Sie sind auf Seite 1von 81

Experimental Designs for Screening

Two-Level Fractional Factorial Designs


!
!
!
!
!
!
!
!
!
!

Motivation
Fractionating a Design
The Defining Relation
Confounding Pattern
Design Resolution
Catalog of Defining Relations
Interactions
Saturated Design
Foldover Design
Power and Sample Size

N.J. Burn & Associates Inc.

581

Experimental Designs for Screening


Two-Level Fractional Factorial Designs Motivation
!

It is apparent that as k, the number of experimental factors that are


varied in a two-level full factorial experiment, becomes larger than
five, the number of runs required, 2k, becomes prohibitive.
Fractional factorial designs make more efficient use of full factorial
designs by confounding potentially unimportant pieces of
information with important pieces of information.

N.J. Burn & Associates Inc.

582

Experimental Designs for Screening


Fractionating a Design
!

As the number of experimental factors increases in a two-level full


factorial experiment, so do the number and order of interaction
terms that are estimable in the linear model.
For example, in a five factor experiment, there are:

one constant
five main effects
ten second-order interactions
ten third-order interactions
five fourth-order interactions
one fifth-order interactions

that can be estimated. This adds up to a total of 32 effects that are


estimable from a 32-run experiment.
N.J. Burn & Associates Inc.

583

Experimental Designs for Screening


Fractionating a Design
!

A question that needs to be posed is this, is it very likely that the


interaction effect between all five factors is significant, especially in
comparison to the main effects? For that matter, are the
interactions between three and four factors likely to be that
important, or even likely to arise at all given the physical aspects of
the process or product under investigation?
If the answers to these questions is no, then we are assuming that
we are really only interested in estimating the main effects, and
perhaps the second-order interaction effects between pairs of
factors. In this example then, we are executing 32 experimental run
conditions to estimate only 15 effects at most, plus the constant. Is
this the most efficient design?

N.J. Burn & Associates Inc.

584

Experimental Designs for Screening


Fractionating a Design
!

If we are willing to sacrifice information by mixing presumed


unimportant effects with important ones, then we can gain efficiency
by reducing the number of runs necessary to estimate the same
number of effects.
In our example, we have a total of 16 model parameters to estimate
(potentially, as we are not compelled to include all or any interaction
terms). Can we select an appropriate subset or fraction of the
original 32 runs that will still give us this information with the best
mixture of unimportant and important information?
The answer is yes but which 16 runs do we choose from the
original 32? The solution is held in the defining relation for fractional
factorial designs.

N.J. Burn & Associates Inc.

585

Experimental Designs for Screening


Fractionating a Design Example
!
!

Let us look at a simpler experiment that demonstrates how mixtures


of pieces of information arise in fractional factorial designs.
Suppose there are four operating variables of interest in a screening
study and restricted resources permit only eight tests to be carried
out. A full 24 design requires 16 runs and it is decided to select the
eight runs for our experiment from this design.
There are 12870 different subsets of eight tests that can be selected
from sixteen, each subset yielding different mixtures of information.
Which subset should we then select?
The particular subset chosen for this demonstration is that for which
the four factor interaction term x1x2x3x4 has the value 1.

N.J. Burn & Associates Inc.

586

Experimental Designs for Screening


Fractionating a Design Example

N.J. Burn & Associates Inc.

587

Experimental Designs for Screening


Fractionating a Design Example
!

The eight selected runs can be re-ordered into the more familiar
format:
x1

x2

x3

x4

1 1 1 1
1 1 1
1 1 1

1
1

1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1
N.J. Burn & Associates Inc.

The pattern of -1s and +1s follows the same full


factorial pattern as a 23 two-level full factorial
design for the first three factors, x1, x2 and x3.
What about the settings for x4? Is there an easier
way to determine its values other than writing out
the corresponding 24 full factorial design and
selecting an appropriate number of runs that meet
some criteria such as x1x2x3x4 = 1?
The pattern can be generated from the defining
relation of the fractional factorial experiment.

588

Experimental Designs for Screening


The Defining Relation
!

The defining relation for a fractional factorial design can be used for
two purposes:
Generating the pattern of -1s and +1s for additional factors included
in the experiment beyond the k factors in a 2k full factorial design.
Generating the confounding pattern of the design. This will be explained
in more detail later.

The defining relation for the previous example is I = x1x2x3x4 where I


is used to denote a column of 1s.

N.J. Burn & Associates Inc.

589

Experimental Designs for Screening


The Defining Relation Operating Rules
1.

Multiplication of any factor by I leaves the column of values for


that factor unchanged. For example,
x1 I = x1

2.

Multiplication of any factor by itself produces a column of 1s or I.


For example,
x3 x3 = (x3)2 = I

3.

4.
5.

x2x3 I = x2x3

(x1x3x4)2 = I

Any operation, such as multiplication, that is performed on one


side of the defining relation equality, must be performed on the
other side.
There can be more than one defining relation.
Multiplication of two or more defining relations results in another
defining relation.

N.J. Burn & Associates Inc.

590

Experimental Designs for Screening


The Defining Relation Example
!

We can use the defining relation from our example, I = x1x2x3x4, to


generate the x4 column.

So, column x4 is the result of the product of columns x1, x2 and x3.

N.J. Burn & Associates Inc.

591

Experimental Designs for Screening


Confounded Effects
!
!
!
!

From the previous example with the defining relation we were left
with x4 = x1x2x3.
This means that the x4 column is identical to that for the x1x2x3
column.
Thus, we cannot independently estimate the effects of both x4 and
x1x2x3 with this experimental design.
If we include columns in the calculation matrix for both of these
terms, the resulting matrix will be singular.
Singular matrices are a problem when software is used for DOE
analysis. The software algorithm will blow up.

N.J. Burn & Associates Inc.

592

Experimental Designs for Screening


Confounded Effects
!

Since we cannot independently estimate these two effects, we say


that they are confounded or aliased with each other. The
implications of this confounding is that if, in the analysis of the
experimental results, it is determined that the x4/ x1x2x3 column has
a significant effect, we cannot be sure whether the effect is due
solely to x4 or x1x2x3 or a mixture of the two.
This is the loss of information (the price you have to pay) for
fractionating a design. A smaller number of runs is required, but all
effects cannot be independently estimated in the final analysis.
However, if third and higher order interaction effects which are not
likely to be important are confounded with main effects and secondorder interactions, we can then assume that any effect observed is
due mostly to the main or second-order effect. If we assume that the
third-order interaction x1x2x3 is zero, then we can attribute any
significance solely to x4.

