See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/323705387

Design of experiments application, concepts, examples: State of the art

Article  in  Periodicals of Engineering and Natural Sciences (PEN) · December 2017

DOI: 10.21533/pen.v5i3.145

CITATIONS READS

170 27,477

1 author:

Benjamin Durakovic
International University of Sarajevo
44 PUBLICATIONS   544 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

All-season PCM glazing system View project

All content following this page was uploaded by Benjamin Durakovic on 14 April 2021.

The user has requested enhancement of the downloaded file.

Periodicals of Engineering and Natural Scinces ISSN 2303-4521
Vol 5, No 3, December 2017, pp. 421‒439
Available online at: http://pen.ius.edu.ba

Design of Experiments Application, Concepts, Examples:

State of the Art

Benjamin Durakovic
Industrial Engineering, International University of Sarajevo

Article Info ABSTRACT

Article history: Design of Experiments (DOE) is statistical tool deployed in various types of
system, process and product design, development and optimization. It is
Received Aug 3th, 2018 multipurpose tool that can be used in various situations such as design for
Revised Oct 20th, 2018 comparisons, variable screening, transfer function identification, optimization
Accepted Dec 1th, 2018 and robust design. This paper explores historical aspects of DOE and provides
state of the art of its application, guides researchers how to conceptualize, plan
and conduct experiments, and how to analyze and interpret data including
Keyword:
examples. Also, this paper reveals that in past 20 years application of DOE have
Design of Experiments, been grooving rapidly in manufacturing as well as non-manufacturing
full factorial design; industries. It was most popular tool in scientific areas of medicine, engineering,
fractional factorial design; biochemistry, physics, computer science and counts about 50% of its
product design; applications compared to all other scientific areas.
quality improvement;
Corresponding Author:
Benjamin Durakovic
International University of Sarajevo
Hrasnicka cesta 15
7100 Sarajevo, Bosnia
Email: [email protected]

1. Introduction
Design of Experiments (DOE) mathematical methodology used for planning and conducting experiments as
well as analyzing and interpreting data obtained from the experiments. It is a branch of applied statistics that is
used for conducting scientific studies of a system, process or product in which input variables (Xs) were
manipulated to investigate its effects on measured response variable (Y).
Over past two decades, DOE was a very useful tool traditionally used for improvement of product quality and
reliability [1]. The usage of DOE has been expanded across many industries as part of decision-making process
either along a new product development, manufacturing process and improvement. It is not used only in
engineering areas it has been used in administration, marketing, hospitals, pharmaceutical [2], food industry [3],
energy and architecture [4] [5], and chromatography [6]. DOE is applicable to physical processes as well as
computer simulation models [7].

1.1. Historical perspective

One Factor At a Time (OFAT) was very popular scientific method dominated until early nineteen century. In
this method one variable/factor is tested at a time while the other variables are constrained except the
investigated one. Testing multiple variables at a time is better especially in cases where data must be analyzed

DOI: 10.21533/pen.v5i3.145 421

B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

carefully. In the 1920s and 1930s Ronald A. Fisher conducted a research in agriculture with the aim of
increasing yield of crop in the UK. Getting data and was challenging e.g. if he relayed on his traditional method
ANOVA (F-test, means Fisher - test) he may plant a crop in spring and get results in fall which is too long for
getting data. Finally, he came up with design of experiment and officially he was the first one who started using
DOE. In 1935, he wrote a book on DOE, in which he explained how valid conclusion could be drawn from the
experiment in presence of nuisance factors. He analyzed presence of nuisance factors with fluctuation of weather
conditions (temperature, rainfall, soil condition). Credit for Response Surface Method (RSM) belongs to
George Box who is also from the UK. He was concerned with experimental design procedures for process
optimization. In 1550s, W. Edwards Deming was concerned with design of experiment as well as statistical
methods. Genichi Taguchi was Japanese statistician concerned with quality improvement methods. He
contributed to statistic by introducing Loss function and experiments extending with an "outer array" in DOE
as an advanced method in the Six Sigma initiatives [8].

1.2. The main uses of DOE

Design of experiment is multipurpose tool that can be used in various situations for identification of important
input factors (input variable) and how they are related to the outputs (response variable). Therefore, DOE mainly
uses "hard tools" as it was reported in [9]. In addition, DOE is bassically regression analysis that can be used in
various situations. Commonly used design types are the following [10]:
1. Comparison ‒ this is one factor among multiple comparisons to select the best option that uses t‒test,
Z‒test, or F‒test.
2. Variable screening ‒ these are usually two-level factorial designs intended to select important factors
(variables) among many that affect performances of a system, process, or product.
3. Transfer function identification ‒ if important input variables are identified, the relationship between
the input variables and output variable can be used for further performance exploration of the system,
process or product via transfer function.
4. System Optimization ‒ the transfer function can be used for optimization by moving the experiment
to optimum setting of the variables. On this way performances of the system, process or product can be
improved.
5. Robust design ‒ deals with reduction of variation in the system, process or product without elimination
of its causes. Robust design was pioneered by Dr. Genichi Taguchi who made system robust against
noise (environmental and uncontrollable factors are considered as noise). Generally, factors that cause
product variation can be categorized in three main groups:
• external/environmental (such as temperature, humidity and dust);
• internal (wear of a machine and aging of materials);
• Unit to unit variation (variations in material, processes and equipment).

1.3. DOE application in research and procedure

Even though DOE tool are not new techniques but its application has expended rapidly over the scientific areas
including product/process quality improvement [11], product optimization [12] and services in the past two
decades. Trainings and recent user-friendly commercial and non-commercial statistical software packages
contributed significantly to DOE expansion in the research in this period. DOE application over the globe and
across various scientific areas in the period from 1920 to 2018 is shown in Figure 1.

