0.

Statistical Design and Analysis of Experiments List of contents, cont.

Part Two
6.1: Factorial experiments - introduction
Lecture notes 6.7: Blocking in factorials
Fall semester 2007
6.9: montgomery example p. 164

Henrik Spliid 6.11: Interaction plot

Informatics and Mathematical Modelling

6.14: Normal probability plot for residuals

Technical University of Denmark 7.1: Factorial experiments with two level factors

7.7: A (very) small example of a 2×2 design

1 2

0.2 0.3

7.10: Yates algorithm by an example 8.20: Generalization of partial confounding calculations

7.12: Numerical example with 3 factors 8.21: Example of partially confounded design
8.1: Block designs for two level factorials 9.1: Fractional designs for two level factorials
8.5: How-to-do blocking by confounding 9.3: Alternative method of construction (tabular method)
8.6: Yates algorithm and blocking 9.7: Generator equation and alias relations
8.7: The confounded block design (what happens?) 9.8: Analysis of data and the underlying factorial
8.9: Construction of block design by the tabular method 9.12: 5 factors in 8 measuremets
8.12: A few generalizations on block designs 9.13: Alias relations for model without high order interactions
8.14: The tabular method for 2× blocks (example) 9.14: Construction of 1/4times25 design (tabular method)
8.15: Partially confounded two level experiments

3 4
 0.4 6.1

Factorial experiments - introduction

10.1: A large example on 2 level factorials
10.11: Summary of analyses (example) Design with two factors
6 measurements
10.12: Combining main effects and interaction estimates B=1 y y
5 6

B=0 y y y3 y4
1 2
A=0 A=1

The estimate of the A-effect based on y:

y = [(y3 + y4 ) − (y1 + y2 )]/2

d
A

5 6

6.2 6.3

One-factor-at-the-time or factorial design

Design with two factors
4 measurements
d d
B=1 z3 z4 Are Ay and Az equivalent ?
d
VarAy = ?
d
VarAz = ?
B=0 z z
1 2
A=0 A=1
Additive model:

Response = µ + A + B + residual
The estimate of the A-effect based on z:

z = [(z2 + z4 ) − (z1 + z3 )]/2

d
A Can it always be applied?

7 8
 6.4 6.5

Factorial designs and interaction

More complicated model:

Design with two factors Response and A−effect
Response = µ + A + B + AB + residual 4 measurements at the two B−levels
B=1 z =30 z =52
3 4
80
Is it more needed for factorial designs than for block designs, for example, where 60 B=1
40
B=0
additivity is often assumed? 20
B=0 z =20 z =40 0
1 2
A=0 A=1 A=0 A=1
If interaction is present, then: which design is best ?

Usage of measurements: which design is best ? The change in the response when factor A is changed is the same at both B-levels
⇐⇒ no interaction
In general: How should a factorial experiment be carried out ?

9 10

6.6 6.7

Design with two factors Response and A−effect

Blocking in factorials: Two alternative factorial designs
4 measurements at the two B−levels
B=1 z =40 z4=12
3
80
60 B=0
40 Complete randomization, 19th and 20th October
20
z =20 z2=50 B=1
Additive Temperature
B=0 1 0
10oC 20oC 30oC 40oC
A=0 A=1 A=0 A=1
5% yy yy yy yy
10% yy yy yy yy

The change in the response when factor A is changed depends on the B-level ⇐⇒ Yijk = µ + ai + cj + acij + Eijk
interaction

A completely randomized 2×4 factorial with two measurements per factor combi-
The second situation is often the case in factorial experiments
nation conducted over, say, two days. The design is one block of size 16.
Never use one-factor-at-the-time designs. There exist better alternatives in all
situations.

11 12
 6.8 6.9

Replication 1, October 19th Example from Montgomery p 164

Additive Temperature
10oC 20oC 30oC 40oC
5% y y y y Material Temperature
10% y y y y type 15oF 70oF 125oF
1 130 155 34 40 20 70
Replication 2, October 20th Data 74(?) 180 80 75 82 58
Additive Temperature Averages y=134.75 y=57.25 y=57.50
10oC 20oC 30oC 40oC 2 150 188 136 122 25 70
5% y y y y Data 159 126 106 115 58 45
10% y y y y Averages y=155.75 y=119.75 y=49.50
3 138 110 174 120 96 104
Yijk = µ + ai + cj + acij + Dayk + Zijk Data 168 160 150 139 82 60
Averages y=144.00 y=145.75 y=85.50
A completely randomized 2×4 factorial with one measurement per factor combi-
nation, but replicated twice, one replication per day, i.e. two blocks of size 8.
Yijk = µ + mi + tj + mtij + Eijk
Never use the first design. Why ?

13 14

6.10 6.11

A better design
Response and temperature
Round I effects for 3 materials
Material Temperature 175
type 1 15oF 70oF 125oF 150
1 yy yy yy 125
2 yy yy yy 100
3 yy yy yy m=3
75
m=1
Round II 50 m=2
Material Temperature
15 70 125
type 15oF 70oF 125oF
1 yy yy yy
2 yy yy yy The figure indicates a possible interaction between materials and temperature.
3 yy yy yy
It is a common case that different ’materials’ react differently to fx temperature
Yijk = µ + mi + tj + mtij + Rk + Zijk treatments.

Give (at least) three reasons why this design is to be preferred.

15 16
 6.12 6.13

ANOVA and estimation in factorial design

Estimates of parameters Estimates of parameters
for full model for additive model
Yijk = µ + mi + tj + mtij + Eijk
c
µ = Y ... c
µ = Y ...
ANOVA for battery data
Source of var. SSQ d.f. s2 EMS F-test d
m i = Y i.. − Y ... d
m i = Y i.. − Y ...
Materiel 10684 2 5342 σ 2 + 12φm 7.91
Temperature 39119 2 19559 σ 2 + 12φt 28.97
tbj = Y .j. − Y ... tbj = Y .j. − Y ...
Interaction 9614 4 2403 σ 2 + 4φmt 3.56
Residual 18231 27 675.2 σ2
ij = Y ij. − Y i... − Y .j. + Y ...
d d
mt mtij = 0 (not in model)
Total 77647 35
c2
σ = s2resid c2
σ = (SSQresid + SSQmt )/(fresid + fmt )
E E
F (4, 27)0.05 = 2.73 =⇒ all parameters in the model are significant at the 5% level
of significance.

