Académique Documents
Professionnel Documents
Culture Documents
I have prepared supplemental text material to accompany the 6th edition of Introduction to Statistical Quality
Control. This material consists of (1) additional background reading on some aspects of statistics and
statistical quality control and improvement, (2) extensions of and elaboration on some textbook topics and (3)
some new topics that I could not easily find a home for in the text without making the book much too long.
Much of this material has been prepared in at least partial response to the many excellent and very helpful
suggestions that have been made over the years by textbook users. However, sometimes there just wasnt any
way to easily accommodate their suggestions directly in the book. Some of the supplemental material is also
in response to FAQs or frequently asked questions from students. I have also provided a list of references
for this supplemental material that are not cited in the textbook.
Feedback from my colleagues indicates that this book is used in a variety of ways. Most often, it is used as
the textbook in an upper-division undergraduate course on statistical quality control and improvement.
However, there are a significant number of instructors that use the book as the basis of a graduate-level course,
or offer a course taken by a mixture of advanced undergraduates and graduate students. Obviously the topical
content and depth of coverage varies widely in these courses. Consequently, I have included some
supplemental material on topics that might be of interest in a more advanced undergraduate or graduate-level
course.
There is considerable personal bias in my selection of topics for the supplemental material. The coverage is
far from comprehensive.
I have not felt as constrained about mathematical level or statistical background of the readers in the
supplemental material as I have tried to be in writing the textbook. There are sections of the supplemental
material that will require more background in statistics than is required to read the text material. However, I
think that many instructors will be able to use selected portions of this supplement material in their courses
quite effectively, depending on the maturity and background of the students.
Chapter 4
S4-1. Random Samples
S4-2. Expected Value and Variance Operators
1
Chapter 5
S5-1. A Simple Alternative to Runs Rules on the x Chart
Chapter 6
S6-1. s2 is not Always an Unbiased Estimator of 2
S6-2. Should We Use d 2 or d 2* in Estimating via the Range Method?
S6-3. Determining When the Process has Shifted
S6-4. More about Monitoring Variability with Individual Observations
S6-5. Detecting Drifts versus Shifts in the Process Mean
S6-6. The Mean Square Successive Difference as an Estimator of 2
Chapter 7
S7-1. Probability Limits on Control Charts
Chapter 8
S8-1. Fixed Versus Random Factors in the Analysis of Variance
S8-2. More about Analysis of Variance Methods for Measurement Systems Capability Studies
Chapter 9
S9-1. The Markov Chain Approach for Finding the ARLs for Cusum and EWMA Control
Charts
S9-2. Integral Equations versus Markov Chains for Finding the ARL
Chapter 10
S10-1. Difference Control Charts
S10-2. Control Charts for Contrasts
S10-3. Run Sum and Zone Control Charts
S10-4. More about Adaptive Control Charts
2
Chapter 11
S-11.1 Multivariate Cusum Control Charts
Chapter 13
S13-1. Guidelines for Planning Experiments
S13-2. Using a t-Test for Detecting Curvature
S13-3. Blocking in Designed Experiments
S13-4. More about Expected Mean Squares in the Analysis of Variance
Chapter 14
S14-1. Response Surface Designs
S14-2. Fitting Regression Models by Least Squares
S14-3. More about Robust Design and Process Robustness Studies
Chapter 15
S15-1. A Lot Sensitive Compliance (LTPD) Sampling Plan
S15-2. Consideration of Inspection Errors
b.
f ( x1 , x2 )dx1dx2 1
f ( x , x )dx dx
1
The two random variables x1 and x2 are independent if f ( x1 , x2 ) f1 ( x1 ) f 2 ( x2 ) where f1 ( x1 ) and f 2 ( x2 ) are
the marginal probability distributions of x1 and x2, respectively, defined as
f1 ( x1 )
f ( x1 , x2 )dx2 and f 2 ( x2 )
f ( x1 , x2 )dx1
In general, if there are p random variables x1 , x2 ,..., x p then the joint probability density function is
b.
... f ( x , x ..., x
1
)dx1dx2 ...dx p 1
f ( x1 , x2 ,..., x p ) f1 ( x1 ) f 2 ( x2 )... f p ( x p )
where fi ( xi ) are the marginal probability distributions of x1, x2 , , xp, respectively, defined as
n
p( x) p x (1 p) n x
x
n!
p x (1 p) n x , x 0,1, 2,..., n
x !(n x)!
Let np so that p / n . We may now write the binomial distribution as
n(n 1)(n 2) (n x 1) n
p( x)
x!
n n
x
1 2
(1) 1 1
x ! n n
n x
x 1
1
1 1
n n n
1 2 x 1
Let n and p 0 so that np remains constant. The terms 1 , 1 ,..., 1
and
n
n n
x
1 all approach unity. Furthermore,
n
1 e as n
n
n
Thus, upon substitution we see that the limiting form of the binomial distribution is
p ( x)
5
x e
x!
p0 (t t ) (1 t ) p0 (t )
Note that
p0 (t t ) p0 (t )
p0 (t )
t
so consequently
p (t t ) p0 (t )
lim 0
p0 (t )
t 0
t
p0 (t )
px (t t ) px 1 (t ) t (1 t ) px (t )
and
p (t t ) px (t )
lim x
px (t )
t
px 1 (t ) px (t )
t 0
Thus we have a system of differential equations that describe the arrivals or births:
p0 (t ) p0 (t ) for x 0
px (t ) px 1 (t ) px (t ) for x 1, 2,...
6
( t ) x e t
x 0,1, 2,...
x!
2 ( x )2
1
f ( x)
e 2
, x
2
and we stated that and 2 are the mean and variance, respectively, of the distribution. We now
show that this claim is correct.
Note that f ( x) 0 . We first evaluate the integral I
1 z2 / 2
e
dz
2
1 x2 / 2 y 2 / 2
e
dx e
dy
2
1 ( x2 y 2 ) / 2
e
dxdy
2
I2
d
2 1
0
2
2
So we have shown that f ( x ) has the properties of a probability density function.
The integrand obtained by the substitution z ( x ) / is, of course, the standard normal
distribution, an important special case of the more general normal distribution. The standard normal
probability density function has a special notation, namely
( z)
1 z2 / 2
e
, z
2
( z ) (t ) dt
Several useful properties of the standard normal distribution can be found by basic calculus:
1. ( z ) ( z ), for all real z, so ( z ) is an even function (symmetric about 0) of z
2. ( z ) z ( z )
7
3. ( z ) ( z 2 1) ( z )
Consequently, ( z ) has a unique maximum at z = 0, inflection points at z 1 , and both
( z ) 0 and ( z ) 0 as z .
The mean and variance of the standard normal distribution are found as follows:
E ( z ) z ( z )dz
( z )dz
( z ) |
0
and
E ( z 2 ) z 2 ( z )dz
[ ( z ) ( z )]dz
( z ) | ( z )dz
0 1
1
Because the variance of a random variable can be expressed in terms of expectation as
2 E ( z ) 2 E ( z 2 ) 2 , we have shown that the mean and variance of the standard normal
distribution are 0 and 1, respectively.
Now consider the case where x follows the more general normal distribution. Based on the
substitution, we have z ( x ) /
1
2 ( x )2
1
E ( x) x
e 2
dx
( z ) ( z )dz
( z )dz z ( z )dz
(1) (0)
and
1
2 ( x )2
1
E(x ) x
e 2
dx
2
2
( z ) 2 ( z )dz
f ( x)
1
x 2
1
(ln x )2
2 2
x0
E ( x) x e
12 2
V ( x) x2 e2 (e 1)
2
k 12 k 2 2
Like the gamma and Weibull distributions, the lognormal finds application in reliability engineering,
often as a model for survival time of components or systems. Some important properties of the
lognormal distribution are:
1. If x1 and x2 are independent lognormal random variables with parameters ( 1 , 12 ), ( 2 , 22 ) ,
respectively, then y x1 x2 is a lognormal random variable with parameters
1 2 and 12 22 .
2. If x1 , x2 ,..., xk are independently and identically distributed lognormal random variables with
1/ k
k
parameters and , then the geometric mean of the xi, or xi
i 1
2
distribution with parameters and / k .
2
, has a lognormal
3. If x is a lognormal random variable with parameters and 2 , and if a, b, and c are constants
such that b ec , then the random variable y bx a has a lognormal distribution with
parameters c a and a 2 2 .
f ( x)
(r )
( x) r 1 e x , x 0
where r > 0 is a shape parameter and 0 is a scale parameter. The parameter r is called a shape parameter
because it determines the basic shape of the graph of the density function. For example, if r = 1, the gamma
9
distribution reduces to an exponential distribution. There are actually three basic shapes; r 1 or
hyperexponential, r = 1 or exponential, and r > 1 or unimodal with right skew.
The cumulative distribution function of the gamma is
x
(r )
F ( x; r , )
(t ) r 1 e x dt
f ( x) e x , x 0
was introduced in Section 3.3.3 of the text. The exponential distribution is frequently used in reliability
engineering as a model for the lifetime or time to failure of a component or system. Generally, we define the
reliability function of the unit as
R(t ) P{x t}
t
1 f ( x)dx
0
1 F (t )
where, of course, F (t ) is the cumulative distribution function. In biomedical applications, the reliability
function is usually called the survival function. For the exponential distribution, the reliability function is
F (t ) e t
The Hazard Function
The mean and variance of a distribution are quite important in reliability applications, but an additional
property called the hazard function or the instantaneous failure rate is also useful. The hazard function is the
conditional density function of failure at time t, given that the unit has survived until time t. Therefore, letting
X denote the random variable and x denote the realization,
10
f ( x | X x ) h( x )
F ( x | X x)
F ( x x | X x) F ( x | X x)
lim
x
x
F ( x X x x | X x )
lim
x
x
F ( x X x x, X x )
lim
x
xP{ X x}
F ( x X x x)
lim
x
x[1 F ( x)]
f ( x)
1 F ( x)
It turns out that specifying a hazard function completely determines the cumulative distribution
function (and vive-versa).
The Hazard Function for the Exponential Distribution
For the exponential distribution, the hazard function is
h( x )
f ( x)
1 F ( x)
e x
e x
That is, the hazard function for the exponential distribution is constant, or the failure rate is just the
reciprocal of the mean time to failure.
A constant failure rate implies that the reliability of the unit at time t does not depend on its age. This
may be a reasonable assumption for some types of units, such as electrical components, but its
probably unreasonable for mechanical components. It is probably not a good assumption for many
types of system-level products that are made up of many components (such as an automobile).
Generally, an increasing hazard function indicates that the unit is more likely to fail in the next
increment of time than it would have been in an earlier increment of time of the same length. This is
likely due to aging or wear.
Despite the apparent simplicity of its hazard function, the exponential distribution has been an
important distribution in reliability engineering. This is partly because the constant failure rate
assumption is probably not unreasonable over some region of the units life.
S3.7. The Failure Rate for the Weibull Distribution
The instantaneous failure rate or the hazard function was defined in Section S3.6 of the Supplemental Text
Material. For the Weibull distribution, the hazard function is
11
h( x )
f ( x)
1 F ( x)
( / )( x / ) 1 e ( x / )
e ( x / )
Note that if 1 the Weibull hazard function is constant. This should be no surprise, since for 1 the
Weibull distribution reduces to the exponential. When 1 , the Weibull hazard function increases,
approaching as . Consequently, the Weibull is a fairly common choice as a model for components
or systems that experience deterioration due to wear-out or fatigue. For the case where 1 , the Weibull
hazard function decreases, approaching 0 as 0 .
For comparison purposes, note that the hazard function for the gamma distribution with parameters r and
is also constant for the case r = 1 (the gamma also reduces to the exponential when r = 1). Also, when r > 1
the hazard function increases, and when r < 1 the hazard function decreases. However, when r > 1 the hazard
function approaches from below, while if r < 1 the hazard function approaches from above. Therefore,
even though the graph of the gamma and Weibull distributions look very similar, and they can both produce
reasonable fits to the same sample of data, they clearly have very different characteristics in terms of describing
survival or reliability data.
12
The expectation of a random variable is very useful in that it provides a straightforward characterization of the
distribution, and it has a simple practical interpretation as the center of mass, centroid, or mean of the
distribution.
Now suppose that y is a function of the random variable x, say y h( x) . Note that y is also a random variable.
The expectation of h( x ) is defined as
E ( y ) yg ( y )dy
or
13
The name for this theorem comes from the fact that we often apply it without consciously thinking
about whether the theorem is true in our particular case.
Useful Properties of Expectation I:
Let x be a random variable with mean , and c be a constant. Then
1. 1. E (c) c
2. 2. E ( x)
3. 3. E (cx) cE ( x) c
4. 4. E[ch( x)] cE[h( x)]
5. If c1 and c2 are constants and h1 and h2 are functions, then
Now the derivative of E[( x c) 2 ] with respect to c is 2 E ( x) 2c , and this derivative is zero when
c E ( x ) . Therefore, E[( x c) 2 ] is a minimum when c E ( x ) .
V ( x) E[( x )2 ]
2
and we usually call
V ( x) E[( x ) 2 ]
the variance operator. It is straightforward to show that if c is a constant, then
V (cx) c 2 2
The variance is analogous to the moment of inertia in mechanics.
Useful Properties of Expectation II:
Let x1 and x2 be random variables with means 1 and 2 and variances 12 and 22 , respectively, and
let c1 and c2 be constants. Then
1. E( x1 x2 ) 1 2
2. It is possible to show that V ( x1 x2 ) 12 22 2Cov( x1 , x2 ) , where
Cov( x1 , x2 ) E[( x1 1 )( x2 2 )]
14
is the covariance of the random variables x1 and x2. The covariance is a measure of the linear
association between x1 and x2. More specifically, we may show that if x1 and x2 are
independent, then Cov( x1 , x2 ) 0 .
3. V ( x1 x2 ) 12 22 2Cov( x1 , x2 )
4. If the random variables x1 and x2 are independent, V ( x1 x2 ) 12 22
5. If the random variables x1 and x2 are independent, E ( x1 x2 ) E ( x1 ) E( x2 ) 12
6. Regardless of whether x1 and x2 are independent, in general
x
E 1
x2
E ( x1 )
E ( x2 )
V ( x x) 4 2
because Cov( x, x) 2 .
