0% found this document useful (0 votes)
197 views48 pages

2k Factorial Lesson 2

This document describes a 2k factorial design experiment involving k factors with two levels each (often labeled + and -). It defines main effects and interactions, and shows how to calculate them from treatment means. A 22 factorial example is provided to illustrate the calculation of effects, contrasts, sums of squares, and an ANOVA table to analyze the results of such an experiment. SAS code and output for the example is also presented.

Uploaded by

Juan Kakaka
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
197 views48 pages

2k Factorial Lesson 2

This document describes a 2k factorial design experiment involving k factors with two levels each (often labeled + and -). It defines main effects and interactions, and shows how to calculate them from treatment means. A 22 factorial example is provided to illustrate the calculation of effects, contrasts, sums of squares, and an ANOVA table to analyze the results of such an experiment. SAS code and output for the example is also presented.

Uploaded by

Juan Kakaka
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 48

Statistics 514: 2k Factorial Design

Lecture 10: 2k Factorial Design


Montgomery: Chapter 6

Fall , 2005
Page 1
Statistics 514: 2k Factorial Design

2k Factorial Design
• Involving k factors
• Each factor has two levels (often labeled + and −)
• Factor screening experiment (preliminary study)
• Identify important factors and their interactions
• Interaction (of any order) has ONE degree of freedom
• Factors need not be on numeric scale
• Ordinary regression model can be employed

y = β0 + β1 x1 + β2 x2 + β12 x1 x2 + 

Where β1 , β2 and β12 are related to main effects, interaction effects defined
later.

Fall , 2005
Page 2
Statistics 514: 2k Factorial Design

22 Factorial Design
Example:

factor replicate

A B treatment 1 2 3 mean

− − (1) 28 25 27 80/3

+ − a 36 32 32 100/3

− + b 18 19 23 60/3

+ + ab 31 30 29 90/3

• Let ȳ(A+ ), ȳ(A− ), ȳ(B+ ) and ȳ(B− ) be the level means of A and B.
• Let ȳ(A− B− ), ȳ(A+ B− ), ȳ(A− B+ ) and ȳ(A+ B+ ) be the treatment
means

Fall , 2005
Page 3
Statistics 514: 2k Factorial Design

Define main effects of A (denoted again by A ) as follows:

A = m.e.(A) = ȳ(A+ ) − ȳ(A− )

= 12 (ȳ(A+ B+ ) + ȳ(A+ B− )) − 12 (ȳ(A− B+ ) + ȳ(A− B− ))


= 12 (ȳ(A+ B+ ) + ȳ(A+ B− ) − ȳ(A− B+ ) − ȳ(A− B− ))
= 12 (−ȳ(A− B− ) + ȳ(A+ B− ) − ȳ(A− B+ ) + ȳ(A+ B+ ))
=8.33

• Let CA =(-1,1,-1,1), a contrast on treatment mean responses, then


m.e.(A)= 12 ĈA

• Notice that

A = m.e.(A) = (ȳ(A+ ) − ȳ.. ) − (ȳ(A− ) − ȳ.. ) = τ̂2 − τ̂1

Main effect is defined in a different way than Chapter 5. But they are
connected and equivalent.

Fall , 2005
Page 4
Statistics 514: 2k Factorial Design

• Similarly

B = m.e.(B) = ȳ(B+ ) − ȳ(B− )

= 12 (−ȳ(A− B− ) − ȳ(A+ B− )) + ȳ(A− B+ ) + ȳ(A+ B+ ) =-5.00

Let CB =(-1,-1,1,1), a contrast on treatment mean responses, then B =m.e.(B )= 12 ĈB

• Define interaction between A and B


1
AB = Int(AB) = (m.e.(A | B+ ) − m.e.(A | B− ))
2

1 1
= (ȳ(A+ | B+ ) − ȳ(A− | B+ )) − (ȳ(A+ | B− ) − ȳ(A− | B− ))
2 2
= 12 (ȳ(A− B− ) − ȳ(A+ B− ) − ȳ(A− B+ ) + ȳ(A+ B+ )) =1.67

