Wollega University Shambu Campus

Faculty of Technology
Wollega University

Department of Industrial Engineering

Course Title: -Experimental Design

Course Code: IEng 4204

Cr.hr:- 2( 4ECTS)

Instructor: -Refisa W.

Email:[email protected]
©2023 G.C
 Chapter One
Introduction
1. Perception of Quality
Wollega University

 Quality has been defined in different ways by different experts

over a period of time. Some of them are as follows:
 Fitness for use
 Conformance to specifications
 Customer satisfaction/delight
 The totality of features and characteristics that satisfy the
needs of the customer
The simplest definition of high quality is a happy customer.
 The true evaluator of quality is the customer.
 …Cont’d
 The quality of a product is measured in terms of the
Wollega University

features and characteristics that describe the

performance relative to customer requirements or
expectations.
 To satisfy the customer, the product must be delivered
in right quality at right time at right place and provide
right functions for the right period at right price.
 Taguchi’s Definition of Quality

 The quality of a product is defined as the loss imparted by the product to

society from the time the product is shipped to the customer.
Wollega University

 The loss may be due to failure, repair, variation in performance, pollution,

noise, etc.
 A truly high quality product will have a minimal loss to society.
 The following are the types of loss:
 Product returns
 Warranty costs
 Customer complaints and dissatisfaction
 Time and money spent by the customer
 Eventual loss of market share and growth
 Under warranty, the loss will be borne by the manufacturer and after
warranty it is to be paid by the customer.
 Whoever pays the loss, ultimately it is the loss to the society.
 Taguchi Quality Loss Function
 The loss which we are talking is the loss due to functional
Wollega University

variation/process variation.
 Taguchi quantified this loss through a quality loss function.
 The quality characteristic is the object of interest of a product
or process.
 Generally, the quality characteristic will have a target. There
are three types of targets.
1. Nominal–the best: When we have a characteristic with bi-lateral
tolerance, the nominal value is the target. That is, if all parts are made to
this value, the variation will be zero and it is the best.
 For example: A component with a specification of 10 0.01 mm has the
nominal value of 10 mm. Similarly, if the supply voltage has a
specification of 230 10 V. Here the nominal value is 230V
 …Cont’d
2. Smaller–the better: It is a non negative measurable
Wollega University

characteristic having an ideal target as zero.

 For example: Tyre wear, pollution, process defectives, etc.
3. Larger–the better: It is also a non negative measurable
characteristic that has an ideal target as infinity (ꝏ).
• For example: Fuel efficiency, strength values, etc.
 For each quality characteristic, there exist some function which
uniquely defines the relation between economic loss and the
deviation of the quality characteristic from its target.
 Taguchi defined this relation as a quadratic function termed
Quality Loss Function (QLF).
 Quality Loss Function (Nominal-the best)
 When the quality characteristic is of the type Nominal–the best, the quality
Wollega University

loss function is given by:

where,
Y = value of the quality characteristic (e.g., length, force,
diameter etc.)
L(Y) = loss in per product when the quality characteristic is
equal to Y
T = target value of Y
K = proportionality constant or cost coefficient which depends
on the cost at the specification limits and the width of the
specification
 …CONT’D
Wollega University

Figure 1.1: shows the QLF for nominal-the best case

 This loss is attributed to the variation in quality of performance (functional
variation).
 Thus, confirmation to specification is an inadequate measure of quality.
 …Cont’d
 When Y = T, the loss is ‘0’. From Figure 1.1, it can be observed
Wollega University

that, as Y deviates from the target T, the quality loss increases

on either side of T.
 Note that a loss of A0 is incurred even at the specification limit
or consumer tolerance
 The quality loss is due to customer dissatisfaction.
 It can be related to the product quality characteristics. and QLF
is a better method of assessing quality loss compared to the
traditional method which is based on proportion of defectives.
 …Cont’d
 Suppose the supply voltage for a compressor is the quality
Wollega University

characteristic (Y).
 The specification for voltage is 230 10 V.
 Target voltage = 230 V
 Consumer tolerance = 10 V.
 Let A0 = average cost of repair/adjustment at a deviation of
At the specification,

