doe - design of experiments - case study
TRANSCRIPT
Design Of ExperimentsExample: case study I
(7.1)DOE - KVL(031023)
Day 1 1. 13:00-13:45h - Introduction; what was experimental design again? 2. 14:00-14:45h - Design in one formula: y = X.b3. 15:00-15:45h - Statistical inference and ANOVA
Day 2 4. 09:00-12:00h - Hands-on DOE: computer exercise I *)
5. 13:00-13:45h - Data inspection and plotting (+ some PCA) 6. 14:00-14:45h - Miscellaneous subjects 7. 15:00-15:45h - Example: case study I
Day 3 8. 09:00-12:00h - Hands-on DOE: computer exercise II *)
9. 13:00-13:45h - Blocking and split-plot10. 14:00-14:45h - Example: case study I11. 15:00-15:45h - Introduction to mixed linear models
*) Participants can choose to take part in the theory (3 afternoons) sessions or theory + computer exercise (2 mornings) sessions.
Case study IDesign Of Experiments
(7.2)DOE - KVL
Case study IColor change during meat storage
(7.3)DOE - KVL
Design Of ExperimentsDifferent research areas
(7.4)
• Replicates are ‘perfect clones’• Experiments are cheap• Simulations give ‘perfect statistics’
DOE - KVL
• Motivation for a lot of developments• Reasonably reproducible• Sparse designs• THE example in literature
• Natural variation• Experiments are expensive • More replicates• More input of the experimenter
Case study IColor change during meat storage
(7.5)DOE - KVL
Loss of redness of pork-meat during storage time under different packaging conditions
© Lisbeth Dahlgård Nannerup, MLI/LMC Kvali-MAP project, Department of Food Chemistry, KVL, Denmark
• Response variable is the color (redness) of the meat
• Measured as a so-called a-value
• Higher redness is more attractive towards consumers (optimization)
Case study IColor change during meat storage
(7.6)DOE - KVL
The effects investigated are packaging conditions
• Storage days (12 levels)
• Percentage oxygen (4)
• Product/Headspace ratio (3)
• Temperature (3)
• Light exposure (3)
8 replicates ( )
Case study IColor change during meat storage
(7.7)DOE - KVL
The experiment is a five (or six if you count replicates) dimensional factorial designDays x Oxygen x Product/Headspace x Temperature x Light (x Animals)
1D 2D 3D … 6D
?
Case study IColor change during meat storage
(7.8)DOE - KVL
0.0 0.5 1.0 1.5 O2 (%)
1:1
P/H 1:1.5
1:3
10T(°C) 8
5
L (Lux)0 600 1200
t (days)0 21 34
• 4 x 3 x 3 x 3 x 8 = 864 samples• 864 x 12 = 10368 measurements
Color loss in porkQuality of the response
(7.9)DOE - KVL
• Each measurement outcome is based on 5 readings• Minolta color measurement (a, b and L-value)• 8mm diameter serves area• a-value is used to express redness of meat
0 5 10 15 20 25 30 350
10
day
a-va
lue
Standard deviation over readings (σ)
Color loss in porkQuality of the response
(7.10)DOE - KVL
0
3
σ Sam
ples
0 5 10 15 20 25 30 351
2
day
σ poo
led
σpooled all samples = 1.59
Measurement error :1.58/√5 = 0.7 a-values
10 (randomly selected) samples are used to computethe pooled standard deviation
Color loss in porkReplicates
(7.11)DOE - KVL
-5
0
5
10
O2 = 0.0% P/H = 1:1 T = 5°C L = 0Lux
0 5 10 15 20 25 30 35-5
0
5
10 O2 = 1.5% P/H = 1:3 T = 10°C L = 1200Lux
day
a-va
lue
a-va
lue
Two design points
x 8
Color loss in porkReplicates?
(7.12)DOE - KVL
First experimental run
2
8
0 352
8
0 35day
a-va
lue
day
a-va
lue
Second experimental run
x2 per production runMean a-value over
all design points
Storage (days)
Color loss in porkExpected trends
(7.13)DOE - KVL
Mean a-value overall design/time points
0.0 0.5 1.0 1.52
8
1:1 1:1.5 1:3
5 102
8
0 600 12008
Light (Lux)
a-va
lue
a-va
lue
Product/Headspace
Temperature (°C)
Oxygen (%)
Color loss in porkResponse surface model for day 21
(7.14)DOE - KVL
Storage time in the setup of this experiment requires special methods.We will treat the data from day 21 as single design (108 points x 8 replicates)
0.0 0.5 1.0 1.5 O2 (%)
1:1
P/H 1:1.5
1:3
10T(°C) 8
5
L (Lux)0 600 1200 Day 21
Color loss in porkResponse surface model for day 21
(7.15)DOE - KVL
An ANalysis Of Variance shows that the temperature as main effect and interactions are not important.
√√X√√X√X√XX√XXX
Color loss in porkResponse surface model for day 2, 21 and 34
(7.16)DOE - KVL
An ANalysis Of Variance shows that the temperature as main effect and interactions are not important.
Day 2 Day 21 Day 34
Color loss in porkResponse surface model for day 21
(7.17)DOE - KVL
An ANalysis Of Variance shows that the temperature as main effect and interactions are not important.
Color loss in porkResponse surface model for day 21
(7.18)DOE - KVL
a-value = b0 + b1 xO2 + b2 xP/H + b3 xL + b4 xO2 xP/H + b5 xO2 xL + b6 xP/H xL + b7 xO2.P/H xL
-0.8x10-3-1.6x10-3-1.2x10-3b7 (O2.P/H.L)
-0.0x10-3-0.7x10-3-0.4x10-3b6 (P/H.L)
2.2x10-30.8x10-31.5x10-3b5 (O2.L)
0.20-0.34-0.07b4 (O2.P/H)
0.5x10-3-0.8x10-3-0.2x10-3b3 (L)
-0.05-0.56-0.30b2 (P/H)
0.48-0.63-0.08b1 (O2)
11.8910.8511.37b0
Confidence interval
=y = X.b b = (X’.X)-1.X.y=
Color loss in porkModel residuals for day 21
(7.19)DOE - KVL
-6 -4 -2 0 2 4 -6 -4 -2 0 2 40.001
0.01
0.05 0.10 0.25
0.50
0.75 0.90 0.95
0.99
0.999
Residuals (mildly) skewed towards low values, but for now we assume ANOVA is ‘robust’ against this non-normality
Color loss in porkModel residuals for day 21
(7.20)DOE - KVL
0 10-10
0
10
0 10-10
0
10
Product/Headspace1:11:1.51:3
Light0
6001200
Color loss in porkResponse surface for day 21
(7.21)DOE - KVL
1:1 1:30
0.5
1
1.5
P/H
O2
(%)
Light = 0Lux
6.2
6.4
6.4
6.4
6.6
6.6
6.6
6.8
6.8
6.8
7
1:1.5
Color loss in porkResponse surface for day 21
(7.22)DOE - KVL
3.63.84
4.2
4.44.4
4.6
4.6
4.8
4.85
5
55.2
5.2
5.25.4
5.4
5.45.6
5.65.6
5.65.8
5.85.8
5.86
66
6.26.2
6.2
6.46.4
6.4
6.66.6
6.6
1:1 1:30
0.5
1
1.5
P/H
O2
(%)
Light = 600Lux
1:1.5
Color loss in porkResponse surface for day 21
(7.23)DOE - KVL
11.21.41.61.82
2.2
2.42.4
2.62.6
2.82.8
33
3.23.2
3.43.4
3.6
3.6
3.63.8
3.8
3.84
4
44.2
4.2
4.24.4
4.4
4.44.6
4.6
4.64.8
4.84.8
4.85
55
55.2
5.25.2
5.4
5.4
5.4
5.65.6
5.6
5.85.8
5.8
66
6
6.26.2
6.2
6.4
6.4
6.4
6.6
1:1 1:30
0.5
1
1.5
P/H
O2
(%)
Light = 1200Lux
1:1.5
Color loss in porkResponse surface for day 21
(7.24)DOE - KVL
1:1 1:1.51:3 0.0
0.51.0
1.51
8
Light = 1200Lux
Light = 600Lux
Light = 0Lux
a-va
lue
Oxygen (%)
Product/Headspace
Color loss in porkResponse surface for day 2
(7.25)DOE - KVL
1:1 1:1.51:3 0.0
0.51.0
1.51
8
Light = 1200Lux
Light = 600Lux
Light = 0Lux
a-va
lue
Oxygen (%)
Product/Headspace
Color loss in porkResponse surface for day 34
(7.26)DOE - KVL
1:1 1:1.51:3 0.0
0.51.0
1.51
8
Light = 1200Lux
Light = 600Lux
Light = 0Lux
a-va
lue
Oxygen (%)
Product/Headspace
Color loss in porkResponse surface for days 2, 21 and 34
(7.27)DOE - KVL
Day 2 Day 21 Day 34
A workable model of for all the design factors, supported by inspecting the collecting (raw) data, and…
‘Models are to be used, not believed’ (Henri Theil)
Data transformationsNon-normal residuals?