Let CAB = (1, −1, −1, 1), a contrast on treatment means, then
AB =Int(AB )= 12 ĈAB

Fall , 2005
Page 5
Statistics 514: 2k Factorial Design

Effects and Contrasts

factor effect (contrast)

A B total mean I A B AB

− − 80 80/3 1 -1 -1 1

+ − 100 100/3 1 1 -1 -1

− + 60 60/3 1 -1 1 -1

+ + 90 90/3 1 1 1 1

• There is a one-to-one correspondence between effects and constrasts, and


constrasts can be directly used to estimate the effects.

• For a effect corresponding to contrast c = (c1 , c2 , . . . ) in 22 design


1X
effect = ci ȳi
2 i

where i is an index for treatments and the summation is over all treatments.

Fall , 2005
Page 6
Statistics 514: 2k Factorial Design

Sum of Squares due to Effect

• Because effects are defined using contrasts, their sum of squares can also be
calculated through contrasts.

• Recall for contrast c = (c1 , c2 , . . . ), its sum of squares is


P 2
( ci ȳi )
SSContrast = P 2
ci /n
So

(−ȳ(A− B− ) + ȳ(A+ B− ) − ȳ(A− B+ ) + ȳ(A+ B+ ))2


SSA = = 208.33
4/n

(−ȳ(A− B− ) − ȳ(A+ B− ) + ȳ(A− B+ ) + ȳ(A+ B+ ))2


SSB = = 75.00
4/n

(ȳ(A− B− ) − ȳ(A+ B− ) − ȳ(A− B+ ) + ȳ(A+ B+ ))2


SSAB = = 8.33
4/n

Fall , 2005
Page 7
Statistics 514: 2k Factorial Design

Sum of Squares and ANOVA

P 2
2
y...
• Total sum of squares: SST = i,j,k yijk − N

• Error sum of squares: SSE = SST − SSA − SSB − SSAB

• ANOVA Table
Source of Sum of Degrees of Mean

Variation Squares Freedom Square F0


A SSA 1 MSA
B SSB 1 MSB
AB SSAB 1 MSAB
Error SSE N −3 MSE
Total SST N −1

Fall , 2005
Page 8
k
Statistics 514: 2 Factorial Design

SAS file and output


option noncenter
data one;
input A B resp;
datalines;
-1 -1 28
-1 -1 25
-1 -1 27
1 -1 36
1 -1 32
1 -1 32
-1 1 18
-1 1 19
-1 1 23
1 1 31
1 1 30
1 1 29
;
proc glm;
calss A B;

Fall , 2005
Page 9
Statistics 514: 2k Factorial Design

model resp=A|B;
run;
---------------------------------------------------
Sum of
Source DF Squares Mean Square F Value Pr > F
Model 3 291.6666667 97.2222222 24.82 0.0002
Error 8 31.3333333 3.9166667
Cor Total 11 323.0000000

A 1 208.3333333 208.3333333 53.19 <.0001


B 1 75.0000000 75.0000000 19.15 0.0024
A*B 1 8.3333333 8.3333333 2.13 0.1828

Fall , 2005
Page 10
Statistics 514: 2k Factorial Design

Analyzing 22 Experiment Using Regresson Model


2
Because every effect in 2 design, or its sum of squares, has one degree of freedom, it
can be equivalently represented by a numerical variable, and regression analysis can be
directly used to analyze the data. The original factors are not necessasrily continuous.
Code the levels of factor A and B as follows

A x1 B x2
- -1 - -1
+ 1 + 1

Fit regression model

y = β0 + β1 x1 + β2 x2 + β12 x1 x2 + 

The fitted model should be

A B AB
y = ȳ.. + x1 + x2 + x1 x2
2 2 2
i.e. the estimated coefficients are half of the effects, respectively.

