Design-Expert version 71 What’s New in Design-Expert version 7 Factorial and RSM Design Pat...

50
Design-Expert vers ion 7 1 What’s New in Design-Expert version 7 Factorial and RSM Design Pat Whitcomb November, 2006

Transcript of Design-Expert version 71 What’s New in Design-Expert version 7 Factorial and RSM Design Pat...

Design-Expert version 7 1

What’s New inDesign-Expert version 7

Factorial and RSM Design

Pat WhitcombNovember, 2006

Design-Expert version 7 2

What’s New

General improvements Design evaluation Diagnostics Updated graphics Better help Miscellaneous Cool New Stuff

Factorial design and analysis

Response surface design

Mixture design and analysis

Combined design and analysis

Design-Expert version 7 3

Two-Level Factorial Designs

2k-p factorials for up to 512 runs (256 in v6) and 21 factors (15 in v6). Design screen now shows resolution and updates with

blocking choices. Generators are hidden by default. User can specify base factors for generators. Block names are entered during build.

Minimum run equireplicated resolution V designs for6 to 31 factors.

Minimum run equireplicated resolution IV designs for 5 to 50 factors.

Design-Expert version 7 4

2k-p Factorial DesignsMore Choices

Need to “check” box to see factor generators

Design-Expert version 7 5

2k-p Factorial DesignsSpecify Base Factors for Generators

Design-Expert version 7 6

MR5 Designs Motivation

Regular fractions (2k-p fractional factorials) of 2k designs often contain considerably more runs than necessary to estimate the [1+k+k(k-1)/2] effects in the 2FI model.

For example, the smallest regular resolution V design for k=7 uses 64 runs (27-1) to estimate 29 coefficients.

Our balanced minimum run resolution V design for k=7 has 30 runs, a savings of 34 runs.

“Small, Efficient, Equireplicated Resolution V Fractions of 2k designs and their Application to Central Composite Designs”, Gary Oehlert and Pat Whitcomb, 46th Annual Fall Technical Conference, Friday, October 18, 2002.

Available as PDF at: http://www.statease.com/pubs/small5.pdf

Design-Expert version 7 7

MR5 DesignsConstruction

Designs have equireplication, so each column contains the same number of +1s and −1s.

Used the columnwise-pairwise of Li and Wu (1997) with the D-optimality criterion to find designs.

Overall our CP-type designs have better properties than the algebraically derived irregular fractions.

Efficiencies tend to be higher.

Correlations among the effects tend be lower.

Design-Expert version 7 8

MR5 DesignsProvide Considerable Savings

k 2k-p MR5 k 2k-p MR5

6 32 22 15 256 122

7 64 30 16 256 138

8 64 38 17 256 154

9 128 46 18 512 172

10 128 56 19 512 192

11 128 68 20 512 212

12 256 80 21 512 232

13 256 92 25 1024 326

14 256 106 30 1024 466

Design-Expert version 7 9

MR4 DesignsMitigate the use of Resolution III Designs

The minimum number of runs for resolution IV designs is only two times the number of factors (runs = 2k). This can offer quite a savings when compared to a regular resolution IV 2k-p fraction.

32 runs are required for 9 through 16 factors to obtain a resolution IV regular fraction.

The minimum-run resolution IV designs require 18 to 32 runs, depending on the number of factors.

• A savings of (32 – 18) 14 runs for 9 factors.

• No savings for 16 factors.

“Screening Process Factors In The Presence of Interactions”, Mark Anderson and Pat Whitcomb, presented at AQC 2004 Toronto. May 2004. Available as PDF at: http://www.statease.com/pubs/aqc2004.pdf.

Design-Expert version 7 10

MR4 DesignsSuggest using “MR4+2” Designs

Problems: If even 1 run lost, design becomes resolution IIIIII –

main effects become badly aliased.

Reduction in runs causes power loss – may miss significant effects.

Evaluate power before doing experiment.