Therefore,

If A0 = 500 at a deviation of 10V, then the cost coefficient

Wollega University

…Cont’d
 Estimation of Average Quality Loss
 The quality loss we obtain from Eq. (1.1) is for a single
Wollega University

product.
 The quality loss estimate should be an average value and
should be estimated from a sample of parts/products.
 Suppose we take a sample of n parts/products. Then we will
have:
 Y1, Y2, ..., Yi values where, i = 1, 2, 3, ..., n
 Let the average of this sample = . This may or may not
be equal to the target value (T).
 The quality loss is modelled as follows:
 Estimation of Average Quality Loss
Wollega University

 The term in the parenthesis is the average of all values of (Yi – T)2 and is
called Mean Square Deviation (MSD).
 Therefore, L(Y) = K(MSD)

On simplification
 …Cont’d
Wollega University

where,
= population standard deviation
= bias of the sample from the target.
For practical purpose we use sample standard deviation ( –1)
Since, L(Y) = K(MSD)
……….. (e)
 Note that the loss estimated by Equation (E) is the average
loss per part/product
Total loss = (number of parts/products)
 Quality Loss Function (smaller-the better case)

 For this case the target value T = 0

Wollega University

……………..(f)
and
……………..(g)

 Figure 1.2 shows the loss function for smaller—the better type
of quality characteristic.
 …Cont’d
 The average loss per part/product
can be obtained by substituting T
Wollega University

= 0 in Eq. (e).

……………(h)

Figure 1.2: QLF For smaller – the better

case
 Quality Loss Function (Larger-the better case)

 Mathematically, a larger–the better characteristic can be considered as the

Wollega University

inverse of a smaller– the better characteristic.

 Therefore, we can obtain L(Y) for this case from Eq. (f). That is
……………………………(i)

………………………………(j)
 Figure 1.3 shows the quality loss for the
case of larger–the better type of quality
characteristic.
 The loss per part/product averaged from
many sample parts/products is given by:

Figure 1.3: QLF for larger-the better type of quality characteristics

 Quality Loss Function (Larger-the better case)

………………………………(k)
Wollega University

On simplification of Eq. (k), the average loss per part/product will be given by:

………………………………(l)
 Illustration 1.1
Nominal–the Best Case
 A company manufactures coolers of various capacities in which the compressor
Wollega University

forms the critical item. The specification for the voltage of the compressor of a
specific capacity is 230 20 V. The consumer loss at the specification was
estimated to be 3000. This includes repair/replacement cost and other related
cost based on customer complaints. The company purchases two types of
compressors (Brand A and Brand B). The output voltage data collected for 20
compressors are given in Table 1.1.
 …Cont’d
 Determine which brand of compressors should be purchased
Wollega University
 …Cont’d
 The mean and standard deviation values computed for the
Wollega University

two brands of compressors and the average loss per

compressor are given in Table 1.2.
Table 1.2: Average loss per compressor

 Since the average loss per compressor is less for Brand B, it

is recommended to purchase Brand B compressor.
 Illustration 1.2
Smaller–the Better Case
 The problem is concerned with the manufacturing of speedometer
Wollega University

cable casings using the raw material supplied by two different

suppliers. The output quality characteristic is the shrinkage of the cable
casing.

The data collected on 20 samples from

the two suppliers is given in Table 1.3.
 …Cont’d
Wollega University

The average loss per casing is computed from:

The mean, variance, and the average loss per cable is given in Table 1.4.
 …Cont’d
Wollega University

The average loss per casing is less with Supplier 2. Hence, Supplier 2 is preferred.
 Illustration 1.3
 Larger–the Better Case
Wollega University

 This problem is concerned with the comparison of the life of electric bulbs
of same wattage manufactured by two different companies. The quality
characteristic (Y) is the life of the bulb in hours

The data collected is given in Table 1.5

Determine the average loss per bulb for the two companies.
 …Cont’d
Wollega University

The average loss per product is computed from Equation given below

The computed values are summarized in Table 1.6

Since the average loss is less with Company 2, Company 2 is preferred .

