Professional Documents
Culture Documents
DESIGN OF EXPERIMENTS
DESIGN OF EXPERIMENTS
DESIGN OF EXPERIMENTS
TERMINOLOGY
1. Design of an experiment
2. Experimental unit
These are objects upon which measurements are taken i.e on which variable understudy is
measured. In an agricultural field experiment, the plot of land will be the experimental unit.
In a feeding experiment of cows, the cow is the experimental unit, etc.
3. Factors
4. Level
This is the intensity setting of a factor e.g the 3 temperatures 100 , 200 , 300 represent 3
levels of the quantitative factor temperature. Similarly Benz, V/W, BMW, Corolla represent 4
levels of the qualitative factor type.
5. Treatment
This is the specific combination of factor levels. The experiment may involve only a single
factor and will therefore have one-way ANOVA. The experiment could also be composed of
levels of 2 or more factors in which case we have two-way ANOVA......... n-way ANOVA
where n is the number of factors. In other words what one does to the experimental unit that
makes it differ from one population to another is called a treatment e.g for one-way ANOVA,
no of treatments = no of levels of a factor.
For a two-way ANOVA, the no of treatments = no of combinations (cells) = no. rows x no. of
columns
1
Variety Fertiliser
F1 F2 F3 F4
V1 * * * *
V2 * * * *
V3 * * * *
No of treatments = 12
6. Experimental error
This is the unexplained random part of the total variation and it is caused by a number of
factors, most important of which are the following;
The experimental error provides a basis for the confidence to be placed in the conclusions of
inferences about the population so it is important to estimate and control the experimental
error.
7. Replication
Advantages of replicating
1. It makes the experiment more precise.
also increases.
2. Replication gives an estimate of the error.
3. It guards against accidents and with many replications, the investigator is in position
to identify outliers.
8. Randomisation
Each individual experimental unit has a known probability of being subjected to each
treatment and the advantages include;
2
2. It helps to increase the induction scope of the experiment i.e the conclusions are made
more powerful.
3. We can generate random variables whose distribution we know and it is possible to
use the F-test for all practical purposes.
It should be noted that in the planning stages of an experiment, the experimenter must decide
whether the levels of factors considered are to be set at fixed values or are to be chosen at
random from many possible levels. This will depend on the objectives of the experiment and
the question to be asked is:
Are the results to be judged for those levels alone fixed or are they to be extended to more
levels of which those in the experiment are a random sample.
In the case of quantitative factors such as time, weight, etc it is usually desirable to pick fixed
levels, some at the extremes and some at intermediate points because a random choice might
not cover the range in which the experimenter is interested e.g
0 . Other factors such as days of the week, halls of
residence, locations, etc may often be a small sample of all possible days, possible halls of
residence, etc. In such cases the particular day or particular hall may not be very important.
What is important is whether or not days, halls, etc. in general increase the variability of the
experiment.
Once the decision has been made as to whether to consider the levels random or fixed, if
random levels are to be used then they must be chosen from all possible levels by a random
process. When all levels are fixed, the mathematical model of the process is called a fixed
model. When all levels are chosen at random, then we have a random model. When more
than one factor is involved, some factors may be at fixed levels and others at random levels
and the model would be a mixed model. Suppose we have a single factor experiment, the
model is;
th th
Where; is the observation in the treatment
µ is the general mean about which the observations are supposed to fluctuate.
th
is the effect of the treatment
3
Fixed effects model Random effects model
1. are fixed constants and 1. They are random
variables and are normally and
independently distributed with mean
a - no of levels or treatments 0, and variance which represents
the variance among the or among
the treatment means and the will
_ _ _ _ average to zero if averaged over all
_ _ _ _ the possible levels but for a levels
_ _ _ _ will not average to zero.
Total X1 X2 X3 X4
Mean µ1o µ2o µ3o µ4o
e.g
These effects add up to zero because
we are looking at only those
treatments which are being
considered. 2. Analysis is the same as for fixed
model
2. Analysis is the same.
3. Hypothesis
3. Hypotheses Ho : =0
Ho: α1 = α2 =………. = αa = 0
HA: Some αi ≠ 0 HA : ≠0
OR
Ho: µ1o = µ2o = ….... = µa = 0 (say)
HA: Some µi ≠ µ
4
experiment is properly designed then there will exist an appropriate way of analysing the
data. The application of the technique of ANOVA is appropriate if the data conforms to the
basic set up of the ANOVA. The lay out and the method of analysis are coordinated in the
design of experiments.