Fall , 2005
Page 11
k
Statistics 514: 2 Factorial Design

SAS Code and Output


option noncenter
data one;
input x1 x2 resp;
x1x2=x1*x2;
datalines;
-1 -1 28
-1 -1 25
-1 -1 27
........
1 1 31
1 1 30
1 1 29
;
proc reg;
model resp=x1 x2 x1x2;
run

Fall , 2005
Page 12
Statistics 514: 2k Factorial Design

Analysis of Variance
Sum of Mean
Source DF Squares Square F Value Pr > F

Model 3 291.66667 97.22222 24.82 0.0002


Error 8 31.33333 3.91667
Corrected Total 11 323.00000

Parameter Estimates

Parameter Standard
Variable DF Estimate Error t Value Pr > |t|

Intercept 1 27.50000 0.57130 48.14 <.0001


x1 1 4.16667 0.57130 7.29 <.0001
x2 1 -2.50000 0.57130 -4.38 0.0024
x1x2 1 0.83333 0.57130 1.46 0.1828

Fall , 2005
Page 13
Statistics 514: 2k Factorial Design

23 Factorial Design
Bottling Experiment:

factor response

A B C treatment 1 2 total

− − − (1) -3 -1 -4

+ − − a 0 1 1

− + − b -1 0 -1

+ + − ab 2 3 5

− − + c -1 0 -1

+ − + ac 2 1 3

− + + bc 1 1 2

+ + + abc 6 5 11

Fall , 2005
Page 14
Statistics 514: 2k Factorial Design

factorial effects and constrasts

Main effects:

A = m.e.(A) = ȳ(A+ ) − ȳ(A− )

= 41 (ȳ(− − −) + ȳ(+ − −) − ȳ(− + −) + ȳ(+ + −) − ȳ(− − +)


+ȳ(+ − +) − ȳ(− + +) + ȳ(+ + +))
=3.00
The contrast is (-1,1,-1,1,-1,1,-1,1)

B : (−1, −1, 1, 1, −1, −1, 1, 1), B = 2.25


C : (−1, −1, −1, −1, 1, 1, 1, 1), C = 1.75

2-factor interactions:
AB : A × B componentwise, AB=.75
AC : A × C componentwise, AC=.25
BC : B × C componentwise, BC=.50
Fall , 2005
Page 15
Statistics 514: 2k Factorial Design

3-factor interaction:

1
ABC = int(ABC) = (int(AB | C+) − int(AB | C−))
2
= 14 (−ȳ(− − −) + ȳ(+ − −) + ȳ(− + −) − ȳ(+ + −)
+ȳ(− − +) − ȳ(+ − +) − ȳ(− + +) + ȳ(+ + +))
=.50

The contrast is (-1,1,1,-1,1,-1,-1,1)= A × B × C.

Fall , 2005
Page 16
Statistics 514: 2k Factorial Design

Contrasts for Calculating Effects in 23 Design

factorial effects

A B C treatment I A B AB C AC BC ABC
− − − (1) 1 -1 -1 1 -1 1 1 -1

+ − − a 1 1 -1 -1 -1 -1 1 1

− + − b 1 -1 1 -1 -1 1 -1 -1

+ + − ab 1 1 1 1 -1 -1 -1 1

− − + c 1 -1 -1 1 1 -1 -1 1

+ − + ac 1 1 -1 -1 1 1 -1 -1

− + + bc 1 -1 1 -1 1 -1 1 -1

+ + + abc 1 1 1 1 1 1 1 1

Fall , 2005
Page 17
Statistics 514: 2k Factorial Design

Estimates:
P
ȳi.
grand mean:
23
P
ci ȳi.
effect :
23−1
Contrast Sum of Squares:
P
( ci ȳi. )2 2
SSeffect = = 2n(effect)
23 /n
Variance of Estimate

σ2
Var(effect) =
n23−2
t-test for effects (confidence interval approach)

effect ± tα/2,2k (n−1) S.E.(effect)