Moments
Although we do not make much use of the notion of the moments of a random variable in the book,
for completeness we give the definition. Let the function of the random variable x be
h( x ) x k
where k is a positive integer. Then the expectation of h( x) x k is called the kth moment about the
origin of the random variable x and is given by
xik p ( xi ), xi is a discrete random variable
all xi
E( xk )
Note that the first origin moment is just the mean of the random variable x. The second origin
moment is
E( x2 ) 2 2
Moments about the mean are defined as
( xi ) k p( xi ), xi is a discrete random variable
all xi
E[( x ) k ]
The second moment about the mean is the variance 2 of the random variable x.
S4.3. Proof That E ( x ) and E ( s 2 ) 2
It is easy to show that the sample average x and the sample variance s2 are unbiased estimators of the
corresponding population parameters and 2 , respectively. Suppose that the random variable x
15
has mean and variance 2 , and that x1 , x2 ,..., xn is a random sample of size n from the population.
Then
1 n
E ( x ) E xi
n i 1
1 n
E ( xi )
n i 1
1 n
n i 1
because the expected value of each observation in the sample is E ( xi ) . Now consider
n
2
( xi x )
E ( s 2 ) E i 1
n 1
n
1
E ( xi x ) 2
n 1 i 1
It is convenient to write
(x x ) x
2
i 1
i 1
2
i
nx 2 , and so
n
n
E ( xi x )2 E ( xi2 ) E (nx 2 )
i 1
i 1
n
Now
E( x )
i 1
2
i
2 and E (x 2 ) 2 2 / n . Therefore
E (s 2 )
1 n
( 2 2 ) n( 2 2 / n)
n 1 i 1
1
n 2 n 2 n 2 2
n 1
(n 1) 2
n 1
2
Note that:
a. These results do not depend on the form of the distribution for the random variable x. Many
people think that an assumption of normality is required, but this is unnecessary.
b. Even though E ( s 2 ) 2 , the sample standard deviation is not an unbiased estimator of the
population standard deviation. This is discussed more fully in section S3-5.
16
L( ) f ( x1 ; ) f ( x2 ; )
f ( xn ; )
The maximum likelihood estimator of is the value of that maximizes the likelyhood function
L( ).
Example 1 The Exponential Distribution
To illustrate the maximum likelihood estimation procedure, set x be exponentially distributed with
parameter . The likelihood function of a random sample of size n, say x1 , x2 ,..., xn , is
n
L( ) e xi
i 1
xi
i 1
Now it turns out that, in general, if the maximum likelihood estimator maximizes L( ), it will also
maximize the log likelihood, ln L ( ) . For the exponential distribution, the log likelihood is
n
ln L( ) n ln xi
i 1
Now
d ln L( ) n n
xi
d
i 1
Equating the derivative to zero and solving for the estimator of we obtain
n
n
x
i 1
1
x
Thus the maximum likelihood estimator (or the MLE) of is the reciprocal of the sample average.
Maximum likelihood estimation can be used in situations where there are several unknown
parameters, say 1 , 2 , , p to be estimated. The maximum likelihood estimators would be found
17
simply by equating the p first partial derivatives L(1 , 2 , , p ) / i , i 1, 2,..., p of the likelihood
(or the log likelihood) equal to zero and solving the resulting system of equations.
Example 2 The Normal Distribution
Let x be normally distributed with the parameters and 2 unknown. The likelihood function of a
random sample of size n is
1 xi
1
L( , )
e 2
i 1 2
n
1
(2 )
2 n/2
1
2
( xi )2
i 1
n
1 n
ln L( , 2 ) ln(2 2 ) 2 ( xi )2
2
2 i 1
Now
ln L( 2 ) 1
2
(x )
i 1
ln L( 2 )
n
1
2 4
2
2
2
(x )
i 1
1 n
xi x
n i 1
1 n
( xi x )2
n i 1
Generally, we like the method of maximum likelihood because when n is large, (1) it results in
estimators that are approximately unbiased, (2) the variance of a MLE is as small as or nearly as small
as the variance that could be obtained with any other estimation technique, and (3) MLEs are
approximately normally distributed. Furthermore, the MLE has an invariance property; that is, if
is the MLE of , then the MLE of a function of , say h( ) , is the same function
h() of the MLE . There are also some other nice statistical properties that MLEs enjoy; see a
book on mathematical statistics, such as Hogg and Craig (1978) or Bain and Engelhardt (1987).
The unbiased property of the MLE is a large-sample or asymptotic property. To illustrate, consider the
MLE for 2 in the normal distribution of example 2 above. We can easily show that
E ( 2 )
n 1 2
n 1 2
2
2
n
n
Notice that the bias in estimating 2 goes to zero as the sample size n . Therefore, the MLE is
an asymptotically unbiased estimator.
18
M k
k
i
i 1
, k 1, 2,..., p
and the first p moments around the origin of the random variable x are just
k E( xk ), k 1, 2,..., p
Example 3 The Normal Distribution
For the normal distribution the first two origin moments are
1
2 2 2
and the first two sample moments are
M 1 x
M 2
1 n 2
xi
n i 1
x
2 2
1 n 2
xi
n i 1
x
2
1 n
( xi x )2
n i 1
The method of moments often yields estimators that are reasonably good. For example, in the above
example the moment estimators are identical to the MLEs. However, generally moment estimators
are not as good as MLEs because they dont have statistical properties that are as nice. For example,
moment estimators usually have larger variances than MLEs.
Least Squares Estimation
The method of least squares is one of the oldest and most widely used methods of parameter
estimation. Section 4.6 gives an introduction to least squares for fitting regression models. Unlike
the method of maximum likelihood and the method of moments, least squares can be employed when
the distribution of the random variable is unknown.
To illustrate, suppose that the simple location model can describe the random variable x:
xi i , i 1, 2,..., n
where the parameter is unknown and the i are random errors. We dont know the distribution of
the errors, but we can assume that they have mean zero and constant variance. The least squares
19
estimator of is chosen so the sum of the squares of the model errors i is minimized. The least
squares function for a sample of n observations x1 , x2 ,..., xn is
n
L i2
i 1
n
( xi ) 2
i 1
Differentiating L and equating the derivative to zero results in the least squares estimator of :
x
In general, the least squares function will contain p unknown parameters and L will be minimized by
solving the equations that result when the first partial derivatives of L with respect to the unknown
parameters are equated to zero. These equations are called the least squares normal equations. See
Section 4.6 in the textbook.
The method of least squares dates from work by Karl Gauss in the early 1800s. It has a very welldeveloped and indeed quite elegant theory. For a discussion of the use of least squares in estimating
the parameters in regression models and many illustrative examples, see Section 4.6 and
Montgomery, Peck and Vining (2007), and for a very readable and concise presentation of the theory,
see Myers and Milton (1991).
S4.5. Proof That E ( S )
In Section S4.4 of the Supplemental Text Material we showed that the sample variance is an unbiased
estimator of the population variance; that is, E ( s 2 ) 2 , and that this result does not depend on the
form of the distribution. However, the sample standard deviation is not an unbiased estimator of the
population standard deviation. This is easy to demonstrate for the case where the random variable x
follows a normal distribution.
Let x have a normal distribution with mean and variance 2 , and let x1 , x2 ,..., xn be a random
sample of size n from the population. Now the distribution of
(n 1) s 2
2
is chi-square with n 1 degrees of freedom, denoted n21 . Therefore the distribution of s2 is
2 /(n 1) times a n21 random variable. So when sampling from a normal distribution, the expected
value of s2 is
2 2
E (s 2 ) E
n 1
n 1
2
n 1
2
n 1
E ( n21 )
(n 1)
2
because the mean of a chi-square random variable with n 1 degrees of freedom is n 1. Now it
follows that the distribution of
20
(n 1) s
is a chi distribution with n 1 degrees of freedom, denoted n 1 . The expected value of S can be
written as
E (s) E
n 1
n 1
n 1
E ( n 1 )
(n / 2)
[(n 1) / 2]
E ( s)
2
(n / 2)
n 1 [(n 1) / 2]
c4
The constant c4 is given in Appendix table VI.
While s is a biased estimator of , the bias gets small fairly quickly as the sample size n increases.
From Appendix table VI, note that c4 = 0.94 for a sample of n = 5, c4 = 0.9727 for a sample of n =
10, and c4 = 0.9896 or very nearly unity for a sample of n = 25.
21
Yield
95
94
93
92
91
90
89
-1.0
-0.5
0.0
Catalyst
0.5
1.0
yij 0 1 xij ij
where 0 and 1 are the intercept and slope, respectively, of the regression line and the regressor or
predictor variable is x1 j 1 and x2 j 1 . The method of least squares can be used to estimate the
slope and intercept in this model. Assuming that we have equal sample sizes n for each factor level
the least squares normal equations are:
2
2n 0 yij
i 1 j 1
n
j 1
j 1
2n 1 y2 j y1 j
0 y
1
2
1 ( y2 y1 )
Note that the least squares estimator of the intercept is the average of all the observations from both
samples, while the estimator of the slope is one-half of the difference between the sample averages
at the high and low levels of the factor x. Below is the output from the linear regression procedure
in Minitab for the catalyst data.
22
Predictor
Coef
SE Coef
Constant
92.4938
0.6752
136.98
0.000
Catalyst
0.2387
0.6752
0.35
0.729
S = 2.70086
R-Sq = 0.9%
R-Sq(adj) = 0.0%
Analysis of Variance
Source
DF
SS
MS
0.912
0.912
0.13
0.729
Residual Error
14
102.125
7.295
Total
15
103.037
Regression
Notice that the estimate of the slope (given in the column labeled Coef and the row labeled Catalyst above)
1
1
( y2 y1 ) (92.7325 92.255) and the estimate of the intercept is 92.4938
2
2
1
1
( y2 y1 ) (93.7325 92.255) . Furthermore, notice that the t-statistic associated with the slope is
2
2
is
0.2387
equal to 0.35, exactly the same value (apart from sign, because we subtracted the averages in the reverse order)
we gave in the text. Now in simple linear regression, the t-test on the slope is actually testing the hypotheses
H0 : 1 0
H0 : 1 0
and this is equivalent to testing H0 : 1 2 .
It is easy to show that the t-test statistic used for testing that the slope equals zero in simple linear regression
is identical to the usual two-sample t-test. Recall that to test the above hypotheses in simple linear regression
the t-statistic is
t0
1
2
S xx
where Sxx
(x
ij
x ) 2 is the corrected sum of squares of the xs. Now in our specific problem,
i 1 j 1
x 0, x1 j 1 and x2 j 1, so S xx 2n. Therefore, since we have already observed that the estimate of
is just sp,
23
t0
2
S xx
1
( y2 y1 )
y y
2
2 1
1
2
sp
sp
2n
n
This is the usual two-sample t-test statistic for the case of equal sample sizes.
Most regression software packages will also compute a table or listing of the residuals from the model. The
residuals from the Minitab regression model fit obtained above are as follows:
Obs
Catalyst
Yield
Fit
SE Fit
Residual
St Resid
-1.00
91.500
92.255
0.955
-0.755
-0.30
-1.00
94.180
92.255
0.955
1.925
0.76
-1.00
92.180
92.255
0.955
-0.075
-0.03
-1.00
95.390
92.255
0.955
3.135
1.24
-1.00
91.790
92.255
0.955
-0.465
-0.18
-1.00
89.070
92.255
0.955
-3.185
-1.26
-1.00
94.720
92.255
0.955
2.465
0.98
-1.00
89.210
92.255
0.955
-3.045
-1.21
1.00
89.190
92.733
0.955
-3.543
-1.40
10
1.00
90.950
92.733
0.955
-1.783
-0.71
11
1.00
90.460
92.733
0.955
-2.273
-0.90
12
1.00
93.210
92.733
0.955
0.477
0.19
13
1.00
97.190
92.733
0.955
4.457
1.76
14
1.00
97.040
92.733
0.955
4.307
1.70
15
1.00
91.070
92.733
0.955
-1.663
-0.66
16
1.00
92.750
92.733
0.955
0.017
0.01
The column labeled Fit contains the predicted values of yield from the regression model, which just turn out
to be the averages of the two samples. The residuals are in the sixth column of this table. They are just the
differences between the observed values of yield and the corresponding predicted values. A normal probability
plot of the residuals follows.
24
99
95
90
Percent
80
70
60
50
40
30
20
10
5
-7.5
-5.0
-2.5
0.0
Residual
2.5
5.0
Notice that the residuals plot approximately along a straight line, indicating that there is no serious
problem with the normality assumption in these data. This is equivalent to plotting the original yield
data on separate probability plots as we did in Chapter 3.
S4.7. Expected Mean Squares in the Single-Factor Analysis of Variance
In section 4.5.2 we give the expected values of the mean squares for treatments and error in the single-factor
analysis of variance (ANOVA). These quantities may be derived by straightforward application of the
expectation operator.
Consider first the mean square for treatments:
E ( MSTreatments ) E
FG SS IJ
H a 1 K
Treatments
SSTreatments
1 a 2 1 2
yi. an y..
n i 1
yij i ij
RSi 1,2,, a
T j 1,2,, n
25
E ( SSTreatments ) E (
1 a 2
1
yi . ) E ( y..2 )
n i 1
an
Consider the first term on the right hand side of the above expression:
1 a 2 1 a
E ( yi . ) E (n n i i . ) 2
n i 1
n i 1
Squaring the expression in parentheses and taking expectation results in
E(
a
1 a 2 1
2
2
y
)
[
a
(
n
n
i2 an 2 ]
i. n
n i 1
i 1
a
an 2 n i2 a 2
i 1
because the three cross-product terms are all zero. Now consider the second term on the right hand side of
E ( SSTreatments ) :
FG 1 y IJ 1 E (an n
H an K an
a
2
..
.. ) 2
i 1
since
1
E (an .. ) 2
an
i 1
FG 1 y IJ 1 [(an)
H an K an
2
..
an 2 ]
an 2 2
since the expected value of the cross-product is zero. Therefore,
E ( SSTreatments ) E (
1 a 2
1
yi . ) E ( y..2 )
n i 1
an
a
an 2 n i2 a 2 (an 2 2 )
i 1
2 (a 1) n i2
i 1
26
E ( MSTreatments ) E
FG SS IJ
H a 1 K
Treatments
2 (a 1) n i2
i 1
a 1
a
n i2
i 1
a 1
SS E
E ( MS E ) E
N a
a n
E ( yij yi. ) 2
N a i 1 j 1
a
n
a
1
1
E yij2 yi2.