N.J. Burn & Associates Inc.

593

Experimental Designs for Screening


Confounding Pattern
!

The fact that x4 is confounded with x1x2x3 in our example is not the
only instance of confounding. When a design is fractionated, all
effects which can be estimated from the corresponding full factorial
design are mixed up with each other and the pattern of confounding
can be derived from the defining relation. For example, if I =
x1x2x3x4, we have the following confounding pattern:

N.J. Burn & Associates Inc.

594

Experimental Designs for Screening


Confounding Pattern
!

Thus each main effect is confounded with a third-order interaction


(which is probably okay) and the second-order interactions are
confounded with each other (which may or may not be okay).
The degree to which main effects and second-order interactions are
confounded with each other or with higher order interactions is
defined by the resolution of the design which will be presented
shortly.

N.J. Burn & Associates Inc.

595

Experimental Designs for Screening


Confounding and the Model
!

In the above example, the linear model that can be fit to response
data can be written as follows:

Y = 0 + 1 x1 + 2 x2 + 3 x3 + 4 x4 + 12 x1 x2 + 13 x1 x3 + 14 x1 x4
!

Each parameter in the model, i, i = 1, 2, , 8 is really estimating


the mixed effects:

1 = 1 + 234
~

0 = 0 + 1234

2 = 2 + 134

12 = 12 + 34

3 = 3 + 124

13 = 13 + 24

4 = 4 + 123

14 = 14 + 23

N.J. Burn & Associates Inc.

596

Experimental Designs for Screening


Nomenclature
!
!

!
!
!

The number of runs required in two-level full factorial designs is a


power of two.
The degree to which a full factorial design is reduced by
fractionating it is also a power of two. For example, half fractions,
quarter fractions, eighth fractions and so on are taken off of full
factorial designs.
The order of reduction can be denoted by q where 1/2q represents
the degree of fractionation.
For example, 24-1 denotes a half fraction of a 24 full factorial design.
Algebraically, 24-1 = 23 = 8 experimental runs, but a 24-1 fractional
factorial design is not the same as a 23 full factorial design which
also has 8 experimental runs.

N.J. Burn & Associates Inc.

597

Experimental Designs for Screening


Degree of Fractionation
!

A full factorial design can only be fractionated to the extent that

2 k q k + 1
!

For example, a fractional factorial design involving seven


experimental factors can be carried out in eight runs as a sixteenth
fraction of a 27 full factorial design since 27-4 = 23 = 8 7 + 1.
However, the effects of seven experimental factors cannot be
estimated in four experimental runs as a 1/32 fraction since
27-5 = 22 = 4 < 7 + 1.
There are simply not enough degrees of freedom to estimate so
many effects with so few runs.

N.J. Burn & Associates Inc.

598

Experimental Designs for Screening


Design Resolution
!

!
!

The resolution of a fractional factorial design describes the extent to


which main effects and second-order interactions are confounded
with each other and with third and higher order interactions.
As the term "resolution" suggests, the higher the resolution the
better in the sense that potentially important effects are not
confounded with each other.
The important effects can be resolved or estimated without worrying
about the effects they are confounded with.
Three common classes of resolution for 2k-q fractional factorial
designs are defined.

N.J. Burn & Associates Inc.

599

Experimental Designs for Screening


Design Resolution
!

Resolution III designs are 2k-q designs for which


(i) no individual operating variable, such as x1, is confounded with
any other individual operating variable, such as x2 and
(ii) at least one individual operating variable is confounded with a
two variable interaction.

An example of a resolution III design is the 23-1 design with defining


relation I = x1x2x3. This is often denoted as:

31
III

N.J. Burn & Associates Inc.

600

Experimental Designs for Screening


Design Resolution
!

Resolution IV designs are 2k-q designs for which


(i) no individual operating variable is confounded with any other
individual operating variable or with any two variable interaction and
(ii) at least one two variable interaction is confounded with another
two variable interaction.

An example of a resolution IV design is the 24-1 design with defining


relation I = x1x2x3x4. This is often denoted as:

2 4IV1
N.J. Burn & Associates Inc.

601

Experimental Designs for Screening


Design Resolution
!

Resolution V designs are 2k-q designs for which


(i) no individual operating variable is confounded with any other
individual operating variable or with any two variable interaction and
(ii) no two variable interaction is confounded with another two
variable interaction and
(iii) at least one two variable interaction is confounded with a three
variable interaction.

An example is the 25-1 design with defining relation I = x1x2x3x4x5.


This is often denoted as:

51
V

N.J. Burn & Associates Inc.

602

Experimental Designs for Screening


Catalog of Defining Relations
!
!

Most fractional factorial experiments are 4, 8 or 16 run designs.


Work has been done by people like George Box to identify the best
confounding patterns for a variety of fractional factorial designs that
achieve the highest resolution.
The following series of slides provide the defining relations that can
be used to generate 4, 8 and 16 run fractional factorial designs.

N.J. Burn & Associates Inc.

603

Experimental Designs for Screening


Catalog of Defining Relations
!

4 Run Fractional Factorial Designs


Design
2 3III1

N.J. Burn & Associates Inc.

Defining Relations
I = x1 x 2 x3

Generators
x3 = x1 x2

604

Experimental Designs for Screening


Catalog of Defining Relations
!

8 Run Fractional Factorial Designs


Design
2 4IV1

Defining Relations
I = x1 x 2 x3 x 4

Generators
x 4 = x1 x 2 x3

2 5III2

I = x1 x 2 x 4

x 4 = x1 x 2

= x1 x3 x5

x5 = x1 x3

I = x1 x 2 x 4

x 4 = x1 x 2

= x1 x3 x5

x5 = x1 x3

= x 2 x3 x 6
I = x1 x 2 x 4

x 6 = x 2 x3
x 4 = x1 x 2

= x1 x3 x5

x5 = x1 x3

= x 2 x3 x 6

x 6 = x 2 x3

= x1 x 2 x3 x7

x7 = x1 x 2 x3

2 6III3

2 7III4

N.J. Burn & Associates Inc.