 Estimation of Quality Loss

There are two types of estimation of quality loss. They are:

Wollega University

1) Traditional Method
 Usually in manufacturing companies, the quality loss is estimated by
considering the number of defects/defectives. In this, it is assumed that all
parts with in the specification will not have any quality loss.
 Loss by defect = Proportion out of specification x number of products x
cost per product
2) Quality Loss Function Method
 This method is based on dispersion (deviation from target). Depending on
the type of quality characteristic, quality loss is estimated using an
appropriate quality loss function. Suppose we have a nominal–the best
type of quality characteristic.
 …Cont’d
Loss by dispersion = X number of products
Wollega University

Example: Comparison of Quality Loss

 A company manufactures a part for which the specification is 40 4 mm (i.e., T =
40 and = 4). The cost of repairing or resetting is 30 (i.e., A0 = 30).
 The process average is centered at the target T = 40 with a standard deviation of
1.33. The annual production is 50,000 parts.
 The proportion out of specification can be estimated from normal distribution as
follows. The Upper Specification Limit (USL) is 44 and the Lower Specification
Limit (LSL) is 36

 Therefore, proportion out of specification above USL is 0.00135. Similarly,

proportion falling below LSL is 0.00135. Thus, total proportion of product
falling out of specification limits is 0.0027 (0.27%).
 …Cont’d
Loss by defect = Proportion out of specification X number of parts X cost
Wollega University

per part
= 0.0027 X 50,000 X 30
= 4,050.

 Note the difference between the two estimates. So, to reduce the quality
loss we need to reduce the dispersion (variation) there by improving
quality.
 Basic Design Principles
 The purpose of statistically designed experiments is to collect appropriate
data which shall be analyzed by statistical methods resulting in valid and
Wollega University

objective conclusions.
 The two aspects of experimental problem are as follows:
1. The design of the experiment, and
2. The statistical analysis of the data.
 The following basic principles are used in planning/designing an
experiment.
1. Replication: Replication involves the repetition of the experiment and
obtaining the response from the same experimental set up once again on
different experimental unit (samples).
 The purpose of replication is to obtain the magnitude of experimental error.
 This error estimate (error variance) is used for testing statistically the
observed difference in the experimental data.
 …Cont’d
 Replications also permit the experimenter to obtain a precise estimate of
the effect of a factor studied in the experiment.
Wollega University

 Replication is not a repeated measurement.

2. Randomization
 The use of statistical methods requires randomization in any experiment.
 Statistical methods require that the observations (errors) are independently
distributed random variables. It meets this requirement.
 It also assists in averaging out the effects of extraneous factors that may
be present during experimentation.
 When complete randomization is not possible, appropriate statistical
design methods shall be used to tackle restriction on randomization.
 3. Blocking

 Blocking is a design technique used to improve the precision

Wollega University

of the experiment.
 Blocking is used to reduce or eliminate the effect of nuisance
factors or noise factors.
 A block is a portion of the experimental material that should
be more homogeneous than the entire set of material or a
block is a set of more or less homogeneous experimental
conditions.
 It is also a restriction on complete randomization.
 The three basic principles of experimental design,
randomization, replication and blocking are part of all
experiments.
 Terminology used in Design Experiment

 Factor: A variable or attribute which influences or is suspected of

influencing the characteristic being investigated.
Wollega University

 All input variables which affect the output of a system are factors. Factors
are varied in the experiment. They can be controlled at fixed levels.
 They can be varied or set at levels of our interest. They can be qualitative or
quantitative. These are also called independent variables.
 Levels of a factor: the values of a factor/independent variable being
examined in an experiment.
 If the factor is an attribute, each of its state is a level.
 For example, setting of a switch on or off are the two levels of the factor
switch setting.
 If the factor is a variable, the range is divided into required number of levels.
 For example, the factor temperature ranges from 1000 to 1200°C and it is to
be studied at three values say 1000°C, 1100°C and 1200°C, these three
values are the three levels of the factor temperature.
 The levels can be fixed or random.
 …Cont’d
 Treatment: One set of levels of all factors employed in a given experimental
Wollega University