N a i 1 j 1
n i 1
Substituting the model into this last expression, we obtain
2
a n
1
1 a n
2
E ( MS E )
E ( i ij ) ( i ij )
N a i 1 j 1
n i 1 j 1
E ( MS E )
a
a
1
2
2
2
2
N
n
i2 a 2
N a
i 1
i 1
27
370
350
370
0.2
308
277
271
0.4
200
150
142
0.6
120
79
73
0.8
72
44
40
44
26
23
6.3
4.6
4.3
2.4
2.4
28
the process is in-control with mean 0 and standard deviation for the first t observations, but
between xt and xt+1 an assignable cause occurs that results in a sustained shift in the process mean to
a new level 0 and the mean remains at this new level for the remaining sample
observations xt 1 , , xm . Under these conditions, Woodall and Montgomery (2000-01) show that
E (s 2 ) 2
t (m t )
( ) 2 .
m(m 1)
(S6.1)
In fact, this result holds for any case in which the mean of t of the observations is 0 and the mean of the
remaining observations is 0 , since the order of the observations is not relevant in computing s2. Note
that s2 is biased upwards; that is, s2 tends to overestimate 2. Furthermore, the extent of the bias depends on
the magnitude of the shift in the mean (), the time period following which the shift occurs (t), and the number
of available observations (m). For example, if there are m = 25 observations and the process mean shifts from
0 to
0 (that is, 1)
between the 20th and the 21st observation (t = 20), then s2 will overestimate
2 by 16.7% on average. If the shift in the mean occurs earlier, say between the 10th and 11th observations, then
s2 will overestimate 2 by 25% on average.
The proof of Equation S6.1 is straightforward. Since we can write
s2
1 m 2
xi mx 2
m 1 i 1
then
E (s 2 ) E
Now
29
1 m 2
1 m
2
x
mx
E ( xi2 ) mE ( x 2 )
m 1 i 1
m 1 i 1
1
m1
FG E ( x )IJ 1 FG E ( x ) E ( x )IJ
H
K m 1H
K
1
ct ( ) (m t )( ) (m t ) h
m1
1
ct (m t )( ) m h
m1
m
2
i
i 1
2
i
2
i
i 1
i t 1
2
0
2
0
and
LMF
MNGH
FG
H
IJ IJ
K K
1
m
mt
mE ( x 2 )
0
m1
m1
m
OP
m PQ
Therefore
2
2
1 2
m t
2
2
t 0 (m t )( 0 ) m m 0
E (s )
m 1
m
m
2
1 2
mt
2
t 0 (m t )( 0 ) 2 m 0
m 1
m
1
(m t ) 2
2
2
(
m
t
)(
( ) 2
m 1
m
1
(m t )
(m t )( ) 2 1
m 1
m
t (m t )
( ) 2
m(m 1)
i 1
(S6.2)
1 R / d2
(S6.3)
is widely used after the application of control charts to estimate process variability and to assess process
capability. In Chapter 4 we report the relative efficiency of the range estimator given in Equation (S6.3) to the
sample standard deviation for various sample sizes. For example, if n = 5, the relative efficiency of the range
estimator compared to the sample standard deviation is 0.955. Consequently, there is little practical difference
between the two estimators. Equation (S6.3) is also frequently used to determine the usual 3-sigma limits on
the Shewhart x chart in statistical process control. The estimator
2 R / d 2*
(S6.4)
is more often used in gauge R & R studies and in variables acceptance sampling. Here d 2 represents a constant
whose value depends on both m and n. See Chrysler, Ford, GM (1995), Military Standard 414 (1957), and
Duncan (1986).
Patnaik (1950) showed that R / is distributed approximately as a multiple of a - distribution. In particular,
R / is distributed approximately as
d2* /
1
1
5
d 2* d 2 1
.
2
128 3
4 32
(S6.5)
It has been pointed out by Duncan (1986), Wheeler (1995), and Luko (1996), among others, that 1 is an
unbiased estimator of and that
22 is an unbiased estimator of 2.
For
22 to be an unbiased estimator of 2,
d 2* was required.
He showed that
(S6.6)
where Vn is the variance of the sample range with sample size n from a normal population with unit variance.
It is important to note that
Vn d32 , so Equation (S5-6) can be easily used to determine values of d 2* from the
*
widely available tables of d2 and d3. Thus, a table of d 2 values, such as the ones given by Duncan (1986),
Wheeler (1995), and many others, is not required so long as values of d2 and d3 are tabled, as they usually are
(once again, see Appendix Table VI). Also, use of the approximation
d 2 d 2 1
4
31
d 2* values given by Duncan (1986) is the most frequently recommended. If a table is required, the
d 2* approaches d2.
*
2 approaches
Sometimes use of Equation (S6.4) is recommended without any explanation. See, for example, the AIAG
measurement systems capability guidelines [Chrysler, Ford, and GM (1995)]. The choice between 1 and 2
has often not been explained clearly in the literature. It is frequently stated that the use of Equation (S6.3)
requires that R be obtained from a fairly large number of individual ranges. See, for example, Bissell (1994,
p. 289). Grant and Leavenworth (1996, p. 128) state that Strictly speaking, the validity of the exact value of
the d2 factor assumes that the ranges have been averaged for a fair number of subgroups, say, 20 or more.
When only a few subgroups are available, a better estimate of is obtained using a factor that writers on
statistics have designated as
d 2* .
Nelson (1975) writes, If fewer than a large number of subgroups are used,
Equation (S6.3) gives an estimate of which does not have the same expected value as the standard deviation
estimator. In fact, Equation (S6.3) produces an unbiased estimator of regardless of the number of samples
m, whereas the pooled standard deviation does not (refer to Section S4.5 of the Supplemental Text Material).
The choice between 1 and 2 depends upon whether one is interested in obtaining an unbiased estimator of
or 2. As m increases, both estimators (S6.3) and (S6.4) become equivalent since each is a consistent
estimator of .
It is interesting to note that among all estimators of the form
c d 2 / (d 2* ) 2 .
The derivation of this result is in the proofs at the end of this section. If we let
d
3 *2 2 R
(d 2 )
then it is shown in the proofs below that
FG
H
MSE ( 3 ) 1
d 22
(d 2* ) 2
IJ
K
Luko (1996) compared the mean squared error of 2 in estimating to that of 1 and recommended
2 on the basis of uniformly lower MSE values. By definition, 3 leads to further reduction in MSE.
It is shown in the proofs at the end of this section that the percentage reduction in MSE using 3
instead of 2 is
32
FG d d IJ
H d K
*
2
50
*
2
Values of the percentage reduction are given in Table S6.1. Notice that when both the number of
subgroups and the subgroup size are small, a moderate reduction in mean squared error can be
obtained by using 3 .
Table S6.1.
Percentage Reduction in Mean Squared Error from using
3 instead of 2
Subgroup
Size, n
Number of Subgroups, m
1
20
5.7269
4.0231
3.1291
2.5846
2.2160
1.9532
1.7536
10
1.5963
Result 1: Let
Proof:
MSE ( ) E[(cR ) 2 ]
[c 2 R 2 2cR 2 ]
c 2 E ( R 2 ) 2cE ( R ) 2
Now E ( R ) Var ( R ) [ E ( R )] d 3 / m (d 2 ) . Thus
33
15
Proofs
10
MSE ( ) c 2 d 32 2 / m c 2 d 22 2 2c (d 2 ) 2
2 [c 2 (d 32 / m d 22 ) 2cd 2 1]
2 [c 2 (d 2* ) 2 2cd 2 1]
Result 2: The value of c that minimizes the mean squared error of estimators of the form cR in estimating
is
d2
.
(d2* ) 2
Proof:
MSE ( ) 2 [c 2 (d 2* ) 2 2cd 2 1]
dMSE ( )
2 [2c(d 2* ) 2 2d 2 ] 0
dc
d
c *2 2 .
(d 2 )
Result 3: The mean square error of 3
FG
H
IJ
K
d2
d
R is 2 1 *2 2 .
* 2
(d 2 )
(d 2 )
Proof:
LM d (d ) 2 d
(d )
N (d )
L d 2 d 1OP
= M
N (d ) (d ) Q
F d IJ
G1
H (d ) K
MSE ( 3 ) 2
2
2
2
* 4
2
* 2
2
2
2
* 2
2
2
* 2
2
OP
Q
d 2 1 (from result 1)
2
2
* 2
2
2
2
* 2
2
Result 4: Let 2
LM
N
OP
Q
R
d
in mean square error using the minimum mean square error estimator instead of
Luko (1996)], is
FG d d IJ
H d K
50
Proof:
34
*
2
*
2
R
[as recommended by
d 2*
2 2 (d2* d2 )
, therefore
d2*
MSE ( 2 ) MSE ( 3 )
FG
H
2 2 (d 2* d 2 )
d 22
2
d 2*
(d 2* ) 2
LM 2(d d ) (d ) d OP
(d )
N d
Q
LM 2(d d ) (d d )(d d ) OP
(d )
N d
Q
(d d ) F
d d I
2
G
d
H d JK
(d d ) F d d I
GH d JK
d
*
2
*
2
*
2
*
2
*
2
*
2
*
2
*
2
* 2
2
*
2
*
2
*
2
*
2
*
2
(d 2* d 2 ) 2
(d 2* ) 2
Consequently
* 2
2
2
2
* 2
2
*
2
IJ
K
*
2
*
2
* 2
2
*
2
*
2
*
2
Control charts monitor a process to determine whether an assignable cause has occurred. Knowing
when the assignable cause has occurred would be very helpful in its identification and eventual
removal. Unfortunately, the time of occurrence of the assignable cause does not always coincide with
the control chart signal. In fact, given what is known about average run length performance of control
charts, it is actually very unlikely that the assignable cause occurs at the time of the signal. Therefore,
when a signal occurs, the control chart analyst should look earlier in the process history to determine
the assignable cause.
But where should we start? The Cusum control chart provides some guidance simply search
backwards on the Cusum status chart to find the point in time where the Cusum last crossed zero
(refer to Chapter 8). However, the Shewhart x control chart provides no such simple guidance.
Samuel, Pignatiello, and Calvin (1998) use some theoretical results by Hinkley (1970) on changepoint problems to suggest a procedure to determine the time of a shift in the process mean following
a signal on the Shewhart x control chart. They assume the standard x control chart with in-control
0 . Suppose that the chart signal at subgroup average xT . Now the incontrol subgroups are x1 , x2 ,..., xt , and the out-of-control subgroups are xt 1 , xt 2 ,..., xT , where
value of the process mean
obviously t T . Their procedure consists of finding the value of t in the range 0 t T that
maximizes
Ct (T t )( xT ,t 0 )2
35
where xT ,t
1 T
xi is the reverse cumulative average; that is, the average of the T t most recent
T t i t 1
subgroup averages. The value of t that maximizes Ct is the estimator of the last subgroup that was
selected from the in-control process.
You may also find it interesting and useful to read the material on change point procedures for process
monitoring in Chapter 10.
S6.4. More about Monitoring Variability with Individual Observations
As noted in the textbook, when one is monitoring a process using individual (as opposed to
subgrouped) measurements, a moving range control chart does not provide much useful additional
information about shifts in process variance beyond that which is provided by the individuals control
chart (or a Cusum or EWMA of the individual observations). Sullivan and Woodall (1996) describe
a change-point procedure that is much more effective that the individuals (or Cusum/EWMA) and
moving range chart.
Assume that the process is normally distributed. Then divide the n observations into two partitions
of n1 and n2 observations, with n1 = 2, 3, , n 2 observations in the first partition and n n1 in the
second. The log-likelihood of each partition is computed, using the maximum likelihood estimators
for and in each partition. The two log-likelihood functions are then added. Call the sum of
the two log-likelihood functions La. Let L0 denote the log-likelihood computed without any partitions.
Then find the maximum value of the likelihood ratio r = 2(La L0). The value of n1 at which this
maximum value occurs is the change point; that is, it is the estimate of the point in time at which a
change in either the process mean or the process variance (or both) has occurred.
2
Sullivan and Woodall show how to obtain a control chart for the likelihood ratio r. The control limits
must be obtained either by simulation or by approximation. When the control chart signals, the
quantity r can be decomposed into two components; one that is zero if the means in each partition are
equal, and another that is zero if the variances in each partition are equal. The relative size of these
two components suggests whether it is the process mean or the process variance that has shifted.
S6.5. Detecting Drifts versus Shifts in the Process Mean
In studying the performance of control charts, most of our attention is directed towards describing
what will happen on the chart following a sustained shift in the process parameter. This is done
largely for convenience, and because such performance studies must start somewhere, and a sustained
shift is certainly a likely scenario. However, a drifting process parameter is also a likely possibility.
Aerne, Champ, and Rigdon (1991) have studies several control charting schemes when the process
mean drifts according to a linear trend. Their study encompasses the Shewhart control chart, the
Shewhart chart with supplementary runs rules, the EWMA control chart, and the Cusum. They design
the charts so that the in-control ARL is 465. Some of the previous studies of control charts with
drifting means did not do this, and different charts have different values of ARL0, thereby making it
difficult to draw conclusions about chart performance. See Aerne, Champ, and Rigdon (1991) for
references and further details.
They report that, in general, Cusum and EWMA charts perform better in detecting trends than does
the Shewhart control chart. For small to moderate trends, both of these charts are significantly better
than the Shewhart chart with and without runs rules. There is not much difference in performance
between the Cusum and the EWMA.
36
MSSD
n
1
( x1 xi 1 )2
2(n 1) i 1
n from a population with mean and variance 2 . Without any loss of generality, we may take the mean to be
zero. Then
n
1
E ( MSSD) E
( xi xi 1 ) 2
2(n 1) i 2
n
1
E ( xi2 xi21 2 xi xi 1 )
2(n 1) i 2
1
[(n 1) 2 (n 1) 2 ]
2(n 1)
2(n 1) 2
2(n 1)
Therefore, the mean square successive difference is an unbiased estimator of the population variance.
37
0.00
0.0007
11.00
0.9531
1.00
0.0060
12.00
0.9776
2.00
0.0258
13.00
0.9901
3.00
0.0744
14.00
0.9959
4.00
0.1632
15.00
0.9984
5.00
0.2914
16.00
0.9994
6.00
0.4443
17.00
0.9998
7.00
0.5988
18.00
0.9999
8.00
0.7340
19.00
1.0000
9.00
0.8380
20.00
1.0000
10.00
0.9092
If the lower control limit is at zero and the upper control limit is at 0.147, then any sample with 15 or more
defective items will plot beyond the upper control limit. The above table shows that the probability of
obtaining 15 or more defectives when the process is in-control is 1 0.9959 = 0.0041. This is about 50%
greater than the false alarm rate on the normal-theory three-sigma limit control chart (0.0027). However, if
we were to set the lower control limit at 0.01 and the upper control limit at 0.15, and conclude that the process
is out-of-control only if a control limit is exceeded, than the false alarm rate is 0.0007 + 0.0016 = 0.0023,
which is very close to the advertised value of 0.0027. Furthermore, there is a nonzero LCL, which can be very
useful in practice. Notice that the control limits are not symmetric around the center line. However, the
distribution of p is not symmetric, so this should not be too surprising.