605

Experimental Designs for Screening


Catalog of Defining Relations
!

16 Run Fractional Factorial Designs


Design
2V51

Defining Relations
I = x1 x 2 x3 x 4 x5

Generators
x5 = x1 x 2 x3 x 4

2 6IV 2

I = x1 x 2 x3 x5

x5 = x1 x 2 x3

= x 2 x3 x 4 x 6

x 6 = x 2 x3 x 4

I = x1 x 2 x3 x5

x5 = x1 x 2 x3

= x 2 x3 x 4 x 6

x 6 = x 2 x3 x 4

= x1 x3 x 4 x7
I = x1 x 2 x3 x5

x7 = x1 x3 x 4
x5 = x1 x 2 x3

= x 2 x3 x 4 x 6

x 6 = x 2 x3 x 4

= x1 x3 x 4 x7

x7 = x1 x3 x 4

= x1 x 2 x 4 x8

x8 = x1 x 2 x 4

2 7IV3

2 8IV 4

N.J. Burn & Associates Inc.

606

Experimental Designs for Screening


Catalog of Defining Relations
!

16 Run Fractional Factorial Designs


Design
955
2 9III
IV

6
210
III

Defining Relations
I = x1 x 2 x3 x5

Generators
x5 = x1 x 2 x3

= x 2 x3 x 4 x 6

x6 = x 2 x3 x 4

= x1 x3 x 4 x7

x7 = x1 x3 x 4

= x1 x 2 x 4 x8

x8 = x1 x 2 x 4

= x1 x 2 x3 x 4 x9
I = x1 x 2 x3 x5

x9 = x1 x 2 x3 x 4
x5 = x1 x 2 x3

= x 2 x3 x 4 x 6
= x1 x3 x 4 x7

N.J. Burn & Associates Inc.

x6 = x 2 x3 x 4
I = x1 x 2 x10

x7 = x1 x3 x 4

= x1 x 2 x 4 x8

x8 = x1 x 2 x 4

= x1 x 2 x3 x 4 x9

x9 = x1 x 2 x3 x 4

x10 = x1 x 2

607

Experimental Designs for Screening


Catalog of Defining Relations
!

16 Run Fractional Factorial Designs


Design
7
211
III

Defining Relations
I = x1 x 2 x3 x5
= x 2 x3 x 4 x 6
= x1 x3 x 4 x7
= x1 x 2 x 4 x8

8
212
III

I = x1 x 2 x10
= x1 x3 x11

= x1 x 2 x3 x 4 x9
I = x1 x 2 x3 x5

x6 = x 2 x3 x 4
x7 = x1 x3 x 4
x8 = x1 x 2 x 4

x10 = x1 x 2
x11 = x1 x3

x9 = x1 x 2 x3 x 4
x5 = x1 x 2 x3

= x 2 x3 x 4 x 6

I = x1 x 2 x10

x6 = x 2 x3 x 4

x10 = x1 x 2

= x1 x3 x 4 x7

= x1 x3 x11

x7 = x1 x3 x 4

x11 = x1 x3

= x1 x 2 x 4 x8

= x1 x 4 x12

x8 = x1 x 2 x 4

x12 = x1 x 4

= x1 x 2 x3 x 4 x9
N.J. Burn & Associates Inc.

Generators
x5 = x1 x 2 x3

x9 = x1 x 2 x3 x 4
608

Experimental Designs for Screening


Catalog of Defining Relations
!

16 Run Fractional Factorial Designs


Design
9
213
III

Defining Relations
I = x1 x 2 x3 x5
= x 2 x3 x 4 x 6
= x1 x3 x 4 x7
= x1 x 2 x 4 x8

10
214
III

= x1 x 2 x3 x 4 x9
I = x1 x 2 x3 x5

N.J. Burn & Associates Inc.

I = x1 x 2 x10
= x1 x3 x11
= x1 x 4 x12
= x 2 x3 x13
I = x1 x 2 x10

Generators
x5 = x1 x 2 x3
x6 = x 2 x3 x 4
x7 = x1 x3 x 4
x8 = x1 x 2 x 4
x9 = x1 x 2 x3 x 4
x5 = x1 x 2 x3

x10 = x1 x 2
x11 = x1 x3
x12 = x1 x 4
x13 = x 2 x3
x10 = x1 x 2

= x 2 x3 x 4 x 6

= x1 x3 x11

x6 = x 2 x3 x 4

x11 = x1 x3

= x1 x3 x 4 x7

= x1 x 4 x12

x7 = x1 x3 x 4

x12 = x1 x 4

= x1 x 2 x 4 x8

= x 2 x3 x13

x8 = x1 x 2 x 4

x13 = x 2 x3

= x1 x 2 x3 x 4 x9

= x 2 x 4 x14

x9 = x1 x 2 x3 x 4

x14 = x 2 x 4
609

Experimental Designs for Screening


Catalog of Defining Relations
!

16 Run Fractional Factorial Designs


Design
11
215
III

Defining Relations
I = x1 x 2 x3 x5
= x 2 x3 x 4 x 6
= x1 x3 x 4 x7
= x1 x 2 x 4 x8
= x1 x 2 x3 x 4 x9

N.J. Burn & Associates Inc.

Generators
I = x1 x 2 x10
= x1 x3 x11
= x1 x 4 x12
= x 2 x3 x13
= x 2 x 4 x14
= x3 x 4 x15

x5 = x1 x 2 x3
x6 = x 2 x3 x 4
x7 = x1 x3 x 4
x8 = x1 x 2 x 4
x9 = x1 x 2 x3 x 4

x10 = x1 x 2
x11 = x1 x3
x12 = x1 x 4
x13 = x 2 x3
x14 = x 2 x 4
x15 = x3 x 4

610

Experimental Designs for Screening


Minitab Exercise Fractional Factorial Designs
!

Click on Stat DOE Factorial Create Factorial Design


Click on the Display Available Designs button

N.J. Burn & Associates Inc.

611

Experimental Designs for Screening


Defining Relations and Confounding Patterns
!
!

Rule 5 governing defining relations states that a new defining


relation can be determined by multiplying defining relations together.
For example, a resolution III 25-2 design has defining relations
I = x1x2x4 = x1x3x5. A third defining relation can be found by
multiplying these two together.