422
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

90000

80000

70000

60000
Publications

50000

40000

30000

20000

10000

Year

Figure 1. DOE application in scientific research [13] 1

Application of DOE in research started in 1920s with Fisher's research in agriculture. Over four upcoming
decades later its application in the research was negligible. Significant use of DOE in the research project was
noticed in the late 1960s and 1970s. Thus, it took about 50 years for the DOE to achieve significant application
in the research. Since in this period there were no software packages that would foster its application DOE had
not signified a strong expansion. Thanks to edication and software development in 1990s and later, the use of
DOE in research over various scientific areas has risen sharply. A linear model that represents a rapid increase
in the use of DOE in the research projects is shown in Figure 2 and represented by a mathematical linear model.

100000
90000 y = 2960.3x - 6E+06
80000 R² = 0.9708

70000
Publications

60000
50000
40000
30000
20000
10000
0

Year

Figure 2. Progressive use of DOE as scientific method over past two decades [13]2

1 Data obtained from Scopus for search “design of experiments" OR "experimental design" OR "DOE" in Title ‒ abstact ‒ Key words
2 Data obtained from Scopus for search “design of experiments" OR "experimental design" OR "DOE" in Title ‒ abstact ‒ Key words
423
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

DOE application in the future can be predicted with linear regression model, which is based on past date over
18 years. Therefore, it can be expected that DOE usage expansion will continue in the future including its
application in the existing and new scientific areas. The state of the art of DOE application in certain scientific
areas is shown in Figure 3.

Medicine 18%
Engineering 10%
Biochemistry & Genetics 10%
Physics & Astronomy 7%
Computer Science 6%
Social Science 5%
Materials 4%
Agricultural 4%
Chemistry 4%
Mathematics 4%
Earth & Planetary 3%
Environmental Science 3%
Arts & Humanities 3%
Neuroscience 2%
Immunology & Microbiology 2%
Chemical Engineering 2%
Pharmacology 2%
Psychology 2%
Business Management 2%
Economics Econometrics 2%
Energy 1%
Nursing 1%
Health Professions 1%
Decision Sciences 1%
Multidisciplinary 1%
Veterinary 0.4%
Dentistry 0.3%
Undefined 0.2%

0 100000 200000 300000 400000 500000 600000

publications

Figure 3. DOE application per scientific area [13]3

DOE as scientific method was most popular in scientific areas of medicine, engineering, biochemistry, physics
and computer science. Its application in these areas counts about 50% compared to all other scientific areas.
Only medicine participates about 18%, while engineering and biochemistry together participate with 20%; and
physics and computer together science participate with 13%.
General practical steps and guidelines for planning and conducting DOE are listed below:

1. State the objectives ‒ it is a list of problems that are going to be investigated.

2. Response variable definition ‒ this is measurable outcome of the experiment that is based on
defined objectives.
3. Determine factors and levels ‒ selection of independent variable (factors) that case change in the
response variable. To identify factors that may affect the response variable fishbone diagram might
be used.
4. Determine Experimental design type ‒ e. g. a screening design is needed for significant factors
identification; or for optimization factor-response function is going to be planned, number of test
samples determination.
5. Perform experiment using design matrix.

3 Data obtained from Scopus for search “design of experiments" OR "experimental design" OR "DOE" in Title ‒ abstact ‒ Key words
424
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

6. Data analysis using statistical methods such as regression and ANOVA.

7. Practical conclusions and recommendations ‒ including graphical representation of the results
and validation of the results.
Most difficult part of DOE is to plan experiment in term of selecting appropriate factors for testing (x‒variables),
what ranges of x's to select, how many replicates is supposed to be used, and is center point required?

1.4. DOE Software

DOE can be quickly designed and analyzed with the help of suitable statistical software. For this purpose, there
are some commercial and freeware statistical packages. The well-known commercial packages include: Minitab,
Statistica, SPSS, SAS, Design-Expert, Statgraphics, Prisma, etc. The well-known freeware packages are: R and
Action for Microsoft Excel [14].
The most popular commercial packages Minitab and Statistica are equipped with user friendly interface and
very good graphics output. Freeware package Action has suitable graphics output and utilizes R platform
together with Excel
Also, DOE design and analysis can be done easily in Microsoft Excel, using the procedure and formulas
described in the following paragraphs. To perform any DOE as it was described earlier, understanding of
analysis of variance (ANOVA) and linear regression as statistical methods is required. Therefore, more details
about these two statistical methods are provided below.

2. Statistical Background
2.1. Analysis of variance (ANOVA)
In cases that there are more than two test samples ANOVA is used to determine whether there are statistically
significant differences between the means the samples (treatments). In cases that experiment contains two
samples only, then t-test is good enough to check whether there are statistically significant differences between
the means of treatments. In this case it is tested hypothesis assuming that a least one mean treatment value (μ)
differs from the others. Therefore, null and alternative hypotheses can be express as [15]:

H0: μ1 = μ1 = ... = μk = 0
(1)
H1: μ j ≠ 0 for at least one j different than zero.

The procedure of test involves an analysis of variance (ANOVA) and performing F-test. Observed value is
calculated as the ratio between treatment mean squares (𝑀𝑆𝑇𝑟 ) and error mean squares 𝑀𝑆𝐸 (error variance):

𝑆𝑆𝑇𝑟
𝑀𝑆𝑇𝑟 𝑎−1
𝐹𝑜 = = (2)
𝑀𝑆𝐸 𝑆𝑆𝐸
𝑎(𝑛 − 1)

where, 𝑆𝑆𝑇𝑟 is sum of squares of treatment, 𝑆𝑆𝐸 is sum of squares of error, (𝑎 − 1) represents treatments degrees
of freedom, 𝑎(𝑛 − 1) represents error degrees of freedom, a is number of treatments (number of samples), n is
number of observation for particular treatment. Total sum of squares (𝑆𝑆𝑇 ) is addition of sum of squares of
treatment and sum of squares of error and it is calculated as:
𝑎 𝑛 𝑎 𝑎 𝑛
2 2 2
𝑆𝑆𝑇 = 𝑆𝑆𝑇𝑟 + 𝑆𝑆𝐸 ; 𝑜𝑟 ∑ ∑ (𝑦𝑖𝑗 − 𝑦̿) = 𝑛 ∑(𝑦̅ 𝑖 − 𝑦̿) + ∑ ∑ (𝑦𝑖𝑗 − 𝑦̅ 𝑖 ) (3)
𝑖=1 𝑗=1 𝑖=1 𝑖=1 𝑗=1