17 18

6.14 7.1

Model control based on residuals Factorial experiments with two-level factors

3 0.999
Standard−normal fractiles

(i−0.5)/n 0.99 The simplest example: 2 factors at 2 levels.

2
0.9
1 0.8
1. factor is called A (can be a temperature fx) (the supposedly most important
0.7
0.6
factor)
0 0.5
0.4
0.3
−1 0.2 2. factor is called B (can be a concentration of an additive)(the supposedly next
0.1
most important factor)
−2
0.01

−3 0.001
For each factor combination r measurements are carried out (completely random-
−100 −50 0 50 100 ized):
Residuals

19 20
 7.2 7.3

22 factorial design Restrictions on parameters: fx A0 + A1 = 0 =⇒

B=0 B=1
Y001 Y011
A=0 : : A0 = - A1 and B0 = - B1 and
Y00r Y01r
Y101 Y111
AB00 = - AB10 = - AB01 = AB11
A=1 : :
Y10r Y11r
All parameters have only one numerical value, positive or negative, depending on
the factor level(s).
Yijk = µ + Ai + Bj + ABij + Eijk

Both indices i and j can take the values ’0’ or ’1’.

µ, Ai, Bj and ABij are the parameters of the model

21 22

7.4 7.5

Effects. Special concept for 2 level factors Special notation for 2k design
Two factors, k = 2
Effect = change in response when the factor is changed from level ’0’ to ’1’, thus
Standard notation for cell sums
B=1 b ab
A-effect: A = A1 - A0 = 2A1 (main effect)
B-effect: B = B1 - B0 = 2B1 (main effect) sum of r measurements
B=0 (1) a
AB-effect: AB = AB11 - AB10 = 2AB11 (interaction) A=0 A=1

B=0 B=1
In General:
A=0 (1) b
k factors at 2 levels: A 2k factorial experiment A=1 a ab

Fx ’a’ = Prk=1 Y10k , the sum in the cell where the factor A is at level ’1’ while factor
B is at level ’0’.

23 24
 7.6 7.7

Parameters, effects and estimation A (very) small numerical example

d Y = response = purity in solution after 48 hours

A-parameter : A 1 = ( - (1) + a - b + ab)/4r
d
B-parameter : B 1
d
= ( - (1) - a + b + ab)/4r A = 1. factor = temperature (4oC , 20oC)
AB-parameter : AB 11 = (+(1) - a - b + ab)/4r
B = 2. factor = concentration of additive (5%, 10%)
c d
A-effect : A = ( - (1) + a - b + ab)/2r = 2A1
c d
B-effect : B = ( - (1) - a + b + ab)/2r = 2B 1 B=0 B=1
d d
AB-effect : AB = (+(1) - a - b + ab)/2r = 2AB 11 12.1 19.8
A=0
14.3 21.0
17.9 24.3
A=1
19.1 23.4

(1)=26.4 b=40.8
a=37.0 ab=47.7

25 26

7.8 7.9

Standard ANOVA table for example Estimation in detail

Source of var. SSQ d.f. s2 F-value (1) a b ab
26.4 37.0 40.8 47.7
A: temp 38.28 1 38.28 35.75
B: conc 78.75 1 78.75 73.60
AB: interaction 1.71 1 1.71 1.60 c
µ = (+26.4 + 37.0 + 40.8 + 47.7)/(22 · 2) = 18.99
Residual 4.27 4 1.07 c
1 = ( - 26.4 + 37.0 - 40.8 + 47.7)/(2 · 2) =
2
A 2.19
Total 123.01 7 c
A0 = - A 1 = −2.19
c c c c
A = A 1 - A0 = 2A1 = 4.38

c
1 = ( - 26.4 - 37.0 + 40.8 + 47.7)/(2 · 2) =
2
Critical F-value: F (1, 4)0.05 = 7.71 =⇒ B 3.14
c
B 0 = - B 1 = - 3.14
c c c c
B = B 1 - B 0 = 2B 1 = 6.28
main effects (highly) significant
interaction not significant c2
σ = (SSQAB + SSQresid )/(1 + 4) = 1.196 ' 1.12
(pooled estimate)

27 28
 7.10 7.11

Yates algorithm, testing and estimation Explanation:

Cell sums: Organized in ’standard order’: (1), a, b, ab
Yates algorithm for k = 2 factors Column I:
Cell sums I II = contrasts SSQ Effects 63.4 = +26.4+37.0 (sum of two first in previous column)
(1) = 26.4 63.4 151.9 = [I] − µ
c = 18.99 88.5 = +40.8+47.7 (sum of two next)
d
a = 37.0 88.5 17.5 = [A] 38.25 A = 4.38 10.6 = - 26.4+37.0 (reverse difference of two first)
d
b = 40.8 10.6 25.1 = [B] 78.75 B = 6.28 6.9 = - 40.8+47.7 (reverse difference of two next)
d
ab = 47.7 6.9 - 3.7 = [AB] 1.71 AB = - 0.93
Column II: Same procedure as for column I (63.4+88.5=151.9)
The important concept about Yates’ algorithm is that is SSQA: [A]2/(2k · 2) = 38.25 (k=2) and likewise for B and AB
represents the transformation of the data to the contrasts - d
A-Effect: A = [A]/(2k−1 · 2) = 4.38 and likewise for B and AB
and subsequently to the estimates and the sums of squares!
The procedure for column I is repeated k times for the 2k design
The sums of squares and effects appear in the ’standard order’

