Chapter 2

Statistical Process Control

In this chapter, we will discuss the basics of Statistical Process Control

(SPC). Concepts that are used in subsequent chapters will be defined. In
Section 2.1, the meaning of the word “quality” is explained. In Section 2.2,
we will focus on the relation between quality and variation. Two impor-
tant concepts, namely, “common causes of variation” and “special causes of
variation”, will be discussed. Furthermore, we will argue that the philoso-
phy behind SPC is not only applicable to manufacturing environments, it is
useful in all areas of an organization. In Section 2.3, the Shewhart control
chart for independent observations is introduced. In subsequent chapters,
we want to study the effect of serial correlation on the behavior of the con-
trol chart. In Section 2.4, we review some of the time series models that
are used throughout this thesis.

2.1 Quality defined

The word “quality” is used by a lot of people in different contexts. In
most companies it is recognized that quality of a service or a product is
very important. However, it is very hard to give a perfect definition that
discriminates products or services of bad quality from products or services
of high quality. Various authors have attempted to formulate such a defi-
nition.
Shewhart (1931) argues that there are two aspects of quality. Firstly,
there is an objective concept of quality, resulting in quantatively measurable
physical characteristics, that are independent of a second, subjective, aspect
of quality. The latter has to do with what we think, feel or sense. Shewhart
recognizes that the subjective side of quality is commercially interesting,

13
14 CHAPTER 2. STATISTICAL PROCESS CONTROL

but that it is necessary to establish standards of quality in a quantitative

(objective) manner. Deming (1982) also stresses the subjective side of
quality:
Quality can be defined only in terms of the agent
(p. 168).
Furthermore, he emphasizes that impressions of quality are not static. They
change over time. This creates problems in defining quality, since it is
difficult to translate future needs of the user into measurable characteristics.
In a vivid case history starting on page 41, Crosby (1979) defines quality
as “conformance to requirements”. In this view, a product is of good quality
if it meets its specifications. This definition encloses an important part of
what customers perceive as “quality”. A manufacturer will most certainly
receive a lot of ‘quality complaints’ if a large part of his production exceeds
tolerances that were agreed upon with the customers. On the other hand,
the definition is too narrow. For example, the specifications themselves
are part of what is perceived as “quality”. A product that is produced in
conformance with specifications that are not popular with customers will
not be ranked as a high quality product.
In Juran and Gryna (1988), a broader definition is given: “quality is
fitness for use”. In this view, “quality” is defined as a relative notion.
Different usages of the product will result in different requirements with
regard to the product. Let us consider shoes as an example. The require-
ments of a person looking for high quality jogging shoes will differ from
the requirements of the same person looking for high quality elegant shoes.
Montgomery (1996) considers ‘conformance to requirements’ as one of two
aspects of ‘fitness for use’. The other aspect is ‘quality of design’, which em-
phasizes the intentional design differences between types of a product. The
conformance aspect is how well the product conforms to the specifications
required by the design.
Nowadays, these definitions of “quality” are labeled as traditional. It
is recognized that the quality of a product or a service is not a single,
identifiable characteristic. Garvin (1987) distinguishes the following eight
dimensions of quality, which, when taken together, incorporate more as-
pects than the ‘traditional’ definitions of quality.

1. Performance is one of the traditional measures of quality. It

refers to the basic functioning of a product. For example, for an
automobile, performance would include acceleration, handling,
cruising speed, and comfort;
2.1. QUALITY DEFINED 15

2. Features that are added to the basic functioning of a product

attribute to a higher quality;

3. Reliability is another traditional measure of quality. A re-

liable product rarely fails. This aspect, which is sometimes
reformulated as ‘being free of deficiencies’ is a very important
dimension of quality;

4. Conformance is related to reliability. It refers to the degree

to which a product meets pre-established requirements. This
dimension of quality is very important in situations where prod-
ucts are used as the components in a more complex assem-
bly. Specifications on the individual components are usually
expressed as a target and a tolerance. If each of the compo-
nents is just slightly too big or too small, a tight fit is unlikely,
and the final product may not perform as intended by the de-
signer, or may wear out early;

5. Durability is a measure of product life, either economically

(expected cost of repair exceeds current product value) or phys-
ically (repair is impossible). A product that lasts longer is usu-
ally viewed as being of higher quality;

6. Serviceability relates to the time and effort that is needed to

repair a product. The breaking down of a product is usually
viewed as an annoyance, but a prompt repair may relieve part
of the irritation;

7. Aesthetics is a subjective dimension of quality. It refers to

the look, feel, sound, taste or smell of a product. It is greatly
influenced by the preferences of the individual customer. On
this dimension of quality it is usually not possible to meet the
needs of every customer;

8. Perceived quality is also a very subjective dimension. When

customers do not have full information about a product they
may base their quality image on past experiences, the reputa-
tion of the manufacturer, the quality of other products from
the same manufacturer, or the name of the product.

