Statistical Design and Analysis of Experiments Part Two Lecture notes Fall semester 2007 Henrik Spliid Informatics and Mathematical Modelling Technical University of Denmark 1 0.1 List of contents, cont. 6.1: Factorial experiments - introduction 6.7: Blocking in factorials 6.9: montgomery example p. 164 6.11: Interaction plot 6.14: Normal probability plot for residuals 7.1: Factorial experiments with two level factors 7.7: A (very) small example of a 2×2 design 2 0.2 7.10: Yates algorithm by an example 7.12: Numerical example with 3 factors 8.1: Block designs for two level factorials 8.5: How-to-do blocking by confounding 8.6: Yates algorithm and blocking 8.7: The confounded block design (what happens?) 8.9: Construction of block design by the tabular method 8.12: A few generalizations on block designs 8.14: The tabular method for 2× blocks (example) 8.15: Partially confounded two level experiments 3 0.3 8.20: Generalization of partial confounding calculations 8.21: Example of partially confounded design 9.1: Fractional designs for two level factorials 9.3: Alternative method of construction (tabular method) 9.7: Generator equation and alias relations 9.8: Analysis of data and the underlying factorial 9.12: 5 factors in 8 measuremets 9.13: Alias relations for model without high order interactions 9.14: Construction of 1/4times2 5 design (tabular method) 4
21
Embed
Statistical Design and Analysis of Experiments · Statistical Design and Analysis of Experiments Part Two ... Blocking in factorials ... How-to-do blocking by confounding 8.6: Yates
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Statistical Design and Analysis of Experiments
Part Two
Lecture notes
Fall semester 2007
Henrik Spliid
Informatics and Mathematical Modelling
Technical University of Denmark
1
0.1
List of contents, cont.
6.1: Factorial experiments - introduction
6.7: Blocking in factorials
6.9: montgomery example p. 164
6.11: Interaction plot
6.14: Normal probability plot for residuals
7.1: Factorial experiments with two level factors
7.7: A (very) small example of a 2×2 design
2
0.2
7.10: Yates algorithm by an example
7.12: Numerical example with 3 factors
8.1: Block designs for two level factorials
8.5: How-to-do blocking by confounding
8.6: Yates algorithm and blocking
8.7: The confounded block design (what happens?)
8.9: Construction of block design by the tabular method
8.12: A few generalizations on block designs
8.14: The tabular method for 2× blocks (example)
8.15: Partially confounded two level experiments
3
0.3
8.20: Generalization of partial confounding calculations
8.21: Example of partially confounded design
9.1: Fractional designs for two level factorials
9.3: Alternative method of construction (tabular method)
9.7: Generator equation and alias relations
9.8: Analysis of data and the underlying factorial
9.12: 5 factors in 8 measuremets
9.13: Alias relations for model without high order interactions
9.14: Construction of 1/4times25 design (tabular method)
4
0.4
10.1: A large example on 2 level factorials
10.11: Summary of analyses (example)
10.12: Combining main effects and interaction estimates
5
6.1
Factorial experiments - introduction
Design with two factors6 measurements
y1 y
2y
3 y
4
y5 y
6B=1
B=0
A=0 A=1
The estimate of the A-effect based on y:
Ay = [(y3 + y4) − (y1 + y2)]/2
6
6.2
Design with two factors4 measurements
z1 z
2
z3 z
4 B=1
B=0
A=0 A=1
The estimate of the A-effect based on z:
Az = [(z2 + z4) − (z1 + z3)]/2
7
6.3
One-factor-at-the-time or factorial design
Are Ay and Az equivalent ?
Var Ay = ?
Var Az = ?
Additive model:
Response = µ + A + B + residual
Can it always be applied?
8
6.4
More complicated model:
Response = µ + A + B + AB + residual
Is it more needed for factorial designs than for block designs, for example, whereadditivity is often assumed?
If interaction is present, then: which design is best ?
Usage of measurements: which design is best ?
In general: How should a factorial experiment be carried out ?
9
6.5
Factorial designs and interaction
Design with two factors4 measurements
z1=20 z
2=40
z3=30 z
4=52 B=1
B=0
A=0 A=1
Response and A−effectat the two B−levels
B=0
B=1
0
20
40
60
80
A=0 A=1
The change in the response when factor A is changed is the same at both B-levels⇐⇒ no interaction
10
6.6
Design with two factors4 measurements
z1=20 z
2=50
z3=40 z
4=12 B=1
B=0
A=0 A=1
Response and A−effectat the two B−levels
B=0
B=1 0
20
40
60
80
A=0 A=1
The change in the response when factor A is changed depends on the B-level ⇐⇒interaction
The second situation is often the case in factorial experiments
Never use one-factor-at-the-time designs. There exist better alternatives in allsituations.