trial. For example, an experiment conducted using temperature T1 and

pressure P1 would constitute one treatment.
 In the case of single factor experiment, each level of the factor is a treatment.
 Experimental unit: Facility with which an experimental trial is conducted
such as samples of material, person, animal, plant, etc.
 Response: The result/output obtained from a trial of an experiment. This is
also called dependent variable. Examples are yield, tensile strength, surface
finish, number of defectives, etc.
 Effect: Effect of a factor is the change in response due to change in the level
of the factor.
 Experimental error: It is the variation in response when the same
experiment is repeated, caused by conditions not controlled in the
experiment. It is estimated as the residual variation after the effects have
been removed.
 Quiz-1
1. Write the two aspects of experimental problem
Wollega University

2. What are the three basic principles are used in

planning/designing an experiment?
3. What are the three types of targets of quality
characteristic?
4. Write the two types of estimation of quality loss.
5. Write the types of quality loss
 Normal Probability Plot
 One of the assumptions in all the statistical models used in
Wollega University

design of experiments is that the experimental errors

(residuals) are normally distributed.
 This normality assumption can be verified by plotting the
residuals on a normal probability paper.
 If all the residuals fall along a straight line drawn through
the plotted points, it is inferred that the residuals/errors follow
normal distribution.
 The sample data is arranged in ascending order of the value of
the variable (X1, X2,…, Xj, …, Xn).
 The ordered data (Xj) is plotted against their observed
cumulative frequency (j – 0.5)/n on the normal probability
paper.
 …Cont’d
 If all the data points fall approximately along a straight line,
Wollega University

it is concluded that the data (variable) follows normal

distribution.
 The mean is estimated as the 50th percentile value on the
probability plot. And the standard deviation is given by the
difference between 84th percentile and the 50th percentile
value.
Example
 The following data represent the hardness of 10 samples of
certain alloy steel measured after a heat treatment process.
How to obtain a normal probability plot on a normal
probability paper is illustrated below:
 Example
275, 258, 235, 228, 265, 223, 261, 200, 276, 237
 Table 1.6 shows the computations required for normal probability plot.
Wollega University

The data arranged in the ascending order of value is shown in the

second column of Table 1.6. The first column indicates the rank of the
ordered data. Third column gives the cumulative frequency of the
ordered data. The last column (Zj) is the standardized normal score.
Table 1.6: Computation for normal Probability Plot
 Analysis of Variance

 The total corrected Sum of Squares (SS) is used as a measure

Wollega University

of overall variability in the data.

……………………….(k)
 It is partitioned into two components.
 Total variation = variation between treatments + variation
within the treatment or error
 SS Total = SS due to treatments + SS due to error
i.e. SS Total = SST + SSe
A typical format used for ANOVA computations is shown in Table 1.7
 …Cont’d
Table 1.7: ANOVA: Single Factor
Experiment
Wollega University

Mean square is also called variance.

 Computation of Sum of squares
Let T = grand total of all observations/response (Y)
Wollega University

N = total number of observations

n = number of replications/number of observations under the ith treatment
SS = sum of squares
CF = correction factor
Ti. = ith treatment total
 …Cont’d
 A manufacturing engineer wants to investigate the effect of
Wollega University

feed rate (mm/min) on the surface finish of a milling

operation. He has selected three different feed rates, i.e., 2, 4
and 6 mm/min for study and decided to obtain four
observations at each feed rate. Thus, this study consists of 12
experiments (3 levels 4 observations). Since the order of
experimentation should be random, a test sheet has to be
prepared as explained below. The 12 experiments are serially
listed in Table 1.8.
Table 1.8: Experiment run numberc
 …Cont’d
Table 1.8: Experiment run number
Wollega University

 Then the number 7th experiment (observation) with feed rate 4

mm/min is run. This procedure is repeated till all the
experiments are scheduled randomly. Table 1.8 gives the test
sheet with the order of experimentation randomized.
 Table 1.9: Test sheet for illustration
Wollega University
 Table 1.9: Test sheet for illustration
 Suppose that the engineer has conducted the experiments in the random
Wollega University

order given in Table 1.9.