There are several other interesting approaches to setting probability-type limits on attribute control charts.
Refer to Ryan (2000), Acosta-Mejia (1999), Ryan and Schwertman (1997), Schwertman and Ryan (1997), and
Shore (2000).
38
yij i ij
but now the treatment effects i are random variables, because the treatment levels actually used in the
experiment have been chosen at random. The population of treatments is assumed to be normally and
independently distributed with mean zero and variance 2 . Note that the variance of an observation is
V ( yij ) V ( i ij )
2 2
We often call 2 and 2 variance components, and the random model is sometimes called the components
of variance model. All of the computations in the random model are the same as in the fixed effects model,
but since we are studying an entire population of treatments, it doesnt make much sense to formulate
hypotheses about the individual factor levels selected in the experiment. Instead, we test the following
hypotheses about the variance of the treatment effects:
H 0 : 2 0
H1 : 2 0
The test statistic for these hypotheses is the usual F-ratio, F = MSTreatments/MSE. If the null hypothesis is not
rejected, there is no variability in the population of treatments, while if the null hypothesis is rejected, there is
significant variability among the treatments in the entire population that was sampled. Notice that the
conclusions of the ANOVA extend to the entire population of treatments.
The expected mean squares in the random model are different from their fixed effects model counterparts. It
can be shown that
E ( MSTreatments ) 2 n 2
E ( MS E ) 2
Frequently, the objective of an experiment involving random factors is to estimate the variance components.
A logical way to do this is to equate the expected values of the mean squares to their observed values and solve
the resulting equations. This leads to
MSTreatments MS E
n
2
MS E
A typical application of experiments where some of the factors are random is in a measurement systems
capability study, as discussed in Chapter 7. The model used there is a factorial model, so the analysis and the
expected mean squares are somewhat more complicated than in the single factor model considered here.
39
The tabular approach does not lend itself to constructing confidence intervals on the variance
components or functions of the variance components of interest. For that reason we do not
recommend the tabular approach for general use. There are three general approaches to constructing
these confidence intervals: (1) the Satterthwaite method, (2) the maximum likelihood large-sample
method, and (3) the modified large sample method. Montgomery (2001) gives an overview of these
different methods. Of the three approaches, there is good evidence that the modified large sample
approach is the best in the sense that it produces confidence intervals that are closest to the stated
level of confidence.
Hamada and Weerahandi (2000) show how generalized inference can be applied to the problem of determining
confidence intervals in measurement systems capability studies. The technique is somewhat more involved
that the three methods referenced above. Either numerical integration or simulation must be used to find the
desired confidence intervals. Burdick, Borror, and Montgomery (2003) discuss this technique.
While the tabular method should be abandoned, the control charting aspect of measurement systems capability
studies should be used more consistently. All too often a measurement study is conducted and analyzed via
some computer program without adequate graphical analysis of the data. Furthermore, some of the advice in
various quality standards and reference sources regarding these studies is just not very good and can produce
results of questionable validity. The most reliable measure of gauge capability is the probability that parts are
misclassified.
40
The Cusum control chart statistic C (or C ) form a Markov process with a continuous state space. By
discretizing the continuous random variable C (or C ) with a finite set of values, approximate ARLs can
be obtained from Markov chain theory. For the upper one-sided Cusum with upper decision interval H, the
intervals
are
defined
as
follows:
(, w / 2],[ w / 2,3w / 2], ,[( k 1/ 2) w, ( k 1/ 2) w], ,[( m 3 / 2) w, H ],[ H , ) where m + 1 is the
number of states and w = 2H/(2m- 1). The elements of the transition probability matrix of the Markov chain
P [ pij ] are
pi 0
w/ 2
pij
f ( x iw k )dx, i 0,1,..., m 1
( j i / 2) w
( j i / 2) w
i 0,1,..., m 1
f ( x iw k )dx
j 1, 2,..., m 1
pmj 0, j 0,1,..., m 1
pmm 1
The absorbing state is m and f denotes the probability density function of the variable that is being monitored
with the Cusum.
From the theory of Markov chains, the expected first passage times from state i to the absorbing state are
m 1
i 1 pij j , i 0,1,..., m 1
j 0
Thus, i is the ARL given that the process started in state i. Let Q be the matrix of transition probabilities
obtained by deleting the last row and column of P. Then the vector of ARLs is found by computing
I Q 1
where 1 is an m1 vector of 1s and I is the m m identity matrix.
When the process is out of control, this procedure gives a vector of initial-state (or zero-state) ARLs. That is,
the process shifts out of control at the initial start-up of the control chart. It is also possible to calculate steadystate ARLs that describe performance assuming that the process shifts out of control after the control chart has
been operating for a long period of time. There is typically very little difference between initial-state and
steady-state ARLs.
41
Let P ( n, i ) be the probability that run length takes on the value n given that the chart started in state i. Collect
these quantities into a vector say
p1 (I Q)11
pn Qpn1 , n 2,3,...
This technique can be used to calculate the probability distribution of the run length, given the control chart
started in state i. Some authors believe that the distribution of run length or its percentiles is more useful that
the ARL, since the distribution of run length is usually highly skewed and so the ARL may not be a typical
value in any sense.
S9.2. Integral Equations Versus Markov Chains for Finding the ARL
Two methods are used to find the ARL distribution of control charts, the Markov chain method and
an approach that uses integral equations. The Markov chain method is described in Section S9.1 of
the Supplemental Text Material. This section gives an overview of the integral equation approach
for the Cusum control chart. Some of the notation defined in Section S9.1 will be used here.
Let P (n, u ) and R (u ) be the probability that the run length takes on the value n and the ARL for the Cusum
when the procedure begins with initial value u. For the one-sided upper Cusum
P(1, u ) 1
f ( x u k )dx
w/ 2
m 1
f ( x u k )dx
j 1
( j 1/ 2) w
( j 1/ 2) w
f ( x u k )dx
and
P(n, u ) P(n 1, 0)
P(n 1, 0)
m 1
j 1
( j 1/ 2) w
( j 1/ 2) w
P(n 1, 0)
m 1
f ( x u k )dx P (n 1, y ) f ( x u k )dx
0
f ( x u k )dx
P(n 1, y ) f ( x u k )dx
P(n 1, y ) f ( x u k )dx
f ( x u k )dx P(n 1, 0 )
w/ 2
P(n 1, j )
j 1
w/ 2
( j 1/ 2) w
( j 1/ 2) w
f ( x u k )dx
f ( x u k )dx
42
m 1
m 1
But these last equations are just the equations used for calculating the probabilities of first-passage
times in a Markov chain. Therefore, the solution to the integral equation approach involves solving
equations identical to those used in the Markov chain procedure.
Champ and Rigdon (1991) give an excellent discussion of the Markov chain and integral equation
techniques for finding ARLs for both the Cusum and the EWMA control charts. They observe that
the Markov chain approach involves obtaining an exact solution to an approximate formulation of the
ARL problem, while the integral equation approach involves finding an approximate solution to the
exact formulation of the ARL problem. They point out that more accurate solutions can likely be
found via the integral equation approach. However, there are problems for which only the Markov
chain method will work, such as the case of a drifting mean.
43
A2 R12 R22 , where R12 and R22 are the average ranges for the reference samples (1) and the current
production samples (2) used to establish the control limits.
The second type of difference control chart was suggested by Ott (1947), who considered the situation where
differences are observed between paired measurements within each subgroup (much as in a paired t-test), and
the average difference for each subgroup is plotted on the chart. The center line for this chart is zero, and the
control limits are at A2 R , where R is the average of the ranges of the differences. This chart would be
useful in instrument calibration, where one measurement on each unit is from a standard instrument (say in a
laboratory) and the other is from an instrument used in different conditions (such as in production).
44
Furnace
Location1
3
7
1 5
4
Furnace
Location2
Furnace
Location3
Furnace
Location4
Figure S10.1. Diagram of a Furnace where four wafers are simultaneously processed and nine quality
measurements are performed on each wafer.
The most widely used approach to monitoring these processes is to first consider the average of all mn
observations from a run as a single observation and to use a Shewhart control chart for individuals to monitor
the overall process mean. The control limits for this chart are usually found by applying a moving range to
the sequence of averages. Thus, the control limits for the individuals chart reflect run-to-run variability, not
variability within a run. The variability within a run is monitored by applying a control chart for s (the standard
deviation) or s 2 to all mn observations from each run. It is interesting to note that this approach is so widely
used that at least one popular statistical software package (Minitab) includes it as a standard control charting
option (called the between within procedure in Minitab). This procedure was illustrated in Example 5-11.
Runger and Fowler (1999) show how the structure of the data obtained on these processes can be
represented by an analysis of variance model, and how control charts based on contrasts can be
designed to detect specific assignable causes of potential interest. Below we briefly review their
results and relate them to some other methods. Then we analyze the average run performance of the
contrast charts and show that the use of specifically designed contrast charts can greatly enhance the
ability of the monitoring scheme to detect assignable causes. We confine our analysis to Shewhart
charts, but both Cusum and EWMA control charts would be very effective alternatives, because they
are more effective in detecting small process shifts, which are likely to be of interest in many of these
applications
Contrast Control Charts
We consider the oxidation process in Figure S10.1, but allow m wafers in each run with n measurements or
sites per wafer. The appropriate model for oxide thickness is
yij ri s j ij
45
(S10.1)
where yij is the oxide thickness measurement from run i and site j, ri is the run effect, sj is the site effect, and
ij is a random error component. We assume that the site effects are fixed effects, since the measurements are
generally taken at the same locations on all wafers. The run effect is a random factor and we assume it is
2
distributed as NID(0, r ) . We assume that the error term is distributed as NID(0, 2 ) . Notice that equation
c cy t
where the elements of the vector c sum to zero and, for convenience, we assume that the
contrast vector has unit length. That is,
c1 0 and cc 1
Any contrast vector is orthogonal to the vector that generates the mean, since the mean can be written
as
yt 1y t / mn
Thus, a contrast generates information that is different from the information produced by the overall
mean from the current run. Based on the particular problem, the control chart analyst can choose the
elements of the contrast vector c to provide information of interest to that specific process.
For example, suppose that we were interested in detecting process shifts that could cause a difference
in mean thickness between the top and bottom of the furnace. The engineering cause of such a
difference could be a temperature gradient along the furnace from top to bottom. To detect this
disturbance, we would want the contrast to compare the average oxide thickness of the top wafer in
the furnace to the average thickness of the bottom wafer. Thus, if m = 4, the vector c has mn = 36
components, the first 9 of which are +1, the last 9 of which are 1, and the middle 18 elements are
zero. To normalize the contrast to unit length we would actually use
c [1,1,...,1,0,0,..., 0,1,1,...,1] / 18
One could also divide the elements of c by nine to compute the averages of the top and bottom wafers,
but this is not really necessary.
In practice, a set of k contrasts, say
c1 , c2 ,..., ck
can be used to define control charts to monitor a process to detect k assignable causes of interest.
These simultaneous control charts have overall false alarm rate , where
k
1 (1 i )
(S10.2)
i 1
and i is the false alarm rate for the ith contrast. If the contrasts are orthogonal, then Equation (S9-2)
holds exactly, while if the contrasts are not orthogonal then the Bonferroni inequality applies and the
in Equation (S10.2) is a lower bound on the false alarm rate.
46
Related Procedures
Several authors have suggested related approaches for process monitoring when non-standard
conditionss relative to rational subgrouping apply. Yashchin (1994), Czitrom and Reese (1997), and
Hurwicz and Spagon (1997) all present control charts or other similar techniques based on variance
components. The major difference in this approach in comparison to these authors is the use of an
analysis-of-variance type partitioning based on contrasts instead of variance components as the basis
of the monitoring scheme. Roes and Does (1995) do discuss the use of contrasts, and Hurwicz and
Spagon discuss contrasts to estimate the variance contributed by sites within a wafer. However, the
Runger and Fowler model is the most widely applicable of all the techniques we have encountered.
Even though the methodology used to monitor specific differences in processing conditions has been
studied by all these authors, the statistical performance of these charts has not been demonstrated.
We now present some performance results for Shewhart control charts.
Average Run Length Performance of Shewhart Charts
In this section we assume that the process shown in Figure S10.1 is of interest. The following
scenarios are considered:
Significant changes between the outside and the inside of each wafer.
c1
c2
c3
[1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1,1,1,1]
16
[0,0,0,0,0,1,1,1,1,0,0,0,0,0,1,1,1,1,0,0,0,0,0,1,1,1,10,0,0,0,0,1,1,1,1]
32
[0,1,1,1,1,1,1,1,1,0,1,1,1,1,1,1,1,1,0,1,1,1,1,1,1,1,1,0,1,1,1,1,1,1,1,1]
A comparison of the ARL values obtained using these contrasts and the traditional approach an
individuals control chart for the mean of all 36 observations- is presented in Tables S10.1, S10.2, and
S10.3. From inspection of these tables, we see that the charts for the orthogonal contrasts, originally
with the same in-control ARL as the traditional chart, are more sensitive to changes at specific
locations, thus improving the chances of early detection of an assignable cause. Notice that the
improvement is dramatic for small shifts, say on the order of 1.5 standard deviations or less.
A similar analysis was performed for a modified version of the process shown in figure S10.1. In
this example, there are seven measurements per wafer for a total of 28 measurements in a run. There
are still three measurements at the center of the wafer, but now there are only measurements at the
perimeter; one in each corner. The same types of contrasts used in the previous example (top versus
bottom, left versus right and edge versus center) were analyzed and the ARL results are presented in
Tables S10.4, S10.6, and S10.6.
47
Table S10.1. Average Run Length Performance of Traditional and Orthogonal Contrast Charts for a
shift in the Edges of all Wafers. In this chart m = 4 and n = 9.
Size of Shift
In Multiples of
Traditional Chart
0.5
11.7
13.6
1.9
2.2
1.5
1.1
1.1
2.5
Table S10.2. Average Run Length Performance of Traditional and Orthogonal Contrast Charts for a
shift in the Top Wafer. In this chart m = 4 and n = 9.