Solution: To reduce chance of resolution loss and increase

power, consider adding some padding:

New Whitcomb & Oehlert “MR4+2” designs

Design-Expert version 7 11

MR4 DesignsProvide Considerable Savings

k 2k-p MR4+2 k 2k-p MR4+2

6 16 14 16 32 34*

7 16 16* 17 64 36

8 16 18* 18 64 38

9 32 20 19 64 40

10 32 22 20 64 42

11 32 24 21 64 44

12 32 26 22 64 46

13 32 28 23 64 48

14 32 30 24 64 50

15 32 32* 25 64 52* No savings

Design-Expert version 7 12

Two-Level Factorial Analysis

Effects Tool bar for model section tools.

Colored positive and negative effects and Shapiro-Wilk test statistic add to probability plots.

Select model terms by “boxing” them.

Pareto chart of t-effects.

Select aliased terms for model with right click.

Better initial estimates of effects in irregular factions by using “Design Model”. Recalculate and clear buttons.

Design-Expert version 7 13

Two-Level Factorial AnalysisEffects Tool Bar

New – Effects Tool on the factorial effects screen makes all the options obvious.

New – Pareto Chart

New – Clear Selection button

New – Recalculate button (discuss later in respect to irregular fractions)

Design-Expert version 7 14

Design-Expert® SoftwareFiltration Rate

Shapiro-Wilk testW-value = 0.974p-value = 0.927A: TemperatureB: PressureC: ConcentrationD: Stir Rate

Positive Effects Negative Effects

Half-Normal Plot

Ha

lf-N

orm

al %

Pro

ba

bili

ty

|Standardized Effect|

0.00 5.41 10.81 16.22 21.63

0102030

50

70

80

90

95

99

A

CD

AC

AD

Two-Level Factorial AnalysisColored Positive and Negative Effects

Design-Expert version 7 15

Two-Level Factorial AnalysisSelect Model Terms by “Boxing” Them.

Half-Normal Plot

Ha

lf-N

orm

al %

Pro

ba

bili

ty

|Standardized Effect|

0.00 5.41 10.81 16.22 21.63

0102030

50

70

80

90

95

99

A

CD

AC

AD

Half-Normal Plot

Ha

lf-N

orm

al %

Pro

ba

bili

ty

|Standardized Effect|

0.00 5.41 10.81 16.22 21.63

0102030

50

70

80

90

95

99Warning! No terms are selected.

Design-Expert version 7 16

Two-Level Factorial AnalysisPareto Chart to Select Effects

The Pareto chart is useful for showing the relative size of effects, especially to non-statisticians.

Problem: If the 2k-p factorial design is not orthogonal and balanced the effects have differing standard errors, so the size of an effect may not reflect its statistical significance.

Solution: Plotting the t-values of the effects addresses the standard error problems for non-orthogonal and/or unbalanced designs.

Problem: The largest effects always look large, but what is statistically significant?

Solution: Put the t-value and the Bonferroni corrected t-value on the Pareto chart as guidelines.

Design-Expert version 7 17

Two-Level Factorial AnalysisPareto Chart to Select Effects

Pareto Chartt-

Va

lue

of

|Eff

ect

|

Rank

0.00

2.82

5.63

8.45

11.27

Bonferroni Limit 5.06751

t-Value Limit 2.77645

1 2 3 4 5 6 7

C

AC

A

Design-Expert version 7 18

Two-Level Factorial AnalysisSelect Aliased terms via Right Click

Design-Expert version 7 19

DESIGN-EXPERT Plotclean

A: water temp B: cy cle timeC: soapD: sof tener

Half Normal plot

Half

Norm

al %

pro

bability

|Effect|

0.00 14.83 29.67 44.50 59.33

0

20

40

60

70

80

85

90

95

97

99

A

B

C

AC

Two-Level Factorial AnalysisBetter Effect Estimates in Irregular Factions

Design-Expert version 6 Design-Expert version 7Design-Expert® Softwareclean

Shapiro-Wilk testW-value = 0.876p-value = 0.171A: water temp B: cycle timeC: soapD: softener