11
6.7
Blocking in factorials: Two alternative factorial designs
Complete randomization, 19th and 20th October
Additive Temperature
10oC 20oC 30oC 40oC
5% y y y y y y y y
10% y y y y y y y y
Yijk = µ + ai + cj + acij + Eijk
A completely randomized 2×4 factorial with two measurements per factor combi-nation conducted over, say, two days. The design is one block of size 16.
12
6.8
Replication 1, October 19th
Additive Temperature
10oC 20oC 30oC 40oC
5% y y y y
10% y y y y
Replication 2, October 20th
Additive Temperature
10oC 20oC 30oC 40oC
5% y y y y
10% y y y y
Yijk = µ + ai + cj + acij + Dayk + Zijk
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.
Never use the first design. Why ?
13
6.9
Example from Montgomery p 164
Material Temperature
type 15oF 70oF 125oF
1 130 155 34 40 20 70
Data 74(?) 180 80 75 82 58
Averages y=134.75 y=57.25 y=57.50
2 150 188 136 122 25 70
Data 159 126 106 115 58 45
Averages y=155.75 y=119.75 y=49.50
3 138 110 174 120 96 104
Data 168 160 150 139 82 60
Averages y=144.00 y=145.75 y=85.50
Yijk = µ + mi + tj + mtij + Eijk
14
6.10
A better design
Round I
Material Temperature
type 1 15oF 70oF 125oF
1 y y y y y y
2 y y y y y y
3 y y y y y y
Round II
Material Temperature
type 15oF 70oF 125oF
1 y y y y y y
2 y y y y y y
3 y y y y y y
Yijk = µ + mi + tj + mtij + Rk + Zijk
Give (at least) three reasons why this design is to be preferred.
15
6.11
Response and temperatureeffects for 3 materials
m=1m=2
m=3
50
75
100
125
150
175
15 70 125
The figure indicates a possible interaction between materials and temperature.
It is a common case that different ’materials’ react differently to fx temperaturetreatments.
16
6.12
ANOVA and estimation in factorial design
Yijk = µ + mi + tj + mtij + Eijk
ANOVA for battery data
Source of var. SSQ d.f. s2 EMS F-test
Materiel 10684 2 5342 σ2 + 12φm 7.91
Temperature 39119 2 19559 σ2 + 12φt 28.97
Interaction 9614 4 2403 σ2 + 4φmt 3.56
Residual 18231 27 675.2 σ2
Total 77647 35
F (4, 27)0.05 = 2.73 =⇒ all parameters in the model are significant at the 5% levelof significance.
17
6.13
Estimates of parametersfor full model
µ = Y ...
mi = Y i.. − Y ...
tj = Y .j. − Y ...
mtij = Y ij. − Y i... − Y .j. + Y ...
σ2E = s2
resid
Estimates of parametersfor additive model
µ = Y ...
mi = Y i.. − Y ...
tj = Y .j. − Y ...
mtij = 0 (not in model)
σ2E = (SSQresid + SSQmt)/(fresid + fmt)
18
6.14
Model control based on residuals
−100 −50 0 50 100−3
−2
−1
0
1
2
3
Sta
nd
ard
−n
orm
al fr
actile
s
Residuals
0.001
0.01
0.10.20.30.40.50.60.70.80.9
0.99
0.999
(i−0.5)/n
19
7.1
Factorial experiments with two-level factors
The simplest example: 2 factors at 2 levels.
1. factor is called A (can be a temperature fx) (the supposedly most importantfactor)
2. factor is called B (can be a concentration of an additive)(the supposedly nextmost important factor)
For each factor combination r measurements are carried out (completely random-ized):
20
7.2
22 factorial design
B=0 B=1
Y001 Y011
A=0 : :
Y00r Y01r
Y101 Y111
A=1 : :
Y10r Y11r
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
7.3
Restrictions on parameters: fx A0 + A1 = 0 =⇒
A0 = - A1 and B0 = - B1 and
AB00 = - AB10 = - AB01 = AB11
All parameters have only one numerical value, positive or negative, depending onthe factor level(s).
22
7.4
Effects. Special concept for 2 level factors
Effect = change in response when the factor is changed from level ’0’ to ’1’, thus
A-effect: A = A1 - A0 = 2A1 (main effect)
B-effect: B = B1 - B0 = 2B1 (main effect)
AB-effect: AB = AB11 - AB10 = 2AB11 (interaction)
In General:
k factors at 2 levels: A 2k factorial experiment
23
7.5
Special notation for 2k design
Two factors, k = 2
Standard notation for cell sums
(1) a
sum of r measurements
b abB=1
B=0A=0 A=1
B=0 B=1
A=0 (1) b
A=1 a ab
Fx ’a’ = ∑rk=1 Y10k , the sum in the cell where the factor A is at level ’1’ while factor
B is at level ’0’.
