2-level factorials
TRANSCRIPT
Lecture 6: 2-level Factorials
SpanosEE290H F05
1
Treatments and BlocksTwo-Way Analysis of Variance - MANOVAThe Importance of TransformationTwo-Level Factorials
Two-Level Factorials
Lecture 6: 2-level Factorials
SpanosEE290H F05
2
What is a factorial?
• So far, we analyzed data based on the type of treatment they received.
• This analysis is known as "one-way" analysis of variance.
• Now we discuss data that can be classified in more than one "ways". Each of these ways is known as a "factor".
• In the beginning we will call the first classification a "treatment" and the second classification a "block".
Lecture 6: 2-level Factorials
SpanosEE290H F05
3
y ti = μ + β i + τ t + ε ti yti = y + (yi - y) + (yt - y) + (yti - y t - y i + y)
Randomized Block Designs
• In general we "block" the effect that we want to eliminate so that we can see the effects of the "treatment".
Lecture 6: 2-level Factorials
SpanosEE290H F05
4
tube 1
tube 2
Gas A Gas B Gas C
Example: Particle Contamination Study
• We have two LPCVD tubes and three gas suppliers. We are interested in finding out if the choice of the gas supplier makes any difference in terms of average particle counts on the wafers.
• Experiment: we run three full loads on each tube, one for each gas. We report the average particle adders for each load (excluding the first wafer in the boat…)
Lecture 6: 2-level Factorials
SpanosEE290H F05
5
A B C1 7 36 2 152 13 44 18 25
10 40 10 20
7 36 2 20 20 20 -5 -5 -5 -10 20 -10 2 1 -313 44 18 20 20 20 5 5 5 -10 20 -10 -2 -1 3
S = SA + SB + ST + SR
3,778 = 2,400 + 150 + 1,200 + 286 = 1 + 1 + 2 + 2
2,400 150 600 14
Treatment (Gas)
Block (Tube)
Decompose according to the equation: yti = μ + βi + τt + ε ti
yti = y + (yi - y) + (yt - y) + (yti - yt - yi + y)
A Simple Two-Way ANOVA for particles...
Sums of squaresDegrees of freedomMean squares
Lecture 6: 2-level Factorials
SpanosEE290H F05
6
yti = μ + βi + τt + εti
yti = y + (yi - y) + (yt - y) + (yti - yt - yi + y)
Decompose particle counts to global average, deviationsfrom that average due to machine choice and gas source choice. What remains is the residual.
The Two-Way ANOVA
moments
estimates
Lecture 6: 2-level Factorials
SpanosEE290H F05
7
Source of Sum DFs Mean sq Variance of sq
Average SA 1bet. blocks SB n-1 sB
2
bet. treatm ST k-1 sT2
residuals SR (n-1)(k-1) sR2
total S nk
Assumptions: - additivity- IIND of residuals
Expected values: s2B : σ2 + k Σ β2i / (n-1)s2T : σ2 + n Σ τ2t / (k-1)
Two-Way ANOVA Table
Lecture 6: 2-level Factorials
SpanosEE290H F05
8
Analysis of Variance
SourceModelErrorC Total
DF325
Sum of Squares1350.00
28.001378.00
Mean Square450.0014.00
F Ratio32.14Prob > F
0.03
Effect Test
SourceBlock Treatment
Nparm12
DF12
Sum of Squares150.00
1200.00
F Ratio10.7142.86
Prob > F0.080.02
MANOVA for our Example
Lecture 6: 2-level Factorials
SpanosEE290H F05
9
This model assumes additivity as well as IIND residuals:
yti = μ + βi +τt + εti
yti = μ + βi +τt
Note that this model “predicts” 20 values using 4+5+1=10 parameters.
The Model of the two-way ANOVA
Lecture 6: 2-level Factorials
SpanosEE290H F05
10
D = Y - A
D = B + T + R
Easy to prove that R ⊥ T and A ⊥ R.
D
B T
R
Geometric Interpretation of the two-way ANOVA
Lecture 6: 2-level Factorials
SpanosEE290H F05
11
ytij = μti + ε tij
y tij = yti + (ytij - yti)
μti = μ + τt + βi + ωti
yti = y + (yt - y) + (yi - y) + (yti - yt - yi + y)
A Two-Way Factorial Design
• When both effects are of interest, then the same arrangement is called a factorial design.
• The only addition to our model is the interaction term.
moments
estimates
Lecture 6: 2-level Factorials
SpanosEE290H F05
12
Source Sum DFs Mean sq of Var of sq
bet. bake SB n-1 sB2
bet. etch ST k-1 sT2
interaction SI (n-1)(k-1) sI2
residuals Se nk(m-1) sE2
total S nkm-1
Assumptions: residuals IIND
Expected values: sB2 : σ2 + mk Σ β2i / (n-1)
sT2 : σ2 + mn Σ τ2t / (k-1)
sI2 : σ2 + mΣΣ ω2ti / (n-1)(k-1)
ANOVA Table for Two-Way Factorials
Lecture 6: 2-level Factorials
SpanosEE290H F05
13
What are your conclusions?
Effect Test
SourceBlock TreatmentBlock *Treatment
DF122
Sum of Squares150.00
1200.0028.00
F-Ratio•••
Prob > F•••
Analysis of VarianceSourceModelErrorC Total
DF505
Sum of Squares1378.00
0.001378.00
Mean Square275.60
•
F Ratio•
Prob > F•
Factorial for our Example
Lecture 6: 2-level Factorials
SpanosEE290H F05
14
Etch Rec Bake DLeff1 1 0.312 1 0.823 1 0.434 1 0.451 1 0.452 1 1.103 1 0.454 1 0.711 1 0.462 1 0.883 1 0.634 1 0.661 1 0.432 1 0.723 1 0.764 1 0.621 2 0.362 2 0.923 2 0.444 2 0.561 2 0.292 2 0.613 2 0.354 2 1.02
Etch Rec Bake DLeff1 2 0.402 2 0.493 2 0.314 2 0.711 2 0.232 2 1.243 2 0.404 2 0.381 3 0.222 3 0.303 3 0.234 3 0.301 3 0.212 3 0.373 3 0.254 3 0.361 3 0.182 3 0.383 3 0.244 3 0.311 3 0.232 3 0.293 3 0.224 3 0.33
Two way factorial experiment for DLeff.
Lecture 6: 2-level Factorials
SpanosEE290H F05
15
Effect TestSourceEtch RecipeBake ProcEtch Rec*Bake Proc
DF326
Sum of Squares0.921.030.25
F Ratio13.8123.221.87
Prob>F0.000.000.11
Analysis of VarianceSourceModelErrorC Total
DF113647
Sum of Squares2.200.803.01
Mean Square0.200.02
F Ratio9.01
Prob>F0.00
Anova Table for 2-way DLeff Factorial
Lecture 6: 2-level Factorials
SpanosEE290H F05
16
ResidualDLeff
-0.4
-0.2
0
0.2
0.4
0.6
0.8
0.25 0.35 0.45 0.55 0.65 0.75
Predicted DLeff
ResidualDLeff
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Bake ProcA B C
Diagnostics for DLeff
Lecture 6: 2-level Factorials
SpanosEE290H F05
17
-3
-2
-1
0
1
2
3
-0.4 -0.2 0.0 0.2 0.4 0.6 0.8
Residual Statistics for DLeff
MaxMin 75%50%25%
Est Mean and its 95% conf. interval
Normal Probability Plot (must be a straight line)
Lecture 6: 2-level Factorials
SpanosEE290H F05
18
σy prop μα
Y = yλ
λ = 1 - α
Variance Stabilization
If the variance is a function of μ, then the appropriate transformation is needed to correct it. In general:
Need for Transformation
where σ can be determined empirically.Non-additivityAn empirically selected transformation can be tested using a formal Anova-based test.