Look at;
Introduction
Consider the problem of determining whether or not different types of tyres exhibit different
amounts of tread loss after 20,000km of driving.
A manager wishes to consider 4 tyres that are available and make some decision about which
type or brand might show the least amount of tread wear or loss.
The brands to be considered are A, B, C, and D and she wants to try these 4 brands under
actual driving conditions. The variable to be measured is the difference in maximum tread
thickness on the tyre between the time it is mounted on the wheel of a car and after it has
completed 20,000km on a car. The measured variable is this difference in thickness in
cm.
The single factor of interest is brands/tyre types ( , = 1, 2 3, 4 )
Since the tyres must be tried on cars and since some measure of error is necessary
(replication) more than one tyre of each brand must be used and a set of 4 of each brand
would seem quite practical. A car normally uses 4 tyres.
This means that we are going to have 16 experimental units (tyres), 4 each for the 4 brand
and if we designate each of the cars as , one might put the brand A tyres on car
brand B on car and so on with the design as shown in the table below.
5
Table 1: Design 1 of the tyre-type problem
I II III IV
A B C D
A B C D
A B C D
A B C D
If one looks at design 1, there is a problem since averages for brands are also averages for
cars. If the cars travel over different terrains using different drivers, any apparent brand
differences are also car differences and therefore this design is called a completely
confounded one since we cannot distinguish between brands and cars in this analysis.
A second attempt at the design may be to try the CRD.
This design applied to the above example will imply assigning the 16 tyres to the 4 cars in a
completely random manner and might give results as in table 2 below.
I II III IV
6
4. It allows the maximum number of degrees of freedom for the error sum of squares.
5. It does not require an experimental unit to participate under more than one treatment
level therefore the sample size is maximised.
Disadvantages
1. Effects of differences among subjects are controlled by random assignment of
subjects to treatment levels. For this to be more effective, subjects should be relatively
homogeneous or a large number of subjects should be used (as many replications as
possible).
2. When many treatment levels are included in the experiment, the required sample size
may be prohibitive especially from the point of view of costs.
3. This design does not offer the possibility of evaluating the interaction effect (you can
find out the interaction effect if you have more than one factor).
Analysis
To carry out ANOVA for one factor experiment, the total sums of squares (SST) is
partitioned or broken down into sums of squares due to the treatments which we donate as
SSB and the error sums of squares (SSE)
SST = SSB + SSE
Randomisation and lay out of the CRD with equal number of observations/treatments
1 2 3 ....... a
The first column represents a random sample of size n for treatment 1.........., ath column
represents a random sample of size n for treatment a
yio and io are the total and mean respectively of the observations in the ith treatment.
yoo and oo are the grand total and mean of all the N = an observations
oo =
7
i.e Total variation = Variation between the treatments + Variation due to error within
treatments
SST = SSB + SSE
SSE = SST – SSB
We can compute
2
SST =
= (prove)
2
SSB = n n is the number of observations per treatment
= (prove)
uncorrected SSB
correction factor
SSE =
OR
SSE = SST – SSB
ANOVA TABLE
Source of variation df ss ms F-ratio
Between treatments SSB MSB = Fc =
Total SST _ _
Tabulated value
FT = F0.05,
– degrees of freedom for the factor of interest.
– degrees of freedom for the error term.
Rejection criteria
8
Fc ≥ FT , Reject Ho => Treatment effects are not equal to zero/Treatment means are
not equal
Fc < FT , Accept Ho=> Treatment effects are equal to zero/Treatment means are equal
Example 1
The data in the table below gives the number of hours of pain relief provided by 4 different
types of headache tablets administered to 24 people. The 24 experimental units were
randomly divided into 4 groups and each group was treated with a different brand/type. Do
the different drug types give significantly different hours of pain relief?