I = x1 x 2 x 4 (x1 x3 x5 )
= x12 x 2 x3 x 4 x5
= I (x 2 x3 x 4 x5 )

There can be many defining relations


that determine the complete
confounding pattern for highly
fractionated designs (e.g. 215-11).

= x 2 x3 x 4 x5
N.J. Burn & Associates Inc.

612

Experimental Designs for Screening


Defining Relations and Confounding Patterns
!

From the previous 25-2 example where we have


I = x1x2x4 = x1x3x5 = x2x3x4x5, we have the following confounding (or
aliasing) pattern:
x1 with x2x4 and x3x5
x2 with x1x4
x3 with x1x5
x4 with x1x2
x5 with x1x3
x2x3 with x4x5
x2x5 with x3x4

This confounding pattern accounts


for the relationships between all
five main effects and the ten
second-order interactions.

We are usually only interested in the confounding pattern up to


second-order interactions.

N.J. Burn & Associates Inc.

613

Experimental Designs for Screening


Defining Relations and Confounding Patterns
!

If we fit the following proposed model for a resolution III 25-2 design:

Y = 0 + 1 x1 + 2 x2 + 3 x3 + 4 x4 + 5 x5 + 23 x2 x3 + 25 x2 x5
!

Each parameter in the model, i, i = 1, 2, , 8 is really estimating


the mixed effects:

1 = 1 + 24 + 35
~

0 = 0 + 124 + 135 + 2345

2 = 2 + 14

23 = 23 + 45

3 = 3 + 15

25 = 25 + 34

4 = 4 + 12
~

5 = 5 + 13
N.J. Burn & Associates Inc.

614

Experimental Designs for Screening


Interactions
!

For resolution V designs, all of the main effects and second-order


interactions can be estimated independently in that they are
confounded with third and higher-order interactions that are
assumed to be insignificant.
However, it is possible in some situations to obtain independent
estimates of second-order interactions in lower resolution designs.
It depends on the extent of prior knowledge the experimenter has with
the system/process/product/design.

N.J. Burn & Associates Inc.

615

Experimental Designs for Screening


Interactions Example
!

Lets use the previous 25-2 example where we have


I = x1x2x4 = x1x3x5 = x2x3x4x5. Ignoring second-order interactions, the
main effects model that can be fit is:

~ ~
~
~
~
~
Y = 0 + 1 x1 + 2 x2 + 3 x3 + 4 x4 + 5 x5
!

The alias pattern for the main effects is:

1 = 1 + 24 + 35
~

2 = 2 + 14
~

3 = 3 + 15
~

4 = 4 + 12
~

5 = 5 + 13
N.J. Burn & Associates Inc.

616

Experimental Designs for Screening


Interactions Example
!

Note that we still have the freedom to include two more terms in the
proposed model corresponding to the confounded pairs of secondorder interactions (23, 45) and (25, 34).
Suppose that the five factors in the experiment are nitrogen content
in lawn fertilizer, amount of lawn watering, lawn aeration, type of
grass in the lawn and grade of top soil.
The response of interest is the total weight of grass clippings after a
season of mowing. More grass clippings is suggestive of a healthier,
thicker lawn.
As an experienced gardener, you strongly suspect there to be a
significant interaction between nitrogen content and amount of
watering.

N.J. Burn & Associates Inc.

617

Experimental Designs for Screening


Interactions Example
!

The design initially proposed to you, based on the defining relations,


is as follows:
x1
1

x2
1

x3 x 4 = x1 x 2 x5 = x1 x3
1
1
1

x3 = grade of top soil

x4 = amount of water

N.J. Burn & Associates Inc.

x1 = nitrogen content
x2 = aeration

x5 = type of grass

618

Experimental Designs for Screening


Interactions Example
!

!
!
!

You see right away that the second-order interaction of interest,


between nitrogen content and amount of watering, corresponds to
the x1x4 interaction the way the design is presently defined.
From examining the confounding pattern, you notice that the x1x4
interaction is confounded with the main effect for x2.
Hence, as currently defined, the proposed design will not meet your
modeling objective.
However, by judicially reassigning the experimental factors, you can
make use of one of the two pairs of second-order interaction pairs
that are not in the model yet.

N.J. Burn & Associates Inc.

619

Experimental Designs for Screening


Interactions Example
!

You propose the following changes:


x1
1

x2
1

x3 x 4 = x1 x 2 x5 = x1 x3
1
1
1

x3 = amount of water

x4 = grade of top soil

N.J. Burn & Associates Inc.

x1 = aeration
x2 = nitrogen content

x5 = type of grass

620

Experimental Designs for Screening


Interactions Example
!

Now the interaction of interest corresponds to x2x3 which can be


added to the model without having it confounded with any of the five
main effects:

Y = 0 + 1 x1 + 2 x2 + 3 x3 + 4 x4 + 5 x5 + 23 x2 x3
!

Depending on your prior knowledge about the process under


investigation, it may still be possible to obtain "independent"
estimates of main effects and some second-order interactions, even
with low resolution designs, if care is taken in the design stage to
obtain a desirable confounding pattern.
The experimental design and subsequent confounding relationships
can always be evaluated before committing to run the experiment.

N.J. Burn & Associates Inc.

621

Experimental Designs for Screening


Saturated Designs
!
!
!
!

In the previous example, it was possible to use extra degrees of


freedom to include one or more interaction terms in the model.
There are a group of resolution III designs known as saturated
designs that do not afford this freedom.
With a design from this group, k operating variables can be
investigated simultaneously in k+1 tests where k+1 is a power of 2.
Examples of saturated two-level fractional factorial designs are:
26
23III1 , 27III 4 , 215III11 , 231
III

In these designs, 3, 7, 15 and 31 factors are investigated in 4, 8, 16


and 32 runs respectively.

N.J. Burn & Associates Inc.

622

Experimental Designs for Screening


Saturated Designs Example
!

!
!

Construction of a resolution III 27-4 design is accomplished by first


writing a 23 design in three of the seven operating variables, X1, X2
and X3.
Each of the remaining operating variables, X4, X5, X6 and X7 is
confounded with an interaction among X1, X2 and X3.
If the aliases X4 = X1X2, X5 = X1X3, X6 = X2X3 and X7 = X1X2X3 are
used, then the resulting design is that shown below.

N.J. Burn & Associates Inc.

x1

x2

x3

x4

x5

x6

x7

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

623

Experimental Designs for Screening


Saturated Designs Example
!