29 30

7.12 7.13

Numerical example with three factors (coded data) Yates algorithm for k = 3 factors
Cell sums I II III = contrasts SSQ Effects
(1) =-4 -3 1 16 = [I] - c
µ = 1.00
B=0 B=1 B=0 B=1 c
a =1 4 15 24 = [A] 36.00 A = 3.00
A=0 - 3, - 1 - 1, 0 - 1, 0 1, 1 b =-1 2 11 18 = [B] 20.25 Bc
= 2.25
A=1 0, 1 2, 3 2, 1 6, 5 ab =5 13 13 6 = [AB] d
2.25 AB = 0.75
C=0 C=1 c =-1 5 7 14 = [C] 12.25 Cc
= 1.75
d
ac =3 6 11 2 = [AC] 0.25 AC = 0.25
d
bc =2 4 1 4 = [BC] 1.00 BC = 0.50
(1) = - 4 b = - 1 c = - 1 bc = +2 abc = 11 9 5 4 = [ABC] d
1.00 ABC = 0.50
a = +1 ab = +5 ac = +3 abc = 11
SSQresid = [((−3)2 + (−1)2) − (−3 − 1)2/2] + ....
= 2.00 + . . . + 0.50 = 5.00, s2resid = SSQresid /8 = 0.625
(variation within cells, r - 1 = 2 - 1 degrees of freedom per cell)
c c c c
SSQA = [A]2/(r · 2k ), Effect A = [A]/(r · 2k−1 ), parameter A1 = [A]/(r · 2 ), A0 = −A1 , and
k
c c
A = 2A1 .

31 32
 8.1 8.2

Block designs, principles and construction Discussion af the randomized plan

Problem
The total time needed to carry out the plan is 1 hour for C0 treatments and 2 hours
Example: factors A, B and C:
for C1 treatments: 2 + 1 + 1 + 2 + 1 + 1 + 2 + 2 = 12 hours.
Recipes A0 ⇔ temp = 20oC A1 ⇔ temp = 28oC
B0 ⇔ conc = 1% B1 ⇔ conc = 2%
C0 ⇔ time = 1 hour C1 ⇔ time = 2 hours Suggestion
Distribute the 8 experiments randomly over two days with 6 hours per day:
Day 1 ∼ 6 hours Day 2 ∼ 6 hours
The treatments are (1) a b ab c ac bc abc bc a b c (1) ab abc ac

Is it balanced with respect to factors and days ?

A randomized (with respect to days) plan bc a b c (1) ab abc ac
Is this a good design ? What can go wrong ? What kind of variable is ’Days’ ?

33 34

8.3 8.4

An experiment with no influence from days The same experiment if D1 and D2 (days) in fact are different

Day 1: D1 = 0 bc = 21 a = 23 b = 16 c = 20 Day 1: D1 = +8 bc = 29 a = 31 b = 24 c = 28
Day 2: D2 = 0 (1)=14 ab = 25 abc = 25 ac = 29 Day 2: D2 = +2 (1) = 16 ab = 27 abc = 27 ac = 31
Cell sums I II III = contrasts SSQ Effects Cell sums I II III = contrasts SSQ Effects Day effect
(1) = 14 37 78 173 = [I] - c
µ = 21.625 c
(1) = 16 47 98 213 = [I] - µ = 26.625 yes
c c
a = 23 41 95 31 = [A] 120.125 A = 7.75 a = 31 51 115 19 = [A] 45.125 A = 4.75 yes
c c
b = 16 49 18 1 = [B] 0.125 B = 0.25 b = 24 59 18 1 = [B] 0.0125 B = 0.25
d
ab = 25 46 13 -5 = [AB] 3.125 AB = −1.25 ab = 27 56 1 -17 = [AB] d
36.125 AB = −4.25 yes
c c
c = 20 9 4 17 = [C] 36.125 C = 4.25 c = 28 15 4 17 = [C] 36.125 C = 4.25
d
ac = 29 9 -3 -5 = [AC] 3.125 AC = −1.25 ac = 31 3 -3 -17 = [AC] d
36.125 AC = −4.25 yes
d
bc = 21 9 0 -7 = [BC] 6.125 BC = −1.75 bc = 29 3 -12 -7 = [BC] d
6.125 BC = −1.75
d
abc = 25 4 -5 -5 = [ABC] 3.125 ABC = −1.25 abc = 27 -2 -5 7 = [ABC] d
6.125 ABC = 1.75 yes
D1 and D2 are contributions from the two days (none here). The experimentor cannot know (or estimate) the difference between days.
What happens if D1 and D2 are in fact not identical (there is a day-today effect) ? The difference between days contaminates the results.

35 36
 8.5 8.6
How can we place the 8 measurements on the two days in
such a way that the influence from days is under control ? Yates algorithm - schematically - once again:

(1) a b ab c ac bc abc
Answer: Let ’Days’ (blocks) follow one of the effects in the model: [I] = +1 +1 +1 +1 +1 +1 +1 +1
[A] = -1 +1 -1 +1 -1 +1 -1 +1
Yijk = µ + Ai + Bj + ABij + Ck + ACik + BCjk + ABCijk + Error + Day` [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
Which term could be used ? Not a main effect, but some higher order [AC] = +1 -1 +1 -1 -1 +1 -1 +1
[BC] = +1 +1 -1 -1 -1 -1 +1 +1
term, for example ABC (why ABC ?):
[ABC] = -1 +1 +1 -1 +1 -1 -1 +1
We want the confounding Blocks = ABC .
Note that any two rows are ’orthogonal’ (product sum = zero).
We say : defining relation I = ABC . . . but how do we do it?
Thus [A] and [B], for example, are orthogonal contrasts.
Look at how contrasts for effects are calculated : The ’index’ for ABCijk is i · j · k if indices are - 1 or + 1 like in Yates’ algoritm.
Choose ` = i · j · k =+1 for a b c abc og -1 for (1) ab ac bc =>
the two blocks wanted.