Brands
1 2 3 4
12.2 4.9 8.0 4.6
9.5 10.6 12.1 6.1
11.6 7.0 5.7 5.0
13.0 8.3 8.6 3.8
10.1 5.5 7.2 8.2
9.6 11.7 12.4 7.7
66.0 48.0 54.0 36.0 =204
11.0 8.0 9.0 6.0
Steps
1. Model
=µ+ +
Where - jth observation under the ith drug type
µ - general mean
- ith treatment effect (ith drug type effect)
- error term
2. Hypothesis
Ho: α1 = α2 = α3 = α4 = 0 OR µ1 = µ2 = µ3 = µ4 = µ( say)
HA: Some αi ≠ 0 Some µio ≠ µ
3. Rejection region
Reject Ho if Fc ≥ FT ( F0.05, 3, 20 ) = 3.10
Accept Ho if Fc < F0.05, 3, 20 = 3.10
4. Computations
9
Correction factor = = = 1734.0
SST =
= [ 12.22 + 9.52 + ........... + 7.72 ] − Correction factor
= 1912.7 – 1734.0
= 178.7
4. Conclusion
Fc = 5.1638 > FT = 3.10 ; Reject Ho implying;
Drug types are significantly different in terms of the hours of pain relief they give.
Note: The sums of squares are NEVER negative.
1 2 3 ......... a
N=
10
SSB =
SST and SSE remain the same as before.
Example 2
Four groups of students were subjected to different teaching techniques and tested at the end
of a specified period of time. The table below gives the performance in percentages. Are the
teaching techniques significantly different judging from the performance of the students?
Teaching techniques
1 2 3 4
65 75 59 94
87 69 78 89
73 83 67 80
79 81 62 88
81 72 83
69 79 76
90
ni 6 7 6 4
io 454 549 425 351
N = 23 N=
= 1779
= = 77.34
1. Model
=µ+ +
Where - jth observation under the ith teaching technique
µ - grand mean
- ith treatment effect (ith teaching technique effect)
- error term
2. Hypothesis
Ho: α1 = α2 = α3 = α4 = 0 OR µ1 = µ2 = µ3 = µ4 = µ( say)
Ho: Some αi ≠ 0 Some µio ≠ µ
11
3. Rejection region
Reject Ho if Fc ≥ FT ( F0.05, 3, 19 ) = 3.13
Accept Ho if Fc < F0.05, 3, 19 = 3.13
4. Computations
Correction factor = = = 137,601.8
SST =
= [ 652 + 872 + ........... + 882 ] − Correction factor
= 139511 – 137,601.8
= 1909.2
= – C.F
= 138314.4 – 137601.8
= 712.6
5. Conclusion
Fc = 3.769 > FT = 3.13 ; Reject Ho implying;
The teaching techniques are significantly different judging from the performance of
the students.
Estimation of effects
Given the data table e.g example 2 and the ANOVA table, we can make the following
inferences/conclusions concerning the population from which the data was obtained.
1) The grand mean is an unbiased estimator of the population mean.
2) The treatment mean is an unbiased estimator of the mean of the treatment
population. E.g = 75.67
12
3) The difference between any 2 treatment effects is estimated unbiasedly by the
difference between the treatment means i.e
4) The factor effect is estimated unbiasedly by the difference between its treatment
means and the grand mean i.e and thus can be positive or negative.
5) The mean square error (MSE) value is an unbiased estimator of the common variance
σ2. e.g σ2 = 63 in example 2.
Assumptions of ANOVA
In applying the ANOVA techniques, certain assumptions should be kept in mind and these
include;
1. The process is repeatable (can be replicated)
2. The population distribution being sampled is normal.
3. The variances of all the a levels of a factor are homogeneous.
Exercise
1. Analyse the tyre type effect in the Table 2 design.
2. In a biological experiment, 4 concentrations of a certain chemical are used to enhance
the growth of a certain type of plant over a specified period of time.
The following growth data in cm were recorded for the plants that survived.
Concentration
1 2 3 4
8.2 7.7 6.9 6.8
8.7 8.4 5.8 7.3
9.4 8.6 7.2 6.3
9.2 8.1 6.8 6.9
8.0 7.4 7.1
6.1
Is there a significant difference in the average growth of these plants for the different
concentrations of the chemical? Use α = 0.01 and α = 0.05. Estimate the concentration
effect.
13
RANDOMISED BLOCK DESIGN (RBD)
Amore careful examination of design 2 in table 2 will reveal some disadvantages of CRD in
this problem.
One thing to be noted is that brand A is never used on car 3 or brand B on car 1 or D on car 2.