From the four basic generators X1X2X4, X1X3X5, X2X3X6 and


X1X2X3X7 arising from the choice of aliases, the following defining
relation for this design can be formed,
I = x1 x2 x4 = x1 x3 x5 = x2 x3 x6 = x1 x2 x3 x7
= x2 x3 x4 x5 = x1 x3 x4 x6 = x3 x4 x7 = x1 x2 x5 x6 = x2 x5 x7 = x1 x6 x7
= x4 x5 x6 = x1 x4 x5 x7 = x2 x4 x6 x7 = x3 x5 x6 x7
= x1 x2 x3 x4 x5 x6 x7

(taking basic generators


one at a time)
(products of two basic
generators)
(products of three basic
generators)
(products of four basic
generators)

Note: These are also referred to as the words of a defining relation.


N.J. Burn & Associates Inc.

624

Experimental Designs for Screening


Saturated Designs Example
!
!

Notice that the smallest order interaction in this defining relation is


three, verifying that the resolution of the design is indeed III.
Again ignoring interactions involving more than two operating
variables, the following eight estimates can be obtained from this
design.
l1 ,

which estimates

l2 , which estimates
l3 , which estimates
l4 , which estimates
l5 , which estimates
l6 , which estimates

(
(
(
(
(
(
(

l7 , which estimates
7
l0 , which estimates 0
N.J. Burn & Associates Inc.

24
14
15
12
13
23
34

+
+
+
+
+
+
+

35
36
26
56
46
45
25

+
+
+
+
+
+
+

67 )
57 )
47 )
37 )
27 )
17 )
16 )

625

Experimental Designs for Screening


Saturated Designs Interpretation of Results
!

Interpretation of results from a saturated design may be ambiguous


because each operating variable is confounded with a number of
two variable interactions.
As will be shown later, ambiguities can be partially resolved by
carrying out another saturated design from the same "family", that is,
a design for which the signs of all values of one or more of the
operating variables are reversed.
Thus, saturated designs are more useful for the first step in a
screening study of several operating variables.

N.J. Burn & Associates Inc.

626

Experimental Designs for Screening


Example Saturated Design
!

This example has been slightly modified from the Course Notes:
Referring to Table 23.5 in the notes, the following changes have been
made:
! x4 = recycle
! x5 = rate of addition of NaOH
! x6 = type of filter cloth
! x7 = holdup time

The standard generators of x4 = x1x2, x5 = x1x3, x6 = x2x3 and x7 = x1x2x3


are used instead of the ones in Equation 23.3.
The columns in the design matrix in Table 23.6 have been accordingly
adjusted.
s are used in stead of s for the model coefficients.
None of these changes affects the resulting analysis and interpretation.
N.J. Burn & Associates Inc.

627

Experimental Designs for Screening


Example Saturated Design
!

During the startup of a new process in a chemical plant, trouble was


encountered in a filtration operation. Filtration was requiring about
70 minutes per batch instead of 40 minutes, the required time for a
similar operation at other plant sites. An investigation was
undertaken to identify the operating variables that affected filtration
time and to determine how these variables might be altered in order
to reduce the filtration time. The operating variables selected for the
initial study are shown in the following table. The low levels
represent the operating conditions prior to this screening study. The
high levels are changes in operating conditions chosen to identify
which, if any, of these seven operating variables affected the
filtration time. It will be noted that four of the operating variables, x1,
x2, x4 and x6 are qualitative variables.

N.J. Burn & Associates Inc.

628

Experimental Designs for Screening


Example Saturated Design
Operating Variable
x1 , water supply
x2 , raw material
x3 , filtration temperature
x4 , recycle
x5 , rate of addition of NaOH
x 6 , type of filter cloth
x7 , holdup time

Level
-1
municipal reservoir

1
well

made on site
low

made at another site


high

included
fast

omitted
slow

new
short

old
long

Model to be fit:

E (Y ) = 0 + 1 x1 + 2 x 2 + 3 x3 + 4 x 4 + 5 x5 + 6 x6 + 7 x7
N.J. Burn & Associates Inc.

629

Experimental Designs for Screening


Example Saturated Design
!

As a first step in the study, a 27-4 resolution III design was employed
because of its economy in tests and its facility for use as a building
block for further tests that might be required.
Basic generators chosen for the design were:
I = x1 x 2 x 4 = x1 x 3 x 5 = x 2 x 3 x 6 = x1 x 2 x 3 x 7
x 4 = x1 x 2
x 5 = x1 x 3
x6 = x 2 x3
x 7 = x1 x 2 x 3

N.J. Burn & Associates Inc.

630

Experimental Designs for Screening


Example Saturated Design
!

The design and the measured steady state filtration time for each
test are:
x1

x2

x3

x4

x5

x6

x7

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

1
1
1
1
1
1
1
1

N.J. Burn & Associates Inc.

y (min.)

68.4
77.7
66.4
81.0
78.6
41.2
68.7
38.7
631

Experimental Designs for Screening


Example Saturated Design
!

When these results were examined, there may well have been a
temptation to conclude that either the sixth test or the eighth test
from the experiment had resolved the problem since both tests
produced filtration times in the order of 40 minutes, the target figure.
As will be shown shortly, a conclusion that changes in x1, x3 and x5
produced this favourable result is only one of several possible
interpretations of these data.
In any case, before making a change in such an important operating
variable as water supply (x1), other interpretations would have to be
assessed.

N.J. Burn & Associates Inc.

632

Experimental Designs for Screening


Example Minitab Output
!

Fitted coefficients are:


Estimated Effects and Coefficients for Filt. (coded units)
Term
Constant
Water Su
Raw Mate
Filt. Te
Recycle
NaOH Add
Type Fil
Holdup

Effect
-10.87
-2.77
-16.58
3.17
-22.83
-3.42
0.53

Coef
65.09
-5.44
-1.39
-8.29
1.59
-11.41
-1.71
0.26

The fitted model is then:


y = 65.09 5.44 x1 1.39 x 2 8.29 x3 + 1.59 x 4 11.41x5 1.71x6 + 0.26 x7

N.J. Burn & Associates Inc.

633

Experimental Designs for Screening


Example Saturated Design
!