 The data collected on the surface roughness are given in Table 1.9 .

 The data can be analyzed through analysis of variance.

 The hypothesis to be tested is
against H1: at least one mean is different.
 …Cont’d
Let Ti. = treatment totals
T.. = grand total
Wollega University
 …Cont’d
 The analysis of variance is given in Table 1.11. The error sum
Wollega University

of squares is obtained by substruction.

 The inference is that the treatment means are significantly different at 5%

level of significance.
 That is, the treatment (feed rate) has significant effect on the surface
finish.
 Linear Regressions
 If the study involves only one dependent variable (response),
Wollega University

one independent variable (factor) and if the relation between

them is linear, it is called simple linear regression.
 If the response is related linearly with more than one
independent variable, it is called multiple linear regression.
 Suppose the yield (Y) of a chemical process depends on the temperature (X).
If the relation between Y and X is linear, the model that describes this
relationship is:
…………………(1a)
 where and are the constants called regression coefficients and e is a
random error
is also called the intercept and is the slope of the line, that is fitted to the data.
 If we have n pairs of data (X1, Y1), (X2, Y2), … (Xn, Yn), we can estimate
the parameters by the method of least squares.
 …Cont’d
The model with the sample data Eq. (1b) is of the same form as above Eq. (1a).
Wollega University

…………………….(1b)
 where b0 and b1 are the estimates of βo and β1 respectively. Applying the
method of least squares we get the following two normal equations which
can be solved for the parameters.
…………………….(1c)

…………………….(1d)

…………………….(1e)

…………………….(1f)
 …Cont’d
Wollega University

 From the normal Eqs. (1c) and (1d), the parameters b0 and b1
can also be expressed in terms of sums of squares as follows:
……………………………………….(1g)
……………………………………….(1h)

…………………………………(1i)
…………………………………(1j)

……… (1k)
 SSX and SSY are called corrected sum of squares. And βo and β1 are estimated
as follows:
 ….Cont’d
Wollega University

Error sum of squares (SSe) for the regression as follows:

 The ANOVA equation for linear regression is:

 Total corrected sum of squares = sum of squares due to regression + sum of
squares due to error
Wollega University

….Cont’d
 …Cont’d
 Test for significance of regression is to determine if there is a linear
Wollega University

relationship between X and Y.

 Rejection of H0 implies that there is significant relationship between the variable X

and Y. we can also test the coefficients βo and β1 using one sample t test with n – 2
degrees of freedom.
 …Cont’d
Wollega University

Example 1: Simple Linear Regression

 A software company wants to find out whether their profit is related to the
investment made in their research and development. They have collected the
following data from their company records.
 …Cont’d
i. Develop a simple linear regression model to the data and estimate the
Wollega University

profit when the investment is 7 million rupees.

ii. Test the significance of regression using F-test.
iii. Test significance of β1 .
SOLUTION:
 In this problem, the profit depends on investment on R&D.
 Hence, X = Investment in R&D and Y = Profit
 The summation values of various terms needed for regression are given in
Table below
Wollega University

…Cont’d
Wollega University

…Cont’d
Wollega University
 Chapter Two
Factorial Experiments
Wollega University

Single-Factor Experiments
 In single-factor experiments only one factor is investigated.
 The factor may be either qualitative or quantitative.
 If the levels of a factor are qualitative (type of tool, type of
material, etc.), it is called qualitative factor.
 If the levels of a factor are quantitative (temperature, pressure,
velocity, etc.), it is called quantitative factor.
 The levels of a factor can be fixed (selecting specific levels) or
random (selecting randomly).
 Some examples of a single-factor experiment are:
 Studying the effect of type of tool on surface finish of a
machined part
 …Cont’d
 Effect of type of soil on yield
Wollega University

 Effect of type of training program on the performance of

participants
 Effect of temperature on the process yield
 Effect of speed on the surface finish of a machined part
 If the levels are fixed, the associated statistical model is called
fixed effects model.
 Each level of the factor considered to be treatment.
 In a single-factor experiment if the order of experimentation as
well as allocation of experimental units (samples) is completely
random, it is called completely randomized design.
 MULTI-FACTOR FACTORIAL EXPERIMENTS

 When the number of factors involved in the experiment is more

Wollega University

than one, we call it a factorial experiment.