Positive Effects Negative Effects

Half-Normal Plot

Ha

lf-N

orm

al %

Pro

ba

bili

ty

|Standardized Effect|

0.00 17.81 35.62 53.44 71.25

0102030

50

70

80

90

95

99

A

C

AC

Design-Expert version 7 20

Two-Level Factorial AnalysisBetter Effect Estimates in Irregular Factions

ANOVA for Selected Factorial ModelAnalysis of variance table [Partial sum of squares]

Sum of Mean FSource Squares DF Square Value Prob > F

Model 38135.17 4 9533.79 130.22 < 0.0001A 10561.33 1 10561.33 144.25 < 0.0001B 8.17 1 8.17 0.11 0.7482C 11285.33 1 11285.33 154.14 < 0.0001

AC 14701.50 1 14701.50 200.80 < 0.0001Residual 512.50 7 73.21Cor Total 38647.67 11

Design-Expert version 7 21

Main effects only model: [Intercept] = Intercept - 0.333*CD - 0.333*ABC - 0.333*ABD [A] = A - 0.333*BC - 0.333*BD - 0.333*ACD [B] = B - 0.333*AC - 0.333*AD - 0.333*BCD [C] = C - 0.5*AB [D] = D - 0.5*AB

Main effects & 2fi model: [Intercept] = Intercept - 0.5*ABC - 0.5*ABD [A] = A - ACD [B] = B - BCD [C] = C [D] = D [AB] = AB [AC] = AC - BCD [AD] = AD - BCD [BC] = BC - ACD [BD] = BD - ACD [CD] = CD - 0.5*ABC - 0.5*ABD

Two-Level Factorial AnalysisBetter Effect Estimates in Irregular Factions

Design-Expert version 7 22

Two-Level Factorial AnalysisBetter Effect Estimates in Irregular Factions

Design-Expert version 6 calculates the initial effects using sequential SS via hierarchy.

Design-Expert version 7 calculates the initial effects using partial SS for the “Base model for the design”.

The recalculate button (next slide) calculates the chosen (model) effects using partial SS and then remaining effects using sequential SS via hierarchy.

Design-Expert version 7 23

Two-Level Factorial AnalysisBetter Effect Estimates in Irregular Fractions

Irregular fractions – Use the “Recalculate” key when selecting effects.

Design-Expert version 7 24

General Factorials

Design:

Bigger designs than possible in v6.

D-optimal now can force categoric balance (or impose a balance penalty).

Choice of nominal or ordinal factor coding.

Analysis:

Backward stepwise model reduction.

Select factor levels for interaction plot.

3D response plot.

Design-Expert version 7 25

General Factorial DesignD-optimal Categoric Balance

Design-Expert version 7 26

General Factorial DesignChoice of Nominal or Ordinal Factor Coding

Design-Expert version 7 27

Categoric FactorsNominal versus Ordinal

The choice of nominal or ordinal for coding categoric factors has no effect on the ANOVA or the model graphs. It only affects the coefficients and their interpretation:

1. Nominal – coefficients compare each factor level mean to the overall mean.

2. Ordinal – uses orthogonal polynomials to give coefficients for linear, quadratic, cubic, …, contributions.

Design-Expert version 7 28

Nominal contrasts – coefficients compare each factor level mean to the overall mean.

Name A[1] A[2] A1 1 0 A2 0 1 A3 -1 -1

The first coefficient is the difference between the overall mean and the mean for the first level of the treatment.

The second coefficient is the difference between the overall mean and the mean for the second level of the treatment.

The negative sum of all the coefficients is the difference between the overall mean and the mean for the last level of the treatment.