Size of Shift
In Multiples of
48
Traditional Chart
0.5
23.4
47
3.9
10
1.5
1.5
3.4
1.1
1.7
2.5
1.2
Table S10.3. Average Run Length Performance of Traditional and Orthogonal Contrast Charts for a
shift in the Left side of all Wafers. In this chart m = 4 and n = 9.
Size of Shift
In Multiples of
Traditional Chart
0.5
26.7
57.2
4.6
13.6
1.5
1.7
4.6
1.1
2.2
2.5
1.4
1.1
Table S10.4. Average Run Length Comparison between Traditional and Orthogonal Contrast Charts for
a shift in the Edge of all Wafers. In this chart m = 4 and n = 7.
Size of Shift
In Multiples of
49
Traditional Chart
0.5
26.7
46.4
4.6
9.8
1.5
1.7
3.3
1.1
1.7
2.5
1.2
Table S10.5. Average Run Length Comparison between Traditional and Orthogonal Contrast Charts for
a change in the Top Wafer. In this chart m = 4 and n = 7.
Size of Shift
In Multiples of
Traditional Chart
0.5
30.8
57.9
5.5
13.8
1.5
4.7
1.2
2.2
2.5
1.4
1.1
Table S10.6. Average Run Length Performance of Traditional and Orthogonal Contrast Charts
for a shift in the left side of all Wafers. In this chart m = 4 and n = 7.
Size of Shift
Traditional Chart
In Multiples of
0.5
26.7
46.4
4.6
9.8
1.5
1.7
3.3
1.1
1.7
2.5
1.2
Decreasing the number of measurements per wafer has increased the relative importance of the changes in the
mean of a subset of the observations and the traditional control charts signal the shift faster than in the previous
example. Still, note that the control charts based on orthogonal contrasts represent a considerable improvement
over the traditional approach.
[ 0 Ai 0 / n , 0 Ai 1 0 / n ), i 0,1, 2,..., a
50
These authors show that a well-designed two-state VSI control chart will perform nearly as well as
the optimal chart, so that in practical use, there is little to be gained in operational performance by
using more than two states. Zimmer, Montgomery and Runger (1998) consider the VSS control chart
and show that two states are not optimal, although the performance improvements when using more
that two states are modest, and mostly occur when the interest is in detecting small process shifts.
Zimmer, Montgomery and Runger (2000) summarize the performance of numerous adaptive control
chart schemes, and offer some practical guidelines for their use. They observe that, in general,
performance improves more quickly from adapting the sample size than from adapting the sampling
interval.
Tagaras (1998) also gives a nice literature review of the major work in the field up through about
1997. Baxley ((1995) gives an interesting account of using VSI control charts in nylon
manufacturing. Park and Reynolds (1994) have presented an economic model of the VSS control
chart, and Prabhu, Montgomery, and Runger (1997) have investigated economic-statistical design of
VSS/SI control charting schemes.
52
Ci (Si 1 Xi ) 1 (Si 1 Xi )
1/ 2
where
0, if Ci k
Si
(Si 1 Xi )(1 k / Ci ), if Ci k
with S0 = 0, and k > 0. An out of control signal is generated when
Yi (Si 1Si )1/ 2 H
where k and H the reference value and decision interval for the procedure, respectively.
Two different forms of the multivariate cusum were proposed by Pignatiello and Runger (1990).
Their best-performing control chart is based on the following vectors of cumulative sums:
Di
j i li 1
Xj
and
MCi max{0,(Di 1Di )1/ 2 kli }
53
54
Response variables come to mind easily for most experimenters. When there is a
choice, one should select continuous responses, because generally binary and ordinal
data carry much less information and continuous responses measured on a welldefined numerical scale are typically easier to analyze. On the other hand, there are
many situations where a count of defectives, a proportion, or even a subjective
ranking must be used as a response.
Measurement precision is an important aspect of selecting the response variables in
an experiment. Insuring that the measurement process is in a state of statistical
control is highly desirable. That is, ideally there is a well-established system of
insuring both accuracy and precision of the measurement methods to be used. The
amount of error in measurement imparted by the gauges used should be understood.
If the gauge error is large relative to the change in the response variable that is
important to detect, then the experimenter will want to know this before conducting
the experiment. Sometimes repeat measurements can be made on each experimental
unit or test specimen to reduce the impact of measurement error. For example, when
measuring the number average molecular weight of a polymer with a gel permeation
chromatograph (GPC) each sample can be tested several times and the average of
those molecular weight reading reported as the observation for that sample. When
measurement precision is unacceptable, a measurement systems capability study may
be performed to attempt to improve the system. These studies are often fairly
complicated designed experiments. Chapter 8 presents an example of a factorial
experiment used to study the capability of a measurement system.
The impeller involved in this experiment is shown in Figure S13.1. Table S13.3 lists
the information about the response variables. Notice that there are three response
variables of interest here.
As with response variables, most experimenters can easily generate a list of candidate
design factors to be studied in the experiment. Coleman and Montgomery call these
control variables. We often call them controllable variables, design factors, or
process variables in the text. Control variables can be continuous or categorical
(discrete). The ability of the experimenters to measure and set these factors is
important. Generally, small errors in the ability to set, hold or measure the levels of
control variables are of relatively little consequence. Sometimes when the
measurement or setting error is large, a numerical control variable such as
temperature will have to be treated as a categorical control variable (low or high
temperature). Alternatively, there are errors-in-variables statistical models that can
be employed, although their use is beyond the scope of this book. Information about
the control variables for the CNC-machining example is shown in Table S13.4.
Held-constant factors are control variables whose effects are not of interest in this
experiment. The worksheets can force meaningful discussion about which factors are
adequately controlled, and if any potentially important factors (for purposes of the
present experiment) have inadvertently been held constant when they should have
55
56
Results of this experiment will be used to determine machine set-up parameters for impeller
machining. A robust process is desirable; that is, on-target and low variability performance
regardless of which tool vendor is used.
57
Figure S13.1. Jet engine impeller (side view). The z-axis is vertical, x-axis is horizontal, y-axis is
into the page. 1 = height of wheel, 2 = diameter of wheel, 3 = inducer blade height, 4 = exducer blade
height, 5 = z height of blade.
Table S13.3. Response Variables
Response variable
Normal operating
Measurement
Relationship of
(units)
precision, accuracy
response variable to
how known?
objective
Blade profile
(inches)
Nominal (target)
E@ 1 X 10 -5 inches
1 X 10-3 inches to
from a coordinate
2 X 10-3 inches at
measurement
all points
machine capability
Estimate mean
absolute difference
from target and
standard deviation
study
Surface finish
Surface defect
count
Smooth to rough
Visual criterion
Should be as smooth
(requiring hand
(compare to
finish)
standards)
Typically 0 to 10
Visual criterion
Must not be
(compare to
excessive in
standards)
number or
as possible
magnitude
58
Proposed settings,
Predicted effects
Control variable
Normal level
setting error-
based on
(for various
(units)
and range
how known?
predicted effects
responses)
x-axis shift*
0-.020 inches
(inches)
y-axis shift*
0-.020 inches
Difference
.001 inches
0, .015 inches
Difference
Difference
(experience)
0-.020 inches
(inches)
Tool vendor
0, .015 inches
(experience)
(inches)
z-axis shift*
.001 inches
.001 inches
(experience)
Internal, external
a-axis shift*
0-.030 degrees
(degrees)
Spindle speed
.001 degrees
0, .030 degrees
Unknown
90%,110%
None?
(guess)
85-115%
1%
(% of
(indicator
nominal)
on control
panel)
Fixture height
0-.025 inches
.002inches
0, .015 inches
Unknown
90%,110%
None?
(guess)
Feed rate (% of 90-110%
nominal)
1%
(indicator
on control
panel)
'The x, y, and z axes are used to refer to the part and the CNC machine. The a axis refers only to the machine.
In the CNC experiment, this worksheet helped the experimenters recognize that the machine had to be fully
warmed up before cutting any blade forgings. The actual procedure used was to mount the forged blanks on
the machine and run a 30-minute cycle without the cutting tool engaged. This allowed all machine parts and
the lubricant to reach normal, steady-state operating temperature. The use of a typical (i.e., mid-level) operator
and the use of one lot of forgings ware decisions made for experimental insurance. Table S13.5 shows the
held-constant factors for the CNC-machining experiment.
59
Measurement
precision-how
known?
How to control
(in experiment)
Anticipated
effects
Type of cutting
fluid
Standard type
None
Temperature of
cutting fluid
(degrees F.)
100F. when
machine is
warmed up
1-2 F. (estimate)
Do runs when
machine has
reached 100F
None
Operator
Several operators
normally work
in the process
Titanium
Forgings
Material
properties may
vary from unit
to unit
None
Slight
Nuisance factors are variables that probably have some effect on the response, but which are of little or no
interest to the experimenter. They differ from held-constant factors in that they either cannot be held entirely
constant, or they cannot be controlled at all. For example, if two lots of forgings were required to run the
experiment, then the potential lot-to-lot differences in the material would be a nuisance variable than could not
be held entirely constant. In a chemical process we often cannot control the viscosity (say) of the incoming
material feed streamit may vary almost continuously over time. In these cases, nuisance variables must be
considered in either the design or the analysis of the experiment. If a nuisance variable can be controlled, then
we can use a design technique called blocking to eliminate its effect. If the nuisance variable cannot be
controlled but it can be measured, then we can reduce its effect by an analysis technique called the analysis of
covariance. Montgomery (2005) gives an introduction to the analysis of covariance.
Table S13.6 shows the nuisance variables identified in the CNC-machining experiment. In this experiment,
the only nuisance factor thought to have potentially serious effects was the machine spindle. The machine has
four spindles, and ultimately a decision was made to run the experiment in four blocks. The other factors were
held constant at levels below which problems might be encountered.
60
Strategy (e.g.,
Nuisance factor
precision-how
randomization,
(units)
known?
blocking, etc.)
Viscosity of
Standard viscosity
cutting fluid
Anticipated effects
Measure viscosity at
None to slight
Ambient
1-2 F. by room
Temperature,F.
thermometer
80F.
hot weather
Spindle
Block or randomize
Spindle-to-spindle
on machine spindle
variation could be
large
Vibration of
machine during
objects in CNC
introduce variation
operation
machine shop
within an impeller
Coleman and Montgomery also found it useful to introduce an interaction sheet. The concept of interactions
among process variables is not an intuitive one, even to well-trained engineers and scientists. Now it is clearly
unrealistic to think that the experimenters can identify all of the important interactions at the outset of the
planning process. In most situations, the experimenters really dont know which main effects are likely to be
important, so asking them to make decisions about interactions is impractical. However, sometimes the
statistically-trained team members can use this as an opportunity to teach others about the interaction
phenomena. When more is known about the process, it might be possible to use the worksheet to motivate
questions such as are there certain interactions that must be estimated? Table S13.7 shows the results of this
exercise for the CNC-machining example.
Table S13.7. Interactions
Control
variable
x shift
y shift
z shift
Vendor
P
a shift
Speed
y shift
z shift
Vendor
a shift
Speed
Height
Height
F,D
-
NOTE: Response variables are P = profile difference, F = surface finish and D = surface defects
61
Feed
Two final points: First, an experimenter without a coordinator will probably fail. Furthermore, if something
can go wrong, it probably will, so he coordinator will actually have a significant responsibility on checking to
ensure that the experiment is being conducted as planned. Second, concerning trial runs, this is often a very
good ideaparticularly if this is the first in a series of experiments, or if the experiment has high significance
or impact. A trial run can consist of a center point in a factorial or a small part of the experimentperhaps
one of the blocks. Since many experiments often involve people and machines doing something they have not
done before, practice is a good idea. Another reason for trial runs is that we can use them to get an estimate
of the magnitude of experimental error. If the experimental error is much larger than anticipated, then this
may indicate the need for redesigning a significant part of the experiment. Trial runs are also a good
opportunity to ensure that measurement and data-acquisition or collection systems are operating as anticipated.
Most experimenters never regret performing trial runs.
response
variable
(units)
normal
operating level
& range
meas. precision,
accuracy
How known?
relationship of
response variable
to
objective
Control Variables
control
variable
(units)
62
normal level
& range
meas.
precision
& setting error
How known?
proposed
settings,
based on
predicted
effects
predicted
effects
(for various
responses)
factor
(units)
desired
experimental
level &
allowable range
measurement
precision
How known?
how to
control (in
experiment)
anticipated
effects
Nuisance Factors
measurement
precision
How known?
nuisance
factor (units)
strategy (e.g.,
randomization,
blocking, etc.)
anticipated effects
Interactions
control var.
1
63
kk 0
kk 0
64
E ( yF )
1
nF
(nF 0 nF 11 nF 22
0 11 22
nF kk )
kk
and
E ( yC )
1
nC
(nC 0 )
Therefore,
E ( yF yC ) 11 22 kk
and so we see that the difference in averages yF yC is an unbiased estimator of the sum of the pure
quadratic model parameters. Now the variance of yF yC is
1
1
V ( yF yC ) 2
nF nC
Consequently, a test of the above hypotheses can be conducted using the statistic
t0
yF yC
1
1
2
nF nC
which under the null hypothesis follows a t distribution with nC 1 degrees of freedom. We would
reject the null hypothesis (that is, no pure quadratic curvature) if | t0 | t / 2,nC 1 .
This t-test is equivalent to the F-test given in the book. To see this, square the t-statistic above:
( yF yC ) 2
1
1
2
nF nC
n n ( y yC ) 2
F C F
(nF nC ) 2
t02
This ratio is identical to the F-test presented in the textbook. Furthermore, we know that the square
of a t random variable with (say) v degrees of freedom is an F random variable with 1 numerator and
v denominator degrees of freedom, so the t-test for pure quadratic effects is indeed equivalent to
the F-test.
A simple method can be used to keep the variability associated with a nuisance variable from impacting
experimental results. One each day (or in general, at each possible level) of the nuisance variable, test all
treatment or factor levels of interest. In our example, this would consist of testing all four hardwood
concentrations that are of interest on a single day. On each day, the four tests are conducted in random order.
This type of experimental design is called a randomized complete block design or RCBD. In the RCBD,
the block size must be large enough to hold all the treatments. If this condition is not satisfied, then an
incomplete block design must be used. These incomplete block designs are discussed in some experimental
design textbooks; for example, see Montgomery (2005).
