Fractional Factorial

Design
Prof. Sayak Roychowdhury
Ref Book: Design and Analysis of Experiments by
D.C. Montgomery
Fractional Factorial
• In 26 full factorial design, there are 64 runs, only 6
dof of 63 dof are main effects, and 15 dofs
correspond to 2 factor interactions
• Rest of the 42 dofs are for 3 factor interactions or
higher.
• If higher order interactions can be reasonably
ignored, then fractional factorial designs are more
economical.
• FF designs are used for screening experiments
Fractional Factorial
• The successful use of fractional factorial design is based
on three key ideas:
• Sparsity of Effects: Out of several variables, the system or
process is likely to be driven by some of the main effects
and low-order interactions
• The projection property: FF designs can be projected
onto larger designs in the subset of significant factors
• Sequential experimentation: Combine two or more FF
designs to construct sequentially larger designs
 Factorial Designs
• Full Factorials: 𝑘 factors, 2 levels (high, low), 2𝑘 runs of each replicate
• We assume response is approximately linear over the range of factor levels.
Fractional Factorials:𝟐𝒌−𝒑 runs, where 𝒑 < 𝒌
• Regular Fractional Factorials (A and D optimal):
• All columns can be obtained by multiplying other columns
• Number of runs 𝑛 = 2𝑚 where 𝑚 is a positive integer (4,8,16,32…)
• Placket Burman Designs (A and D optimal):
• Invented by Placket and Burman (1946) with irregular fractional
factorial designs
• Number of runs are found in multiples of 4. (12,16,20,24…)
• Chosen based on availability
• Rule of Thumb: Pick smallest A-optimal design with enough columns
 𝑘
One-half Fraction 2 Design
• Consider ½ fraction of 23 design

Selected
fraction

• Selected only +ABC, so ABC is called generator

• 𝑰 = 𝑨𝑩𝑪 is called defining relation
 𝑘
One-half Fraction 2 Design
• Consider ½ fraction of 23 design A= BC; B=AC; C=AB

Selected
fraction

• Selected only +ABC, so ABC is called generator

• 𝑰 = 𝑨𝑩𝑪 is called defining relation
 𝑘
One-half Fraction 2 Design
• Estimation of main effects of A, B and C

• Estimation of interactions

• [A]= [BC], [B]=[AC], [C]=[AB] impossible to differentiate the

effects. This is called aliasing. A&BC, B&AC, C&AB are aliases.
• We are really estimating A+BC, B+CA, C+AB
 𝑘
One-half Fraction 2 Design
• In Minitab, the alias structure for the same design is
shown in the following way:
Aliases can be obtained by
multiplying the factor with I=ABC:
𝐴. 𝐼 = 𝐴. 𝐴𝐵𝐶 = 𝐵𝐶 and so on

𝐼 = +𝐴𝐵𝐶 is also called principal fraction

 Resolution of Design
Alias: When the estimate of an effect also include the influence of
one or more other effect. E.g. C=AB in 23−1 design.
A design is of resolution R if no p-factor effect is aliased with another
effect containing less than R - p factors
 Resolution of Design
• A design is of resolution R if no p-factor effect is aliased
with another effect containing less than R - p factors
• In general, the resolution of a two-level fractional factorial
design is equal to the number of letters in the shortest
word in the defining relation
• Fractional designs that have the highest possible
resolution consistent with the degree of fractionation
required are usually deployed
• The higher the resolution, the less restrictive the
assumptions that are required regarding which
interactions are negligible to obtain a unique
interpretation of the results
Construction of One-half Fraction
• One-half fraction of 2𝑘 design of highest resolution maybe
constructed by writing down full 2𝑘−1 factorial
• 𝑘 𝑡ℎ factor is added with + and – sign same as that of the highest
order interaction of ABC…(K-1)
• Below, factor C is added, with identical levels as of AB.
• Another option is to confound highest order interaction and
creating 2 blocks, each will be one-half fraction
Sequences of Fractional Factorials
• It is advisable to run a one-half fraction of experiment
first, analyse the result, and then run the other half
• For example if there are k = 4 factors, and you can afford
24 = 16 runs, it is better to run half fraction 24−1
𝐼𝑉 = 8
runs and then decide on the next runs based on the
results.
• Sequential experimentation lose information only for
highest order interactions
NPP of Effects Example

A B C D E F G Time
1 -1 -1 -1 -1 1 1 0.25
1 1 -1 1 -1 -1 -1 0.6
-1 -1 -1 1 1 1 -1 1.74
-1 -1 1 1 -1 -1 1 1.17
-1 1 -1 -1 1 -1 1 1.3
1 -1 1 -1 1 -1 -1 0.93
1 1 1 1 1 1 1 1.32
-1 1 1 -1 -1 1 -1 0.7
Steps for NPP of Effects