37 38

8.7 8.8

Real data with a certain influence (unknown in practice) from blocks (days):
The confounded block design: Blocks = ABC
Day 1: D1 = +8 ab = 24 bc = 29 (1) = 20 ac = 28
Day 2: D2 = +2 b = 26 a = 30 abc = 36 c = 24
Ideal data without influence from blocks:
Day 1: D1 = 0 ab = 16 bc = 21 (1) = 12 ac = 20 Cell sums I II III = contrasts Day effect
Day 2: D2 = 0 b = 24 a = 28 abc = 34 c = 22 (1) = 20 50 100 217 = [I] yes
a = 30 50 117 19 = [A]
Cell sums I II III = contrasts b = 26 52 8 13 = [B]
(1) = 12 40 80 177 = [I] ab = 24 65 11 -9 = [AB]
a = 28 40 97 19 = [A] c = 24 10 0 17 = [C]
b = 24 42 8 13 = [B] ac = 28 -2 13 3 = [AC]
ab = 16 55 11 - 9 = [AB] bc = 29 4 -12 13 = [BC]
c = 22 16 0 17 = [C] abc = 36 7 3 15 = [ABC] yes
ac = 20 -8 13 3 = [AC]
bc = 21 -2 - 24 13 = [BC]
abc = 34 13 15 39 = [ABC] What has changed and what has not changed? Why?
The effect from days is controlled (not eliminated) only to influence the ABC
What happens if the two days in fact influence the results differently (there is a
interaction term (block confounding).
day-to-day effect) ?

39 40
 8.9 8.10

Construction using the tabular method : Analysis of variance for block confounded design

Arrange data in standard order and use column multiplication :

In the example we imagine that r=2 measurements per factor combination were
Factor levels Block no. = used. The residual SSQ is computed as the variation between these two measure-
Code A B C ABC = A·B·C ments giving a total residual sum of squares with 8 degrees of freedom.
(1) -1 -1 -1 -1
a +1 -1 -1 +1
Correspondingly the responses on slide 8.8 (bottom) are sums of 2 measurements.
b -1 +1 -1 +1
ab +1 +1 -1 -1
c -1 -1 +1 +1
ac +1 -1 +1 -1
bc -1 +1 +1 -1
abc +1 +1 +1 +1

Block no. −1 => one block, Block no. +1 => the other block

41 42

8.11 8.12

ANOVA for block confounded three factor design A few generalizations

I2 = BCD ∼ Days
2 (1) bc d bcd
Effects SSQ d.f. s F-value
A 22.56 1 22.56 9.12 A 24 factorial design in 4 blocks of 4 : I1 = ABC ∼ abd acd ab ac
B 10.56 1 10.56 4.27 Operators a abc b c
AB 5.06 1 5.06 2.05 bd cd ad abcd
C 18.05 1 18.05 7.30
AC 0.56 1 0.56 0.22 The principal block: (1) bc abd acd
BC 10.56 1 10.56 4.26
ABC = Blocks 14.06 1 14.06 not relevant
b× (1) bc abd acd = b b2c ab2d abcd ⇒ b c ad abcd = another block!
Residual 19.80 8 2.475 Multiply any block with an ’element’ that is not in the block, and you get another
Total 101.24 15 block.
F (1, 8)0.05 = 5.32 =⇒ A and C main effects are significant. The B effect is only Total block variation = ABC + BCD + ABC·BCD = ABC + BCD + AD
significant at the 10% level of significance, and so is BC. When analyzing the data from the above 24 design all effects A, B, AB, ... , ABCD
except ABC, BCD and AD can be estimated and tested.
The ABC effect cannot be tested because it is confounded with blocks (days)
(does it seem to be a real problem ?). ABC, BCD and AD are confounded with blocks

43 44
 8.13 8.14

Construction principle : Introduce blocks into factorial by confounding Construction using the tabular method
Effect Confound Factor levels Four different Principal
Level Code A B C D ABC BCD blocks block
A (1) −1 −1 −1 −1 −1 −1 1 : ( −1 , −1) (1)
B a +1 −1 −1 −1 +1 −1 2 : ( +1 , −1)
AB b −1 +1 −1 −1 +1 +1 4 : ( +1 , +1)
AC ab +1 +1 −1 −1 −1 +1 3 : ( −1 , +1)
BC c −1 −1 +1 −1 +1 +1 4 : ( +1 , +1)
ABC = I1 ac +1 −1 +1 −1 −1 +1 3 : ( −1 , +1)
D bc −1 +1 +1 −1 −1 −1 1 : ( −1 , −1) bc
AD <= ABC·BCD abc +1 +1 +1 −1 +1 −1 2 : ( +1 , −1)
BD d −1 −1 −1 +1 −1 +1 3 : ( −1 , +1)
ABD ad +1 −1 −1 +1 +1 +1 4 : ( +1 , +1)
ACD bd −1 +1 −1 +1 +1 −1 2 : ( +1 , −1)
BCD = I2 abd +1 +1 −1 +1 −1 −1 1 : ( −1 , −1) abd
ABCD cd −1 −1 +1 +1 +1 −1 2 : ( +1 , −1)
acd +1 −1 +1 +1 −1 −1 1 : ( −1 , −1) acd
bcd −1 +1 +1 +1 −1 +1 3 : ( −1 , +1)
All effects ABC, BCD and ABC·BCD = AD will be confounded with blocks. abcd +1 +1 +1 +1 +1 +1 4 : ( +1 , +1)

45 46

8.15 8.16

Resolving block confoundings for AB with one more experiment:

Partially confounded 2k factorial experiment Suppose we also want to assess the interaction term AB.
We need an experiment in which AB is not confounded:
B=0 B=1
A=0 (1) b
A=1 a ab (1)2 a2 b2 ab2
Experiment 2 : I=B
batch 3 batch 4
Suppose batches=blocks, block size = 2 : using two new batches.

(1)1 ab1 a 1 b1 Model again: Yijν = µ + Ai + Bj + ABij + Eijν + batches (blocks)

Experiment 1 : I = AB
batch 1 batch 2
The B main effect is confounded with blocks in experiment 2, but AB is not.
Model as usual: Yijν = µ + Ai + Bj + ABij + Eijν + batches (blocks) AB can then be estimated in experiment 2.