425
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

whre, 𝑦𝑖𝑗 is 𝑗-th observation taken in treatment i, 𝑦̿ is an overall average (grand mean) of all observation for
each material group i, 𝑦̅𝑖 is an average value of observation in treatment i, a is number of treatments (groups),
n is number of observations per each treatment. All of them are shown in Table 1 as well.
Table 1. General data matrix for a single-factor ANOVA
Material group Observations Averages
1 𝑦11 𝑦12 … 𝑦1𝑛 𝑦̅1
2 𝑦21 𝑦22 … 𝑦2 𝑦̅2
⋮ ⋮ ⋮ … ⋮ ⋮
𝒂 𝑦𝑎1 𝑦𝑎2 … 𝑦𝑎𝑛 𝑦̅𝑎
∑𝑎𝑖=1 𝑦̅𝑖
𝑦̿ =
𝑎
Having calculated Fo, H0 can be accepted or rejected in the following cases:

Rejected if 𝐹𝑜 > 𝐹𝛼,(𝑎−1),𝑎(𝑛−1)

H0 is { (4)
Accepted if 𝐹𝑜 < 𝐹𝛼,(𝑎−1),𝑎(𝑛−1)

H0 is going to be rejected if observed value of 𝐹𝑜 is grater than its critical value 𝐹𝛼,(𝑎−1),𝑎(𝑛−1) . The critical
value is taken from corresponding statistical table for significance level α, the degrees of freedom for the
numerator (a ‒ 1), and degrees of freedom for the denominator a(n ‒ 1). Also, H0 is going to be accepted if
observed value is lower than its critical value (𝐹𝑜 < 𝐹𝛼,(𝑎−1),𝑎(𝑛−1) ). If p-value approach for the statistics 𝐹𝑜 is
used than H0 is going to be rejected if p < α, and will be accepted if p > α.

2.2. Linear regression

A general multiple linear regression model with k repressor variables is express as:

𝑦 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + ⋯ + 𝛽𝑘 𝑥𝑘 + 𝜀 (5)

where, y is response variable also called dependant variable, βj, βj, j= 0. 1, 2, …, k, are regression coefficients
or parameters and represent expected change in response variable per unit change in input variable, xj, is
regressesor variable or input variable or independent variable, ε is random error, which supposed to be normally
distributed with N (0, σ2) for given values of x and y.
Parameter estimate in multiple linear regression models is done using least squares method. In case that there
are multiple observations (n) on the response variable y1, y2, … yn, and that there are observation at each input
variable xij, (i = 1, 2, …, n) than it can be represented as matrix in Table 2 an written as [16]:

𝑦𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + 𝛽2 𝑥𝑖2 + ⋯ + 𝛽𝑘 𝑥𝑖𝑘 + 𝜀𝑖 (6)

𝑦𝑖 = 𝛽0 + ∑ 𝛽𝑗 𝑥𝑖𝑗 + 𝜀𝑖 ; 𝑖 = 1, 2, … , 𝑛 (7)
𝑗=1

Table 2. Multiple linear regression model data

y x1 x2 … xk
y1 x11 x12 … x1k
y2 x21 x22 … x2k
… … … … …

yn Xn1 Xn2 … Xnk

426
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Referring to Table 2, it can be formulated using the general multivariate linear regression model in matrix form
as follows:

𝑌𝑛𝑥1 = 𝑋𝑛𝑥(𝑘+1) 𝛽(𝑘+1)𝑥1 + 𝜀𝑛𝑥1 (8)

where, Y denotes an (n x 1) vector of dependent variable observation, X is a n x (k+1) matrix of independent

variables, β is a ((k+1) x 1) vector of unobserved parameters of interest, and ε is an (n x 1) random errors vector.
Subscript n is a number of observations, subscript k is a number of independent variables. The general matrix
form of the model becomes as follows:

𝑦1 1 𝑥11 𝑥12 … 𝑥1𝑘 𝛽0 𝜀1

𝑦2 1 𝑥21 𝑥22 … 𝑥2𝑘 𝛽1 𝜀2
[ ⋮ ]=[ ][ ]+ [ ⋮ ] (9)
⋮ ⋮ ⋮ ⋱ ⋮ ⋮
𝑦𝑛 1 𝑥𝑛1 𝑥𝑛2 … 𝑥𝑛𝑘 𝛽𝑘 𝜀𝑛

After slowing the function that minimizes sum of squared errors ε2, the least squared estimator of 𝛽 can be
calculated from the following:

−1
𝛽̂ = (𝑋 ′ 𝑋) 𝑋 ′ 𝑌 (10)

Significance test of regression model is to determine existence of the significant relationship between input
variable and output variable. In this case it is tested hypothesis assuming that a least one 𝛽 parameter differs
from the others.

H0: β1 = β1 = ... = βk = 0
(11)
H1: βj ≠ 0 for at least one j different than zero.

The procedure of test involves an analysis of variance (ANOVA) and performing F-test. Observed value is
calculated as the ratio between regression mean squares (𝑀𝑆𝑅 ) and error mean squares 𝑀𝑆𝐸 (error variance):

𝑆𝑆𝑅
𝑀𝑆𝑅 𝑘
𝐹𝑜 = = (12)
𝑀𝑆𝐸 𝑆𝑆𝐸
𝑛−𝑘−1

whre, 𝑆𝑆𝑅 regression sum of squares, 𝑆𝑆𝐸 is error sum of squares, k is degree of freedom for the numerator, (n
‒ k ‒ 1) is degree of freedom for the denominator. Total sum of squares is addition of regression sum of squares
and error sum of squares and it is calculated as:
𝑛 𝑛

𝑆𝑆𝑇 = 𝑆𝑆𝑅 + 𝑆𝑆𝐸 = ∑(𝑦̂𝑖 − 𝑦̅ )2 + ∑(𝑦𝑖 − 𝑦̂𝑖 )2 (13)

𝑖=1 𝑖=1

whre, 𝑦̂𝑖 is the predicted value for the ith test, 𝑦̅ is the mean of all response variable observations, 𝑦𝑖 is the ith
observed value of the response variable.

