Fractional Factorial Design

(Chapter 6)
Agenda

Chapter 5:
• Short Recap of Two Level Factorial Design
– The Replicated Design
– Regression model

Chapter 6:
• Introduction to Fractional Factorial Design
• The One-Half Fraction of the Design
• Smaller Fractions - Fractional Factorial Designs

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Recap - Example - Factorial Design
Design factors Low level High level (+)
(−)
A: Catalyzing 𝛼 𝛽
substance
B: Concentration 5% 8%
C: Temperature 80℃ 90℃

Experiment # Treatment 𝑨 𝑩 𝑪 Result, 𝒚

Combination
1 (1) − (𝛼) − (5%) − (80℃) 89 g
2 𝑎 + (𝛽) − (5%) − (80℃) 84 g
3 𝑏 − (𝛼) + (8%) − (80℃) 131 g
4 𝑎𝑏 + (𝛽) + (8%) − (80℃) 130 g
5 𝑐 − (𝛼) − (5%) + (90℃) 124 g
6 𝑎𝑐 + (𝛽) − (5%) + (90℃) 121 g
7 𝑏𝑐 − (𝛼) + (8%) + (90℃) 116 g
8 𝑎𝑏𝑐 + (𝛽) + (8%) + (90℃) 113 g

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Short Recap – Two Level Factorial Design
• Design matrix:
Exp.
A B C AB AC BC ABC
No
1    + + +  89
2 +     + + 84
3  +   +  + 131
4 + +  +    130
5   + +   + 124
6 +  +  +   121
7  + +   +  116
8 + + + + + + + 113
-3 18 10 1 0 -26 -1
∑𝒊 𝒚 𝒊 ∑𝒊 𝒚 𝒊
Mean effect
𝒏/𝟐 𝒏/𝟐

• The Replicated Case: Blackboard

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Regression Model for the Result
• The analysis (from our previous example) tells us that the factors B and
C together with the interaction BC are active
• The construction of the contrasts implies that the following regression
model can be used to describe the experimental result
the average effect of the contrast
average result over all experiments
corresponding factor variable that can be high (+1) or low (-1),

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Regression Model for the Result

• Notice that: The regression coefficient is measuring the effect when

is increased by one unit. However, in factorial experiments is
increased from to , i.e., two units. Therefore, we must divide by
2 in the regression model

• Ex: 𝑩 𝑪 𝑩 𝑪

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Fractional (or Reduced) Factorial Design
(Chapter 6)

• Requires fewer experiments but (interaction) effects are

confounded cannot analyze/distinguish all interactions
• May save time and money if it is known that certain
interactions between factors are of no importance!
• Often used as a screening tool to figure out which factors
should be more carefully studied in a complete factorial
design of experiments
• Useful in quality improvement because often the results
are mainly influenced by a small number of factors
(Pareto’s law – the vital few and the trivial many)

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Fractional Factorial Design

• Suppose there are factors (A,B,...,J,K) in an experiment.

All possible factorial effects include
effects of order 1: A, B, ..., K (main effects)
effects of order 2: AB, AC, ....,JK (2-factor interactions)
effects of order 3: … Etc.
• Hierarchical Ordering principle
– Lower order effects are more likely to be important than higher
order effects
– Effects of the same order are equally likely to be important
• Pareto principle
– The number of important effects in a factorial experiment is small

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Example
Suppose you were designing a new forklift:

• Want to consider the following nine factors each with 2 levels

• 1. Engine Size; 2. Lifting chain; 3. Fork; 4. Weight; 5. Automatic vs
Manual; 6. Shape; 7. Tires; 8. Suspension; 9. Battery;
• Only have resources for conducting runs
– If you drop three factors for a full factorial design, those factor
and their interactions with other factors cannot be investigated
– Want investigate all nine factors in the experiment
– A fraction of full factorial design will be used
– Confounding (aliasing) will happen because using a subset
• How to choose (or construct) the fraction?

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
The One-Half Fraction of the Design

• Suppose there are four factors in an experiment (A, B, C

and D), each of 2 levels.
• Suppose the available resource is enough for conducting 8
runs
full factorial design consists of all the 16 level
combinations of the four factors
• We need to choose half of them
• The chosen half is called a fractional factorial design

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
All Combinations in a Full Design
A B C D
- - - -
+ - - -
- + - -
+ + - -
- - + -
+ - + -
- + + -
+ + + -
- - - +
+ - - +
- + - +
+ + - +
- - + +
+ - + +
- + + +
+ + + +

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Design matrix
Construct designs via “confounding” (aliasing)
• Select 3 factors (e.g. A, B, C) to form a full factorial design
• Confound (alias) D with a high order interaction of A, B and C. For
example, D = ABC

A B C AB AC BC D=
ABC

- - - + + + -

+ - - - - + +

- + - - + - +

+ + - + - - -

- - + + - - +

+ - + - + - -

- + + - - + -

+ + + + + + +

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Defining Relation

• The chosen level combinations form a half of the design

• The product of columns A, B, C and D equals 1, i.e.,

which is called the defining relation, and ABCD is

called a defining word (contrast)
is determined by the relation

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Defining Relation

• Multiply the defined equality with the aliased factor:

• That is:

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Design 1: Alias Structure for with

•
•
By the same procedure:

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Design 2: Alias Structure for with

The defining relation generates another fractional

factorial design, denoted by design 2. Its alias structure is given by:

 Recall that design 1 is defined by

 Comparing design 1 and design 2, which one we should
choose?
Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Word Length and Resolution

• Length of a defining word is defined to be the number

of the involved factors
• Resolution of a fractional factorial design is defined
to be the minimum length of the defining words,
usually denoted by Roman numbers, III, IV, V,...