Any variation within brand A may reflect variation between cars 1, 2 and 4. Thus the random
error may not be merely an experimental error but may include variation between cars. Since
the chief objective of experimental design is to reduce the experimental error, a better design
might be one in which the car variation is removed from the error and a design that requires
that each brand be used once on each car is called a Randomised Block Design.
Car
I II III IV
In this design the order in which the 4 brands are placed on a car is random and each car gets
one tyre of each brand. In this way, better comparisons can be made between brands since
they are all driven over the same terrain using the same drivers and so on.
This provides a more homogeneous environment in which to test the 4 brands. In general
these groupings for homogeneity are called blocks and the randomisation is now restricted
within blocks.
The design also allows the car (block) variation to be independently assessed and removed
from the error term and the model for the design is;
=µ+ +
Where - observation from the ith tyre type and the jth car type
µ - general mean
- ith tyre type effect
- jth car type (block) effect
Advantages
1. The chief advantage of the RBD over the CRD is that it makes an attempt to control
the error by identifying a portion of the total variation with the block means.
14
2. This design enables the research to use different techniques to different blocks though
the techniques should be the same within the block.
3. The analysis is straight forward and remains so unless due to accident, data (on an
entire block or treatment) gets missing. If data from individual units is missing, then
we use the Yates missing plot technique to estimate that missing value and then
continue with the test and analysis.
4. Each experimental unit receives each of the treatments assigned in a random
sequence. This cuts down considerably on the number of experimental units or
subjects needed for the experiment (point of view, cost) => one experimental unit can
appear in more than one treatment.
Disadvantages
1. The chief disadvantage is that if the blocks are not internally homogeneous then a
large error term will result.
I II
A D
B C
C A
D B
Layout of RBD
In general, consider a rectangular array of r rows and c columns. i.e we may have a factor A
as the rows with r levels and factor B as the columns with c levels.
15
Layout of a two way classification with one observation/cell
: : : : : : : :
: : : : : : : :
i
: : : : : : : :
: : : : : : : :
R
Totals
Means
1. Model
=µ+ +
Where - ith row effect
- jth column effect
is the observation in the ith row and jth column i.e the ijth cell
e.g observation in the 2nd row and 3rd column
th
and are the total and mean respectively of all the observations in the i row.
and are the total and mean respectively of all observations in the jth column.
and are the grand total and mean respectively of all the rc observations.
N = rc
2. Hypotheses
16
To determine if part of the variation in the observations is due to differences among
the rows or due to factor A, we test the following hypothesis
H01: µ1o = µ2o = µ3o = ....... = µro = µ( say)
HA1: Some µio ≠ µ
OR
Ho1: α1 = α2 = α3 = ...... = αr = 0 Assuming the αi are row effects
HA1: Some αi ≠ 0
Similarly to determine if part of the variation is due to differences among the columns
or due to Factor B
Ho2: µo1 = µo2 = µo3 = ....... = µoc = µ( say)
HA2: Some µoj ≠ µ
OR
Ho2: β1 = β2 = β3 = ...... = βc = 0 Assuming the βj are the column effects
Ho2: Some βj ≠ 0
3. Rejection Criteria
Reject Ho1 if F1c ≥ FTα , (r-1), (r-1) (c-1)
Reject Ho2 if F2c ≥ FTα , (c-1), (r-1) (c-1)
SST = from
SSR =
= (Prove)
SSC = (Prove)
SSE = SST – SSR – SSC
OR
SSE =
Source of Df SS MS F-ratio
variation
Rows ( factor A) r –1 SSR MSR = F1c =
17
Columns (factor B) c –1 SSC MSC = F2c =
Error (r-1) (c-1) SSE MSE = _
Critical region
Reject Ho1 if F1c ≥ FTα , (r-1), (r-1) (c-1)
Reject Ho2 if F2c ≥ FTα , (c-1), (r-1) (c-1)
Example
Qn. Test for the significance of the car type effect and the tyre type effect on tread loss.