AB interaction confounded with blocks in experiment 1. The price paid is that the main effect B can only be estimated in experiment 1
and AB only in experiment 2: Partial confounding.

47 48
 8.17 8.18

Analyze both experiments using contrasts : Use of unconfounded contrasts for effects:
Unconfounded contrasts The two (unconfounded) A-contrasts can be combined into an
[A]1 = −(1)1 + a1 − b1 + ab1 (from experiment 1) estimate of A and a part which expresses uncertainty:
[A]2 = −(1)2 + a2 − b2 + ab2 (from experiment 2)
[A]total = [A]1 + [A]2 (both experiments combined)
[B]1 = −(1)1 − a1 + b1 + ab1 (from experiment 1)

[AB]2 = +(1)2 − a2 − b2 + ab2 (from experiment 2)

[A]difference = [A]1 − [A]2 (both experiments combined)

Confounded contrasts Sums of squares for A and between unconfounded A’s:

[AB]1 = +(1)1 − a1 − b1 + ab1 (from experiment 1)
SSQA = [A]2total/(2 · 22), df = 1
[B]2 = −(1)2 − a2 + b2 + ab2 (from experiment 2)
SSQU ncert,A = ([A]21 + [A]22 )/22 − SSQA, df = 2 − 1
Blocks
Variation calculated as usual: From block totals
Block (batch) totals = T1, T2, T3, T4, and Ttot = T1 + T2 + T3 + T4
SSQblocks = (T12 + T22 + T32 + T42)/2 − Ttot
2
/8, df = 4 − 1 =3.
49 50

8.19 8.20

Estimates of effects: Generalization:

A 2k factorial partially confounded, in principle as above:
d
A = [A]total /(2 · 22−1) Full precision
Fx: RA = number of unconfounded A-contrasts : [A]1, [A]2, . . . , [A]RA
d
B = [B]1/(1 · 22−1) Half precision
Assume r repetitions (most often r = 1) for each response within the blocks.
c
d
AB = [A]2/(1 · 22−1) Half precision [A] = [A]1 + [A]2 + . . . + [A]RA A = [A]/(RA · r · 2k−1)

c c c
1 = −A0 = [A]/(RA · r · 2 ) = σ 2/(RA · r · 2k−2)
k
In general A Var{A}
Estimate = [Contrast]/(R · 2k−1)
R = number of times the effect is unconfounded in the experiment SSQA = [A]2/(RA · r · 2k ) fA = 1

PRA
Here: RA = 2 , RB = 1 , RAB = 1, and r = 1 (is assumed here). SSQUncertainty,A = 2
i=1 [A]i /(r · 2k ) − SSQA fUncertainty,A = RA − 1

This calculation is done for all unconfounded effect-contrasts.

Block variation is calculated as the variation between blocks disregarding the factors.
It contains block effects and confounded factor effects.

51 52
 8.21 8.22

An example Calculations for a twice determined effect

(1)1=15 ab1=7 a1=9 b1=5

[A]1 = - 15+7+9 - 5 = - 4 , [A]2 = - 11+7 - 12+8 = - 8
Exp. 1 : I = AB
batch 1 batch 2
[A] = [A]1 + [A]2 = - 12 , ([A]3 ∼ blocks)
(1)2=11 a2=7 b2=12 ab2=8
Exp. 2 : I=B
batch 3 batch 4 d
A = - 12/(2 · 22−1) = - 3.0 ,
(1)3=9 b3=11 a3=8 ab3=6
Exp. 3 : I=A SSQA = ( - 12)2/(2 · 22) = 18.0 ,
batch 5 batch 6

SSQU ncert,A = [(−4)2 + ( - 8)2]/22 - 18 = 2.0

Model for experiment: Yijν = µ + Ai + Bj + ABij + Eijν +Block effects
Likewise for B and AB.

53 54

8.23 9.1

Completing the ANOVA table Fractional 2k designs

Block totals : T1 = 15 + 7 = 22, T2 = 9 + 5 = 14, ... ,[T6] = 8 + 6 = 14
Example: factors A, B and C (the weights of 3 items) from 1.4 again
SSQblocks = Pi Ti2/2 − (Pi Ti)2/12 =
(222 + 142 + . . . + 142)/2 − (22 + 14 + . . . + 14)2/12 = 28.0 , d.f. = 5
The complete factorial design (1) a b ab c ac bc abc
and it contains blocks and confounded factor effects
The weighing design from 1.4 was (1) ab ac bc

ANOVA table for example d

Estimate, for example: A = [ - (1) +ab +ac - bc]/2
Source d.f. SSQ s2 F-value p-value Precision
A 1 18.0 18.0 2.45 0.22 2/3
B 1 18.0 18.0 2.45 0.22 2/3
AB 1 2.0 2.0 0.27 0.64 2/3
Blocks+(A,B,AB) 5 28.0 5.60 (0.76) (0.63) –
Error 3 22.0 7.33
Total 11 88.0

55 56
 9.2 9.3

Illustration by removing columns from contrast table An alternative method of construction

Contrasts (1) a b ab c ac bc abc Example: the complete 22 factorial
[I] = +1 +1 +1 +1
[A] = -1 +1 +1 - 1 Effects
[B] = -1 +1 - 1 +1 I Level
([AB]) = +1 +1 -1 -1 A A-effect
[C] = -1 -1 +1 +1 B B-effect
([AC]) = +1 -1 +1 - 1 AB AB-interaction
([BC]) = +1 -1 - 1 +1
([ABC]) = -1 -1 -1 -1 Method: Introduce the extra factor, C, by confounding it with a A, B or AB. Which
one of these can (most probably) be 0 ?