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Resolution

• Design 1: is a resolution IV design denoted by

• Design 2: is a resolution III design denoted by
• A design is of resolution if no -factor effect is aliased with another
effect containing less than factors
– Design 1: main effects are not aliased with other main effects and
2-factor interactions
– Design 2: main effects are not aliased with main effects
• Design 1 is better than Design 2, because Design 1 has higher
resolution than Design 2

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Our Example with Design 1
Contrasts
BCD ACD ABD CD BD AD D
A B C AB AC BC ABC
Exp. No y
1    + + +  89
2 +     + + 84
3  +   +  + 131
4 + +  +    130
5   + +   + 124
6 +  +  +   121
7  + +   +  116
8 + + + + + + + 113
𝑬𝒋 -3 18 10 1 0 -26 -1

From the previous analysis statistically significant contrasts are:

• (B+ACD) with average effect = 18
• (BC+AD) with average effect = -26
• Potentially also (C+ABD) with average effect = 10

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Confounding Patterns for Reduced Factorial
Designs with Multiple Generating Relations

• Notation = number of factor levels

= number of factors
= number of generating relations
= shortest “word” in the defining relation
• Use a confounding table for reduced 2-factorial design of experiments
for example available in Table 6.22 (course book)
• Example: Assume that we want to investigate the impact of 7 different
factors with as few experiments as possible
• Need a design matrix with at least 7 contrasts one for each main factor
which implies that at least 8 experiments must be carried out
Requires a reduced factorial design

• The confounding table renders the design:

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Reduced Factorial Design Matrix
• Generating relations from confounding table (Table 6.22)

Contrasts

D E F G
Exp. No A B C AB AC BC ABC y
1    + + + 
2 +     + +
3  +   +  +
4 + +  +   
5   + +   +
6 +  +  +  
7  + +   + 
8 + + + + + + +

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Components of the Defining Relation
• From the generating relations we can directly obtain the initial
components of the defining relation (sometimes referred to as defined
equalities)
• D = AB I = ABD
• E = AC I = ACE
• F = BC I = BCF
• G = ABC I = ABCG
I = ABD = ACE = BCF = ABCG

• The complete Defining Relation DR is obtained by multiplying the initial

components of the defining relation with each other in all possible ways:
– (1) Pairwise: No of combinations

– (2) Three & Three: No of combinations

– (3) Four & Four: No of combinations

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Components of the Defining Relation
• (Multiplication Pairwise)
– (ABD)⋅(ACE) = BCDE
– (ABD) (BCF) = ACDF
– (ABD) (ABCG) = CDG
– (ACE) (BCF) = ABEF
– (ACE) (ABCG) = BEG
– (BCF) (ABCG) = AFG
• (Multiplication Three&Three)
– (ABD)⋅(ACE) (BCF) = DEF
– (ABD)⋅(ACE) (ABCG) = ADEG
– (ABD)⋅(BCF) (ABCG) = BDFG
– (ACE)⋅(BCF) (ABCG) = CEFG
• (Four&Four)
– (ABD)⋅(ACE) (BCF) (ABCG) = ABCDEFG

• The complete Defining set of relations, DR, is then:

DR: I = ABD= ACE= BCF= ABCG= BCDE = ACDF=
= CDG = ABEF = BEG = AFG = DEF = ADEG =
= BDFG = CEFG = ABCDEFG
Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics
Confounded interactions

• The interactions confounded on A can now be obtained by multiplying the

complete Defining Relation, DR, with A:
A⋅DR gives:
A = BD= CE= ABCF= BCG= ABCDE = CDF=
= ACDG = BEF = ABEG = FG = ADEF = DEG =
= ABDFG = ACEFG = BCDEFG
• The interactions confounded on the other main factors and thereby the
confounding patterns for all contrasts can be determined in the same manner.
𝑋 ⋅ 𝐷𝑅 ⇒ all factor interactions confounded on the main factor 𝑋
• By studying the Defining Relation, DR we can conclude:
– The shortest word length is 3 (i.e. at least three way interactions). We say
the experimental design has a resolution of 3
– Main factors will be confounded on no less than 2-way interactions.
𝑋 ⋅ 𝐷𝑅 ⇒ “smallest” interaction is between two factors 𝑋 𝑋

Fredrik Olsson / Lund University / Centre for Mathematical Sciences / Mathematical Statistics