Rejected if 𝐹𝑜 > 𝐹𝛼,𝑘,𝑛−𝑘−1

H0 is { (14)
Accepted if 𝐹𝑜 < 𝐹𝛼,𝑘,𝑛−𝑘−1

427
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

H0 is going to be rejected if observed value of 𝐹𝑜 is grater than its critical value 𝐹𝛼,𝑘,𝑛−𝑘−1 . The critical value is
taken from corresponding statistical table for significance level α, the degrees of fridom for the numerator k,
and degrees of freedom for the denominator (n ‒ k ‒ 1). Also, H0 is going to be accepted if observed value is
lower than its critical value (𝐹𝑜 < 𝐹𝛼,𝑘,𝑛−𝑘−1 ). If p-value approach for the statistics 𝐹𝑜 is used than H0 is going
to be rejected if p < α, and will be accepted if p > α.
Goodness of fit of a model is measured by coefficient of determination R2, which is a measure of how well real
data points are approximated with regression model. It measures how amount of reductions of variability in
response variable caused using input variables in regression model.

𝑆𝑆𝑅
𝑅2 = (15)
𝑆𝑆𝑇

Higher value of R2 means better fit. If R2 = 1 it means that there is perfect fit. However, having a higher value
dos not meat that there is a good fit and that regression model is good one because adding a new variable to the
model (either the variable is significant or not) will increase R2 value, which will lead to poor prediction. To
solve this, an adjusted R2 is introduced, which will not always increase with adding a new variable.

2 𝑆𝑆𝑅
𝑅𝑎𝑑𝑗 = (16)
𝑆𝑆𝑇

2
If non-significant variable is added to the model 𝑅𝑎𝑑𝑗 will decrease and vice versa.

3. Factorial Design
Factorial experiments can be design with one, two, three and more factors. Experiments with only one factor
are often called simple comparative experiments. In these cases, t‒test or ANOVA were used for analysis.
Factorial experiments with two factors (A and B) usually include two level factorial designs for identification
of factor effects on the response variable by investigating all possible combinations of the factor levels. The
factor effect is defined as change in the response variable by changing the level of the factor.
Factorial experiments with multiple factors (A, B, …, K), with two levels ("low" and "high") the complexity of
experimentation might be a problem. The number of possible combinations goes up with the number of factors,
for instance a 2-level design with 8 factor has 256 combination which very set such type of experiments and
analyze data. Multiple factor experiment requires a lot of resources, materials and it is time consuming and
expensive. Additional problem with multiple factorial design is to maintain experimental conditions unchanged
during a huge number of experiments.
Trying to overcome the problems with multiple factor factorial designs and depending from a case to case, it is
possible to be designed as Full Factorial Design 2k or Fractional Factorial Design 2k‒p. In this case number 2
represents number of levels, while k is number of factors and p is the fraction size of the full factorial used.
More details about full factorial and fractional factorial design with examples is provided below [17].

3.1. Full Factorial Design

A full factorial design is convenient for a low number of factors if the resources are available. Conceptual
approach for DOE is explained for two 22 and three 23 factors as well as general 2k factorial design, in which k
represents number of factors while number 2 represents number of levels. Uppercase letters A, B, C… are
usually used for factor designation while lowercase letters are used treatments. Each factor has two levels low
(‒) and high (+). Number of combinations for 22 is four, for 23 is eight and so on. Each combination is called
treatment which is represented with a lowercase letter. The number of test units for each treatment is called the
428
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

number of replicates. For example, if three units / samples were tested at each treatment, the number of
replicates is three.
Design matrix for a 22 factorial design is given in Table 3:
Table 3. design matrix and signs for effects in 22 factorial design

A B AB

‒1 ‒1 ‒1 +1
a +1 ‒1 ‒1
b ‒1 +1 ‒1
ab +1 +1 +1

where, A and B represent factors while AB represents interaction between A and B factor.
This design is an orthogonal. The following linear regression model is used for the analysis the analysis:

𝑦 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + 𝛽12 𝑥1 𝑥2 + 𝜀 (17)

Where, x1 and x2 represent coded variables of factors A and B respectively, while x1x2 cross‒product and
represents AB interaction, xj = ‒1 and xj = +1 represent low and high levels of factors.

1
𝑠𝑠 = (𝑐𝑜𝑛𝑡𝑟𝑎𝑠𝑡)2 (18)
𝑛22

Whre, contrast is obtained from Table 3 as sum of products of signs in corresponding column and the runs listed
in corresponding rows.

3.2. General 2k Factorial Design

Experimental data analysis was done using Design of Experiment (DOE) ‒ full factor factorial design.
Generalized case of a 2k factorial design is introduced and applied in this study, where k is number of factors at
𝑘 𝑘
two levels. Statistical model includes k main effects, ( ) two‒factor interaction, ( ) three‒factor interactions,
2 3
𝑘
…, ( ) one k‒factor interaction. The procedure for a 2 factorial design was the following:
k
𝑘
1. estimated factor effect ‒ effects are estimated and their magnitudes were examined with the aim of
important factors identification;
2. initial model formulation ‒ full model is included that takes in account all main effects and interactions;
3. statistical testing ‒ ANOVA is used to test significance of main effects and interactions;
4. model refinement ‒ non significant factors from initial model are removed;
5. residual analysis ‒ to check adequacy of the model and assumptions;
6. result interpretation ‒ graphical analysis of the results such as main effects, interactions etc.
General form of ANOVA mentioned in step 3 is shown in the following table [17]:

429
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Table 4. ANOVA for a general 2k design

Source of variation Sum of squares Degrees of Mean Square Fo
freedom
k main effects
A SSA 1 𝑆𝑆𝐴 𝑀𝑆𝐴
𝑀𝑆𝐴 = 𝐹𝑜 =
B SSB 1 𝑎−1 𝑀𝑆𝐸
… … …
K SSK 1
𝒌
( ) two‒factor
𝟐
interaction
AB SSAB 1 𝑆𝑆𝐴𝐵 𝑀𝑆𝐴𝐵
𝑀𝑆𝐴𝐵 = 𝐹𝑜 =
AC SSAC 1 (𝑎 − 1)(𝑏 − 1) 𝑀𝑆𝐸
… … …
JK SSJK 1
𝒌
( ) three‒factor
𝟑
interactions
ABC SSABC 1 … …
ABD SSABD 1
… … …
IJK SSIJK 1
… … …
𝒌
( ) k‒factor
𝒌
interaction
ABC…K SSABC…K 1 … …
Error SSE 2k(n‒1)
Total SST n2k‒1

To estimate sum of squares and effects, contrasts associated with the effect were determined first.

𝐶𝑜𝑛𝑡𝑟𝑎𝑠𝑡𝐴𝐵…𝐾 = (𝑎 ± 1)(𝑏 ± 1) … (𝑘 ± 1) (19)

where, a, b, …, k are experiment treatment combinations.

After computation of the contrast, the effects and the sum of squares can be estimated as follows:

2
𝐴, 𝐵 … 𝐾 = (𝐶𝑜𝑛𝑡𝑟𝑎𝑠𝑡𝐴𝐵…𝐾 ) (20)
𝑛2𝑘

1
𝑆𝑆𝐴,𝐵…𝐾 = (𝐶𝑜𝑛𝑡𝑟𝑎𝑠𝑡𝐴𝐵…𝐾 )2 (21)
𝑛2𝑘

where, n represents number of replicates.

430
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Total sum of squares is calculated using the following equation:

𝑎 𝑏 𝑛
2
𝑆𝑆𝑇 = ∑ ∑ ∑(𝑦𝑖𝑗𝑘 − 𝑦̅… ) = 𝑆𝑆𝐴 + 𝑆𝑆𝐵 + ⋯ + 𝑆𝑆𝐸 (22)
𝑖=1 𝑗=1 𝑘=1

where, 𝑦𝑖𝑗𝑘 is an observation for factor A at i-th level (i = 1, 2, … a) and for B factor at j-th level (j = 1, 2, …
b) and for k-th replicate (k = 1, 2, … n).

Sum of squares of error can be obtained from using the following equation:

𝑆𝑆𝐸 = 𝑆𝑆𝑇 − 𝑆𝑆𝑀 (23)

where, 𝑆𝑆𝑀 is model sum of squares and it is equal: 𝑆𝑆𝑀 = 𝑆𝑆𝐴 + 𝑆𝑆𝐵 + ⋯ + 𝑆𝑆𝐾 .

3.3. Fractional Factorial Design

If the number of factors is increased, then the number of test units and treatment combinations (runs) is going
to be increased. For example, for 5 factors, 32 units and treatments are needed (25=32). Since a low number of
main effects and lower order interactions are significant to the response variable, and usually higher order
interactions are not significant to the response variable, than fractional factorial designs are introduced.
Therefore, fractional factorial designs take in account only a low number of main effects and lower order
interactions. The higher order interactions are neglected due to its negligible effects on the response variable.
For instance, 23 = 8, and has 8 treatment combination and 8 test units is required. For some reason he cannot
afford all 8 combinations he decided to run one half factorial design 23‒1 = 4. Therefore, instead of 8 test units
he will test only 4. Now he needs to determine which four treatment combinations to test. As the first step, he
complete 23 full factorial design matrix shown in Table 5 and selected those with positive values in ABC column
as it shown in
Table 6.

Table 5. Full 23 factorial design matrix

I A B AB C AC BC ABC
(1) 1 -1 -1 1 -1 1 1 -1
a 1 1 -1 -1 -1 -1 1 1
b 1 -1 1 -1 -1 1 -1 1
ab 1 1 1 1 -1 -1 -1 -1
c 1 -1 -1 1 1 -1 -1 1
ac 1 1 -1 -1 1 1 -1 -1
bc 1 -1 1 -1 1 -1 1 -1
abc 1 1 1 1 1 1 1 1

431
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Table 6. The 23‒1 factorial design factorial design matrix with the defining relations I = ABC and I = ‒ABC
I A B AB C AC BC ABC
a 1 1 -1 -1 -1 -1 1 1
b 1 -1 1 -1 -1 1 -1 1
c 1 -1 -1 1 1 -1 -1 1
abc 1 1 1 1 1 1 1 1
(1) 1 -1 -1 1 -1 1 1 -1
ab 1 1 1 1 -1 -1 -1 -1
ac 1 1 -1 -1 1 1 -1 -1
bc 1 -1 1 -1 1 -1 1 -1

abc bc

c
+ + ac

C C

b ab
+ +
B B
‒ ‒ ‒ ‒
A a (1) A
(a) (b)
3‒1
Figure 4. The 2 factorial design factorial design orthogonal representation with defining relations: (a)
principal fraction I = ABC; (c) alternative fraction I = ‒ABC

In this example ABC is called the generator. It noticed that some columns in
Table 6 have the same values (they are the same) and they are colored in the same color (I = ABC, A = BC, B =
AC, C = AB), where I is called defining relation, which may be aliased with several effects. The alias can be
determined by using the defining relation I = ABC by multiplying any column by the defining relation yields
the aliases such as: A · I = A · ABC = A2BC. Since square of any column is equal to identity I, then it can be
written as: A = BC. The other aliases may be obtained, B · I = B · ABC = AB2C and C · I = C · ABC = ABC2, and
replacing squared member by I it is obtained B = AC, C = AB. Therefore, when A, B and C is estimated it is
actuality estimated A + BC, B + AC, C + AB.
Resolution a two-level fractional factorial design 23‒1 is called a resolution III design and it is equal to the
number of letters in the shortest word in the defining relation in which Roman numbers in subscript are used to
define resolution (e.g. 23−1 4−1
III design for I = ABC and I = ‒ABC; 2IV design for I = ABCD and I = ‒ABCD etc.).