From the defining relation for this design it can be confirmed that the
eight coefficient estimates are confounded with a number of twofactor interactions. Interactions involving more than two operating
variables have been ignored.
x1
x2
x3
x4
x5
x6
x7

+
+
+
+
+
+
+

N.J. Burn & Associates Inc.

x 2x 4
x 1x 4
x 1x 5
x 1x 2
x 1x 3
x 1x 7
x 1x 6

+
+
+
+
+
+
+

x 3x 5
x 3x 6
x 2x 6
x 3x 7
x 2x 7
x 2x 3
x 2x 5

+
+
+
+
+
+
+

x 6x 7
x 5x 7
x 4x 7
x 5x 6
x 4x 6
x 4x 5
x 3x 4

634

Experimental Designs for Screening


Example Saturated Design
!

Because further tests were carried out in this study, a second


subscript has been added to these estimates to denote that they
arise from the first set of tests.
01 ,
11 ,
21 ,
31 ,
41 ,
51 ,
61 ,
71 ,

an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of

N.J. Burn & Associates Inc.

( 0
) = 65.09
( 1 + 24 + 35 + 67 ) = 5.44
( 2 + 14 + 36 + 57 ) = 1.39
( 3 + 15 + 26 + 47 ) = 8.29
( 4 + 12 + 37 + 56 ) = 1.59
( 5 + 13 + 27 + 46 ) = 11.41
( 6 + 17 + 23 + 45 ) = 1.71
( 7 + 16 + 25 + 34 ) = 0.26
635

Experimental Designs for Screening


Example Saturated Design
!

Because no estimate of the pure error variance is available, one


interpretation of these estimates can be made on the basis of their
relative magnitudes. Among the coefficients of operating variables,
the estimates 5, 3 and 1 are much larger in magnitude than the
other estimates. A number of alternative interpretations are possible.

N.J. Burn & Associates Inc.

636

Experimental Designs for Screening


Example Saturated Design
!

A simple explanation of these three large estimates might be that


only the terms 1x1, 3x3 and 5x5 are important, all two variable
interactions being of negligible size.
Another explanation might be that only operating variables x1 and x3
are affecting the filtration time, their influence being explained by
terms 1x1, 3x3 and 13x1x3.
A third possibility is that only operating variables x1 and x5 are
important, their effect being accounted for by terms 1x1, 5x5 and
15x1x5.
Another alternative is that only operating variables x3 and x5 are
causing the response to change via terms 3x3 , 5x5 and 35x3x5.

N.J. Burn & Associates Inc.

637

Experimental Designs for Screening


Example Saturated Design
01 ,
11 ,
21 ,
31 ,
41 ,
51 ,
61 ,
71 ,

01 ,
11 ,
21 ,
31 ,
41 ,
51 ,
61 ,
71 ,

an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of

an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of

N.J. Burn & Associates Inc.

( 0
) = 65.09
( 1 + 24 + 35 + 67 ) = 5.44
( 2 + 14 + 36 + 57 ) = 1.39
( 3 + 15 + 26 + 47 ) = 8.29
( 4 + 12 + 37 + 56 ) = 1.59
( 5 + 13 + 27 + 46 ) = 11.41
( 6 + 17 + 23 + 45 ) = 1.71
( 7 + 16 + 25 + 34 ) = 0.26

01 ,
11 ,
21 ,
31 ,
41 ,
51 ,
61 ,
71 ,

( 0
) = 65.09
( 1 + 24 + 35 + 67 ) = 5.44
( 2 + 14 + 36 + 57 ) = 1.39
( 3 + 15 + 26 + 47 ) = 8.29
( 4 + 12 + 37 + 56 ) = 1.59
( 5 + 13 + 27 + 46 ) = 11.41
( 6 + 17 + 23 + 45 ) = 1.71
( 7 + 16 + 25 + 34 ) = 0.26

01 ,
11 ,
21 ,
31 ,
41 ,
51 ,
61 ,
71 ,

an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of

an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of

( 0
) = 65.09
( 1 + 24 + 35 + 67 ) = 5.44
( 2 + 14 + 36 + 57 ) = 1.39
( 3 + 15 + 26 + 47 ) = 8.29
( 4 + 12 + 37 + 56 ) = 1.59
( 5 + 13 + 27 + 46 ) = 11.41
( 6 + 17 + 23 + 45 ) = 1.71
( 7 + 16 + 25 + 34 ) = 0.26
( 0
) = 65.09
( 1 + 24 + 35 + 67 ) = 5.44
( 2 + 14 + 36 + 57 ) = 1.39
( 3 + 15 + 26 + 47 ) = 8.29
( 4 + 12 + 37 + 56 ) = 1.59
( 5 + 13 + 27 + 46 ) = 11.41
( 6 + 17 + 23 + 45 ) = 1.71
( 7 + 16 + 25 + 34 ) = 0.26
638

Experimental Designs for Screening


Minitab Exercise Saturated Experiment
!
!

Open the file Topic06SatExp.MTW


Select Stat DOE Factorial Analyze Factorial Design
Select C12 Y as the Response
Click on the Graphs button
! Click on the Four in one radio button
! Click on OK

Click on OK

N.J. Burn & Associates Inc.

639

Experimental Designs for Screening


!

Minitab output
Analysis of Variance
Source
Model
Linear
X1
X2
X3
X4
X5
X6
X7
Error
Total

DF
7
7
1
1
1
1
1
1
1
0
7

Adj SS
1887.53
1887.53
236.53
15.40
549.46
20.16
1041.96
23.46
0.55
*
1887.53

Adj MS
269.65
269.65
236.53
15.40
549.46
20.16
1041.96
23.46
0.55
*

F-Value
*
*
*
*
*
*
*
*
*

P-Value
*
*
*
*
*
*
*
*
*

Note that test statistics


and p-values cannot be
calculated because this
is an exact fit as
exhibited by an R2 of
100%.

Model Summary
S
*

R-sq
100.00%

R-sq(adj)
*

N.J. Burn & Associates Inc.

R-sq(pred)
*

640

Experimental Designs for Screening


!