Answer: AB (if any at all)

Note: [A] = −[BC], [B] = −[AC], [C] = −[AB] and [I] = −[ABC] ⇐⇒
Confounding of factor effects. Confound C = AB ⇐⇒ Defining relation I = ABC

A and B form the complete underlying factorial. The factor C is introduced into
the complete underlying factorial as shown below:

57 58

9.4 9.5

Construction of a 1/2 × 23 design: Tabular method Construction of a 24−1 design

There are two possibilities. The complete underlying factorial is formed by A, B and C - the three first (most
important) factors. Introduce D (the fourth factor):
Construction : C = - AB Construction : C = +AB
22 design A B C = - AB (1/2)23 22 design A B C = +AB (1/2)23 Principle
(1) - - - (1) (1) - - + c 1/2 × 24
a + - + ac a + - - a I
b - + + bc b - + - b A
ab + + - ab ab + + + abc B
AB
C
The two designs are called complementary AC
Together they form the complete 23 factorial BC
ABC = ± D

In general : C = ± AB can be used, i.e. two possibilities

59 60
 9.6 9.7

Introduce factor D ⇒ 1/2 × 24 design: Tabular method Alias relations = Factor confoundings
Choose one of the possibilities, fx
Generator relation: D = +ABC =⇒ I = +ABCD : The defining relation
23 codes A B C D=+ABC 24−1 codes
Alias relations
(1) - - - - (1)
I = +ABCD The design is a resolution IV design
a + - - + ad
A = +BCD
b - + - + bd
B = +ACD Main effects and three-factor
ab + + - - ab
c - - + + cd
AB = +CD interactions confounded (often OK)
C = +ABD
ac + - + - ac
bc - + + - bc
AC = +BD Two-factor interactions are confounded
abc + + + + abcd
BC = +AD with other two-factor interactions
ABC = +D

The 24−1 design contains the data code ’(1)’, and it is called
The principal fraction

61 62

9.8 9.9

Analysis of data and the underlying factorial Analysis of effects based on normal probability plot
The analysis can based on the underlying factorial (A,B,C) The 7 estimated effects are 76/4, 6/4, ... , 66/4, respectively
(forget all about D while you do the computations) :
2
(i−0.5)/n

Standard−normal fractiles
0.9
Yates algorithm for a 24−1 design 1
0.8
Measure- 0.7
0.6
ment Response 1 2 3 Contrast SSQ 0 0.5
(1) . 45 145 255 566 I + ABCD - 0.4
0.3
a d 100 110 311 76 A + BCD 722.0 −1
0.2

b d 45 135 75 6 B + ACD 4.5 0.1

ab . 65 176 1 -4 AB + CD 2.0
−2
c d 75 55 -35 56 C + ABD 392.0 −40 −30 −20 −10 0 10 20 30 40
Effects sorted
ac . 60 20 41 -74 AC + BD 684.5
bc . 80 -15 -35 76 BC + AD 722.0 The present plot does not indicate any particularly small or large effect estimates
abc d 96 16 31 66 ABC + D 544.5
The data are from page 288 (6.ed) (5.ed: 308)
The plot is shown to illustrate the method. With only 7 points it is difficult to
conclude anything
The textbook has many more realistic examples
63 64
 9.10 9.11

Example continued Reduced model analysis of variance computations

Yates algorithm for a 24−1 design - final model
Suppose it is concluded that B=0 =⇒ AB=0, BC=0, BD=0 Measure- Esti-
ment Response 1 2 3 Effect SSQ mate
It could be concluded that also CD=0 (SSQAB+CD is small) (1) . 45 145 255 566 µ - 70.75
a d 100 110 311 76 A 722.0 9.50
From the beginning it was assumed that BCD=0, ACD=0, ABD=0 and ABC=0 b d 45 135 75 6 4.5 -
ab . 65 176 1 -4 2.0 -
(3 factor interactions) and ABCD=0
c d 75 55 -35 56 C 392.0 7.00
ac . 60 20 41 -74 AC 684.5 -9.25
Remove terms corresponding to all these assumptions and conclusions from the bc . 80 -15 -35 76 AD 722.0 9.50
analysis: abc d 96 16 31 66 D 544.5 8.25

Residual variance estimate

c 2 = (4.5 + 2.0)/(1 + 1) = 3.25 ' 1.802

σ

65 66

9.12 9.13

5 factors in 8 measurements Alias relations for model without high order interactions
Without many-factor interactions
Construction Alias relations - all
I
I I = ACD = BCE = ABDE
A = CD
A A = CD = ABCE = BDE
=⇒ B = CE
B B = ABCD = CE = ADE
I1=ACD AB = DE
AB AB = BCD = ACE = DE
I2=BCE C = AD = BE
C C = AD = BE = ABCDE
I1I2=ABDE AC = D
AC = D AC = D = ABE = BCDE
BC = E
BC = E BC = ABD = E = ACDE
(ABC) = BD = AE
ABC ABC = BD = AE = CDE

In defining relation I = ACD = BCE = ABDE at least 3 letters in all terms

(except I) ⇒ Resolution is III.

67 68
 9.14 9.15

Construction of 1/4 × 25 design A 25−2 factorial in 2 blocks of 4

Underlying Introduce × de ⇒
Factorial D E Design Principal Use fx D= - AC and E= - BC and Blocks=ABC :
Code A B C =+AC =+BC 25−2 fraction
(1) - - - + + de (1) Alias relations reduced
a + - - - + ae ad I = - ACD = - BCE +ABDE
b - + - + - bd be A = - CD
ab + + - - - ab abde B = - CE
c - - + - - c cde AB = +DE
ac + - + + - acd ace C = - AD = - BE
bc - + + - + bce bcd AC = -D
abc + + + + + abcde abc BC = -E
In the principal fraction D = - AC and E= - BC, and t here are 4 possible (equally (ABC) = - BD = - AE = Blocks
usefull) designs:
D = ± AC
combined with
E = ± BC