Designs with a resolution less than III are never used. Fractional factorial designs with the highest available
resolution with their defining relations is shown in Table 7.

432
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Table 7. The 23‒1 factorial design factorial design matrix with the defining relations I = ABC and I = ‒ABC
Number of factors Fraction Resolution Runs (treatments) Defining relation
3 23‒1 III 4 I = ABC
4 24‒1 IV 8 I = ABCD
5 25‒1 V 16 I = ABCDE
6 26‒1 VI 32 I = ABCDEF
7 27‒1 VII 64 I = ABCDEFG

4. Application

4.1. The 24 Full Factorial Design

DOE a very suitable tool used for process performance and product quality improvement [18], for solar
technologies [19] etc. In this example DOE is applied on injection-molding process with the aim of improving
product quality such as excessive flash. Factors considered as affecting for flash formation are: pack pressure
(A), pack time (B), injection speed (C), and screw RPM (D), while clamping pressure, injection pressure and
melting temperature were under control. Each factor affecting flash formation is considered at low and high
levels and shown in Table 8.
Table 8. Factors contributing to flash formation
FACTORS low (‒1) High +1
Pack pressure (A) in bar 10 30
Pack time (B), in sec 1 5
Injection speed (C) in mm/sec 12 50
Screw RPM (D) in rpm 100 200

The design matrix for 24 factorial design and single replicate response data for flash size in millimeters is shown
in Table 9.
Table 9. Design matrix for flash formation
Run Run A B C D AB AC AD BC BD CD ABC BC AC AB ABCD Flash
label D D D (mm)
1 (1) -1 -1 -1 -1 1 1 1 1 1 1 -1 -1 -1 -1 1 0.22
2 a 1 -1 -1 -1 -1 -1 -1 1 1 1 1 -1 1 1 -1 6.18
3 b -1 1 -1 -1 -1 1 1 -1 -1 1 1 1 -1 1 -1 0
4 ab 1 1 -1 -1 1 -1 -1 -1 -1 1 -1 1 1 -1 1 5.91
5 c -1 -1 1 -1 1 -1 1 -1 1 -1 1 1 1 -1 -1 6.6
6 ac 1 -1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 1 1 6.05
7 bc -1 1 1 -1 -1 -1 1 1 -1 -1 -1 -1 1 1 1 6.76
8 abc 1 1 1 -1 1 1 -1 1 -1 -1 1 -1 -1 -1 -1 8.65
9 d -1 -1 -1 1 1 1 -1 1 -1 -1 -1 1 1 1 -1 0.46
10 ad 1 -1 -1 1 -1 -1 1 1 -1 -1 1 1 -1 -1 1 5.06
11 bd -1 1 -1 1 -1 1 -1 -1 1 -1 1 -1 1 -1 1 0.55
12 abd 1 1 -1 1 1 -1 1 -1 1 -1 -1 -1 -1 1 -1 4.84
13 cd -1 -1 1 1 1 -1 -1 -1 -1 1 1 -1 -1 1 1 11.55
14 acd 1 -1 1 1 -1 1 1 -1 -1 1 -1 -1 1 -1 -1 9.9
15 bcd -1 1 1 1 -1 -1 -1 1 1 1 -1 1 -1 -1 -1 9.9
16 abcd 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 9.9

433
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Applying Equations (19) and Equation (20) in the design matrix (Table 9) main effects are calculated and shown
in Table 10.
Table 10. Factor effect estimate and sum of squares
Model Effects Coefficients = SS Percent contribution
term (effect / 2)
A 2.5575 1.27 26.16323 0.12126637
B 0.061875 0.03 0.015314 7.0981E-05
C 5.76125 2.88 132.768 0.6153788
D 1.47125 0.73 8.658306 0.04013119
AB 0.466125 0.23 0.86909 0.00402823
AC -2.6345 -1.31 27.76236 0.12867835
AD -0.748 -0.374 2.238016 0.01037319
BC 0.218625 0.10 0.191188 0.00088615
BD -0.507375 -0.25 1.029718 0.00477273
CD 1.82325 0.91 13.29696 0.06163133
ABC 0.556875 0.27 1.240439 0.00574943
BCD -0.598125 -0.29 1.431014 0.00663274
ACD 2.22045E-16 1.11E-16 1.97E-31 9.1409E-34
ABD -0.130625 -0.06 0.068252 0.00031635
ABCD -0.067375 -0.03 0.018158 8.416E-05
Mean4 5.78
In this example there is a problem to apply F-test. Since there is only one replicate of the response variable F-
test is not possible to get done because error degrees of freedom is zero (MS E = SSE/0; ‒ division by zero is
undefined). Potential solution can be a graphical solution such as Normal probability diagram of effects 5, or
dropping entire factors from the model. Therefore based on Table 10, a graphical solution is used to represent
main effects and identification of those significant, which is shown in Figure 5.

120%

100% C
Normal % probability

A
80% CD
D
60%

40%

20%
AC
0%
-4 -2 0 2 4 6 8
effrcts

Figure 5. Normal probability plot of the effects

4
Mean response, calculated as average flash value.

434
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

From normal probability diagram of effects it is observed that A, C, D, are significant factors and AC, and CD
are significant interactions. Plots of the main effects of factor A, C, D as well as AC and CD interactions are
shown in Figure 6.