In general, for a RCBD with a treatments we will run a complete replicate of these treatments in each of b
blocks. The order in which the runs are made in each block is completely random. The statistical model for
the RCBD is
i 1, 2,..., a
yij i j ij
j 1, 2,..., b
where is an overall mean,
random error term, taken to be NID(0, 2 ). We will think of the treatments as fixed factors. Defining
the treatment and block effects as deviations from an overall mean leads to the test on equality of the
treatment
means
being
equivalent
to
a
test
of
the
hypotheses
that
The analysis of variance (ANOVA) can be adapted to the analysis of the RCBD. The fundamental
ANOVA equality becomes
a
i 1
j 1
i 1 j 1
or
The null hypothesis of no difference in factor level or treatment means is tested by the statistic
SS Factor /(a 1)
MS Factor
SS E /(a 1)(b 1)
MS E
To illustrate the procedure, reconsider the tensile strength experiment data in Table 4-6, and suppose
that the experiment was run as a RCBD. Now the columns of Table 4-6 (currently labeled
Observations) would be labeled blocks or days. Minitab will perform the RCBD analysis. The
output follows.
ANOVA: strength versus concentration, day (or blocks)
Factor
conc
fixed
10
15
20
day
fixed
66
DF
SS
MS
conc
382.792
127.597
25.03
0.000
day
53.708
10.742
2.11
0.121
Error
15
76.458
5.097
Total
23
512.958
Notice that the concentration factor is significant; that is, there is evidence that changing the
hardwood concentration affects the mean strength. The F-ratio for blocks or days is small, suggesting
that the variability associated with the blocks was small. There are some technical problems
associated with statistical testing of block effects see the discussion in Montgomery (2005, Chapter
5).
The blocking principal can be extended to experiments with more complex treatment structures. For
example, in Section 13-5.5, we observe that in a replicated factorial experiment, each replicate can
be run in a single block. Thus a nuisance factor can be accommodated in a factorial experiment.
As an illustration, consider the router experiment (Example 13-6 in the text). Suppose that each of
the replicates were run on a single printed circuit board. Considering boards (or replicates) as blocks,
we can analyze this experiment as a 22 factorial in four blocks. The Minitab analysis follows. Notice
that both main effects and the interaction are important. There is also some indication that the block
effects are significant.
Effect
Coef
SE Coef
23.831
0.4359
54.67
0.000
1.744
0.7550
2.31
0.046
1.494
0.7550
1.98
0.079
-2.156
0.7550
-2.86
0.019
Constant
Block
16.638
8.319
0.4359
19.08
0.000
7.538
3.769
0.4359
8.65
0.000
A*B
8.712
4.356
0.4359
9.99
0.000
DF
Seq SS
Adj SS
Adj MS
Blocks
44.36
44.36
14.787
4.86
0.028
Main Effects
1334.48
1334.48
667.241 219.48
0.000
2-Way Interactions
303.63
303.63
303.631
0.000
Residual Error
27.36
27.36
3.040
15
1709.83
Total
67
99.88
R| ij 11,,22,,,,ab
S|
Tk 1,2,, n
given as Equation (13.2) in the textbook. It is relatively easy to develop the expected
mean squares from direct application of the expectation operator.
For an illustration, consider finding the expected value for one of the main effect
mean squares, say
E ( MS A ) E
FG SS IJ 1 E (SS )
H a 1K a 1
A
where SSA is the sum of squares for the row factor. Since
SS A
1 a 2
y...2
y
i.. abn
bn i 1
E ( SS A )
FG IJ
H K
a
1
y2
E yi2.. E ...
bn i 1
abn
Recall that . 0, . 0,( ). j 0,( )i . 0, and ( ).. 0 , where the dot subscript
implies summation over that subscript. Now
b
yi .. yijk bn bn i n . n( )i . i ..
j 1 k 1
bn bn i i ..
and
a
a
1
1
E yi2..
E (bn ) 2 (bn) 2 i2 i2.. 2(bn) 2 i 2bn i .. 2bn i i ..
bn i 1
bn i 1
LM
N
a
1
a (bn ) 2 (bn) 2 i2 abn 2
bn
i 1
a
abn 2 bn i2 a 2
i 1
so
68
OP
Q
1
1
E ( y...2 )
E (abn ... ) 2
abn
abn
1
(abn ) 2 abn 2
abn
abn 2 2
Therefore
E ( MS A ) E
FG SS IJ
H a 1K
A
1
E ( SS A )
a 1
a
1
abn 2 bn i2 a 2 (abn 2 2 )
a 1
i 1
LM
N
LM
N
a
1
2 (a 1) bn i2
a 1
i 1
OP
Q
OP
Q
bn i2
i 1
a 1
The other expected mean squares are derived similarly. In general, for the fixed
effects model, the expected value of the mean squares for main effects and interaction
are equal to the error variance 2 plus a term involving the corresponding fixed
effect. The fixed effect term will be zero if the treatment means are zero or if the
interaction effects are negligible. The expected value of the error mean square is 2 ,
so the ratio of the model term mean square to the error mean square results in a onesided upper-tail test. The use of the F-distribution as the reference distribution
follows from the normality assumption on the response variable.
The Random Effects Model
In Section 8.6.2 we discuss briefly the use of analysis of variance methods for
measurement systems capability studies. The two-factor factorial random effects
model is assumed to be appropriate for the problem. The model is
yij i j ( ) ij ijk
R| ij 11,,22,,,,ab
S|
Tk 1,2,, n
as given as Equation (8.23) in the textbook. We list the expected mean squares for
this .model in Equation (8.26), but do not formally develop them. It is relatively easy
to develop the expected mean squares from direct application of the expectation
operator.
For example, consider finding
69
E ( MS A ) E
FG SS IJ 1 E (SS )
H a 1K a 1
A
where SSA is the sum of squares for the row factor. Recall that the model components
i , j and ( )ij are normally and independently distributed with means zero and
variances 2 , 2 , and 2 respectively. The sum of squares and its expectation are
defined as
SS A
1 a 2
y2
yi .. ...
bn i 1
abn
E ( SS A )
FG IJ
H K
a
1
y2
E yi2.. E ...
bn i 1
abn
Now
b
yi .. yijk bn bn i n . n( )i . i ..
j 1 k 1
and
a
a
1
1
E yi2..
E (bn ) 2 (bn) 2 i2 i2.. 2(bn) 2 i 2bn i .. 2bn i i ..
bn i 1
bn i 1
1
2
We can now collect the components of the expected value of the sum of squares for
factor A and find the expected mean square as follows:
FG SS IJ
H a 1K
1 L1
M E y
a 1 N bn
E ( MS A ) E
2
i ..
i 1
F y IJ OP
EG
H abn K Q
2
...
1
2
2 (a 1) n(a 1)
bn 2
a 1
2
2 n
bn 2
70
There are situations where there are only a specific set of operators that perform the
measurements in a gage R & R study, so we cannot think of the operators as having
been selected at random from a large population. Thus an ANOVA model involving
parts chosen at random and fixed operator effects would be appropriate. This is a
mixed model ANOVA. For details of using mixed models in measurement systems
capability studies, see Montgomery (2005), Burdick, Borror, and Montgomery
(2003), and Dolezal, Burdick, and Birch (1998).
71
nC 1 center points.
This
results in a design with at least 25 runs, while the second-order model in k = 4 design factors has only 15
parameters. Obviously, there are situations where it would be desirable to reduce the number of required runs.
One approach is to use a fractional factorial in the cube. However, the fractional must be either of resolution
V, or it must be of resolution III* (main effects aliased with two-factor interactions, but no two-factor
interactions aliased with each other). A resolution IV design cannot be used because that results in two-factor
interactions aliased with each other, so the cross-product terms in the second-order model cannot be estimated.
A small composite design is a CCD with a resolution III* fraction in the cube. For the in k = 4 design factor
example, this would involve setting the generator D = AB for the one-half fraction (the standard resolution IV
half-fraction uses D = ABC). This results in a small composite design with a minimum of 17 runs. Hybrid
designs are another type of small response surface design that are in many ways superior to the small
composite. These and other types of response surface designs are supported by several statistics software
packages. These designs are discussed extensively in Myers and Montgomery (2002).
1.
2.
3.
Taguchi called this the robust parameter design problem. In Chapter 14 we extend the idea somewhat to
include not only robust product design but process robustness studies.
Taguchi defined meaningful engineering problems and the philosophy that he recommended is sound.
However, he advocated some novel methods of statistical data analysis and some approaches to the design of
experiments that the process of peer review revealed were unnecessarily complicated, inefficient, and
sometimes ineffective. In this section, we will briefly overview Taguchi's philosophy regarding quality
engineering and experimental design. We will present some examples of his approach to robust parameter
design, and we will use these examples to highlight the problems with his technical methods. It is possible to
combine his sound engineering concepts with more efficient and effective experimental design and analysis
based on response surface methods, as we did in the process robustness studies examples in Chapter 14.
73
A key component of Taguchi's philosophy is the reduction of variability. Generally, each product or process
performance characteristic will have a target or nominal value. The objective is to reduce the variability
around this target value. Taguchi models the departures that may occur from this target value with a loss
function. The loss refers to the cost that is incurred by society when the consumer uses a product whose
quality characteristics differ from the nominal. The concept of societal loss is a departure from traditional
thinking. Taguchi imposes a quadratic loss function of the form
L(y) = k (y - T)2
shown in Figure S14.1 below. Clearly this type of function will penalize even small departures of y from the
target T. Again, this is a departure from traditional thinking, which usually attaches penalties only to cases
where y is outside of the upper and lower specifications (say y > USL or y < LSL in Figure S14.1. However,
the Taguchi philosophy regarding reduction of variability and the emphasis on minimizing costs is entirely
consistent with the continuous improvement philosophy of Deming and Juran.
In summary, Taguchi's philosophy involves three central ideas:
1.
Products and processes should be designed so that they are robust to external sources of variability.
2.
Experimental design methods are an engineering tool to help accomplish this objective.
These are sound concepts, and their value should be readily apparent. Furthermore, as we have seen in the
textbook, experimental design methods can play a major role in translating these ideas into practice.
We now turn to a discussion of the specific methods that Taguchi recommends for applying his concepts in
practice. As we will see, his approach to experimental design and data analysis can be improved.
Taguchis Technical Methods
An Example
We will use a connector pull-off force example to illustrate Taguchis technical methods. For more
information about the problem, refer to the original article in Quality Progress in December 1987 (see "The
Taguchi Approach to Parameter Design," by D. M. Byrne and S. Taguchi, Quality Progress, December 1987,
pp. 19-26). The experiment involves finding a method to assemble an elastomeric connector to a nylon tube
that would deliver the required pull-off performance to be suitable for use in an automotive engine application.
The specific objective of the experiment is to maximize the pull-off force. Four controllable and three
uncontrollable noise factors were identified. These factors are shown in Table S14.1 below. We want to find
the levels of the controllable factors that are the least influenced by the noise factors and that provides the
maximum pull-off force. Notice that although the noise factors are not controllable during routine operations,
they can be controlled for the purposes of a test. Each controllable factor is tested at three levels, and each
noise factor is tested at two levels.
74
In the Taguchi parameter design methodology, one experimental design is selected for the controllable factors
and another experimental design is selected for the noise factors. These designs are shown in Table S14.2.
Taguchi refers to these designs as orthogonal arrays, and represents the factor levels with integers 1, 2, and
3. In this case the designs selected are just a standard 23 and a 34-2 fractional factorial. Taguchi calls these the
L8 and L9 orthogonal arrays, respectively.
The two designs are combined as shown in Table S14.3 below. This is called a crossed or product array
design, composed of the inner array containing the controllable factors, and the outer array containing the
noise factors. Literally, each of the 9 runs from the inner array is tested across the 8 runs from the outer array,
for a total sample size of 72 runs. The observed pull-off force is reported in Table S14.3.
Table S14.1. Factors and Levels for the Taguchi Parameter Design Example
Controllable Factors
Levels
A=
Interference
Low
Medium
High
B=
Thin
Medium
Thick
C=
Insertion,depth
Shallow
Medium
Deep
D=
Percent adhesive in
Low
Medium
High
connector pre-dip
Uncontrollable Factors
Levels
E = Conditioning time
24 h
120 h
F = Conditioning temperature
72F
150F
25%
75%
75
E
1
1
2
2
1
1
2
2
F ExF
1
1
1
2
2
1
2
2
2
1
2
2
1
1
1
2
Table S14.3.
___________________________________________________________________________________
Outer Array (L8)
E
Responses
.
Run
SNL
15.6
9.5
16.9
19.9
19.6
19.6
20.0
19.1 17.525
24.025
15.0
16.2
19.4
19.2
19.7
19.8
24.2
21.9 19.475
25.522
16.3
16.7
19.1
15.6
22.6
18.2
23.3
20.4 19.025
25.335
18.3
17.4
18.9
18.6
21.0
18.9
23.2
24.7 20.125
25.904
19.7
18.6
19.4
25.1
25.6
21.4
27.5
25.3 22.825
26.908
16.2
16.3
20.0
19.8
14.7
19.6
22.5
24.7 19.225
25.326
16.4
19.1
18.4
23.6
16.8
18.6
24.3
21.6 19.8
25.711
14.2
15.6
15.1
16.8
17.8
19.6
23.2
24.2 18.338
24.852
16.1
19.9
19.3
17.3
23.1
22.7
22.6
28.6 21.200
26.152
__________________________________________________________________________________
y2
SN T 10 log 2
S
2. Larger the better:
1 n 1
SN L 10 log 2
n i 1 yi
3. Smaller the better:
1 n 2
SN L 10 log yi
n i 1
Notice that these SN ratios are expressed on a decibel scale. We would use SNT if the objective is to reduce
variability around a specific target, SNL if the system is optimized when the response is as large as possible,
76
and SNS if the system is optimized when the response is as small as possible. Factor levels that maximize the
appropriate SN ratio are optimal.
In this problem, we would use SNL because the objective is to maximize the pull-off force. The last two
columns of Table S14.3 contain
and SNL values for each of the nine inner-array runs. Taguchi-oriented
practitioners often use the analysis of variance to determine the factors that influence y and the factors that
influence the signal-to-noise ratio. They also employ graphs of the "marginal means" of each factor, such as
the ones shown in Figures S14.2 and S14.3. The usual approach is to examine the graphs and "pick the winner."