Minitab output
Coded Coefficients
Term
Constant
X1
X2
X3
X4
X5
X6
X7

Effect
-10.875
-2.775
-16.575
3.175
-22.82
-3.425
0.5250

Coef
65.09
-5.437
-1.387
-8.288
1.587
-11.41
-1.712
0.2625

SE
Coef
*
*
*
*
*
*
*
*

T-Value
*
*
*
*
*
*
*
*

P-Value
*
*
*
*
*
*
*
*

VIF
1.00
1.00
1.00
1.00
1.00
1.00
1.00

Regression Equation in Uncoded Units


Y = 65.09 - 5.437 X1 - 1.387 X2 - 8.288 X3 + 1.587 X4 - 11.41 X5 - 1.712 X6
+ 0.2625 X7

N.J. Burn & Associates Inc.

641

Experimental Designs for Screening


!

Minitab output
Aliases
I
A
B
C
D
E
F
G

+
+
+
+
+
+
+
+

ABD + ACE
BD + CE +
AD + CF +
AE + BF +
AB + CG +
AC + BG +
AG + BC +
AF + BE +

+ AFG + BCF + BEG + CDG + DEF


FG + BCG + BEF + CDF + DEG
EG + ACG + AEF + CDE + DFG
DG + ABG + ADF + BDE + EFG
EF + ACF + AEG + BCE + BFG
DF + ABF + ADG + BCD + CFG
DE + ABE + ACD + BDG + CEG
CD + ABC + ADE + BDF + CEF

* NOTE * Could not graph the specified residual type because MSE = 0 or the
degrees of freedom for error = 0.

N.J. Burn & Associates Inc.

642

Experimental Designs for Screening


Example Foldover Design
!

It is well known that low resolution fractional factorial designs (III and
IV) confound main effects with second-order interactions (III) and
second-order interactions with each other (IV).
This sometimes makes the interpretation of a DOE analysis difficult.

Foldover designs can be used to increase the resolution of a low


resolution design and help to resolve lingering questions from the
initial design.
Foldover designs are a nice sequential strategy to employ when
resources are limited.

Because of the ambiguity in interpreting the results of these tests, a


second foldover set of eight tests was conducted using a 27-4
resolution III design formed from the first design by reversing the
signs of all values for all seven operating variables.
This second design is shown on the next slide along with the
measured filtration times obtained for these additional tests.

N.J. Burn & Associates Inc.

643

Experimental Designs for Screening


Example Foldover Design
x1

x2

x3

x4

x5

x6

x7

y (min.)

1
1

1
1

1
1

1
1

1
1

1
1

1
1

66.7
65.0

86.4

61.9

47.8

59.0

42.6

67.6

N.J. Burn & Associates Inc.

644

Experimental Designs for Screening


Example Foldover Design
!

Switching the signs of all values for one operating variable xi in a


27-4 resolution III design is equivalent to replacing xi with -xi.
Because of the manner in which this second design has been
constructed from the first design, its defining relation can be
obtained by replacing every operating variable xi in the defining
relation for the first design, by -xi. The resulting defining relation is
then:
I = x1 x 2 x 4 = x1 x 3 x 5 = x 2 x 3 x 6 = x1 x 2 x3 x 7
x 4 = x1 x 2
x 5 = x1 x 3
x6 = x 2 x3
x 7 = x1 x 2 x3

N.J. Burn & Associates Inc.

645

Experimental Designs for Screening


Example Minitab Output
!

Minitab was used to obtain the following results for the second
design on its own:
Estimated Effects and Coefficients for Filt. (coded units)
Term
Constant
Water Su
Raw Mate
Filt. Te
Recycle
NaOH Add
Type Fil
Holdup

N.J. Burn & Associates Inc.

Effect
-2.500
-5.000
15.750
2.250
-15.600
3.300
-9.150

Coef
62.125
-1.250
-2.500
7.875
1.125
-7.800
1.650
-4.575

646

Experimental Designs for Screening


Example Foldover Design
!

From the defining relation for the foldover design it can be confirmed
that the eight coefficient estimates are confounded in the following
matter. Interactions involving more than two operating variables
have been ignored.
x1
x2
x3
x4
x5
x6
x7

x 2x 4
x 1x 4
x 1x 5
x 1x 2
x 1x 3
x 1x 7
x 1x 6

N.J. Burn & Associates Inc.

x 3x 5
x 3x 6
x 2x 6
x 3x 7
x 2x 7
x 2x 3
x 2x 5

x 6x 7
x 5x 7
x 4x 7
x 5x 6
x 4x 6
x 4x 5
x 3x 4

647

Experimental Designs for Screening


Example Foldover Design
!

So, we have:
02 ,
12 ,
22 ,
32 ,
42 ,
52 ,
62 ,
72 ,

an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of

( 0
) = 62.125
( 1 24 35 67 ) = 1.25
( 2 14 36 57 ) = 2.50
( 3 15 26 47 ) = 7.875
( 4 12 37 56 ) = 1.125
( 5 13 27 46 ) = 7.80
( 6 17 23 45 ) = 1.65
( 7 16 25 34 ) = 4.575

which creates even more possibilities.


N.J. Burn & Associates Inc.

648

Experimental Designs for Screening


Example Foldover Design
!

However, we can combine the results from the original design with
those from the foldover design in the following manner:

( 01 + 02 ) 2,
( 11 + 12 ) 2,
( 21 + 22 ) 2,
( 31 + 32 ) 2,
( 41 + 42 ) 2,
( 51 + 52 ) 2,
( 61 + 62 ) 2,
( 71 + 72 ) 2,
N.J. Burn & Associates Inc.

an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of

0 = 63.6
1 = 3.34
2 = 1.94
3 = 0.21
4 = 1.36
5 = 9.61
6 = 0.03
7 = 2.16

649

Experimental Designs for Screening


Example Foldover Design
( 41 42 ) 2 ,
( 51 52 ) 2 ,
( 21 22 ) 2 ,
( 31 32 ) 2 ,
( 71 72 ) 2 ,
( 61 62 ) 2 ,
( 11 12 ) 2 ,

an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of
an estimate of

( 01 02 ) 2 ,

N.J. Burn & Associates Inc.

( 12 + 37 + 56 ) = 0.231
( 13 + 27 + 46 ) = 1.81
( 14 + 36 + 57 ) = 0.56
( 15 + 26 + 47 ) = 8.08
( 16 + 25 + 34 ) = 2.42
( 17 + 23 + 45 ) = 1.68
( 24 + 35 + 67 ) = 2.09

an estimate of the block effect , = 1.5

650

Experimental Designs for Screening


Example Foldover Design
!
!