Main effect of A Main effect of C Main effect of D

8 10 8

6 8 6
6
4 4
4
2 2
2
0 0 0
0 1 2 3 0 1 2 3 0 1 2 3

a) b) c)
10 D‒ D+
C‒ C+
8 8

6 6
4
4
2
2
0
0 1 2 3 0
A 0 2
A 4

d) e)
Figure 6. Main effects and interactions

Refereeing to Figure 6 it is observed that all three main effects are positive indicating that the molding process
supposed to be performed on low level to minimize flash size. Alongside main effects there are significant
interactions. Since main effect are involved with significant interactions than they have no much meaning.
Screw RPM and injection speed have little meaning at high pack pressure but has better performances in
reducing flash size at low level. Therefore, the best setup for getting better product quality (reducing flash size),
is obtained when all factors A, C, D are at the low level.

Since factor B is not significant it is dropped from the model. All associated interactions with B factor are
dropped from as well. Therefore, the design becomes a 23 factorial design with two replicates so called hidden
replication (still has 16 responses). In this the mean case sum of squares of error (MS E) can be calculated
because degrees of freedom are different than zero (2k(n‒1) = 23(2‒1) = 8). ANOVA result for 23 factorial design
with two replicates is shown in Table 11.

435
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Table 11. Factor effect estimate and sum of squares

SS DF MSS Percent F0 Fcr = F0 > Fα=5%,ν1,ν2
contribution Fα=5%,ν1,ν2
A* 26.16 1 26.16 0.1212 29.47 5.32 Sign.
C* 132.76 1 132.76 0.6153 149.57 5.32 Sign.
D* 8.65 1 8.65 0.0401 9.75 5.32 Sign.
AC* 27.76 1 27.76 0.1286 31.27 5.32 Sign.
CD* 13.29 1 13.29 0.0616 14.97 5.32 Sign.
Model 208.64
(SSM)
Error (SSE) 7.10 8 0.88
Total (SST) 215.75 15 14.38
Where, total sum of squares (SST) is obtained using Equation (22), model sum of squares is obtained as SSM =
Σ SSA,B,C…K = SSA + SSC + SSD + SSAD + SSCD, while error sum of squares is obtained as SSE = SST ‒ SSM.
Regression model that represent predicted/fitted value for flash formation is given below, and includes only
significant factor and interactions.

𝑦̂ = 5.78 + 1.3𝑥1 + 2.9𝑥3 + 0.7𝑥4 − 1.3𝑥1 𝑥3 + 0.9𝑥3 𝑥4 (24)

Based on Equation (24), it is possible to perform diagnostic check if it is applied to the residuals.
Table 12. Factor effect estimate and sum of squares
Run label y ̂
𝒚 ̂
𝒆 = 𝒚−𝒚
(1) 0.22 -0.427625 0.647625

a 6.1875 4.764375 1.423125

b 0 -0.427625 0.427625
ab 5.9125 4.764375 1.148125
c 6.6 7.968125 -1.368125
ac 6.05 7.891125 -1.841125
bc 6.765 7.968125 -1.203125
abc 8.657 7.891125 0.765875
d 0.462 1.043625 -0.581625
ad 5.06 6.235625 -1.175625
bd 0.55 1.043625 -0.493625
abd 4.84 6.235625 -1.395625
cd 11.55 9.439375 2.110625
acd 9.9 9.362375 0.537625
bcd 9.9 9.439375 0.460625
abcd 9.9 9.362375 0.537625

Normal probability plot of residuals and residuals versus predicted value of flash size is shown in Figure 7.

436
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

100% 2.5
2
Normal % probability

80% 1.5
1
60%

Residuals
0.5
40% 0
-0.5 40 60 80 100
20% -1
-1.5
0%
-2
-4.00 -2.00 0.00 2.00 4.00
-2.5
Residuals Predicterd

Figure 7. Model adequacy checking

Residuals versus predicted value plot shows randomly dispersed data around the horizontal axis (predicted
values), therefore the variance has no tendency to change along predicted value axis indicating that linear
regression model is appropriate for the data.

4.2. The 24‒1 Fractional Factorial Design

Applying the 24‒1 fractional factorial design on the previous example with the highest possible resolution (IV)
and using defining relation I=ABCD it is obtained the 23 design, which is shown in Table 13.
Table 13. The 24−1
IV factorial design with the defining relation I = ABCD

Basic Design Flas Residual

h s
RU Run A B C D=AB AB AC A y 𝑦̂ Σ(yi ‒𝑦)2 𝑒
N label C D = 𝑦 − 𝑦̂
alias BC AC AB ABC BC BD CD
D D D
1 (1) -1 -1 -1 -1 1 1 1 0.22 -0.8975 30.4704 1.1175
2 ab 1 1 -1 -1 1 -1 -1 5.9 4.7375 0.0256 1.1625
3 ac 1 -1 1 -1 -1 1 -1 6.05 7.2125 0.0961 -1.1625
4 bc -1 1 1 -1 -1 -1 1 6.7 7.8175 0.9216 -1.1175
5 ad 1 -1 -1 1 -1 -1 1 5.1 6.2625 0.4096 -1.1625
6 bd -1 1 -1 1 -1 1 -1 0.55 1.6675 26.9361 -1.1175
7 cd -1 -1 1 1 1 -1 -1 11.5 10.382 33.1776 1.1175
5
8 abcd 1 1 1 1 1 1 1 9.9 8.7375 17.3056 1.1625
mean 5.74

Referring to Table 13 factors A, B and C, represent basic design factors that are aliased with three factor
interactions BCD, ACD and ABD respectively, while factor D is aliased with ABC three factor interaction.
The two factor interactions AB, AC and AD are re aliased with the other two factor interactions BC, BD and
CD respectively. Therefore, the relationships are: A = BCD, B = ACD, C = ABD, D = ABC, AB = BC, AC =
BD and AD = CD. Using Equation (20) and multiplying corresponding column with measured values (flash)
and summing up, a contrast is obtained for the certain column. Using Eq. (21), contrast is divided by 𝑛2𝑘 and
main effects are obtained for number of replicates n = 1, and number of factors k = 3 and shown in Table 14.