Realizing that quality is not a one-dimensional characteristic of a product

and that quality is determined at various levels in the production process,
16 CHAPTER 2. STATISTICAL PROCESS CONTROL

Garvin stresses that manufacturers should not strive to be first on all eight
dimensions of quality. Rather, he should select a number of dimensions on
which to compete.
Traditionally, the quality control departments in factories compete with
regard to the conformance dimension of quality. It is their responsibility
to ensure that requirements set on a quality characteristic are met. Such
requirements are usually stated in the form of specification limits. All
parts within limits are classified as conforming. The objective then is to
produce “zero defects”. Sullivan (1984), argued that this conformance-to-
specification-limits approach effectively prevents ongoing quality improve-
ment. As long as all outcomes of a production process are within specifi-
cation limits, a process engineer would have great difficulty in convincing
his plant manager to make any investment for the improvement of quality.
Sullivan advocates defining quality as “uniformity around the target”. In
this more modern point of view, which was put forward by, among oth-
ers, Deming and Taguchi, any deviation from target reduces reliability and
increases costs, in the form of plant and customer loss. Operational objec-
tives directed towards achieving ongoing quality improvement should not
be stated in terms of specification limits such as “zero defects”. The atten-
tion for quality improvement will diminish as soon as the manufacturing
process is able to produce amply within specs. A more continuous drive
for ongoing quality improvement will be obtained if the aim is to reduce
variation around the target.
This relation between “quality” and “variation” is summarized in the
following phrase which can be found in Montgomery (1996):

“Quality is inversely proportional to variability”.

This statement demonstrates the contribution of statistical methods to

quality improvement projects. By acknowledging that variation is present
in process outcomes, and that they are to some extent uncertain, it be-
comes necessary to employ methods which take uncertainty explicitly into
account.
In this thesis, techniques are described that are aimed at improving
quality through reduction of variability. We realize that this only covers
one of the many facets of quality, but nevertheless, it can make an important
contribution.
2.2. REDUCING VARIABILITY 17

2.2 Reducing variability

In the previous section, an attempt was made to define the use of the word
“quality”. As a result of the close relation between quality and variation, it
was natural to use terminology that expresses uncertainty in process out-
comes such as “rarely”, “degree to which”, “probability”, and “variation”.
It was also argued that quality improvement can be obtained by reduction
of variability. The approach towards reduction of variability is based on
the idea that the phenomena causing variation in process outcomes can
be classified in two groups: common causes of variation and special causes
of variation. In Subsection 2.2.1, the difference between these classes is
discussed. Moreover, we will concern ourselves with the question how this
distinction can be utilized to improve the quality of process outcomes. In
many practical cases, the implementation of this approach is limited to
manufacturing processes. However, the basic ideas are applicable in all ar-
eas of the organization, from production to sales. It is important that the
whole organization is aware of the opportunities of quality improvement by
reduction of variation. This concept, which is sometimes called “Statistical
Thinking”, is discussed in Subsection 2.2.2.
The information concerning the presence of one or both types of varia-
tion is in practice based on samples from a process. These samples should
be taken with care. This is discussed in Subsection 2.3.1.