In this case, factors A and C have larger effects than do B and D. In terms of maximizing SNL we would select
AMedium, CDeep, BMedium, and DLow. In terms of maximizing the average pull-off force y , we would choose
AMedium, CMedium, BMedium and DLow. Notice that there is almost no difference between CMedium and CDeep. The
implication is that this choice of levels will maximize the mean pull-off force and reduce variability in the pulloff force.
Taguchi advocates claim that the use of the SN ratio generally eliminates the need for examining specific
interactions between the controllable and noise factors, although sometimes looking at these interactions
improves process understanding. The authors of this study found that the AG and DE interactions were large.
Analysis of these interactions, shown in Figure S14.4, suggests that AMedium is best. (It gives the highest pulloff force and a slope close to zero, indicating that if we choose AMedium the effect of relative humidity is
minimized.) The analysis also suggests that DLow gives the highest pull-off force regardless of the conditioning
time.
Figure S14.3. The Effects of Controllable Factors on the Signal to Noise Ratio
77
When cost and other factors were taken into account, the experimenters in this example finally decided to use
AMedium, BThin, CMedium, and Dlow. (BThin was much less expensive than BMedium, and CMedium was felt to give
slightly less variability than CDeep.) Since this combination was not a run in the original nine inner array trials,
five additional tests were made at this set of conditions as a confirmation experiment. For this confirmation
experiment, the levels used on the noise variables were ELow, FLow, and GLow. The authors report that good
results were obtained from the confirmation test.
2 7III 4
fractional
1511
factorial, the L9 is a 3III fractional factorial, the L12 is a Plackett-Burman design, the L16 is a 2 III fractional
factorial, and so on. Box, Bisgaard, and Fung (1988) trace the origin of these designs. Some of these designs
have very complex alias structures. In particular, the L12 and all of
the designs that use three-level factors will involve partial aliasing of two-factor interactions with main
effects. If any two-factor interactions are large, this may lead to a situation in which the experimenter does
not get the correct answer. For more details on aliasing in these types of designs, see Montgomery (2005).
Taguchi argues that we do not need to consider two-factor interactions explicitly. He claims that it is possible
to eliminate these interactions either by correctly specifying the response and design factors or by using a
sliding setting approach to choose factor levels. As an example of the latter approach, consider the two
factors pressure and temperature. Varying these factors independently will probably produce an interaction.
However, if temperature levels are chosen contingent on the pressure levels, then the interaction effect can be
minimized. In practice, these two approaches are usually difficult to implement unless we have an unusually
high level of process knowledge. The lack of provision for adequately dealing with potential interactions
between the controllable process factors is a major weakness of the Taguchi approach to parameter design.
Instead of designing the experiment to investigate potential interactions, Taguchi prefers to use three-level
factors to estimate curvature. For example, in the inner and outer array design used by Byrne and Taguchi, all
four controllable factors were run at three levels. Let x1, x2, x3 and x4 represent the controllable factors and let
78
z1, z2, and z3 represent the three noise factors. Recall that the noise factors were run at two levels in a complete
factorial design. The design they used allows us to fit the following model:
y 0 j x j jj x
j 1
j 1
j z j
j 1
i j
j 2
z z j ij zi x j
ij i
i 1 j 1
Notice that we can fit the linear and quadratic effects of the controllable factors but not their two-factor
interactions (which are aliased with the main effects). We can also fit the linear effects of the noise factors
and all the two-factor interactions involving the noise factors. Finally, we can fit the two-factor interactions
involving the controllable factors and the noise factors. It may be unwise to ignore potential interactions in
the controllable factors.
This is a rather odd strategy, since interaction is a form of curvature. A much safer strategy is to identify
potential effects and interactions that may be important and then consider curvature only in the important
variables if there is evidence that the curvature is important. This will usually lead to fewer experiments,
simpler interpretation of the data, and better overall process understanding.
Another criticism of the Taguchi approach to parameter design is that the crossed array structure usually leads
to a very large experiment. For example, in the foregoing application, the authors used 72 tests to investigate
only seven factors, and they still could not estimate any of the two-factor interactions among the four
controllable factors.
There are several alternative experimental designs that would be superior to the inner and outer method used
in this example. Suppose that we run all seven factors at two levels in the combined array design approach
72
discussed in the textbook. Consider the 2 IV fractional factorial design. The alias relationships for this design
are shown in the top half of Table S14.4. Notice that this design requires only 32 runs (as compared to 72).
In the bottom half of Table S14.4, two different possible schemes for assigning process controllable variables
and noise variables to the letters A through G are given. The first assignment scheme allows all the interactions
between controllable factors and noise factors to be estimated, and it allows main effect estimates to be made
that are clear of two-factor interactions. The second assignment scheme allows all the controllable factor main
effects and their two-factor interactions to be estimated; it allows all noise factor main effects to be estimated
clear of two-factor interactions; and it aliases only three interactions between controllable factors and noise
factors with a two-factor interaction between two noise factors. Both of these arrangements present much
cleaner alias relationships than are obtained from the inner and outer array parameter design, which also
required over twice as many runs.
In general, the crossed array approach is often unnecessary. A better strategy is to use the combined array
design discussed in the textbook. This approach will almost always lead to a dramatic reduction in the size of
the experiment, and at the same time, it will produce information that is more likely to improve process
understanding. For more discussion of this approach, see Myers and Montgomery (2002). We can also use a
combined array design that allows the experimenter to directly model the noise factors as a complete quadratic
and to fit all interactions between the controllable factors and the noise factors, as demonstrated in Chapter 14
of the textbook.
Another possible issue with the Taguchi inner and outer array design relates to the order in which the runs are
performed. Now we know that for experimental validity, the runs in a designed experiment should be
conducted in random order. However, in many crossed array experiments, it is possible that the run order
wasnt randomized. In some cases it would be more convenient to fix each row in the inner array (that is, set
the levels of the controllable factors) and run all outer-array trials. In other cases, it might be more convenient
to fix the each column in the outer array and the run each on the inner array trials at that combination of noise
factors.
79
AF = BCD
CG = EF
AG = BDE
DE = ABG
C = EFG
BC = ADF
DF = ABC
BD = ACF = AEG
DG = ABE
E = CFG
BE = ADG
ACE = AFG
F = CEG
BF = ACD
ACG = AEF
G = CEF
BG = ADE
BCE = BFG
AB = CDF = DEG
CD = ABF
BCG = BEF
AC = BDF
CE = FG
CDE = DFG
AD = BCF = BEG
CF = ABD = EG
CDG = DEF
AF = BDG
Controllable factors are assigned to the letters C, E, F, and G. Noise factors are assigned to the letters A, B, and D. All interactions between
controllable factors and noise factors can be estimated. and all controllable factor main effects can be estimated clear of two-factor interactions.
2.
Controllable factors are assigned to the letters A, B, C, and D. Noise factors are assigned to the letters E, F. and G. All controllable factor main
effects and two-factor interactions can be estimated; only the CE, CF, and CG interactions are aliased with interactions of the noise factors.
Exactly which strategy is pursued probably depends on which group of factors is easiest to change, the
controllable factors or the noise factors. If the tests are run in either manner described above, then a split-plot
structure has been introduced into the experiment. If this is not accounted for in the analysis, then the results
and conclusions can be misleading. There is no evidence that Taguchi advocates used split-plot analysis
methods. Furthermore, since Taguchi frequently downplayed the importance of randomization, it is highly
likely that many actual inner and outer array experiments were inadvertently conducted as split-plots, and
perhaps incorrectly analyzed. Montgomery (2005) discusses split-plot designs and their analysis. Box and
Jones give a good discussion of split-plot designs in process robustness studies.
A final aspect of Taguchi's parameter design is the use of linear graphs to assign factors to the columns of the
orthogonal array. A set of linear graphs for the L8 design is shown in Figure S14.5. In these graphs, each
number represents a column in the design. A line segment on the graph corresponds to an interaction between
the nodes it connects. To assign variables to columns in an orthogonal array, assign the variables to nodes
first; then when the nodes are used up, assign the variables to the line segments. When you assign variables
to the nodes, strike out any line segments that correspond to interactions that might be important. The linear
graphs in Figure 5 imply that column 3 in the L8 design contains the interaction between columns 1 and 2,
column 5 contains the interaction between columns 1 and 4, and so forth. If we had four factors, we would
assign them to columns 1, 2, 4, and 7. This would ensure that each main effect is clear of two-factor
interactions. What is not clear is the two-factor interaction aliasing. If the main effects are in columns 1, 2, 4,
and 7, then column 3 contains the 1-2 and the 4-7 interaction, column 5 contains the 1-4 and the 2-7 interaction,
and column 6 contains the 1-7 and the 2-4 interaction. This is clearly the case because four variables in eight
runs is a resolution IV plan with all pairs of two-factor interactions aliased. In order to understand fully the
two-factor interaction aliasing, Taguchi would refer the experiment designer to a supplementary interaction
table.
80
Taguchi (1986) gives a collection of linear graphs for each of his recommended orthogonal array
designs. These linear graphs seem -to have been developed heuristically. Unfortunately, their use
can lead to inefficient designs. For examples, see his car engine experiment [Taguchi and Wu (1980)]
and his cutting tool experiment [Taguchi (1986)]. Both of these are 16-run designs that he sets up as
resolution III designs in which main effects are aliased with two-factor interactions. Conventional
methods for constructing these designs would have resulted in resolution IV plans in which the main
effects are clear of the two-factor interactions. For the experimenter who simply wants to generate a
good design, the linear graph approach may not produce the best result. A better approach is to use
a table that presents the design and its full alias structure. These tables are easy to construct and are
routinely displayed by several widely available and inexpensive computer programs.
Table S14.5. Data for the "Marginal Means" Plots in Figure S12-6
Factor A
Factor B
10
10
13
11.00
10
14
9.67
10
8.33
8.00
9.67
11.67
A Averages
B Averages
Figure S14.6. Marginal Means Plots for the Data in Table S14.5
Figure S14.7. The AB Interaction Plot for the Data in Table S14.5.
Consider first the signal to noise ratio for the target is best case
y2
SN T 10 log 2
S
This ratio would be used if we wish to minimize variability around a fixed target value. It has been suggested
by Taguchi that it is preferable to work with SNT instead of the standard deviation because in many cases the
process mean and standard deviation are related. (As gets larger, gets larger, for example.) In such cases,
82
he argues that we cannot directly minimize the standard deviation and then bring the mean on target. Taguchi
claims he found empirically that the use of the SNT ratio coupled with a two-stage optimization procedure
would lead to a combination of factor levels where the standard deviation is minimized and the mean is on
target. The optimization procedure consists of (1) finding the set of controllable factors that affect SNT, called
the control factors, and setting them to levels that maximize SNT and then (2) finding the set of factors that
have significant effects on the mean but do not influence the SNT ratio, called the signal factors, and using
these factors to bring the mean on target.
Given that this partitioning of factors is possible, SNT is an example of a performance measure independent
of adjustment (PERMIA) [see Leon et al. (1987)]. The signal factors would be the adjustment factors. The
motivation behind the signal-to-noise ratio is to uncouple location and dispersion effects. It can be shown that
the use of SNT is equivalent to an analysis of the standard deviation of the logarithm of the original data. Thus,
using SNT implies that a log transformation will always uncouple location and dispersion effects. There is no
assurance that this will happen. A much safer approach is to investigate what type of transformation is
appropriate.
Note that we can write the SNT ratio as
y2
SN T 10 log 2
S
2
10 log( y ) 10 log( S 2 )
If the mean is fixed at a target value (estimated by y ), then maximizing the SNT ratio is equivalent to
minimizing log (S2). Using log (S2) would require fewer calculations, is more intuitively appealing, and would
provide a clearer understanding of the factor relationships that influence process variability - in other words,
it would provide better process understanding. Furthermore, if we minimize log (S2) directly, we eliminate the
risk of obtaining wrong answers from the maximization of SNT if some of the manipulated factors drive the
mean y upward instead of driving S2 downward. In general, if the response variable can be expressed in terms
of the model
y ( x d , x a ) ( x d )
where xd is the subset of factors that drive the dispersion effects and xa is the subset of adjustment factors that
do not affect variability, then maximizing SNT will be equivalent to minimizing the standard deviation.
Considering the other potential problems surrounding SNT , it is likely to be safer to work directly with the
standard deviation (or its logarithm) as a response variable, as suggested in the textbook. For more discussion,
refer to Myers and Montgomery (2002).
The ratios SNL and SNS are even more troublesome. These quantities may be completely ineffective in
identifying dispersion effects, although they may serve to identify location effects, that is, factors that drive
the mean. The reason for this is relatively easy to see. Consider the SNS (smaller-the-better) ratio:
1 n 2
SN S 10log yi
n i 1
The ratio is motivated by the assumption of a quadratic loss function with y nonnegative. The loss function
for such a case would be
1 n 2
L C yi
n i 1
83
1 n 2
log L log C log yi
n i 1
and
SNS = 10 log C - 10 log L
so maximizing SNS will minimize L. However, it is easy to show that
2
2
1 n 2
1 n 2
y
yi ny
n i 1
n i 1
2
n 1 2
y
S
n
Therefore, the use of SNS as a response variable confounds location and dispersion effects.
The confounding of location and dispersion effects was observed in the analysis of the SNL ratio in the pulloff force example. In Figures S14.2 and S14.3 notice that the plots of y and SNL versus each factor have
approximately the same shape, implying that both responses measure location. Furthermore, since the SNS
and SNL ratios involve y2 and 1/y2, they will be very sensitive to outliers or values near zero, and they are not
invariant to linear transformation of the original response. We strongly recommend that these signal-to-noise
ratios not be used.
A better approach for isolating location and dispersion effects is to develop separate response surface models
for y and log(S2). If no replication is available to estimate variability at each run in the design, methods for
analyzing residuals can be used. Another very effective approach is based on the use of the response model,
as demonstrated in the textbook and in Myers and Montgomery (2002). Recall that this allows both a response
surface for the variance and a response surface for the mean to be obtained for a single model containing both
the controllable design factors and the noise variables. Then standard response surface methods can be used
to optimize the mean and variance.
Finally, we turn to some of the applications of the analysis of variance recommended by Taguchi. As an
example for discussion, consider the experiment reported by Quinlan (1985) at a symposium on Taguchi
methods sponsored by the American Supplier Institute. The experiment concerned the quality improvement
of speedometer cables. Specifically, the objective was to reduce the shrinkage in the plastic casing material.
15-11
(Excessive shrinkage causes the cables to be noisy.) The experiment used an L16 orthogonal array (the 2 III
design). The shrinkage values for four samples taken from 3000-foot lengths of the product manufactured at
each set of test conditions were measured and the responses y and SNS computed.