437
B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Table 14. Estimate of effects and aliases

Effect estimate Aliased structure
[A] = 1.995 [A] → A + BCD
[B] = 0.045 [B] → B + ACD
[C] = 5.595 [C] → C + ABD
[D] = 2.045 [D] → D + ABC
[AB] = 2.28 [AB] → AB + BC
[AC] = -3.12 [AC] → AC + BD
[AD] = -0.52 [AD] → AD + CD
Plot of main effects is shown in Figure 8.

1.00
0.90
C
0.80
AB
0.70
D
0.60
0.50 A
0.40
0.30

AC 0.20
0.10
0.00
-4 -2 0 2 4 6 8

Figure 8. Normal Probability plot of effects

From normal probability diagram of effects, it is observed that A, C, and D, are significant factors and AC, and
AB are significant interactions. Based on the above analysis it is possible to predict flash size using the following
regression model:
𝑦̂ = 5.74 + 1.99𝑥1 + 5.59𝑥3 + 2.04𝑥4 + 2.28𝑥1 𝑥2 − 3.12𝑥1 𝑥3 (25)

5. Conclusion
Thanks to software development, the use of DOE as scientific tool has increased rapidly in past 20 years in
manufacturing and non‒manufacturing industries over the world. It was most popular tool used by scientist in
medicine (with 18%), engineering and biochemistry (with 20%), physics and computer science (with 13%),
providing about 50% participation of these scientific areas compared to the all other scientific areas. The trend
of use of DOE is rapidly growing and it is expected to slow down in the near future for current scientific areas,
but also will expanded over new scientific areas and have rapid grow there.
Implementing valid and efficient factorial experiments provide quantitative data that can be used as support for
decision making during system, process and product design or improvement. Full factorial design and fractional
factorial design with examples of application of DOE in product quality improvement is presented with sep‒
by‒step procedures and result interpretation. This can be a very useful guide to experimenter how to design and
conduct experiments, and how to analyze and interpret data.

438
 B. Durakovic PEN Vol. 5, No. 3, December 2017, pp. 421–439

Refferences
[1] B. Duraković and H. Bašić, "Continuous Quality Improvement in Textile Processing by Statistical
Process Control Tools: A Case Study of Medium-Sized Company," Periodicals of Engineering and
Natural Sciences, vol. 1, no. 1, pp. 36-46, 2013.
[2] F. Paulo and L. Santos, "Design of experiments for microencapsulation applications: A review,"
Materials Science and Engineering: C, vol. 77, no. August, pp. 1327-1340, 2017.
[3] P. Yu, M. Y. Low and W. Zhou, "Design of experiments and regression modelling in food flavour and
sensory analysis: A review," Trends in Food Science & Technology, vol. 71, no. January, pp. 202-215,
2018.
[4] A. Schlueter and P. Geyer, "Linking BIM and Design of Experiments to balance architectural and
technical design factors for energy performance," Automation in Construction, vol. 86, no. February ,
pp. 33-43, 2018.
[5] B. Durakovic and M. Torlak, "Simulation and experimental validation of phase change material and
water used as heat storage medium in window applications," J. of Mater. and Environ. Sci., vol. 8, no. 5,
pp. 1837-1746, 2017.
[6] D. B. Hibbert, "Experimental design in chromatography: A tutorial review," Journal of
Chromatography B, vol. 910, no. Dec, pp. 2-13, 2012.
[7] S. S. Garud, I. A. Karimi and M. Kraft, "Design of computer experiments: A review," Computers &
Chemical Engineering, vol. 106, no. November, pp. 71-95, 2017.
[8] J. K. Telford, "A Brief Introduction to Design of Experiments," Johns Hopkins APL Technical Digest,
vol. 27, no. 3, pp. 224-232, 2007.
[9] B. Durakovic, H. Bašić and H. Muhič, "The Interrelationships between quality managment practicies
and their effects on innovation," in Trends in the Development of Machinery and Associated
Technology, Budapest, 2014.
[10] H. Guo and A. Mettas, "Design of Experiments and Data Analysis," in 2012 Annual Reliability and
Maintainability Symposium.
[11] B. Duraković and H. Bašić, "Textile Cutting Process Optimization Model Based On Six Sigma
Methodology In A Medium-Sized Company," Journal of Trends in the Development of Machinery and
Associated Technology, vol. 16, no. 1, pp. 107-110, 2012.
[12] B. Durakovic and M. Torlak, "Experimental and numerical study of a PCM window model as a thermal
energy storage unit," International Journal of Low-Carbon Technologies, vol. 12, no. 3, p. 272–280,
2017.
[13] "Scopus," 2017. [Online]. Available: www.scopus.com. [Accessed September 2017].
[14] D. Granato, V. M. d. A. Calado and B. Jarvis, "Observations on the use of statistical methods in Food
Science and Technology," Food Research International, vol. 55, p. 137–149, 2014.
[15] D. C. Montgomery, Statistical Quality Control, John Wiley, 2009.
[16] D. C. Montgomery and G. C. Runger, Applied Statistics and Probability for Engineers, Willey, 2005.
[17] D. C. Montgomery, Design and Analysis of Experiments, 5th edition, John Wiley and Sons, Inc. New
York, 2001., 8 ed., John Wiley & Sons, Inc, 2013.
[18] H. Bašić, B. Duraković and A. Softić, "Six Sigma Model Testing In Optimizing Medium-Sized
Company Production Process," Journal of Trends in the Development of Machinery and Associated
Technology, vol. 16, no. 1, pp. 103-106, 2012.
[19] V. Okatia, A. Behzadmehra and S. Farsad, "Analysis of a solar desalinator (humidification–
dehumidification cycle) including a compound system consisting of a solar humidifier and subsurface
condenser using DoE," Desalination, vol. 397, no. November, pp. 9-21, 2016.

439

View publication stats