69 70

9.16 10.1

Construction using the tabular method Example 8-6 p 308, design construction

23 A B C D= - AC E= - BC 25−2 ABC=Block Construction of design by introducing the factors F, G and H into the complete factorial defined
by A, B, C, D and E. The design is carried out in 4 blocks.
(1) - - - - - (1) - ∼1
a + - - + - ad + ∼2 I
A
b - + - - + be + ∼2 B
ab + + - + + abde - ∼1 AB
C
c - - + + + cde + ∼2 AC
BC
ac + - + - + ace - ∼1 ABC = +F
bc - + + + - bcd - ∼1 D
AD
abc + + + - - abc + ∼2 BD
ABD = +G
CD
ACD
Design: BCD
ABCD
= Blocks

E
Block 1 Block 2 AE
(1) abde ace bcd ad be cde abc BE
ABE = Blocks
Even ABC Uneven ABC CE
ACE
BCE
ABCE
DE

71 72
 10.2 10.3

ADE Description and options given by user :

BDE Confoundings:(F=ABC,G=ABD,H=BCDE,Blocks=ABE=ACDE),
ABDE Options:(LaTeX Dispersion)
CDE
ACDE = Blocks ( BCD*ABE = ACDE ) The treatments of the experiment were :
BCDE = +H h afghj bfgij abi
ABCDE cfi acgij bcghj abcfh
dgi adfij bdfhj abdgh
cdfgh acdhj bcdij abcdfgi
ej aefg befghi abehij
Example 8-6 p 308, statistical analyses cefhij aceghi bceg abcefj
deghij adefhi bdef abdegj
Analysis of 2**k complete and fractional factorial designs cdefgj acde bcdehi abcdefghij

This program was prepared by Henrik Spliid Data ordering in relation to standard order is :
Informatics and Mathematical Modelling (IMM) 1 2 3 4
Technical University of Denmark (DTU) 5 6 7 8
Lyngby, DK-2800, Denmark. ([email protected]) 9 10 11 12
Version: 25/08/99 13 14 15 16
17 18 19 20
file=Montg_ex8-6.dat, Edited September 1, 2001. Course F-343, DFH. 21 22 23 24
Input for this problem was read from file Montg_ex9-6.dat 25 26 27 28
Output was written to file Montg_ex9-6.out 29 30 31 32

The 10 factors A - J are treated in the design From the treatments given above the following confoundings have been computed.
The 5 factors A - E define the complete underlying factorial structure Interactions between factors and blocks assumed = zero
The 3 factors F - H are embedded in the underlying factorial structure Max. factorial interaction order considered = 3
The 2 factors I - J define the blocking

73 74

Alias relations to interaction order 3 :

10.4 From the treatments given above the following confoundings have been computed.
Interactions between factors and blocks assumed = zero
10.5
* Max. factorial interaction order considered = 2
A = +BCF = +BDG Alias relations to interaction order 2 :
B = +ACF = +ADG *
AB = +CF = +DG A
C = +ABF = +DFG B
AC = +BF = +EGH AB = +CF = +DG
BC = +AF = +DEH C =
ABC = +F = +CDG AC = +BF
D = +ABG = +CFG BC = +AF
AD = +BG = +EFH ABC = +F
BD = +AG = +CEH D =
ABD = +CDF = +G AD = +BG
CD = +FG = +BEH BD = +AG
ACD = +BDF = +BCG = +AFG ABD = +G
BCD = +ADF = +ACG = +BFG = +EH = Blocks CD = +FG
ABCD = +DF = +CG = +AEH ACD
E BCD = +EH = Blocks
AE = +DFH = +CGH ABCD = +DF = +CG
BE = +CDH = +FGH E
ABE = +CEF = +DEG = Blocks AE
CE = +BDH = +AGH BE
ACE = +BEF = +GH ABE = Blocks
BCE = +AEF = +DH CE
ABCE = +EF = +ADH = +BGH ACE = +GH
DE = +BCH = +AFH BCE = +DH
ADE = +BEG = +FH ABCE = +EF
BDE = +AEG = +CH DE
ABDE = +EG = +ACH = +BFH ADE = +FH
CDE = +EFG = +BH BDE = +CH
ACDE = +ABH = +CFH = +DGH = Blocks ABDE = +EG
BCDE = +H CDE = +BH
ABCDE = +DEF = +CEG = +AH ACDE = Blocks
Note: Design has resolution IV BCDE = +H
============================== ABCDE = +AH