Lecture 6: 2-level Factorials
SpanosEE290H F05
19
Effect TestSourceEtch RecipeBake ProcEtch Rec*Bake Pro
DF326
Sum of Squares3.565.240.40
F Ratio21.9348.431.22
Prob>F0.000.000.32
Analysis of VarianceSourceModelErrorC Total
DF113647
Sum of Squares9.191.95
11.14
Mean Square0.840.05
F Ratio15.45Prob>F0.00
Anova Table for 2-way ln(DLeff) Factorial
Lecture 6: 2-level Factorials
SpanosEE290H F05
20
-0.6-0.5-0.4-0.3-0.2-0.10.00.10.20.30.40.5
-1.5 -1.0 -0.5 0.0
lnDLeff Predicted
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
1.0 1.5 2.0 2.5 3.0
Bake Proc
Residual ln(DLeff) Residual ln(DLeff)
Diagnostics for ln(DLeff)
Lecture 6: 2-level Factorials
SpanosEE290H F05
21
-3
-2
-1
0
1
2
3
-0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0.0 0.1 0.2 0.3 0.4 0.5
Residual Statistics for ln(DLeff)
Lecture 6: 2-level Factorials
SpanosEE290H F05
22
DLeff
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4
DLeff Predicted
lnDLeff
DLeff
-1.5
-1.0
-0.5
0.0
-1.5 -1.0 -0.5 0.0lnDLeff Predicted
ln(DLeff)
Improvement due to Transformation
Lecture 6: 2-level Factorials
SpanosEE290H F05
235 x 5 Latin Square design.
<- treatmentsblocks
Other Blocked Arrangements• Sometimes we might have to deal with more than one
blocking variables.• Example: testing the effect of a develop recipe on wafers
and resists that come from different vendors.• Attach two types of labels, A,B.. and I, II... and combine
them to balance the blocking effects:
Lecture 6: 2-level Factorials
SpanosEE290H F05
24
y ti = μ+ τ t + β i + ω ti
y = f(x1,x2...,xn)
Measuring the Effect of Variables
• The first step is always to find whether any variable has an effect on the outcome.
• Nothing more can be done for qualitative (categorical) variables such as recipe type, vendor name etc.
• Some times we deal with quantitative variables (temperature, pressure).
• Once the effect has been confirmed, the next step is to build empirical quantitative models of the process:
Lecture 6: 2-level Factorials
SpanosEE290H F05
25
they can be used to visualize, control, design, diagnose etc.
How can such a model be "extracted" from the process?
Quantitative Models of the Process
Lecture 6: 2-level Factorials
SpanosEE290H F05
26
x
y
x1 x2
level (-1) level (+1)
y1
y2
Estimated model:y = Avgy +1/2 Effectx x'
Estimated model:y = Avgy +1/2 Effectx x'^
“True” model y = η + 1/2 Ε x’
Effectx = y2 - y1
y = (y2 + y1) / 2
x' = 2 (x - Avgx) / ( x2 - x1 )Avgx = (x2 + x1) / 2
A Simple 1-Factor, 2-Level Factorial
Lecture 6: 2-level Factorials
SpanosEE290H F05
27
(4)
(1) (2)
(3)
x1
x2
A Simple 2-Factor 2-Level Factorial
y
2y
3y4y
1y
Lecture 6: 2-level Factorials
SpanosEE290H F05
28
Calculations for the 2 Factor Factorial
• Average response• Effect for x1
• Effect for x2
• Interaction of x1 and x2
• Model
Lecture 6: 2-level Factorials
SpanosEE290H F05
29
Note on Economy in experimentation:
(1) (2)
(3) (4)
(5) (6)
(7) (8)
Simple, few runs, can be augmented.2k
l1 × l2 × ...lk
General Factorial Designs at Two Levels
Lecture 6: 2-level Factorials
SpanosEE290H F05
30
(1) (2)
(3) (4)
x1
x2
(1)
(2)
(3)
(4)
x1
x2
(5)
How many runs do we need for the same precision?