(7.28)DOE - KVL
-6 -4 -2 0 2 40.001
0.01 0.05 0.10 0.25 0.50 0.75 0.90 0.95 0.99
0.999
21)constantlog(
xyxyx
y
xy
sbsandxbaythen
bxayif
xy
===
+=
=
+=+=
Log-normal distributions are often fond in nature (weight, size, etc.), time-series andgrowth models
y = b0exp(b1x) ln(y) = ln(b0) + b1xy’ = b0’ + b1x
(linear transformation leave the residual distributions unattached)
Data transformationVariance stabilizing
(7.29)DOE - KVL
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5λ
SSe
Sometimes a so-called Box-Cox transformation of the form y = (x + constant)λcan help to ‘stabilize’ (make more similar) the model errors
The ‘optimal’ transformation is the one that minimizes the sum-of-squares of residuals as a function of λ(Maximum Likelihood):
( )
( )nxex
xxyxxy
/ln
1
)ln()0(
)1(
∑=
=
−= −λ
λ
λλ
Color loss in porkDay 21 after transformation
(7.30)DOE - KVL
-80 -60 -40 -20 0 20 40 60 80 -80 -60 -40 -20 0 20 40 600.001
0.01
0.05 0.10 0.25
0.50
0.75 0.90 0.95
0.99
0.999
Color loss in porkDay 21 after transformation
(7.31)DOE - KVL
Beforetransformation
Color loss in porkDay 21 after transformation
(7.32)DOE - KVL
0 100 200-100
0
100
200
0 100 200-100
0
100
200
Product/Headspace1:11:1.51:3
Light0
6001200
Color loss in porkDay 21 after transformation
(7.33)DOE - KVL
1:1 1:1.51:3 0.0
0.51.0
1.51
8
Light = 1200Lux
Light = 600Lux
Light = 0Lux
a-va
lue
Oxygen (%)
Product/Headspace
GEMANOVAMultiplicative-model analysis of the DOE-data
(7.34)
GEneralized Multiplicative ANalysis Of VAriannce (GEMANOVA)
Classical ANOVA - Additive model:color = b0 + b1 xO2 + b2 xP/H + b3 xL + b4 xO2.P/H + b5 xO2.L + b6 xP/H.L + b7 xO2.P/H.L
GEMANOVA - Multiplicative model:color = c0 cA ct cO2 cP/H cT cL
DOE - KVL
GEMANOVA = PARAFACParallel factor analysis
(7.35)
X E= + +
X E= + +
Principal Component Analysis (PCA)
PARAFAC
A factor model in three (or more!) dimensional space
with scores/loadings in three (or more) directions
DOE - KVL
GEMANOVAOne factor model, all effects free
(7.36)
c0 = 1092color = c0 cA ct cO2 cP/H cT cL
2 4 6 8
0.33
0.37
0 10 20 30
0.28
0.29
0 0.5 1 1.5
0.49
0.52
1:1 1:1.5 1:30.52
0.60
5 8 10
0.577
0.579
600 1200
0.56
0.62
0
Animal
Storage (days)
Prod./Headsp.
Light (Lux)Temperature (°C)
Oxygen (%)
DOE - KVL
GEMANOVAOne factor model, all effects free
(7.37)
color = c0 cA ct cO2 cP/H cT cL + e
Data range = [0.98 – 14.82]; RMSPfit = 1.47 color-values; R2 = 0.41
-6 -4 -2 0 2 4
DOE - KVL
Error
Jackknife re-samplingUncertainty estimation
(7.38)
( )( ) ( )
( ) ( )
( ) ( )θθθθ
θ
θθθ
αθ
θ
θ
αα
αα
ˆˆˆ
ˆˆ
ˆˆ
1)ˆ(
ˆˆ
*
***
^
2/1;
^
2/;
2/12/
−⇔−
==
−≤≤−
−=≤≤
==
=
−
−
x
x
uFt
setset
ssP
uFt
Ft
dfdf
F : Cumulative Distribution Function
x : data; plug-in principle
α-coverage probability
Uncertainty estimate
Estimates from new distributionfound from some re-sampling strategy (*)
The re-sampling assumption!
DOE - KVL
Jackknife re-samplingUncertainty estimation
(7.39)
( ) ( )( )( ) ( )
( ) ( )^
2/1
^
2/
2*(.)
*)(
^
*(.)
^
ˆˆ
ˆˆ1
ˆˆ1
sezsez
nnse
nbias
iJ
J
αα θθθ
θθ
θθ
−−≤≤−
−−
=
−−=
∑
n = #samples (= 8 animals) (.) : re-sampling expectation
Jackknife bias estimate
Jackknife standard error estimate
Assume normal distribution
DOE - KVL
GEMANOVAOne factor model, Jackknife 5% coverage probability
(7.40)
color = c0 cA ct cO2 cP/H cT cL c0 = 1052 - 1092 - 1133
2 4 6 8
0.32
0.38
0 10 20 300.27
0.29
0 0.5 1 1.50.48
0.52
1:1 1:1.5 1:3
0.52
0.60
5 8 10
0.57
0.58
0 600 12000.54
0.62
Anim
al
Storage (days)
Prod./Headsp.
Light (Lux)T (°C)
Oxygen (%)
DOE - KVL
RMSPfit = 1.47R2 = 0.41
No effect!
GEMANOVAOne factor model, Temperature effect eliminated
(7.41)
color = c0 cA ct cO2 cP/H 1T cL c0 = 607 - 631 - 654
2 4 6 8
0.32
0.38
0 10 20 300.27
0.29
0 0.5 1 1.50.48
0.52
1:1 1:1.5 1:3
0.52
0.60
5 8 10
1
0 600 12000.54
0.62
Anim
al
Storage (days)
Prod./Headsp.
Light (Lux)
Oxygen (%)
Temperature (°C)
RMSPfit = 1.47R2 = 0.41(same performance)
DOE - KVL