Battery LifeInterpreting the coefficients

Design-Expert version 7 29

Ordinal contrasts – using orthogonal polynomials the first coefficient gives the linear contribution and the second the quadratic:

Name B[1] B[2] 15 -1 1 70 0 -2

125 1 1

B[1] = linear

B[2] = quadratic

Battery LifeInterpreting the coefficients

Polynomial Contrasts

-3

-2

-1

0

1

2

15 70 125Temperature

Design-Expert version 7 30

General Factorial AnalysisBackward Stepwise Model Reduction

Design-Expert version 7 31

Select Factor Levels for Interaction Plot

Design-Expert version 7 32

General Factorial Analysis3D Response Plot

Design-Expert® Software

wood failure

X1 = A: WoodX2 = B: Adhesive

Actual FactorsC: Applicator = brushD: Clamp = pneumaticE: Pressure = firm

chestnut

red oak

poplar

maple

pine

PRF-ET

PRF-RT

RF-RT

EPI-RT

LV-EPI-RT 38

52.5

67

81.5

96

w

ood

failu

re

A: Wood B: Adhesive

Design-Expert version 7 33

Factorial Design Augmentation

Semifold: Use to augment 2k-p resolution IV; usually as many additional two-factor interactions can be estimated with half the runs as required for a full foldover.

Add Center Points.

Replicate Design.

Add Blocks.

Design-Expert version 7 34

What’s New

General improvements Design evaluation Diagnostics Updated graphics Better help Miscellaneous Cool New Stuff

Factorial design and analysis

Response surface design

Mixture design and analysis

Combined design and analysis

Design-Expert version 7 35

Response Surface Designs

More “canned” designs; more factors and choices. CCDs for ≤ 30 factors (v6 ≤ 10 factors)

• New CCD designs based on MR5 factorials.

• New choices for alpha “practical”, “orthogonal quadratic” and “spherical”.

Box-Behnken for 3–30 factors (v6 3, 4, 5, 6, 7, 9 & 10)

“Odd” designs moved to “Miscellaneous”.

Improved D-optimal design. for ≤ 30 factors (v6 ≤ 10 factors)

Coordinate exchange

Design-Expert version 7 36

MR-5 CCDsResponse Surface Design

Minimum run resolution V (MR-5) CCDs:

Add six center points to the MR-5 factorial design.

Add 2(k) axial points.

For k=10 the quadratic model has 66 coefficients. The number of runs for various CCDs:

• Regular (210-3) = 158

• MR-5 = 82

• Small (Draper-Lin) = 71

Design-Expert version 7 37

MR-5 CCDs (k = 6 to 30)Number of runs closer to small CCD

0

100

200

300

400

500

600

0 5 10 15 20 25 30

k: # of factors

n: #

of r

uns

CCD

MR-5 CCD

SCCD

Design-Expert version 7 38

MR-5 CCDs (k=10, = 1.778)

Regular, MR-5 and Small CCDs

210-3 CCD

158 runs

MR-5 CCD

82 runs

Small CCD

71 runs

Model 65 65 65

Residuals 92 16 5

Lack of Fit 83 11 1

Pure Error 9 5 4

Corr Total 157 81 70

Design-Expert version 7 39

MR-5 CCDs (k=10, = 1.778)

Properties of Regular, MR-5 and Small CCDs

210-3 CCD

158 runs

MR-5 CCD

82 runs

Small CCD

71 runs

Max coefficient SE 0.214 0.227 16.514

Max VIF 1.543 2.892 12,529

Max leverage 0.498 0.991 1.000

Ave leverage 0.418 0.805 0.930

Scaled D-optimality 1.568 2.076 3.824

Design-Expert version 7 40

MR-5 CCDs (k=10, = 1.778)

Properties closer to regular CCD

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

2

4

6

8

10

12

14

16

S

tdE

rr o

f D

esig

n

A: A

B: B

210-3 CCD MR-5 CCD Small CCD158 runs 82 runs 71 runs

different y-axis scale

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

1

2

3

4

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

1

2

3

4

S

tdE

rr o

f D

esig

n

A: A

B: B

A-B slice

Design-Expert version 7 41

210-3 CCD MR-5 CCD Small CCD158 runs 82 runs 71 runs

all on the same y-axis scale

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

1

2

3

4

S

tdE

rr o

f D

esig

n

A: A

C: C

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

1

2

3

4

S

tdE

rr o

f D

esig

n

A: A

C: C

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

1

2

3

4

S

tdE

rr o

f D

esig

n

A: A

C: C

A-C slice

MR-5 CCDs (k=10, = 1.778)

Properties closer to regular CCD

Design-Expert version 7 42

MR-5 CCDsConclusion

Best of both worlds:

The number of runs are closer to the number in the small than in the regular CCDs.