24
7.6
Parameters, effects and estimation
A-parameter : A1 = ( - (1) + a - b + ab)/4r
B-parameter : B1 = ( - (1) - a + b + ab)/4r
AB-parameter : AB11 = (+(1) - a - b + ab)/4r
A-effect : A = ( - (1) + a - b + ab)/2r = 2A1
B-effect : B = ( - (1) - a + b + ab)/2r = 2B1
AB-effect : AB = (+(1) - a - b + ab)/2r = 2 AB11
25
7.7
A (very) small numerical example
Y = response = purity in solution after 48 hours
A = 1. factor = temperature (4oC , 20oC)
B = 2. factor = concentration of additive (5%, 10%)
Yates algorithm for k = 2 factorsCell sums I II = contrasts SSQ Effects(1) = 26.4 63.4 151.9 = [I] − µ = 18.99a = 37.0 88.5 17.5 = [A] 38.25 A = 4.38b = 40.8 10.6 25.1 = [B] 78.75 B = 6.28ab = 47.7 6.9 - 3.7 = [AB] 1.71 AB = - 0.93
The important concept about Yates’ algorithm is that isrepresents the transformation of the data to the contrasts -and subsequently to the estimates and the sums of squares!
29
7.11
Explanation:
Cell sums: Organized in ’standard order’: (1), a, b, ab
Column I:63.4 = +26.4+37.0 (sum of two first in previous column)
88.5 = +40.8+47.7 (sum of two next)
10.6 = - 26.4+37.0 (reverse difference of two first)
6.9 = - 40.8+47.7 (reverse difference of two next)
Column II: Same procedure as for column I (63.4+88.5=151.9)
SSQA: [A]2/(2k · 2) = 38.25 (k=2) and likewise for B and AB
A-Effect: A = [A]/(2k−1 · 2) = 4.38 and likewise for B and AB
The procedure for column I is repeated k times for the 2k design
The sums of squares and effects appear in the ’standard order’
Which term could be used ? Not a main effect, but some higher orderterm, for example ABC (why ABC ?):
We want the confounding Blocks = ABC .
We say : defining relation I = ABC . . . but how do we do it?
Look at how contrasts for effects are calculated :
37
8.6
Yates algorithm - schematically - once again:
(1) a b ab c ac bc abc
[I ] = +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
Note that any two rows are ’orthogonal’ (product sum = zero).
Thus [A] and [B], for example, are orthogonal contrasts.
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.
38
8.7
The confounded block design: Blocks = ABC
Ideal data without influence from blocks:
Day 1: D1 = 0 ab = 16 bc = 21 (1) = 12 ac = 20
Day 2: D2 = 0 b = 24 a = 28 abc = 34 c = 22
Cell sums I II III = contrasts
(1) = 12 40 80 177 = [I]
a = 28 40 97 19 = [A]
b = 24 42 8 13 = [B]
ab = 16 55 11 - 9 = [AB]
c = 22 16 0 17 = [C]
ac = 20 - 8 13 3 = [AC]
bc = 21 - 2 - 24 13 = [BC]
abc = 34 13 15 39 = [ABC]
What happens if the two days in fact influence the results differently (there is aday-to-day effect) ?
39
8.8
Real data with a certain influence (unknown in practice) from blocks (days):
Day 1: D1 = +8 ab = 24 bc = 29 (1) = 20 ac = 28
Day 2: D2 = +2 b = 26 a = 30 abc = 36 c = 24
Cell sums I II III = contrasts Day effect
(1) = 20 50 100 217 = [I] yes
a = 30 50 117 19 = [A]
b = 26 52 8 13 = [B]
ab = 24 65 11 -9 = [AB]
c = 24 10 0 17 = [C]
ac = 28 -2 13 3 = [AC]
bc = 29 4 -12 13 = [BC]
abc = 36 7 3 15 = [ABC] yes
What has changed and what has not changed? Why?
The effect from days is controlled (not eliminated) only to influence the ABCinteraction term (block confounding).
40
8.9
Construction using the tabular method :
Arrange data in standard order and use column multiplication :
Factor levels Block no. =
Code A B C ABC = A·B·C(1) -1 -1 -1 -1
a +1 -1 -1 +1
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
8.10
Analysis of variance for block confounded design
In the example we imagine that r=2 measurements per factor combination wereused. The residual SSQ is computed as the variation between these two measure-ments giving a total residual sum of squares with 8 degrees of freedom.
Correspondingly the responses on slide 8.8 (bottom) are sums of 2 measurements.