Factor Avg. at - Avg. at + Effect Coef. Est. "i " (i -.5)/7 Effect Factor =NORMINV((i -0.5)/7,0,1)
A 1.2275 0.775 -0.4525 -0.22625 1 0.071428571 -0.4525 A -1.4652323
B 1.0225 0.98 -0.0425 -0.02125 2 0.214285714 -0.0425 B -0.7916378
C 0.9725 1.03 0.0575 0.02875 3 0.357142857 0.0025 F -0.3661057
D 0.795 1.2075 0.4125 0.20625 4 0.5 0.0175 G 0
E 0.68 1.3225 0.6425 0.32125 5 0.642857143 0.0575 C 0.36610572
F 1 1.0025 0.0025 0.00125 6 0.785714286 0.4125 D 0.79163783
G 0.9925 1.01 0.0175 0.00875 7 0.928571429 0.6425 E 1.46523234

• Plot the ordered effects on X axis, the normal quantiles

obtained above on the Y axis
• Factors off the line are significant
Normal Probability Plot of Effects

NPP of Effects
2

1.5 E

1
D
0.5
C
0 G
-0.6 -0.4 -0.2 0 F 0.2 0.4 0.6 0.8
-0.5
B
-1

A -1.5

-2
Example
• A chemical product is produced in a pressure vessel. Four
factors are A: Temperature, B: Pressure, C: Concentration
of Formaldehyde, D: Stirring Rate.
• Response: Filtration Rate (to be maximized)
• Additional objective: Reduce formaldehyde concentration
Example 1
𝟐𝟒 Full Factorial Design

Source: Montgomery D.C., Design and Analysis of Experiments

Example 1
• 𝟐𝟒−𝟏
𝑰𝑽 Fractional Factorial Design with defining fraction
𝑰 = 𝑨𝑩𝑪𝑫
Source: Montgomery D.C., Design and Analysis of Experiments

𝑨. 𝑨𝑩𝑪𝑫 = 𝑩𝑪𝑫; 𝑩. 𝑨𝑩𝑪𝑫 = 𝑨𝑪𝑫; 𝑪. 𝑨𝑩𝑪𝑫 = 𝑨𝑩𝑫; 𝑫. 𝑨𝑩𝑪𝑫 = 𝑨𝑩𝑪

𝑨𝑩. 𝑨𝑩𝑪𝑫 = 𝑪𝑫; 𝑩𝑪. 𝑨𝑩𝑪𝑫 = 𝑨𝑫; 𝑪𝑫. 𝑨𝑩𝑪𝑫 = 𝐀𝐁; so on
Example 1

• Calculation of effects
• Normal probability
plot of effects
• ANOVA

𝐶𝑜𝑛𝑡𝑟𝑎𝑠𝑡 2
𝑆𝑆𝐹𝑎𝑐𝑡𝑜𝑟 =
𝑁
2 𝐶𝑜𝑛𝑠𝑡𝑟𝑎𝑠𝑡
𝐸𝑓𝑓𝑒𝑐𝑡 =
𝑁
Where N= number of runs
One-Quarter Fractional Factorial
• For moderately larger number of factors 𝑘 , 2𝑘−2 fractional
factorial is often useful
• This is full factorial with 𝑘 − 2 factors, 2 additional factors
added with appropriately chosen interactions with the first
𝑘 − 2 factors
• 2 generators 𝑃 and 𝑄, 𝐼 = 𝑃 and 𝐼 = 𝑄 are called generating
relations
• Complete defining relation 𝐼 = 𝑃 = 𝑄 = 𝑃𝑄 (𝑃𝑄 is
generalized interaction).
• Elements of defining relation 𝑃, 𝑄, 𝑃𝑄 are also called words
• Aliases of an effect are produced by multiplication of that
factor with each word
One-Quarter Fractional Factorial
• Aliases of an effect are produced by multiplication of that
factor with each word
• Example: In a 26−2 design 𝐼 = 𝐴𝐵𝐶𝐸 and 𝐼 = 𝐵𝐶𝐷𝐹 are
chosen as defining generator
• Complete defining relation 𝐼 = 𝐴𝐵𝐶𝐸 = 𝐵𝐶𝐷𝐹 = 𝐴𝐷𝐸𝐹
• Aliases :
𝐴 + 𝐵𝐶𝐸 + 𝐴𝐵𝐶𝐷𝐹 + 𝐷𝐸𝐹
𝐵 + 𝐴𝐶𝐸 + 𝐶𝐷𝐹 + 𝐴𝐵𝐷𝐸𝐹
etc.
• For 26−2 design, start with full factorial 24 design, then
put 𝐸 = 𝐴𝐵𝐶, 𝐹 = 𝐵𝐶𝐷 columns
 𝑘−𝑝
General 2 Design
• 2𝑘 factorial design containing 2𝑘−𝑝 runs is called 1/2𝑝
fraction of the 2𝑘 design, or 2𝑘−𝑝 fractional factorial
design
• 𝑝 independent generators to be chosen
• 2𝑝 − 𝑝 − 1 generalized interactions
• Each effect has 2𝑝 − 1 aliases
• Care should be taken so that effects of interest should not
be aliased with each other
• A reasonable criterion is the resulting 2𝑘−𝑝 design should
be of highest possible resolution
Plackett – Burman Designs
• 2 level fractional factorial designs developed by Plackett
and Burman to study 𝑘 = 𝑁 − 1 variables in 𝑁 runs
• For 𝑁 = 12, 20, 24, 28, 36 runs. These designs cannot be
represented as cubes, so they are called non-geometric
designs.
• Main effects may be partially aliased with 2 factor
interactions
• PB designs are non-regular designs. In regular designs, all
effects can be estimated distinctly, or completely aliased.
• In non-regular designs, some information on aliased
effects maybe available
Plackett – Burman Designs
• Write the appropriate row in the table as a column
• A second column is then generated from the first by moving the
element of the column down 1 position, and the placing the last
element of column 1 in first position of column 2.
• 3rd column is generated from 2nd in the same way
• Generate 𝑘 columns in this way
• A row of minus signs are then added
• This method is only for 12, 20, 24 and 36 runs
Plackett – Burman Designs
• PB designs have complex alias structure
• In a 12 run design, every main effect is partially aliased
with every 2-factor interactions, not involving itself
• PB is a non-regular design. In a regular design, all effects
can be either estimated independently of other effects or
they are completely aliased.
Choosing Fractional Factorial Designs

Preferred n=#runs # factors Design # columns

* 4 3 Reg. 3
4 PB 3
* 8 4-7 Reg. 7
8 PB 7
16 8-11 Reg. 15
* 12 PB 11
* 16 12-16 Reg. 15
16 PB 15
 Optimal Designs
• Volume of the joint confidence region that contains all the
model regression coefficients is inversely proportional to
the square root of the determinant of 𝑋 ′ 𝑋.
• Minimize the variance of estimates
• 𝑋 ′ 𝑋 – Information matrix
• A-optimal designs: Minimizes trace (sum of the elements
of main diagonal) of inverse of information matrix. It
minimizes the average variance of the estimates of the
regression coefficients
• D-optimal designs: Maximizes |𝑋 ′ 𝑋| , minimizes
generalized variance of the model regression coefficients.
 Optimal Designs
• G-optimal designs: Minimizes the maximum prediction
variance 𝑉(𝑦)
ො

• I-optimal designs: Minimizes average prediction variance.

• 2𝑘 designs are A,D, G and I optimal for first order model

and first order model with interactions.
Saturated and Supersatured
• When 𝑘 = 𝑁 − 1, where 𝑘 is the number of factors, and
𝑁 is the number of runs, it is called a saturated design
• Examples 23−1
𝐼𝐼𝐼 , 27−4
𝐼𝐼𝐼 , 15 factors 16 runs design
• When 𝑘 > 𝑁 − 1, the designs are called supersaturated
designs
• One plausible way of creating supersaturated design is to
use Hadamard Matrices (square orthogonal matrices)
• For more info, check section 8.8 (Montgomery, DoE book)
Resolution III Design
• Resolution III designs can be used up to 𝑘 = 𝑁 − 1 factors
in 𝑁 runs, where 𝑁 is a multiple of 4, e.g. 23−1
𝐼𝐼𝐼 , 27−4
𝐼𝐼𝐼
• 27−4
𝐼𝐼𝐼 is 1/16 th of 27

• Can be constructed first by writing the full factorial with A,

B, C, then associating 4 additional factors with aliases,
such as D=AB, E=AC, F=BC, G=ABC
Resolution III Design
• 27−4
𝐼𝐼𝐼 can be used to obtain other resolution III designs for
studying designs with fewer than 7 factors and 8 runs
• 26−3
𝐼𝐼𝐼 can be obtained just by dropping the column G
• Defining relation of 26−3𝐼𝐼𝐼 can be obtained from defining
relation of 27−4𝐼𝐼𝐼 with any word containing letter G
dropped
• It is possible to have resolution III design with 15 factors in
16 runs, 215−11
𝐼𝐼𝐼
• First write 16 treatment combinations for a full 24 design
with factors A,B,C,D and then equating the remaining 11
factors as 2
Resolution of Design
Resolution of Design
Randomize
Randomized Order: To balance the effects of
extraneous and uncontrollable factors.
Important to make valid statistical
inferences!
 Significant Factors
• Ways:
• Main effects plot, interactions plot
• Normal Probability Plot of Effects
• ANOVA (Analysis of Variance)
Steps for NPP of Effects
2