Quinlan, following the Taguchi approach to data analysis, used SNS as the response variable in an analysis of
variance. The error mean square was formed by pooling the mean squares associated with the seven effects
that had the smallest absolute magnitude. This resulted in all eight remaining factors having significant effects
(in order of magnitude: E, G, K, A, C, F, D, H). The author did note that E and G were the most important.
Pooling of mean squares as in this example is a procedure that has long been known to produce
considerable bias in the ANOVA test results, To illustrate the problem, consider the 15 NID(0, 1)
random numbers shown in column 1 of Table S13.6. The square of each of these numbers, shown in
column 2 of the table, is a single-degree-of-freedom mean square corresponding to the observed
random number. The seven smallest random numbers are marked with an asterisk in column 1 of
84
Table S13.6. The corresponding mean squares are pooled to form a mean square for error with seven
degrees of freedom. This quantity is
MS E
0.5088
0.0727
7
Finally, column 3 of Table S13.6 presents the F ratio formed by dividing each of the eight remaining mean
squares by MSE. Now F0.05,1,7 = 5.59, and this implies that five of the eight effects would be judged significant
at the 0.05 level. Recall that since the original data came from a normal distribution with mean zero, none of
the effects is different from zero.
Analysis methods such as this virtually guarantee erroneous conclusions. The normal probability plotting of
effects avoids this invalid pooling of mean squares and provides a simple, easy to interpret method of analysis.
Box (1988) provides an alternate analysis of
Table S14.6. Pooling of Mean Squares
NID(0,1) Random
Numbers
F0
-08607
0.7408
10.19
-0.8820
0.7779
10.70
0.3608*
0.1302
0.0227*
0.0005
0.1903*
0.0362
-0.3071*
0.0943
1.2075
1.4581
20.06
0.5641
0.3182
4038
-0.3936*
0.1549
-0.6940
0.4816
-0.3028*
0.0917
0.5832
0.3401
0.0324*
0.0010
1.0202
1.0408
14.32
-0.6347
0.4028
5.54
6.63
4.68
the Quinlan data that correctly reveals E and G to be important along with other interesting results not apparent
in the original analysis.
It is important to note that the Taguchi analysis identified negligible factors as significant. This can have
profound impact on our use of experimental design to enhance process knowledge. Experimental design
methods should make gaining process knowledge easier, not harder.
On the other hand, while the Taguchi controversy was in full bloom, many companies reported success with
the use of Taguchi's parameter design methods. If the methods are flawed, why do they produce successful
results? Taguchi advocates often refute criticism with the remark that "they work." We must remember that
the "best guess" and "one-factor-at-a-time" methods will also work-and occasionally they produce good results.
This is no reason to claim that they are good methods. Most of the successful applications of Taguchi's
technical methods have been in industries where there was no history of good experimental design practice.
Designers and developers were using the best guess and one-factor-at-a-time methods (or other unstructured
approaches), and since the Taguchi approach is based on the factorial design concept, it often produced better
results than the methods it replaced. In other words, the factorial design is so powerful that, even when it is
used inefficiently, it will often work well.
As pointed out earlier, the Taguchi approach to parameter design often leads to large, comprehensive
experiments, often having 70 or more runs. Many of the successful applications of this approach were in
industries characterized by a high-volume, low-cost manufacturing environment. In such situations, large
designs may not be a real problem, if it is really no more difficult to make 72 runs than to make 16 or 32 runs.
On the other hand, in industries characterized by low-volume and/or high-cost manufacturing (such as the
aerospace industry, chemical and process industries, electronics and semiconductor manufacturing, and so
forth), these methodological inefficiencies can be significant.
A final point concerns the learning process. If the Taguchi approach to parameter design works and yields
good results, we may still not know what has caused the result because of the aliasing of critical interactions.
In other words, we may have solved a problem (a short-term success), but we may not have gained process
knowledge, which could be invaluable in future problems.
In summary, we should support Taguchi's philosophy of quality engineering. However, we must rely on
simpler, more efficient methods that are easier to learn and apply to carry this philosophy into practice. The
response surface modeling framework that we present in Chapter 14 of the textbook is an ideal approach to
process optimization and as we have demonstrated, it is fully adaptable to the robust parameter design problem
and to process robustness studies.
86
P( B) P( A) P( E2 ) P( A) P( E1 )
By defining the quantities
87
pe p(1 e2 ) (1 p)e1
The OC curve of a sampling plan plots the probability of lot acceptance versus the lot fraction
defective. When inspection error is present, the probability of lot acceptance is
c
n
Pa ( e) ped (1 pe )nd
d 0 d
Thus the OC curve of a sampling plan when inspection error is present can be easily obtained.
Now consider how inspection error impacts AOQ and ATI. For error-free inspection, the AOQ for
single sampling inspection for attributes is
AOQ
( N n) pPa
N
If defective items are replaced, and if the inspection of the replacements also involves inspection
error, then the AOQ becomes
AOQ
The ATI assuming error-free inspection and 100% screening of rejected lots is
ATI n (1 Pa )( N n)
The corresponding ATI when the discovered defectives are replaced and the replacement process is subject to
inspection error is
ATI
n (1 Pa (e) )( N n)
1 pe
ATI n (1 Pa (e) )( N n)
The design of a sampling plan is usually accomplished by choosing a set of parameters for the plan such that
the OC curve passes through two points, generally at the AQL and LTPD levels of quality. When there is
inspection error, the probabilities of acceptance no longer have the values that would have been obtained with
perfect inspection. However, it is possible to modify the design procedure by forcing the actual OC curve to
fit the desired points.
To illustrate this process, suppose that we want the OC curve to match the points (AQL, 1- ) and (LTPD,
). Then in designing the sampling plan we need to use
88
To illustrate, suppose that AQL 0.01,1 0.95, LTPD 0.06, 0.1 . The single sampling plan
without any considerations of inspection error has n 89, c 2 . Now suppose that there is inspection error
and e1 0.01, e2 0.15 . Then we can calculate
89
90
18. Davis, R. B., Homer, A., and Woodall, W. H. (1990), Performance of the Zone Control Chart,
Communications in Statistics: Theory and Methods, Vol. 19, pp. 1581-1587.
19. Dolezal, K. K., Burdick, R. K., and Birch, N. J. (1998), Analysis of a Two-Factor R & R Study
with Fixed Operators, Journal of Quality Technology, Vol. 30, pp. 163-170.
20. Duncan, A.J. (1955), The Use of the Range in Computing Variabilities, Industrial Quality
Control, Vol. 9, No. 5, pp. 18-22. (Clarification and Further Comment, Vol. 11, No. 8, p. 70;
Errata, Vol. 12, No. 2, p. 36).
21. Finison, L. J., Spencer, M., and Finison, K. S. (1993), Total Quality Measurement in Health
Care: Unsing Individuals Charts in Infection Control, pp. 349-359, in ASQC Quality Congress
Transactions, ASQC, Milwaukee, WI.
22. Grant, E. L. and Leavenworth, R. S. (1996), Statistical Quality Control, 7th Edition, McGrawHill, New York.
23. Hahn, G. J. (1977). Some Things Engineers Should Know About Experimental Design, Journal
of Quality Technology, Vol. 9, pp. 13-20.
24. Hahn, G. J. (1984). Experimental Design in a Complex World, Technometrics, Vol. 26, pp. 1931.
25. Hamada, M., and Weerahandi, S. (2000), Measurement Systems Assessment via Generalized
Inference, Journal of Quality Technology, Vol. 32, pp. 241-253.
26. Herath, H. S. B., Park, C. S., and Prueitt, G. C. (1995), Monitoring Projects using Cash Flow
Control Charts, The Engineering Economist, Vol. 41, pp. 27.
27. Hinkley, D. V. (1970), Inference About the Change-Point in a Sequence of Random Variables,
Biometrika, Vol. 57, pp. 1-17.
28. Hogg, R. V., and Craig, A. T. (1978), Introduction to Mathematical Statistics, 4th edition,
Macmillan, New York.
29. Howarth, R. J. (1995), Quality Control Charting for the Analytical Laboratory: Part 1. Univariate
Methods, Analyst, Vol. 120, pp. 1851-1873.
30. Hunter, W. G. (1977). Some Ideas About Teaching Design of Experiments With 25 Examples of
Experiments Conducted by Students, The American Statistician, Vol. 31, pp. 12-17.
31. Hurwicz, A., and Spagon P. D. (1997), Identifying Sources of Variation in a Wafer Planarization
Process, in V. Czitrom and P. D. Spagon (eds), Statistical Case Studies for Industrial Process
Improvement, pp. 105-144, Society for Industrial and Applied Mathematics, Philadelphia, PA.
32. Jaehn, A. H. (1987), Zone Control Charts SPC Made Easy, Quality, Vol. 26 (October), pp.
51-53.
33. Jin, C., and Davis, R. B. (1991), Calculation of Average Run lengths for Zone Control Charts
With Specified Zone Scores, Journal of Quality Technology, Vol. 23, pp. 355-358.
34. Klein, M. (2000), Two Alternatives to the Shewhart x Control Chart, Journal of Quality
Technology, Vol. 32, pp. 427-431.
35. Leon, R. V., A. C. Shoemaker and R. N. Kackar (1987). Performance Measures Independent of
Adjustment. Technometrics, Vol. 29, pp. 253-265
36. Levey, S., and Jennings, E. R. (1992), Historical Perspectives: The use of Control Charts in the
Clinical Laboratory, Archives of Pathology and Laboratory Medicine, Vol. 116, pp. 791-798.
37. Luko, S. N. (1996), Concerning the Estimators R / d 2 and R / d 2* in Estimating Variability in a
Normal Universe, Quality Engineering, Vol. 8, No. 3, pp. 481-487.
91
38. Montgomery, D. C., Peck, E. A., and Vining, G. G. (2006), Introduction to Linear Regression
Analysis, 4th edition, John Wiley & Sons, new York.
39. Myers, R. H., and Milton, J. S. (1991), A First Course in the Theory of Linear Statistical Models,
PWS-Kent, Boston, MA.
40. Nelson, L. S. (1975), Use of the Range to Estimate Variability, Journal of Quality Technology,
Vol. 7, pp. 46-48.
41. Ott, E. R. (1947), An Indirect Calibration of an Electronic Test Set, Industrial Quality Control,
Vol. 3, No. 4, pp. 11-14.
42. Park, C., and Reynolds, M. R., Jr. (1994), Economic Design of a Variable Sample Size X Chart
Communications in Statistics: Simulation and Computation, Vol. 23, pp. 467-483.
43. Patnaik, P. R. (1950), The Use of the Mean Range as an Estimator of Variance in Statistical
Tests, Biometrika, Vol. 37, pp. 78-87.
44. Prabhu, S. S., Montgomery, D. C., and Runger, G. C. (1997), Economic-Statistical Design of an
Adaptive X Chart, International Journal of Production Economics, Vol. 49, pp. 1-15.
45. Quinlan, J. (1985). Product Improvement by Application of Taguchi Methods. Third Supplier
Symposium on Taguchi Methods, American Supplier Institute, Inc., Dearborn, MI.
46. Reynolds, J. H. (1971), The Run Sum Control Chart Procedure, Journal of Quality Technology,
Vol. 3, pp. 23-27.
47. Roberts, S. W. (1966), A Comparison of Some Control Chart Procedures, Technometrics, Vol.
8, pp. 411-430.
48. Rodriguez, R. N. (1996), Health Care Applications of Statistical Process Control: Examples
using the SAS System, pp. 1381-1396, in Proceedings of the 21st SAS Users Group Conference,
SAS Users Group International, Cary, NC.
49. Roes, K.C.B., and Roes, R. J. M. M. (1995), Shewhart-Type Charts in Nonstandard Situations
(with discussion). Technometrics, Vol. 37, pp. 15-40
50. Runger, G. C., and Fowler, J. W. (1998). Run-to-Run Control Charts with Contrasts. Quality
and Reliability Engineering International, Vol. 14, pp. 262-272.
51. Runger, G. C., and Montgomery, D. C. (1993), Adaptive Sampling Enhancements for Shewhart
Control Charts, IIE Transactions, Vol. 25, pp. 41-51.
52. Ryan, T. P, and Schwertman, N. C. (1997), Optimal Limits for Attributes Control Charts,
Journal of Quality Technology, Vol. 29, pp. 86-98.
53. Samuel, T. R., Pignatiello, J. J., Jr., and Calvin, J. A. (1998), Identifying the Time of a Step
Change with x Control Charts, Quality Engineering, Vol. 10, pp. 521-527.
54. Schilling, E. G. (1978), A Lot Sensitive Sampling Plan for Compliance Testing and Acceptance
Inspection, Journal of Quality Technology, Vol. 10, pp. 47-51.
55. Schwertman, N. C., and Ryan, T. P (1997), Implementing Optimal Attributes Control Charts,
Journal of Quality Technology, Vol. 29, pp. 99-104.
56. Shore, H. (2000), General Control Charts for Attributes, IIE Transactions, Vol. 32, pp. 11491160.
57. Sullivan, J. H., and Woodall, W. H. (1996), A Comparison of Control Charts for Individual
Observations, Journal of Quality Technology, Vol. 28, pp. 398-408.
92
58. Tagaras, G. (1998), A Survey of Recent Developments in the Design of Adaptive Control
Charts, Journal of Quality Technology, Vol. 30, pp. 212-231.
59. Tippett, L. H. C. (1925), On the Extreme Individuals and the Range in Samples Taken from a
Normal Population, Biometrika, Vol. 17, pp. 364-387.
60. Wheeler, D. J. (1995), Advanced Topics in Statistical Process Control, SPC Press, Knoxville, TN.
61. Woodall, W. H., and Montgomery, D. C. (2000-01), Using Ranges to Estimate Variability,
Quality Engineering, Vol. 13, pp. 211-217.
62. Yashchin, E. (1994), Monitoring Variance Components Technometrics, Vol. 36, pp. 379-393.
63. Zimmer, L. S., Montgomery, D. C., and Runger, G. C. (1998), A Three-State Adaptive Sample
Size X Control Chart, International Journal of Production Research, Vol. 36, pp. 733-743.
64. Zimmer, L. S., Montgomery, D. C., and Runger, G. C. (2000), Guidelines for the Application of
Adaptive Control Charting Schemes, International Journal of Production Research, Vol. 38, pp.
1977-1992.
93