42
8.11
ANOVA for block confounded three factor design
Effects SSQ d.f. s2 F-value
A 22.56 1 22.56 9.12
B 10.56 1 10.56 4.27
AB 5.06 1 5.06 2.05
C 18.05 1 18.05 7.30
AC 0.56 1 0.56 0.22
BC 10.56 1 10.56 4.26
ABC = Blocks 14.06 1 14.06 not relevant
Residual 19.80 8 2.475
Total 101.24 15
F (1, 8)0.05 = 5.32 =⇒ A and C main effects are significant. The B effect is onlysignificant at the 10% level of significance, and so is BC.
The ABC effect cannot be tested because it is confounded with blocks (days)(does it seem to be a real problem ?).
43
8.12
A few generalizations
A 24 factorial design in 4 blocks of 4 :
I2 = BCD ∼ Days
(1) bc d bcd
I1 = ABC ∼ abd acd ab ac
Operators a abc b c
bd cd ad abcd
The principal block: (1) bc abd acd
b× (1) bc abd acd = b b2c ab2d abcd ⇒ b c ad abcd = another block!
Multiply any block with an ’element’ that is not in the block, and you get anotherblock.
Total block variation = ABC + BCD + ABC·BCD = ABC + BCD + AD
When analyzing the data from the above 24 design all effects A, B, AB, ... , ABCDexcept ABC, BCD and AD can be estimated and tested.
ABC, BCD and AD are confounded with blocks
44
8.13
Construction principle : Introduce blocks into factorial by confounding
Effect Confound
Level
A
B
AB
AC
BC
ABC = I1D
AD <= ABC·BCD
BD
ABD
ACD
BCD = I2ABCD
All effects ABC, BCD and ABC·BCD = AD will be confounded with blocks.
45
8.14
Construction using the tabular method
Factor levels Four different Principal
Code A B C D ABC BCD blocks block
(1) −1 −1 −1 −1 −1 −1 1 : ( −1 , −1) (1)
a +1 −1 −1 −1 +1 −1 2 : ( +1 , −1)
b −1 +1 −1 −1 +1 +1 4 : ( +1 , +1)
ab +1 +1 −1 −1 −1 +1 3 : ( −1 , +1)
c −1 −1 +1 −1 +1 +1 4 : ( +1 , +1)
ac +1 −1 +1 −1 −1 +1 3 : ( −1 , +1)
bc −1 +1 +1 −1 −1 −1 1 : ( −1 , −1) bc
abc +1 +1 +1 −1 +1 −1 2 : ( +1 , −1)
d −1 −1 −1 +1 −1 +1 3 : ( −1 , +1)
ad +1 −1 −1 +1 +1 +1 4 : ( +1 , +1)
bd −1 +1 −1 +1 +1 −1 2 : ( +1 , −1)
abd +1 +1 −1 +1 −1 −1 1 : ( −1 , −1) abd
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)
abcd +1 +1 +1 +1 +1 +1 4 : ( +1 , +1)
46
8.15
Partially confounded 2k factorial experiment
B = 0 B = 1A = 0 (1) bA = 1 a ab
Suppose batches=blocks, block size = 2 :
Experiment 1 :(1)1 ab1 a1 b1
batch 1 batch 2I = AB
Model as usual: Yijν = µ + Ai + Bj + ABij + Eijν+ batches (blocks)
AB interaction confounded with blocks in experiment 1.
47
8.16
Resolving block confoundings for AB with one more experiment:
Suppose we also want to assess the interaction term AB.We need an experiment in which AB is not confounded:
Experiment 2 :(1)2 a2 b2 ab2
batch 3 batch 4I = B
using two new batches.
Model again: Yijν = µ + Ai + Bj + ABij + Eijν+ batches (blocks)
The B main effect is confounded with blocks in experiment 2, but AB is not.AB can then be estimated in experiment 2.
The price paid is that the main effect B can only be estimated in experiment 1and AB only in experiment 2: Partial confounding.
48
8.17
Analyze both experiments using contrasts :
Unconfounded contrasts
[A]1 = −(1)1 + a1 − b1 + ab1 (from experiment 1)
[A]2 = −(1)2 + a2 − b2 + ab2 (from experiment 2)
[B]1 = −(1)1 − a1 + b1 + ab1 (from experiment 1)
[AB]2 = +(1)2 − a2 − b2 + ab2 (from experiment 2)
Confounded contrasts
[AB]1 = +(1)1 − a1 − b1 + ab1 (from experiment 1)
[B]2 = −(1)2 − a2 + b2 + ab2 (from experiment 2)
Blocks
Variation calculated as usual: From block totals
49
8.18
Use of unconfounded contrasts for effects:
The two (unconfounded) A-contrasts can be combined into anestimate of A and a part which expresses uncertainty: