αi – ith treatment effect; ϵij- error term
4. Least square estimators of the parameters
The parameters µ and αi are estimated by the method of least squares. i.e. by
p ni 2 p ni

minimising error sum of squares. L =  

i 1 j 1
ij =  (y
i 1 j 1
ij     i )2
i.e. choose the values of µ and αi say ˆ and ˆ i that minimises L.

Which is obtained by solving p+1 simultaneous equations

L L
0  0 i= 1,2…p
 and  i , -

Differentiating with respect to µ and αi and equating to zero, we obtain p+1

equations called as normal equations as follows,
L
0

p p ni
N   nii   yij
i 1 i 1 j 1

And
L
0
 i

ni
ni   ni i   yij i= 1,2 …p
j 1

These normal equations are not linearly independent, as first equation is equal
to the sum of rest p equations. Hence no unique solution exists for µ and αi, i=
1,2..p. Since we have defined the treatment effects as deviations from overall
p

mean, we add a independent constraint,  n

i 1
i i 0 and solve the

simultaneous normal equations. Solving we get the solutions as

ˆ  y.. and ˆ i  yi.  y.. i=1,2..p

The fitted model after substituting the estimates ˆ and ˆ i in the linear

model , we get

Yij= ̂ + ̂ i + ϵij

Or

Yij= y.. + ( yi.  y.. )+ ( y ij  y i . )

Or
Yij- y.. = ( y i.  y.. )+ ( y ij  y i . ), the error term is chosen that both sides are

balanced
Squaring both sides and summing over all the observations we get
p ni p

  ( yij  y.. )   ni ( yi .  y.. ) 2 + 

p ni


2
( yi j  yi. )2 , the cross
i 1 j 1 i 1 i 1 j 1

product vanishes
Or
SST = SSTR + SSE
Where
p ni p ni
y..2
SST =  (y
i 1 j 1
ij  y.. ) 2
= 
i 1

j 1
y 
2
ij
N ( simplified formula)

p p
yi2. y..2
SSTR = n (y
i 1
i i.  y.. ) 2
= 
i 1 ni

N
p ni

And SSE =  
( yij  yi .. ) 2 computed as SSE = SST-SSTR
i 1 j 1
 Thus total corrected sum of squares can be partitioned into a sum of squares of
the differences between the treatment averages and the grand average
(SSTR),plus a sum of squares of the differences of the observations within
treatments averages(SSE)

5. Statistical Analysis:
We now investigate a formal test of the hypothesis
H0: α1= α2=…=αp=0
Against the alternative
H1: at least one αi≠ αj, for all i,j
We have assumed that the errors ϵij are normally and independently distributes
with mean zero and variance σ2. The observations yij are normally and
independently distributed with mean µ + αi and variance σ2. Thus SST is a
sum of squares in normally distributed random variables hence can be shown
that SST/σ2 is distributes as chi-square with N-1 degrees of freedom. Further we
can show that SSE/σ2 is chi-square N-p degrees of freedom and that SSTR /σ2
is chi - square variate with p-1 degrees of freedom if the null hypothesis i.e. H0 :
αi = 0 is true. It also implies that SSTR/ σ2 and SSE/σ2 independently distributed
chi-square random variables. Therefore if the null hypothesis is true, the ratio

SSTR /( p  1) MSSTR
F   (1)
SSE /( N  p) MSSE
is distributed as F distribution with p-1 and N-p degrees of freedom.
Equation (1) is the test statistics for testing H0: there is no differences in
treatment means.
6. Analysis of variance Table( ANOVA Table)
Sources Degree Sum of Mean F-Value
of of square sum of
variation freedo s squares
m
Treatment p-1 SSTR MSSTR MSSTR
 s = / MSSE
SSTR/(p
-1)
Error N-p SSE MSSE =
SSE/(p-
1) (k-1)
total N-1 TSS

We reject H0 and conclude that there are differences in the treatment means if
F0> Fα, p-1, N-p Where F0 is computed from equation 1 and Fα, p-1, N-p, is the table
value referring to F table at α level of significance corresponding to p-1and N-p
degrees freedom.
7. Advantages and disadvantages of CRD
Advantages:
a) The design is very simple to implement.
b) Any number of treatments can be used with unequal replications and does not
make the statistical hypothesis complicated.
c) The statistical analysis is simple. Even if some/all of the observations for any
treatment is missing , the statistical analysis of the data does not become
complicated
d) The design provides maximum number of degrees of freedom for the estimation
of error variance. For small experiments, the precision increases with increasing
number of error d.f.

