Six Sigma Presentation

Presentation on Relevance of DOE (Design of Experiments)

Presented by :-
Naveen Gupta
Anshul Mathur
Girish Harsha
Rahul Kotai
Deepika Goyal
Ram Kishore Sankhla
 Introduction to DOE

• Design of experiments, or experimental design, is the

design of all information-gathering exercises where
variation is present, whether under the full control of
the experimenter or not.
 History of DOE
• Developed by Mr. R.A. Fisher in 1925
• This is quality control compared to statistical
process control
• Statistical process control is a passive method
in the sense that as the production process
continues, sample of output are taken and
plotted on the control charts to find whether
the process is in control
• Experimental design on the other hand as the
name suggests, is performed on a process to
improve its capability further by conducting a
series of experiments to study the change in
the output variables by bringing intentional
and measured changes in the input variable to
the process
 Benefits of Designed Experiments
Identifying the process variables (X and Y)
influence upon a quality characteristic Z
(Output)
Determining the
Determining the optimal
optimal level of the
level of the influence
influential
controllable process
Benefits of designed controllable process
variable X so as to have the
experiments variables X so as to
quality characteristic Z as
achieve minimal
close as possible to its
variability in the
desire (nominal) value
quality characteristic
Z

Determining the optimal level of the influential

controllable process variables X so that the
effects of uncontrollable process variables Y can
be minimized of the quality characteristic Z
 Factors, Levels, Responses and
Randomisation
• Factors: Variable that we want to study effect of is called factor
– For example exercise is a factor, diet is another factor. In medical
experiments where effect of some drugs are of interest then the drug is
a factor
• Levels: Several values of a factor is usually selected and these
values are called levels
– For example different exercise types are levels of the exercise factor.
• Response: Result of experiment or observations
• Randomisation: Individuals or subjects for experiments are chosen
randomly. I.e. if we have n combination of factor levels and we need
k replications of each combinations then we take from nk individuals
randomly (it could be tricky in many situations) and each of them is
assigned to the factor combination randomly with uniform
distribution
• Orthogonality: If we take level of one of the factors and sum over
other factor levels if the effects of others sum to zero then this
design is called orthogonal design. Orthogonal designs make
analysis of the result simpler
 Design of Experiment
• In general design of experiment should be done carefully. Usual
considerations for design are:
• First the purpose of experiment should be understood carefully
– What is compared against what? Do you know anything about the system under
study? Do you know known variations between different subjects of the system (it
could be either factors you are comparing or subject of experiments). What would
you get as a result of experiment.
• Randomisation
– Care should be exercised when carrying out experiments. If it is not done with
care then there could be variation or similarity just because of the subjects of
study. For example by studying people in one city you cannot make decision
about the whole world.
• Blocking
– To remove variations due to known or suspected effects you may do experiments
in blocks.
• Replication
– Since every experiment has its own error it is always good idea to replicate each
experiment. If this is done with care then effects of random fluctuations due to
experiment can be reduced
 Factorial design
Let us assume that we have two factors and for each of them we set three
levels. For example three exercises and three diets. Exercise is one
factor and different exercises are levels of this factor, diet is the second
factor and different diets are levels of this factor. The number of all
combinations is 3x3=9. Two way crossed or three level, two factor
factorial design.
Ex1 Ex2 Ex2

Diet1 Observations Observations Observations

Diet2 Observations Observations Observations

Diet3 Observations Observation Observations

 Block design
• In many cases before designing the experiment we know that there are
some parameters that affect the result of experiment in the same way
– For example if we are interested in effect of some substance to tree growth
then we know that such factors as time of year (e.g. summer and winter),
location of where tree grows (e.g. England and France) will have some
systematic effect
– If we would plant trees randomly over time and location then variation due to
time and location would mask out the effect of substances we want to study
• When we want block design then block becomes one or more additional
factors
– For example if we want to remove variation due to location we can repeat
exactly same experiment on different location
• Now we have one factor with 4 levels and one parameter for blocking
with 3 levels then for randomised block design we would have 4x3
matrix for the experiment. Randomisation takes place within blocks
• General rule: block whatever you can, randomise the rest
• Blocking should be done with care. If effects of block factors is not as
expected then the results would be less reliable
 Latin square designs
• Latin square are kxk squares where on each column and
each row numbers from 1 to k appear only once. For
example 3x3 Latin square could look like.
1 3 2
2 1 3
3 2 1
• This type of designs are usually used for block designs.
Where rows and columns are parameters we want to block
and each cell is the level of the factor we are interested in.
• There are higher level of extension of Latin squares –
– for three dimension Graeco-Latin and for four dimension Hyper-
graeco-Latin squares
 Interactions

If we are testing two or more factors then it is important to consider

interaction terms first. So the first question we should ask if there is
an interaction between factors. As it was mentioned above the
model for two factor cases will be:
 Interactions
If we are testing two or more factors then it is important to consider interaction
terms first. So the first question we should ask if there is an interaction
between factors. As it was mentioned above the model for two factor cases
will be:
Non-Interaction
Whiteness

A B C
 Interactions
If we are testing two or more factors then it is important to consider interaction
terms first. So the first question we should ask if there is an interaction
between factors. As it was mentioned above the model for two factor cases
will be:
Non-Interaction Interaction

Whiteness
Whiteness

A B C A B C
 Nested (hierarchical) design
•Sometime it happens that levels of one factor has
nothing to do with that of another factor Top
•For example if we want to test performance of
schools then we randomly select several schools
and from each school we choose several classes
and consider results in maths Factor
•Each class in each school was taught by some levels
teacher. It is unlikely that the same teacher
taught in several schools
Sublevels of factor
•So it would not be reasonable to consider effect
levels
of school and classes as additive (class from one
school has nothing to do with that from another
school). This type of experiments are called
nested designs

In factorial designs (crossed designs) we analyse interactions between different factors

first and try to remove them. In nested design we want to know interactions.
 Examples
www.examples.com

Laziz spice products is a prominent player in the branded spice

market in the northern India. It is considering changing the
packaging of its popular chilli powder from plain card paper to
glazed card paper. It is not sure about the effect of this change. It
decides to conduct an experiment by choosing twelve retail outlets
having similar sales figures. The chosen outlets are divided into two
groups of six outlets each. One group of outlets was stocked with
chilli powder having plain packaging, while the other group was
stocked with chilli powder having glazed packaging. All the outlets
were monitored for one month for proper display of chilli powder in
the outlets. At the end of the month, sales figures were collected
from all the outlets and were tabulated as shown in rows 3 to 10 of
the spreadsheet
 Tools for doing Experimental Design
• Annova Test
– One Way Annova
– Two way Annova (With Replication)
– Two way Annova (Without Replication)
Factorial Experiment Design
– 22 Factorial Experimental Design
– Taguchi Method in Experimental Design
2^3 Full Factorial Array (Similar to the Taguchi L8 Orthogo