Fall , 2005
Page 18
Statistics 514: 2k Factorial Design

Regresson Model
Code the levels of factor A and B as follows

A x1 B x2 C x3
- -1 - -1 - -1
+ 1 + 1 + 1

Fit regression model

y = β0 + β1 x1 + β2 x2 + β3 x3 + β12 x1 x2 + β13 x1 x3 + β23 x2 x3 + β123 x1 x2 x3 + 

The fitted model should be

A B C AB AC BC ABC
y = ȳ.. + x1 + x2 + x3 + x1 x2 + x1 x3 + x2 x3 + x1 x2 x3
2 2 2 2 2 2 2

i.e. β̂ = effect
2
, and

σ2 σ2
Var(β̂) = =
n2k n23

Fall , 2005
Page 19
k
Statistics 514: 2 Factorial Design

SAS Code: Bottling Experiment


data bottle;
input A B C devi;
datalines;
-1 -1 -1 -3
-1 -1 -1 -1
1 -1 -1 0
1 -1 -1 1
-1 1 -1 -1
-1 1 -1 0
1 1 -1 2
1 1 -1 3
-1 -1 1 -1
-1 -1 1 0
1 -1 1 2
1 -1 1 1
-1 1 1 1
-1 1 1 1
1 1 1 6
1 1 1 5

Fall , 2005
Page 20
Statistics 514: 2k Factorial Design

;
proc glm;
class A B C; model devi=A|B|C;
output out=botone r=res p=pred;
run;
proc univariate data=botone pctldef=4;
var res; qqplot res / normal (L=1 mu=est sigma=est);
histogram res / normal; run;
proc gplot; plot res*pred/frame; run;

data bottlenew;
set bottle;
x1=A; x2=B; x3=C; x1x2=x1*x2; x1x3=x1*x3; x2x3=x2*x3;
x1x2x3=x1*x2*x3; drop A B C;

proc reg data=bottlenew;


model devi=x1 x2 x3 x1x2 x1x3 x2x3 x1x2x3;

Fall , 2005
Page 21
Statistics 514: 2k Factorial Design

SAS output for Bottling Experiment

ANOVA Model:

Dependent Variable: devi


Sum of
Source DF Squares Mean Square F Value Pr > F
Model 7 73.00000000 10.42857143 16.69 0.0003
Error 8 5.00000000 0.62500000
CorTotal 15 78.00000000

A 1 36.00000000 36.00000000 57.60 <.0001


B 1 20.25000000 20.25000000 32.40 0.0005
A*B 1 2.25000000 2.25000000 3.60 0.0943
C 1 12.25000000 12.25000000 19.60 0.0022
A*C 1 0.25000000 0.25000000 0.40 0.5447
B*C 1 1.00000000 1.00000000 1.60 0.2415
A*B*C 1 1.00000000 1.00000000 1.60 0.2415

Fall , 2005
Page 22
Statistics 514: 2k Factorial Design

Regression Model:

Parameter Standard
Variable DF Estimate Error t Value Pr > |t|

Intercept 1 1.00000 0.19764 5.06 0.0010


x1 1 1.50000 0.19764 7.59 <.0001
x2 1 1.12500 0.19764 5.69 0.0005
x3 1 0.87500 0.19764 4.43 0.0022
x1x2 1 0.37500 0.19764 1.90 0.0943
x1x3 1 0.12500 0.19764 0.63 0.5447
x2x3 1 0.25000 0.19764 1.26 0.2415
x1x2x3 1 0.25000 0.19764 1.26 0.2415

Fall , 2005
Page 23
Statistics 514: 2k Factorial Design

General 2k Design

• k factors: A, B , . . . , K each with 2 levels (+,−)


• consists of all possible level combinations (2k treatments) each with n replicates
• Classify factorial effects:

type of effect label the number of effects

main effects (of order 1) A, B , C , . . . , K k


 
k
2-factor interactions (of order 2) AB , AC , . . . , JK  
2
 
k
3-factor interactions (of order 3) ABC ,ABD,. . . ,IJK  
3
... ... ...
 
k
k-factor interaction (of order k ) ABC · · · K  
k

Fall , 2005
Page 24
Statistics 514: 2k Factorial Design

• In total, how many effects?