Disadvantage:
1) In most circumstances, the experimental units are not homogeneous, particularly,
when a large no. of units involved. The CRD fails to take account of the variation
among the experimental units as it does not use the principle of local control,
This will increase the value of error variance under CRD
2) CRD is appropriate only if all plots are homogeneous. In reality this will never
happen in field experiment, thus the CRD not recommended for field trials.
8. Example: A set of data involving four tropical feed stuffs A,B,C and D tested on
22 chicks is given below. All the twenty two chicks are treated alike in all respect
except the feeding treatment, Analyse the data.
Data:
Wt. gain in chicks fed on different feeding materials composed of tropical feed
stuff
A: 55 49 42 20 62 73
B: 61 112 30 89 63
C: 42 97 81 95 92 102
D: 169 137 169 85 154

Solution:
We have 4 treatments A,B,C and D and we have to compare these 4 treatments
on the response vatiable ( Wt. Gain). Only the treatments (feed stuffs) affecting
the wt. gain other than general effect. Hence we assume the one way anova
model
Yij = μ + αi+ €ij, i=1,23.4, j=1,2,…ni ( Where n1-6, n2-5, n3=6, n4=5)
Where µ general effect
αi – ith treatment effect; ϵij- error term
We test the hypothesis
H0: α1= α2= α2=α4=0
Against the alternative
H1: at least one αi≠ αj, for all i,j, (i≠j) , (i,j: 1,2,3,4)

y..2
Step 1:Calculation of correction factor(CF): = 18792/22= 160.0483
N
 p k
y..2
Step;2: Calculation of Total sum of squares: 
i 1

j 1
y  2
ij
N =

(552+492+…+1542)-CF
=37793.318
p
yi2. y..2
Step;3: Calculation of Treatment sum of squares= i 1 ni N
 =

(3012/6+3552/5+5092/6+7142/5) –CF = 25583.152

( Total(A) =301, Total(B) =355, Total(C) =509, Total(D) =714)
Step 4: Calculation of Error sum of squares
ESS = 37793.318-25583.152=12210.167

Anova Table:
Sources of Degree of Sum of squares Mean sum of F-Value
variation freedom squares
Treatments p-1=3 SSTR= MSSTR = MSSTR/ MSSE
(foodstuffs) 25583.152 SSTR/(p-1)= = 12.571
8527.717
Error N-p=18 SSE=12210.167 MSSE =
SSE/(p-1) (k-1)
678.343
total N-1 =21 TSS
=37793.318
Table value for Fα is F.05, 3,18 = 3.16
Conclusion: Sinc Fcal(12.571) > F.05, 5,15 we reject Hα. Hence we conclude that
all treatment means are not equal
When Ho is rejected we can go for multiple comparison test ( method of least
square difference or critical difference ) if needed to test which pair of
treatment are significantly different . Proceed as follows.
 1 1
Compute yi.  y j. t
and  , n  p
MESS  
n n  . The Table value is
 i j 

t .05,18 is 2.101

Absolute differences OF yi.  y j . given in the following table

Treatments A B C D
A - 25.73333 39.23333* 97.20000*
B - - 13.50000 71.46667*
C - - - 57.96667*

1 1
t , n  p MESS  
Calculation of critical difference n n 
 i j 

Treatments A B C D
A - 34.22 18.51 34.22
B - - 34.22 35.34
C - - - 34.22

The mean difference is significant at 5% level for the pairs (A,C), A,D), (B,D),
and (C.D) which are the pairs marked *

Conclusion: In this session we have defined some terminologies which are

used in design and analysis of experimentation. We have discussed the
principles of experimental design. We have introduced completely randomised
design and discussed its analysis . We have derived the least square estimates
of the parameters of the design. We summarised the complete analysis of CRD
using ANOVA table. We also discussed the advantages and disadvantages of
CRD. We have discussed the complete analysis of CRD using an example.