1. Model
=µ+ +
Columns – tyre types
Rows – car types
Where – observation for the ith car type and the jth tyre type
µ - general mean
- ith car type effect
- jth tyre type effect
– experimental error
2. Hypotheses
Ho1: α1 = α2 = α3 = α4 = 0
HAI: Some αi ≠ 0
OR
Ho2: β1 = β2 = β3 = β4 = 0
HA2: Some βj ≠ 0
18
3. Critical region
Reject Ho1 if F1c ≥ F0.05, 3, 9 = 3.86 F0.01, 3, 9 = 6.99
Reject Ho2 if F2c ≥ F0.05, 3, 9 = 3.86
4. Computations
Correction factor = = = 2328.06
SST =
= [ 172 + 142 + ........... + 92 ] – C.F
= 2409 – 2328.06
= 80.94
SSR =
= C.F
= 2366.75 C.F
= 38.69
SSC =
= C.F
= 2358.75 C.F
= 30.69
5. ANOVA Table
6. Conclusions
19
To test the hypothesis that the average tread loss of all the four car types is the same,
the computed F obtained was 9.9 which is also significant at the 1% level of
significance. This means that the car to car variation is significant since we reject Ho1.
For the case of the tyre types F2c = 7.8 which is also significantly larger than the
corresponding critical F given. i.e The hypothesis of equal tyre type means Ho2 is also
rejected at both 1% and 5% level of significance.
Note that this hypothesis of equal tyre type means could not be rejected using a CRD.
The RBD allows for the removal of the car effect which has reduced the common
variance significantly from 4.2 to 1.3.
Remark
When data are presented in a tabular form, there is usually no way to determine how
the data were collected. Was the randomisation complete over all N observations or
was the experiment run in blocks with randomisation restricted to within the blocks?
To help in signifying the design of the experiment, it is suggested that in the case of a
CRD, no vertical or horizontal lines be drawn as in table 2. When randomisation has
been restricted, either vertical or horizontal lines (as shown in table 3) can be used to
indicate this order of restriction on the randomisation.
Exercise
The table below gives the yields in thousands of kgs/hectare of 3 varieties of beans grown
using 4 different types of fertilisers.
TABLE X
Fertilisers Varieties
V1 V2 V3
t1 64 72 74
t2 55 57 47
t3 59 66 58
t4 58 57 53
20
Since blocking an experiment is a very useful procedure to reduce the experimental error, we
will look at an example that may be helpful in illustrating a situation in which the same
experimental units e.g animals, tyres, etc are measured before an experiment and then again
at the conclusion of the experimental treatment. Our interest is primarily in the effect of the
treatment since the block effects can be removed to reduce experimental error.
These designs are sometimes referred to as repeated measures designs because data is
collected on the same units the 2nd time and the following example will illustrate the
procedure.
Example 4
A study on the physical strength measured in kilograms on 7 subjects before and after a
specified training period gave the following results:
tc =
= 12.05
Since tc > tT, the training was effective. The post test measures were significantly different
from the pre test measures. If we considered the above example as an experiment at two (2)
levels pre and post with 7 blocks which are our subjects, making up an RBD.
Our data would appear as shown in table 6 below.
Table 6: RBD presentation for the data in |Table 5
21
Measures
Subjects Pre test Post test
1 100 115 215
2 110 125 235
3 90 105 195
4 110 130 240
5 125 140 265
6 130 140 270
7 105 125 230
770 880 = 1650
Computations
C.F = = = 19446.29
SST = C.F
= 197,450 – 19446.29
= 2985.71
ANOVA table
Source of variation df ss ms F-ratio
Rows (subjects) 6 2085.71 347.62 F1c =
= 58.42 > FT 0.05.,6,6
Columns (treatment) 1 864.29 864.29 F2c =
= 145.26 > FT 0.05.,1,6
Error 6 35.71 5.95 -
Total 13 2985.71 - -
22
Conclusion
The training has made a significant difference in the physical strength measurements. (The
large F2c indicates a strong treatment effect)
There is also a significant difference in the subjects.
Exercise
1. Adhesive force on gummed material was determined under 3 fixed humidity and 3
fixed temperature conditions. Four (4) readings were made under each set of
conditions (treatments)
The experiment was randomised and the results are set out in an ANOVA table as
follows:
Complete the above table and give conclusions regarding the significance of the
effects of the different factors.
Using a 5% level of significance, indicate whether a soil type effect, fertiliser type
effect and the interaction effect are statistically significant.
3. A group of 10 students was pre tested before and after instruction and the following
achievement scores were obtained.
23
Student Before After
1 14 17
2 12 16
3 20 21
4 8 10
5 11 10
6 15 14
7 17 20
8 18 22
9 9 14
10 7 12
24