• Each effect (main or interaction) has 1 degree of freedom
full model (i.e. model consisting of all the effects) has 2k − 1 degrees of
freedom.

• Error component has 2k (n − 1) degrees of freedom (why?).


• One-to-one correspondence between effects and contrasts:
– For main effect: convert the level column of a factor using − ⇒ −1 and
+⇒1
– For interactions: multiply the contrasts of the main effects of the involved
factors, componentwisely.

Fall , 2005
Page 25
Statistics 514: 2k Factorial Design

General 2k Design: Analysis

• Estimates:
P
ȳi
grand mean :
2k
For effect with constrast C = (c1 , c2 , . . . , c2k ), its estimate is
P
ci ȳi
effect = (k−1)
2
• Variance
σ2
Var(effect) =
n2k−2
what is the standard error of the effect?

• t-test for H0 : effect=0. Using the confidence interval approach,

effect ± tα/2,2k (n−1) S.E.(effect)

Fall , 2005
Page 26
Statistics 514: 2k Factorial Design

Using ANOVA model:

• Sum of Squares due to an effect, using its constrast,


P 2
ci ȳi. k−2 2
SSeffect = = n2 (effect )
2k /n

• SST and SSE can be calculated as before and a ANOVA table including SS due to
the effests and SSE can be constructed and the effects can be tested by F -tests.

Using regression:

• Introducing variables x1 , . . . , xk for main effects, their products are used for
interactsions, the following regression model can be fitted

y = β0 + β1 x1 + . . . + βk xk + . . . + β12···k x1 x2 · · · xk + 

The coefficients are estimated by half of effects they represent, that is,

effect
β̂ =
2

Fall , 2005
Page 27
Statistics 514: 2k Factorial Design

Unreplicated 2k Design
Filtration Rate Experiment

Fall , 2005
Page 28
Statistics 514: 2k Factorial Design

factor
A B C D filtration

− − − − 45
+ − − − 71
− + − − 48
+ + − − 65
− − + − 68
+ − + − 60
− + + − 80
+ + + − 65
− − − + 43
+ − − + 100
− + − + 45
+ + − + 104
− − + + 75
+ − + + 86
− + + + 70
+ + + + 96

Fall , 2005
Page 29
Statistics 514: 2k Factorial Design

Unreplicated 2k Design

• No degree of freedom left for error component if full model is fitted.


• Formulas used for estimates and contrast sum of squares are given in Slides
26-27 with n=1

• No error sum of squares available, cannot estimate σ 2 and test effects in both
the ANOVA and Regression approaches.

• Approach 1: pooling high-order interactions


– Often assume 3 or higher interactions do not occur

– Pool estimates together for error

– Warning: may pool significant interaction

Fall , 2005
Page 30
Statistics 514: 2k Factorial Design

Unreplicated 2k Design

• Approach 2: Using the normal probability plot (QQ plot) to identify significant
effects.

– Recall

σ2
Var(effect) =
2(k−2)
If the effect is not significant (=0), then the effect estimate follows
σ2
N (0, 2(k−2) )
– Assume all effects not significant, their estimates can be considered as a
σ2
random sample from N (0, 2(k−2) )

– QQ plot of the estimates is expected to be a linear line

– Deviation from a linear line indicates significant effects

Fall , 2005
Page 31
k
Statistics 514: 2 Factorial Design

Using SAS to generate QQ plot for effects

goption colors=(none);

data filter;
do D = -1 to 1 by 2;do C = -1 to 1 by 2;
do B = -1 to 1 by 2;do A = -1 to 1 by 2;
input y @@; output;
end; end; end; end;
datalines;
45 71 48 65 68 60 80 65 43 100 45 104 75 86 70 96
;

data inter; /* Define Interaction Terms */


set filter;
AB=A*B; AC=A*C; AD=A*D; BC=B*C; BD=B*D; CD=C*D; ABC=AB*C; ABD=AB*D;
ACD=AC*D; BCD=BC*D; ABCD=ABC*D;

proc glm data=inter; /* GLM Proc to Obtain Effects */


class A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD;
model y=A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD;