Properties of the MR-5 designs are closer to those of the regular than the small CCDs.

• The standard errors of prediction are higher than regular CCDs, but not extremely so.

• Blocking options are limited to 1 or 2 blocks.

Design-Expert version 7 43

Practical alphaChoosing an alpha value for your CCD

Problems arise as the number of factors increase: The standard error of prediction for the face centered

CCD (alpha = 1) increases rapidly. We feel that an alpha > 1 should be used when k > 5.

The rotatable and spherical alpha values become too large to be practical.

Solution: Use an in between value for alpha, i.e. use a practical

alpha value.

practical alpha = (k)¼

Design-Expert version 7 44

Standard Error Plots 26-1 CCDSlice with the other four factors = 0

Face Centered Practical Spherical = 1.000 = 1.565 = 2.449

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

Design-Expert version 7 45

Standard Error Plots 26-1 CCDSlice with two factors = +1 and two = 0

Face Centered Practical Spherical = 1.000 = 1.565 = 2.449

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

Design-Expert version 7 46

Standard Error Plots MR-5 CCD (k=30) Slice with the other 28 factors = 0

Face Centered Practical Spherical = 1.000 = 2.340 = 5.477

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.2

0.4

0.6

0.8

1

S

tdE

rr o

f D

esig

n

A: A

B: B

Design-Expert version 7 47

Standard Error Plots MR-5 CCD (k=30) Slice with 14 factors = +1 and 14 = 0

Face Centered Practical Spherical = 1.000 = 2.340 = 5.477

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.4

0.8

1.2

1.6

2

2.4

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.4

0.8

1.2

1.6

2

2.4

S

tdE

rr o

f D

esig

n

A: A

B: B

-1.00 -0.50

0.00 0.50

1.00

-1.00

-0.50

0.00

0.50

1.00

0

0.4

0.8

1.2

1.6

2

2.4

S

tdE

rr o

f D

esig

n

A: A

B: B

Design-Expert version 7 48

Choosing an alpha value for your CCD

k Practical Spherical k Practical Spherical6 1.5651 2.4495 19 2.0878 4.35897 1.6266 2.6458 20 2.1147 4.47218 1.6818 2.8284 21 2.1407 4.58269 1.7321 3.0000 22 2.1657 4.6904

10 1.7783 3.1623 23 2.1899 4.795811 1.8212 3.3166 24 2.2134 4.899012 1.8612 3.4641 25 2.2361 5.000013 1.8988 3.6056 26 2.2581 5.099014 1.9343 3.7417 27 2.2795 5.196215 1.9680 3.8730 28 2.3003 5.291516 2.0000 4.0000 29 2.3206 5.385217 2.0305 4.1231 30 2.3403 5.477218 2.0598 4.2426

Design-Expert version 7 49

D-optimal DesignCoordinate versus Point Exchange

There are two algorithms to select “optimal” points for estimating model coefficients:

Point exchange

Coordinate exchange

Design-Expert version 7 50

D-optimal Coordinate Exchange*

Cyclic Coordinate Exchange Algorithm

1. Start with a nonsingular set of model points.

2. Step through the coordinates of each design point determining if replacing the current value increases the optimality criterion. If the criterion is improved, the new coordinate replaces the old. (The default number of steps is twelve. Therefore 13 levels are tested between the low and high factor constraints; usually ±1.)

3. The exchanges continue and cycle through the model points until there is no further improvement in the optimality criterion.

* R.K. Meyer, C.J. Nachtsheim (1995), “The Coordinate-Exchange Algorithm for Constructing Exact Optimal Experimental Designs”, Technometrics, 37, 60-69.