Beruflich Dokumente
Kultur Dokumente
Objective:
Comparing (a) treatment averages.
Results of (a x n) = N runs
Run
Treatment 1 2 …. n
ε ij
random variables representing
experimental errors.
2. The experimental errors ε ij are normally
and independently distributed,
NID (0, σ 2)
3. The error variance σ 2 is constant.
The ANOVA Proposition
The total variability in the responses can
be decomposed into:
• Variability between treatment averages
and the grand average (SSTreatment ).
• Variability of individual observations
within each treatment and the treatment
average (SSError ).
Run
Treatment 1 2 …. n Average
Between Treatments
Within Treatment
1 y11 y12 ….. y1n y1.
2 y21 y22 ….. y2n y2.
.. .. .. .. ..
.. .. .. .. ..
a ya1 ya2 ….. yan ya.
1 1
a
SSTreatment = n∑ ( yi. − y.. ) 2
a n 2
2 y..
SS Total = ∑∑ yij −
1 1 N
1 ay 2
SS = ∑y − 2 ..
n N
Treatment i.
1
a y y 2 2
=∑ − i. ..
1
n ∑n i i
Statistical Analysis
Testing the hypothesis that all treatments
have no effect, against the alternative that
at least one treatment has a significant
effect.
i.e.,
H0: τ i = 0 Vs. H1: at least one τ i
≠0
The test statistic:
SSTreatment
(a − 1) MSTreatment
F0 = =
SS Error MS Error
( N − a)
Source SS df MS F0
1 1 N
1 a 2 y..2
SSTreatment = ∑ yi . − = 66870.55
n 1 N
Source SS df MS F0
Treatment 66870.55 3 22290.18 66.8
Total 72209.75 19
At α = 5%,
F0.05,3,16 = 3.24
RF (W) yi. 1 2 3 4 5
160 551.2 23.8 -9.2 -21.2 -12.2 18.8
180 587.4 -22.4 5.6 2.6 -8.4 22.6
200 625.4 -25.4 25.6 -15.4 11.6 3.6
220 707.0 18.0 -7.0 8.0 -22.0 3.0
y.. = 617.75
The Normality
GN-EXPERT Plot
Assumption
Normal Plot of Residuals
Rate
99
Probability
95
Probability
90
80
70
%%
50
Normal
30
20
Normal
10
5
Residuals
-25.4 -12.65 0.1 12.85 25.6
Residual
Outlier t-Plot of Residuals
3.62
Externally Studentized Residuals
1.81
0.00
-1.81
-3.62
1 4 7 10 13 16 19
Run Number
Plot of Residuals Vs. Fitted Values
DESIGN-EXPERT Plot
Etch Rate
Residuals vs. Predicted
25.6
12.85
Residuals
Residuals
0.1
-12.65
-25.4
Predicted
Predicted
value
Plot of Residuals Vs. Run Order
25.6
12.85
Residuals
0.1
-12.65
-25.4
1 4 7 10 13 16 19
Run Number
Plot of Residuals Vs. Factor Levels
DESIGN-EXPERT Plot
Etch Rate
Residuals vs. A
25.6
12.85
Residuals
Residuals
0.1
-12.65
-25.4
1 2 3 4
Factor
A
Level
Positive results could:
µ
ˆ = y..
τˆi = yi. − y..
Example:
For the RF power example, obtain estimates
of the model parameters.
µˆ = y.. = 617.75
τˆ1 = 551.20 − 617.75 = −66.55
τˆ2 = 587.40 − 617.75 = − 30.35
τˆ3 = 625.40 − 617.75 = 7.65
τˆ4 = 707.00 − 617.75 = 89.25
Confidence Intervals
A 100(1-α )% confidence interval
for estimating the ith treatment mean
is given by:
MSError
µ i ∈ yi ± tα 2( N − a )
n
Example:
For the RF power example, construct a 95%
confidence interval on the average etch rate
at 220 W of RF power.
333.70
µ 4 ∈ 707.00 ± t0.025(16)
5
689.68 < µ 4 < 724.32
P.S. The interval does not include the value
of the grand average 617.75.
Simultaneous Estimation
100(1-α )% confidence intervals for the
estimation of r treatment averages are given
by the Bonferroni intervals:
MSError
µ i ∈ yi ± tα 2r ( N −a)
n
Example:
For the RF power example, construct a 90%
confidence interval for estimating the mean
of treatments 3 and 4 simultaneously.
333 .70
µ3 ∈ 625 .4 ± t0.025 ,16
5
333 .70
µ4 ∈ 707 .0 ± t0.025 ,16
5
Graphical Comparison of Means
1 2 3 4
yi
500 550 600 650 700 750
By inspection of the diagram, we can
conclude:
• The four averages could not be
obtained from the same distribution.
• Treatment 4 yields the highest average
etch rate.
• Treatment 1 yields the lowest average
etch rate.
Comparing Pairs of means:
Fisher’s Test
The least significant differences (LSD)
between pairs of treatment averages is given
by:
2 MSE
t , for n = cost.
α 2( N − a ) n
Or,
t MSE ( 1 + 1 ) , for n ≠ n
α 2( N − a ) n n i j
i j
Example:
Fisher’s test for the RF power data, at
α =0.05:
2(333.70)
LSD = t = 24.49
(0.025,16) 5
X = A: A
676.25
Etch Rate
Etch Rate
627.5
578.75
530 RF Power
160 180 200 220
A: A