Fall , 2005
Page 32
Statistics 514: 2k Factorial Design

estimate ’A’ A 1 -1; estimate ’AC’ AC 1 -1;


run;

proc reg outest=effects data=inter; /* REG Proc to Obtain Effects */


model y=A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD;

data effect2; set effects;


drop y intercept _RMSE_;
proc transpose data=effect2 out=effect3;
data effect4; set effect3; effect=col1*2;
proc sort data=effect4; by effect;
proc print data=effect4;

/*Generate the QQ plot */


proc rank data=effect4 out=effect5 normal=blom;
var effect; ranks neff;
proc print data=effect5;
symbol1 v=circle;
proc gplot data=effect5;
plot effect*neff=_NAME_;
run;

Fall , 2005
Page 33
k
Statistics 514: 2 Factorial Design

Ranked Effects

Obs _NAME_ COL1 effect neff


1 AC -9.0625 -18.125 -1.73938
2 BCD -1.3125 -2.625 -1.24505
3 ACD -0.812data filter;
do D = -1 to 1 by 2;do C = -1 to 1 by 2;
do B = -1 to 1 by 2;do A = -1 to 1 by 2;
input y @@; output;
end; end; end; end;
datalines;
45 71 48 65 68 60 80 65 43 100 45 104 75 86 70 96
;5 -1.625 -0.94578
4 CD -0.5625 -1.125 -0.71370
5 BD -0.1875 -0.375 -0.51499
6 AB 0.0625 0.125 -0.33489
7 ABCD 0.6875 1.375 -0.16512
8 ABC 0.9375 1.875 -0.00000
9 BC 1.1875 2.375 0.16512
Fall , 2005
Page 34
Statistics 514: 2k Factorial Design

10 B 1.5625 3.125 0.33489


11 ABD 2.0625 4.125 0.51499
12 C 4.9375 9.875 0.71370
13 D 7.3125 14.625 0.94578
14 AD 8.3125 16.625 1.24505
15 A 10.8125 21.625 1.73938

Fall , 2005
Page 35
Statistics 514: 2k Factorial Design

QQ plot

Fall , 2005
Page 36
Statistics 514: 2k Factorial Design

Filtration Experiment Analysis

Fit a linear line based on small effects, identify the effects which are potentially
significant, then use ANOVA or regression fit a sub-model with those effects.

1. Potentially significant effects: A, AD, C, D, AC .

2. Use main effect plot and interaction plot

3. ANOVA model involving only A, C , D and their interactions (projecting the


original unreplicated 24 experiment onto a replicated 23 experiement)

4. regression model only involving A, C , D , AC and AD .

5. Diagnostics using residuals.

Fall , 2005
Page 37
Statistics 514: 2k Factorial Design

Interaction Plots for AC and AD

* data step is the same.

proc sort; by A C;
proc means noprint;
var y; by A C;
output out=ymeanac mean=mn;

symbol1 v=circle i=join; symbol2 v=square i=join;


proc gplot data=ymeanac; plot mn*A=C;
run;

* similar code for AD interaction plot

Fall , 2005
Page 38
Statistics 514: 2k Factorial Design

Fall , 2005
Page 39
Statistics 514: 2k Factorial Design

Fall , 2005
Page 40
Statistics 514: 2k Factorial Design

ANOVA with A, C and D and their interactions

proc glm data=filter;


class A C D;
model y=A|C|D;
==================================
Source DF Sum Squares Mean Square F Value Pr > F
Model 7 5551.437500 793.062500 35.35 <.0001
Error 8 179.500000 22.437500
Cor Total 15 5730.937500