The block effect is the difference in average response values


between the two sets of eight runs.
Had it been large, it would have indicated the presence of some
other variables, beyond the seven being studied, whose change
between the two sets of tests strongly affected the filtration time.
By combining the original with the foldover experiment, we have
effectively created a 27-3 resolution IV experimental design.

N.J. Burn & Associates Inc.

651

Experimental Designs for Screening


Example Foldover Design
!

Among the above sixteen estimates, the largest is -9.61, an estimate


of 5 and 8.08, an estimate of the linear combination of two variable
interactions 15 + 26 + 47. The next largest estimate is -3.34, an
estimate of 1.
The investigators concluded that operating variables 1 and 5 alone,
the water supply and the rate of addition of caustic soda, affected the
filtration time, and the estimate -8.08 occurred primarily because of the
interaction 15.
Even at this stage, of course, other interpretations are possible. For
example, the estimate -8.08 might have been due to the interaction
x2x6 and/or the interaction x4x7. It is noted, however, that the estimates
of 2, 6, 4, and 7 are all relatively small and, although this does not
necessarily mean that interactions among these variables must also
be small, this is often the case.

N.J. Burn & Associates Inc.

652

Experimental Designs for Screening


Example Foldover Design
!

The investigators interpretation can be summarized conveniently


by the following chart which shows the average filtration time
obtained at each of the four sets of operating conditions of water
supply and rate of addition of NaOH. Evidence of the large negative
interaction between the two variables is very strong.
slow

65.4

42.6

fast

68.5

78.0

rate of addition
of NaOH

reservoir

well
water supply

N.J. Burn & Associates Inc.

653

Experimental Designs for Screening


Example Foldover Design
!

The corrective action implied by these results was to change the


water supply from the municipal reservoir to the well and reduce the
rate of addition of caustic soda. These changes were made and
satisfactory filtration times close to 40 minutes were obtained in
subsequent plant operation.

N.J. Burn & Associates Inc.

654

Experimental Designs for Screening


Exercise Foldover Design
!

The reduced fitted model from this example is:

y = 63.6 3.34 x1 9.61x5 8.08 x1 x5


!

What is the predicted filtration time at the new operating conditions


using well water and a slow rate of NaOH addition?

N.J. Burn & Associates Inc.

655

Experimental Designs for Screening


Minitab Exercise Foldover Experiment
!

Click on Stat DOE Factorial Create Factorial Design


Use the Number of factors drop down menu to select 7
Click on the Designs button
! Select the first row for a 2^7-4 fractional factorial design
! Click OK

Click on the Factors button


! Change the Factor Names from A through G to X1 through X7
! Click OK

Click on the Options button


! Click on the Fold on all factors radio button
! Uncheck the Randomize runs box
! Click OK

Click OK
N.J. Burn & Associates Inc.

656

Experimental Designs for Screening


Minitab Exercise Foldover Experiment
!

Enter the Y response data in the Minitab worksheet


X1

X2

X3

X4

X5

X6

X7

-1

-1

-1

-1

68.4

-1

-1

-1

-1

77.7

-1

-1

-1

-1

66.4

-1

-1

-1

-1

81.0

-1

-1

-1

-1

78.6

-1

-1

-1

-1

41.2

-1

-1

-1

-1

68.7

38.7

-1

-1

-1

66.7

-1

-1

-1

65.0

-1

-1

-1

86.4

-1

-1

-1

61.9

-1

-1

-1

47.8

-1

-1

-1

59.0

-1

-1

-1

42.6

-1

-1

-1

-1

-1

-1

-1

67.6

N.J. Burn & Associates Inc.

657

Experimental Designs for Screening


Minitab Exercise Foldover Experiment
!
!

Select Stat DOE Factorial Analyze Factorial Design


Select C12 Y as the Response
Click on the Graphs button
! Click on the Four in one radio button
! Click on OK

Click on OK

N.J. Burn & Associates Inc.

658

Experimental Designs for Screening


Minitab Exercise Foldover Experiment
!

Coefficients table
Coded Coefficients
Term
Constant
X1
X2
X3
X4
X5
X6
X7
X1*X2
X1*X3
X1*X4
X1*X5
X1*X6
X1*X7
X2*X4
X1*X2*X4

Effect
-6.687
-3.888
-0.4125
2.712
-19.213
-0.06250
-4.313
0.4625
-3.613
1.1125
-16.163
4.838
-3.362
-4.188
2.963

N.J. Burn & Associates Inc.

Coef
63.61
-3.344
-1.944
-0.2062
1.356
-9.606
-0.03125
-2.156
0.2312
-1.806
0.5563
-8.081
2.419
-1.681
-2.094
1.481

SE
Coef
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*

T-Value
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*

P-Value
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*
*

VIF
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00

659

Experimental Designs for Screening


Minitab Exercise Foldover Experiment
!

Reduced model
Start removing higher order terms with the smallest magnitude
coefficient value (highest p-value)
Analysis of Variance
Source
Model
Linear
X1
X5
2-Way Interactions
X1*X5
Error
Total

N.J. Burn & Associates Inc.

DF
3
2
1
1
1
1
12
15

Adj SS
2700.3
1655.4
178.9
1476.5
1044.9
1044.9
467.1
3167.3

Adj MS
900.09
827.69
178.89
1476.48
1044.91
1044.91
38.92

F-Value
23.13
21.27
4.60
37.94
26.85
26.85

P-Value
0.000
0.000
0.053
0.000
0.000
0.000

660

Experimental Designs for Screening


Minitab Exercise Foldover Experiment
Model Summary
S
6.23867

R-sq
85.25%

R-sq(adj)
81.57%

R-sq(pred)
73.78%

Coded Coefficients
Term
Constant
X1
X5
X1*X5

Effect
-6.69
-19.21
-16.16

Coef
63.61
-3.34
-9.61
-8.08

SE Coef
1.56
1.56
1.56
1.56

T-Value
40.78
-2.14
-6.16
-5.18

P-Value
0.000
0.053
0.000
0.000

VIF
1.00
1.00
1.00

Regression Equation in Uncoded Units


Y = 63.61 - 3.34 X1 - 9.61 X5 - 8.08 X1*X5

N.J. Burn & Associates Inc.

661

Das könnte Ihnen auch gefallen