 In factorial experiments, combination of two or more levels of
more than one factor is the treatment. That is, every level of
one factor is combined with every level of other factors.
 When all the possible treatments are studied, we call it a full
factorial experiment.
 If the number of factors is only two, it will be a two-factor
factorial experiment.
 Suppose we want to study the effect of temperature and pressure on the
reaction time of a chemical process. Further we want to investigate the
temperature at two levels (70°C and 90°C) and pressure at two levels (200
MPa and 250 MPa).
 MULTI-FACTOR FACTORIAL EXPERIMENTS

 The two-factor factorial design will be represented as given in

Wollega University

Table 2.1.
Two-factor factorial Design

 Note that there are four treatment combinations (1, 2, 3 and 4)

in the two-factor design as given in Table 2.1.
 Main effect of a factor: The effect of a factor is defined as
the change in response due to a change in the level of the
factor.
 MULTI-FACTOR FACTORIAL EXPERIMENTS

 Figures 2.1(a) and 2.1(b). Usually in a two-level factorial design,

Wollega University

we represent the levels of the factor as low (–) and high (+).

Representation of response of data for two factor experiment

 MULTI-FACTOR FACTORIAL EXPERIMENTS

 Figures 2.1(a) and 2.1(b). Usually in a two-level factorial design, we

Wollega University

represent the levels of the factor as low (–) and high (+).
• The effect due to temperature in this experiment (Figure 2.1(a)) is the
average of the effect at the high level pressure and at the low level of
pressure.
• That is,
 Effect of temperature at the high level of pressure = (100 – 80) = 20
 Effect of temperature at the low level of pressure = (130 – 110) = 20
 Average effect of temperature = 1/2(20 + 20) = 20
 Similarly, we can find the effect of pressure.
 Average effect of pressure = 1/2{(100 – 130) + (80 – 100)} = –30
 Interaction effect: The combined effect due to both the factors
is called the interaction effect.
 Tests on Means
Hypothesis Tests
Wollega University

 A statistical hypothesis is an assumption about the population

being sampled. There are two types of hypothesis.
1. Null hypothesis (H0)
2. Alternative hypothesis (H1)
 A null hypothesis is a claim or statement about a population parameter, that
is assumed to be true.
• For example, a company manufacturing electric bulbs claims that the average life of
their bulbs
 Tests on Means
 An alternative hypothesis is a claim or a statement about a
Wollega University

population parameter, that is true if null hypothesis is false.

 For example, the alternative hypothesis of life of bulbs, is that
the average life of bulbs is less than 1000 hr. It is written as:

 A test of hypothesis is simply a rule by which a hypothesis is

either accepted or rejected.
Tests on a Single Mean
 The null and alternative hypothesis for this test is as follows:
 Tests on a Single Mean
Wollega University

…………………….. (2.1a)

 This test is applicable for a normal population with known

variance or if the population is non-normal but the sample size
is large (n > 30), in which case is replaced by in Eq.
(2.1a).
 P-value Approach
Wollega University

 We reject the null hypothesis if

 We do not reject the null hypothesis if

Example
 The tensile strength of fabric is required to be at least 50 kg/cm2. From past
experience it is known that the standard deviation of tensile strength is 2.5
kg/cm2. From a random sample of 9 specimens, it is found that the mean
tensile strength is 48 kg/cm2.
 Example

i. State the appropriate hypotheses for this experiment and test

Wollega University

the hypotheses α = 0.05. What is your conclusion?

ii. What is your decision based on the p-value?
 Example
Wollega University

ii. p-value approach: From standard normal table, the tail area
under the curve beyond –2.41 is 0.008. So, the p-value for the
test is 0.008.
 Since α – value is more than the p-value we reject the null
hypothesis