75 76
Response: Log-SD 10.6 Effects and aliases no. Sum of Squares Deg.fr. Effect estimates
10.7
Printout of input data:
* 0 52.4032 1 1.2797
Response Code A B C D E F G H I J A 1 .6641 1 .2881
1 (1) = 1.02 h 0 0 0 0 0 0 0 1 0 0 B 2 .3180 1 -.1994
2 a = 1.82 afghj 1 0 0 0 0 1 1 1 0 1 A B 3 .0003 1 -.0056
3 b = .89 bfgij 0 1 0 0 0 1 1 0 1 1 C 4 .0058 1 -.0269
4 ab = 1.39 abi 1 1 0 0 0 0 0 0 1 0 A C 5 .0294 1 -.0606
5 c = .91 cfi 0 0 1 0 0 1 0 0 1 0 B C 6 .0167 1 -.0456
6 ac = 1.78 acgij 1 0 1 0 0 0 1 0 1 1 A B C = + F 7 .0124 1 -.0394
7 bc = .87 bcghj 0 1 1 0 0 0 1 1 0 1 D 8 .0914 1 .1069
8 abc = 1.21 abcfh 1 1 1 0 0 1 0 1 0 0 A D 9 1.1213 1 -.3744
9 d = 1.48 dgi 0 0 0 1 0 0 1 0 1 0 B D 10 .0226 1 .0531
10 ad = 1.41 adfij 1 0 0 1 0 1 0 0 1 1 A B D = + G 11 .1093 1 .1169
11 bd = 1.17 bdfhj 0 1 0 1 0 1 0 1 0 1 C D 12 .0088 1 .0331
12 abd = 1.33 abdgh 1 1 0 1 0 0 1 1 0 0 A C D 13 .0248 1 -.0556
13 cd = 1.67 cdfgh 0 0 1 1 0 1 1 1 0 0 B C D = + I 14 .0102 1 -.0356
14 acd = 1.35 acdhj 1 0 1 1 0 0 0 1 0 1 A B C D 15 .0017 1 -.0144
15 bcd = 1.11 bcdij 0 1 1 1 0 0 0 0 1 1 E 16 .0000 1 -.0019
16 abcd = 1.08 abcdfgi 1 1 1 1 0 1 1 0 1 0 A E 17 .0004 1 .0069
17 e = .97 ej 0 0 0 0 1 0 0 0 0 1 B E 18 .0790 1 .0994
18 ae = 1.70 aefg 1 0 0 0 1 1 1 0 0 0 A B E = + J 19 .0088 1 .0331
19 be = .81 befghi 0 1 0 0 1 1 1 1 1 0 C E 20 .0124 1 .0394
20 abe = 1.45 abehij 1 1 0 0 1 0 0 1 1 1 A C E 21 .0003 1 .0056
21 ce = .94 cefhij 0 0 1 0 1 1 0 1 1 1 B C E 22 .0020 1 .0156
22 ace = 1.68 aceghi 1 0 1 0 1 0 1 1 1 0 A B C E 23 .0026 1 -.0181
23 bce = .75 bceg 0 1 1 0 1 0 1 0 0 0 D E 24 .0026 1 .0181
24 abce = 1.43 abcefj 1 1 1 0 1 1 0 0 0 1 A D E 25 .0063 1 -.0281
25 de = 1.38 deghij 0 0 0 1 1 0 1 1 1 1 B D E 26 .0282 1 .0594
26 ade = 1.18 adefhi 1 0 0 1 1 1 0 1 1 0 A B D E 27 .0215 1 -.0519
27 bde = 1.23 bdef 0 1 0 1 1 1 0 0 0 0 C D E 28 .0011 1 .0119
28 abde = 1.46 abdegj 1 1 0 1 1 0 1 0 0 1 A C D E 29 .0011 1 -.0119
29 cde = 1.49 cdefgj 0 0 1 1 1 1 1 0 0 1 B C D E = + H 30 .0014 1 .0131
30 acde = 1.29 acde 1 0 1 1 1 0 0 0 0 0 A B C D E 31 .0205 1 -.0506
31 bcde = 1.48 bcdehi 0 1 1 1 1 0 0 1 1 0 ------------------------------------------------------------------
32 abcde = 1.22 abcdefghij 1 1 1 1 1 1 1 1 1 1 Total for Effects 2.6247 31

77 78

Normal probability plot for estimates

HS +I--------I---------I---------I---------I---------I+
10.8 10.9
3.0000- -
I I
I I Estimates for final model:
I I ==========================
I (A) X I
2.0000- - Effect Estimate Stand. dev. t-test
I I -------------------------------------------------------------
I X I Grand mean 1.2797 .0202 >99.95%
I X I A .2881 .0404 >99.95%
I X I B -.1994 .0404 >99.95%
1.0000- 2 - D .1069 .0404 98.5 %
I X I A D ( = BG ? ) -.3744 .0404 >99.95%
I X2 I A B D = + G .1169 .0404 99.1 %
I 2 I B C D = + I (blocks) -.0356 .0404 61.3 %
I XX I A B E = + J (blocks) .0331 .0404 57.9 %
.0000- 3 - A C D E (blocks) -.0119 .0404 22.8 %
I 2 I
I XX I Variability Estimate Stand. dev.
I X2 I -----------------------------------------------------
I X I Residual standard dev. .1143 .0169
-1.0000- 2 - Coefficient of variation 8.93 % 1.32 %
I X I
I X I For computing the residual variance and standard deviation the sums
I (B) X I of squares for the significant effects and the blocks are subtracted
I I from the total variation giving the residual sum of squares with
-2.0000- - 32-1-8 = 23 degrees of freedom.
I X (AD) I
I I
I I A model control can be made using the dispersion effects
I I See Montgomery p. 239 and 300.
-3.0000- -
HS +I--------I---------I---------I---------I---------I+ The plot below does not indicate any dispersion effects :
-.3900 -.1050 .1950
-.2550 .0450 .3450

79 80
 Normal probability plot for dispersion effects
HS +I--------I---------I---------I---------I---------I+
10.10 10.11
3.0000- -
I I
I I
I I Summary of analyses
I X I
2.0000- -
I I
I X I Significant effects: A, B, D, AD, G (could be the conclusion)
I X I
I X I
1.0000- X X - Term Levels Effect Parameters
I X I
I XXX I µ constant 1.28 1.28
I 2 I A 0 - 15 × 0.001 inch 0.29 [ - 0.145, +0.145]
I 2 I
.0000- XX X - B 0 - 15 × 0.001 inch - 0.20 [+0.10, - 0.10]
I XX I
I 2 I
D tool vendor 0.11 [ - 0.055, +0.055]
I 2 X I AD interaction - 0.37 [+0.185, - 0.185]
I X I
-1.0000- 2 - G 0 - 15 × 0.001 inch 0.12 [ - 0.06, +0.06]
I X I σ2 residual variance 0.112 0.112
I X I
I X I
I I
-2.0000- -
I X I
I I
I I
I I
-3.0000- -
HS +I--------I---------I---------I---------I---------I+
-1.5600 -.4200 .7800
-1.0200 .1800 1.3800

81 82

10.12

Combining main effect and interaction effect estimates

The estimate of the combined effect af A, D and AD is :

A0 = - 0.145 AD00= - 0.185 AD01=+0.185

A1 =+0.145 AD10=+0.185 AD11= - 0.185
D0= - 0.055 D1=+0.055

giving (- 0.145 - 0.055 - 0.185 = - 0.385, for example) :

A = 0 - 0.385 +0.095
A = 1 +0.275 +0.015
D=0 D=1

The minimum response is wanted (it is log-standard deviation). Choose A=0, D=0,
B=1, G=0 (the combination with lowest estimated response)

Model identified

Yijkl = µ + Ai + Bj + Dk + ADik + Gl + ijkl

83