1.5 E

1
D
0.5

z score
0
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8

-0.5

-1

A -1.5

-2
Effects

More chances to find

insignificant factors for
noise estimation.
A Few Tips for Experimentation
• Use your non-statistical knowledge of the problem
• Especially valuable for choosing factors and levels
• Using statistics is no substitute for thinking about the problem
• Keep the design and analysis as simple as possible
• KISS – Keep it Simple, Stupid!
• more complex = more potential for error
• Complex experiments and analyses are often fraught with errors
• Know the difference between statistical and practical
importance
• Every statistically significant difference is not important or practical to
implement
• Experiments are usually iterative
• Our knowledge increases with time. Expect to use several experiments to
arrive at the optimum process.
• Rule of thumb: don’t use more than 25% of resources in 1st Expt.
P
g
3
Definitions for DOE
• Balanced Design (Orthogonality) – An experiment design where each
level of a factor is repeated the same number of times for all
possible combinations of the levels of the other factors. A balanced
design is said to be orthogonal
• Block – A group of homogeneous experimental runs
• Blocking Variable – A factor, usually a noise variable, that is an
undesirable source of variability. A group of experimental runs at a
single level of the blocking variable is a block.
• Experimental Design – The formal plan for conducting the
experiment. This includes the choices of factors, levels, responses,
blocks, sample sizes, repetitions, replications & randomization.
• Factors (or inputs) – One of the controlled or uncontrolled variables
whose influence on a response (output) is being studied in an
experiment.
• Interaction – When the difference in the response between levels of
one factor varies as the level of the other factors is changed.

P
g
3
 Definitions for DOE – continued
• k1 x k2 x k3 …-- The description of the basic design of a factorial
experiment. The number of “k’s” is the number of factors. The
value of the “k” is the number of levels of interest of that factor. For
example, a 2 x 3 x 3 experiment has three factors; one input has two
levels, and two have three levels. In this case, 2x3x3 = 18
combinations for this full factorial experiment.
• k-way Interaction – an interaction between k number of variables
• Level – Values of the factor being studied in the experiment.
• Main Effect – The change in the average response observed during
the change from one level to another for a single factor
• Test Run (Experimental Run) – A single combination of factor levels
that yields one or more observations of the responses
• Treatment Combination – An experimental run using a set of specific
levels in each input variable.

P
g
3
 Advanced DOE Definitions
• Confounding – One or more effects that cannot be unambiguously
attributed to a single factor or interaction. Usually due to problems
in design.
• Fixed Factor – Factors whose levels are specifically chosen.
Conclusions about fixed factors generalize to only those levels.
Determination of the factor’s effect on the level of the output is
usually the goal of the experiment. (Knob settings, for example)
• Random Factor – Factors whose levels are selected randomly from a
larger population of possible levels. Determination of the factor’s
contribution to the overall variance of the system is the goal of this
experiment. (Selecting 3 machines out of 20, for example)
• Repetitions – consecutive experimental runs using the same
treatment combinations. With no change in setup between runs.
• Replications – experimental runs using the same treatment
combinations that were not run consecutively. Replications are
duplicate runs of the entire experiment. Sample size calculations
apply to replicates, not repetitions.
• http://www.itl.nist.gov/div898/handbook/pri/section7/pri7.htm
P
g
4
Factors, Levels and Responses
▪ Problem Definition
▪ Select the factor levels that make the pizza delicious in taste and
consistent in shape.
▪ Factors, Levels, and Ranges of Interest
Factors Levels (low, high) Responses

Flour Cheap, Expensive Taste

Yeast Less, More Consistency

Olive Oil Virgin, Extra Virgin

Kneading Less, More

Mixture Old, New
Wood Unknown, Hickory
 Including Cost Analysis (Cube Plot Analysis)

What if said the overall task is to minimize cost while satisfying your customers.
• Hickory Wood is not much more expensive than Unknown Wood
• Expensive Flour is significantly higher than cheap flour
• Also, the kneading machine can be set at a high or low level without significantly increasing the energy usage (and
hence, cost).
• Customer satisfaction ratings were consistently high for Taste levels of 8 or above and Crust Consistency levels of 6
or over.
Examine carefully all the model graphs and pick the best recipe for pizza that will minimize your cost while still
ensuring high customer satisfaction. (After Narrowing it Down to 4 Significant factors)
ANOVA- Taste
ANOVA-Consistency