The test statistic is: ……….. (2.2b)

 …Cont’d
Wollega University

Example
 A cement manufacturer claims that the mean settling time of
his cement is not more than 45 minutes. A random sample of
20 bags of cement selected and tested showed an average
settling time of 49.5 minutes with a standard deviation of 3
minutes. Test whether the company’s claim is true.
Use α = 0.05.
 …Cont’d
Wollega University

The inference is that the settling time is greater than 45

minutes.
 Tests on two means
Wollega University
 Tests on two means
Wollega University

……………………. (2.2c)

 If variances are not known and sample size is large (n > 30), σ1
and σ2 in Eq. (2.2c) are replaced by S1 and S2

Example
 A company manufacturing clay bricks claims that their bricks (Brand A) dry
faster than its rival company’s Brand B. A customer tested both brands by
selecting samples randomly and the following results have been obtained
 …Example
Table 2.2 Given Illustration
Wollega University

 Test whether the company’s claim is true at 5% significance level. Also

construct the 95% confidence interval for the difference in the two means.
Wollega University

…Example
 …Example
 From normal table, Z0.05.= 1.65. Hence we reject Ho. That is,
Wollega University

the mean drying time of both the brands is different

 p-value approach: The p-value for the test is 0.0436.
Wollega University

…Example
 Quiz-2
1. If the levels of a factor are type of tool, type of material,
Wollega University

etc., it is called ________________________.

2. If the levels of a factor are temperature, pressure, velocity,
etc., it is called ________________________.
3. What are the two types of hypothesis?
4. When we reject and do not reject null hypothesis?
5. Define p-value approach?
 Chapter Three
Special Designs
K Factorial designs
Wollega University

 Factorial designs with several factors are common in

research, especially at the early stages of process/product
development.
 . A factorial design with k-factors, each at only two levels is
called 2k design, a special case of multi-factor factorial
experiment.
 It provides the smallest number of treatment combinations
with k-factors in a full factorial experiment. These designs are
usually used as factor screening experiments.
 The 2k design assists the experimenter to identify the few
significant factors among them.
 K Factorial designs
Wollega University

 Generally, the following assumptions are made in the 2k

factorial designs.
 The factors are fixed.
 The designs are completely randomized.
 The usual normality assumption holds good.
The 22 of the Factorial Design
 The simplest design in the 2k series is a 22 design; two
factors each at two levels.
 The two levels of the factors are usually referred to as low and
high. Let us study this design with an example.
 Illustration 3.1
Wollega University

 Consider the problem of studying the effect of temperature

and pressure on the yield of a chemical process. Let the two
factors temperature and pressure are denoted by A and B
respectively. Suppose the following results are obtained from
this study. The number of replications is two.
 Illustration 3.1
Wollega University

 The response (yield) total from the two replications for the
four-treatment combinations is shown below.
 Illustration 3.1
Wollega University

 By convention, the effect of a factor is denoted by a capital letter and the

treatment combinations are represented by lower case letters. Thus,
 A refers to the effect of A
 B refers to the effect of B
 AB refers to the effect of AB interaction.
 For any factor, the low level is represented by 0 or – or 1. And the high
level of the factor is represented by 1 or + or 2. Thus, we have three
notations, namely (0, 1), (–, +) and (1, 2) in these 2k designs.
 Under (0, 1) notation, the four treatment combinations are represented as follows:
0 0: Both factors at low level; denoted by (1)
1 0: A high, B low; denoted by a
0 1: A low, B high; denoted by b
1 1: A high, B high: denoted by ab
 Illustration 3.1
Wollega University

 These four treatment combinations can be represented by the

coordinates of the vertices of the square as shown in Figure 3.1.