Source DF Type I SS Mean Square F Value Pr > F


A 1 1870.562500 1870.562500 83.37 <.0001
C 1 390.062500 390.062500 17.38 0.0031
A*C 1 1314.062500 1314.062500 58.57 <.0001
D 1 855.562500 855.562500 38.13 0.0003
A*D 1 1105.562500 1105.562500 49.27 0.0001
C*D 1 5.062500 5.062500 0.23 0.6475
A*C*D 1 10.562500 10.562500 0.47 0.5120
*ANOVA confirms that A, C, D, AC and AD are significant effects

Fall , 2005
Page 41
k
Statistics 514: 2 Factorial Design

Regression Model
* the same date step

data inter; set filter; AC=A*C; AD=A*D;

proc reg data=inter; model y=A C D AC AD;


output out=outres r=res p=pred;

proc gplot data=outres; plot res*pred; run;

===========================
Dependent Variable: y
Analysis of Variance
Sum of Mean
Source DF Squares Square F Value Pr > F
Model 5 5535.81250 1107.16250 56.74 <.0001
Error 10 195.12500 19.51250
Corrected Total 15 5730.93750

Root MSE 4.41730 R-Square 0.9660

Fall , 2005
Page 42
Statistics 514: 2k Factorial Design

Dependent Mean 70.06250 Adj R-Sq 0.9489


Coeff Var 6.30479

Parameter Estimates
Parameter Standard
Variable DF Estimate Error t Value Pr > |t|

Intercept 1 70.06250 1.10432 63.44 <.0001


A 1 10.81250 1.10432 9.79 <.0001
C 1 4.93750 1.10432 4.47 0.0012
D 1 7.31250 1.10432 6.62 <.0001
AC 1 -9.06250 1.10432 -8.21 <.0001
AD 1 8.31250 1.10432 7.53 <.0001

Fall , 2005
Page 43
Statistics 514: 2k Factorial Design

Response Optimization / Best Setting Selection

Use x1 , x3 , x4 for A, C , D ; and x1 x3 , x1 x4 for AC ,AD respectively. The regresson


model gives the following function for the response (fitration rate):

y = 70.06 + 10.81x1 + 4.94x3 + 7.31x4 − 9.06x1 x3 + 8.31x1 x4

Want to maximize the response. Let D be set at high level (x4 = 1)

y = 77.37 + 19.12x1 + 4.94x3 − 9.06x1 x3

Contour plot

goption colors=(none);
data one;
do x1 = -1 to 1 by .1;
do x3 = -1 to 1 by .1;
y=77.37+19.12*x1 +4.94*x3 -9.06*x1*x3 ; output;
end; end;
proc gcontour data=one; plot x3*x1=y;
run; quit;

Fall , 2005
Page 44
Statistics 514: 2k Factorial Design

Contour Plot for Response Given D

Fall , 2005
Page 45
Statistics 514: 2k Factorial Design

Residual Plot

Fall , 2005
Page 46
Statistics 514: 2k Factorial Design

Some Other Issues

• Half normal plot for (xi ), i = 1, . . . , n:


– let x̃i be the absolute values of xi
– sort the (x̃i ): x̃(1) ≤ ... ≤ x̃(n)
n+i
– calculate ui = Φ−1 ( 2n+1 ), i = 1, ..., n
– plot x̃(i) against ui
– look for a straight line

Half normal plot can also be used for identifying important factorial effects

• Other methods to identify significant factorial effects (Lenth method).


Hamada&Balakrishnan (1998) analyzing unreplicated factorial experiments: a
review with some new proposals, statistica sinica.

• Detect dispersion effects


• Experiment with duplicate measurements
– for each treatment combination: n responses from duplicate

Fall , 2005
Page 47
Statistics 514: 2k Factorial Design

measurements

– calculate mean ȳ and standard deviation s.

– Use ȳ and treat the experiment as unreplicated in analysis

Fall , 2005
Page 48

You might also like