Factorial vs the "one-factor-at-a-time" experiment
Lecture 6: 2-level Factorials
SpanosEE290H F05
31
23 Pilot Plant Example
Lecture 6: 2-level Factorials
SpanosEE290H F05
32
Calculation of Main Effects and Interactions
Lecture 6: 2-level Factorials
SpanosEE290H F05
33
oPressure (P) 300-550 mtorrTemperature (T) 605-650 CSilane flow (F) 100-250 sccmDep. Rate R in Å/minP T F R Effects- - -+ - -- + -+ + -- - ++ - +- + ++ + +
94.8111.0214.1255.894.1
145.9286.7340.5
192.940.9
162.847.96.9
11.930.8-5.9
AVGPTFPTPFTFPTF
σ experiment (estimate from replicates) = 9.05 (8) σ mean (estimate from replicates) = 3.20σ effect (estimate from replicates) = 6.40
Two Level Factorial for Polysilicon Deposition
Lecture 6: 2-level Factorials
SpanosEE290H F05
34
s2 = ν1s1
2+ν2s22+...+ νgsg2
ν1 + ν2 +...+ νg
N ( 0, 4N
σ2 )
Sum of squares can be used to estimate σ2
Example from previous page:σ effect (estimate from insignificant effects) = 8.65 (3)
V(effect) = V ( y+ - y-) = ( 14
+ 14
)σ2 = 12
σ2
in general:V(effect) = 4
Nσ2
Variance Estimation via replicated runs:
Via insignificant high order effects:High order effects can be seen as samples from one distribution:
Lecture 6: 2-level Factorials
SpanosEE290H F05
35
Pi = 100(i - 12
)/m i = 1,2, ..., m
Used to identify insignificant effects.
Normal Probability Plots
Lecture 6: 2-level Factorials
SpanosEE290H F05
36
0.0 2.0 4.0 6.0 8.0 10.01.1
15
102030507080909599
99.999.99
Normal Probability Plot - Example
Lecture 6: 2-level Factorials
SpanosEE290H F05
37
Normal Probability Plot - 25 Factorial Example
Lecture 6: 2-level Factorials
SpanosEE290H F05
38
Normal Probability Plot - 25 Factorial Example
Lecture 6: 2-level Factorials
SpanosEE290H F05
39
y = 192.9+20.4P'+81.4T'+23.9F'+15.4T'F'^
86420-100
0
100
200
1086420-20
-10
0
10
20
30
Polysilicon Deposition Example (cont.)
Lecture 6: 2-level Factorials
SpanosEE290H F05
40
Natural log of Deposition Rate R in ln(Å/min)
P T F ln(R) Effects- - -+ - -- + -+ + -- - ++ - +- + ++ + +
5.150.240.900.21-0.060.070.08-0.08
AVGPTFPTPFTFPTF
σ experiment (estimated from replications) = .028 (8) σ mean (estimated from replications) = .010σ effect (estimated from replications) = .020σ effect (estimated from insignificant effects) = .073 (4)
4.554.715.375.544.544.985.665.83
Using a Transformation to improve the model
Lecture 6: 2-level Factorials
SpanosEE290H F05
41
Mean++++++++
8
PT+--++--+
4
PF+-+--+-+
4
TF++----++
4
PTF-++-+--+
4
Avg.4.554.715.375.544.544.985.665.83
T--++--++
4
P-+-+-+-+
4
F----++++
4
Quick Calculation of Effects - Contrasts
Lecture 6: 2-level Factorials
SpanosEE290H F05
42
12345678
Avg.4.554.715.375.544.544.985.665.83
(1)9.26
10.919.52
11.490.160.170.440.17
(2)20.1721.010.330.611.651.970.01
-0.27
(3)41.180.943.62
-0.260.840.280.32
-0.28
d84444444
est 5.150.240.90
-0.060.260.070.08
-0.08
effAVGPTPTFPFTFPTF
T--++--++
P-+-+-+-+
F----++++
Quick Calculation of Effects - Yate's
Lecture 6: 2-level Factorials
SpanosEE290H F05
43
y = 51.15+.12P'+.45T'+.11F'^
86420-0.2
0.0
0.2
0.4
0.6
0.8
1.0
1086420-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Effects, interactions and residuals must be checked...
Lecture 6: 2-level Factorials
SpanosEE290H F05
44
Conclusions
• From One way ANOVA to MANOVA.• From “table” models to continuous models.• The most important concepts are:
– There is a “true” model (with unknown “moments”)– There is an “estimated” model.– Each estimated model parameter is a “statistic”.– Critical underlying assumptions must be checked.– Residuals must be Independently, Identically,
Normally Distributed (IIND).
(See chapter 7 and chapter 10 in BHH, or chapter 12 in class textbook)