Figure 3.1: Graphical Representation of 22

designs
 Fractional Factorial Designs
Wollega University

 For large numbers of treatments (k), the total number of runs

for a full factorial can get very large (2k)
 Fractional factorial designs are helpful when:
 High-order interactions are small/ignorable
 We wish to “screen” many factors to find a small set of
important factors, to be studied more thoroughly later
 Resources are limited
 These fractional factorial designs are used to identify factors
which have large effects.
 When only a fraction of a complete replication is run, the
design is called a fractional replication or fractional
factorial.
 Fractional Factorial Designs
Wollega University

 For running a fractional factorial replication, a confounding

scheme is used to run in two blocks and only one of the blocks
is run. This forms the one-half fraction of 2k design.
 Consider a 23 design with three factors A, B and C. One
complete replication of this design consists of 8 runs.
 Suppose the experimenter cannot run all the 8 runs and can
conduct only 4 runs.
 This becomes a one-half fraction of a 23 design. Since this
design consists of 23 -1 = 4 treatment combinations, a one-half
fraction of 23 design, called 23-1 design, where 3 denotes the
number of factors and 2-1= 1/2 denotes the fraction.
 Fractional Factorial Designs
Wollega University

 Suppose we select the treatment combinations corresponding to the + sign

under ABC as the one-half fraction, shown in the upper half of Table 3.2.
 The resulting design is the one-half fraction of the 23 design and is
called 23 -1 design (Table 3.2). Since the fraction is formed by
selecting the treatments with +ve sign under ABC effect, ABC is
called the generator of this fraction.
 Further, the identity element I is always +ve, so we call
I = +ABC, the defining relation for the design

Table 3.1: One half-

fraction in 23 Design
 Fractional Factorial Designs
Wollega University

 The linear combination of observations used to estimate the

effects of A, B and C are:

 Similarly, the two factors interactions are estimated from:

 From the above, we find that

 Blocking and Confounding
 Sometimes, it is impossible to perform all 2k factorial experiments
Wollega University

under homogeneous condition. A batch of raw material: not large

enough for the required runs
 Blocking is a technique for dealing with controllable nuisance
variables. Blocking is used when resources are not sufficient to
conduct more than one replication and replication is desired.
 There are two cases of blocking, i.e., replicated designs and un
replicated designs
 It also noise reduction technique.
 If there are n replicates, each replicate is treated as a block.
 Each replicate is run in one of the blocks (time periods, batches of
raw materials etc.).
 Runs within the blocks are randomized.
 (Illustration 3.3).
 Consider the chemical process experiment of 22 design discussed in
Wollega University

(Illustration 3.2).
 The two replications of this experiment are given in Table 3.3.

 Suppose one batch of material is just enough to run one replicate. Each
replication is run in one block. Thus, we will have two blocks for this
problem as shown in Figure 3.3.
 From Figure 3.3 the block totals are Block 1 (B1) = 179 and Block 2 (B2)
= 184.
 (Illustration 3.3).
Wollega University

 The analysis of this design is same as that of 22 design. Here, in addition to

effect sum of squares, we will have the block sum of squares.
 Blocking and Confounding
 Confounding is a design technique for arranging a complete
Wollega University

replication of a factorial experiment in blocks.

 In this technique certain treatment effects are confounded
(indistinguishable) with blocks.
 Usually, higher order interactions are confounded with blocks.
 When an effect is confounded with blocks, it is indistinguishable from the blocks.
 In 2k factorial design, 2m incomplete blocks are formed, where
m < k.
 Incomplete block means that each block does not contain all
treatment combinations of a replicate.
 These designs results in two blocks, four blocks, eight blocks
and so on
 Blocking and Confounding
 Confounding is a design technique for arranging a complete
Wollega University

replication of a factorial experiment in blocks.

 In this technique certain treatment effects are confounded
(indistinguishable) with blocks.
 Usually, higher order interactions are confounded with blocks.
 When an effect is confounded with blocks, it is
indistinguishable from the blocks.
 In 2k factorial design, 2m incomplete blocks are formed, where
m < k.
 Incomplete block means that each block does not contain all
treatment combinations of a replicate.
 These designs results in two blocks, four blocks, eight blocks
and so on
 Quiz-3
1. What are the assumptions made in the 2k factorial designs?
Wollega University

2. When Fractional factorial designs are helpful?

3. Define confounding and Blocking
4. What does it mean Incomplete block?
5. What are the two cases of blocking?