2.2.1 Common causes and special causes of variation

Dr. Walter A. Shewhart is unmistakably the founding father of Statisti-
cal Process Control. The concepts he developed in the twenties have been
written down in his pioneering work, “Economic Control of Quality of Man-
ufactured Product” (1931). His ideas are still very relevant and the most
important tool, the control chart (which will be discussed in the next sec-
tion and, in fact, throughout this thesis), is nowadays used world wide in
a virtually unaltered form.
Shewhart compares a manufacturer that tries to make all products con-
form with a certain target to a marksman, aiming at a bull’s-eye. Just as
it is impossible for the marksman to hit the bull’s-eye with every shot, it is
not to be expected that every product will exactly comply with the target.
This is not a problem in itself, since ‘small’ variations around the target
are acceptable for most customers. The problem is in Shewhart’s words:
18 CHAPTER 2. STATISTICAL PROCESS CONTROL

“how much may the quality of a product vary and

yet be controlled? In other words, how much vari-
ation should we leave to chance? ”

It is important to realize that this formulation of the problem explicitly

states that, even if the quality of a certain product is ‘controlled’, some
variability must be allowed for. The marksman may do his utmost to hit
the bull’s-eye with each shot, but he cannot hope to hit the target every
time. A complex of causes of variation exerts its influence on the outcome
of each shot. Both the manufacturer and the marksman are not always
able to do what they want to do and cannot hope to precisely understand
why.
The marksman’s example is just one of many that can be given to
illustrate that in all aspects of our lives, there is some variation that is
considered to be ‘normal’ or ‘acceptable’, and does not call for action. The
underlying sources that are responsible for this type of variation were orig-
inally called chance causes of variation by Shewhart. Nowadays, the term
common causes is used, a reformulation that is due to Deming. An im-
portant aspect of the variation due to common causes is that it is to some
extent predictable: it is possible, based on earlier experiences, to determine
limits that bound the effect of the common causes. In the marksman’s
example, numerous common causes such as wind, trembling of the marks-
man’s hands, and the precision of the rifle, affect the result of each shot.
As long as the hits are within some range of the bull’s-eye, the marksman
leaves the variation to chance.
If the observed variability is such that it exceeds the boundaries of
‘normal’ variation we are inclined to undertake action. The presence of
something out of the ordinary, not within the class of common causes, is
suspected. Such causes of variation belong to the class of special causes
(in Shewhart’s terminology: assignable causes). If a single shot of the
marksman deviates more than normally from the bull’s-eye because he was
startled by some noise of the bystanders, he might want to ask them to
keep quiet or he might want to use earplugs.
In a manufacturing environment, it is, in most cases, not difficult to
visualize a large number of common causes, the joint effect of which causes
variation in the outcomes. This variation is inherently part of the process,
and is always present, from day to day, from hour to hour. Usually, it is
not within the power of an operator to remove common causes of variation;
such variation is ‘left to chance’. If it is necessary to remove common causes
of variation, this requires in most cases a profound revision of the process,
2.2. REDUCING VARIABILITY 19

which is the responsibility of the owner of the process, i.e. management.

Special causes of variation are not part of the process, and occur only
accidentally. However, when a special cause of variation is present, it will
have a large effect on the outcomes of the manufacturing process. If removal
is possible, a special cause can usually be eliminated without revising the
process. In many cases, an operator can be instructed to recognize and
remove special causes of variation, thereby improving the quality of the
outcomes of the process.
It is important to realize that the responsibility for reducing the effect
of special causes of variation lies on a different management level than
the reduction of common causes of variation. Counteracting the effect of
special causes of variation can be delegated to operators, whereas reducing
the effect of common causes of variation is the responsibility of the owner of
the process. It is important for an operator to know whether or not special
causes are present, so that he can undertake action to remove this cause of
variation. But it is even more important to know when to leave the process
alone when only common causes of variation are affecting the outcomes.
What happens in practice, is that operators try to counteract the effect
of common causes of variation as if it was a special cause of variation.
In many cases, this will result in larger variation in the outcomes. This
phenomenon of intervening in a stable process when it would have been
better to do nothing, is called ‘tampering’. It results in frustration, because
of unsuccessful searches for special causes of variation, and in waste of time
and money.
It is therefore of critical importance to be able to distinguish situations
where only common causes of variation affect the outcomes of a process,
from situations where also special causes are present. If only common causes
of variation are present, the manufacturing process is said to be “statisti-
cally in control”. This does not mean that there is no variation, or that
there is small variation. It does mean that the outcomes are predictable,
within statistical limits. In Shewhart’s words:

“... a phenomenon will be said to be controlled

when, through the use of past experience, we can
predict, at least within limits, how the phenomenon
may be expected to vary in the future. Here it is
understood that prediction within limits means that
we can state, at least approximately, the probability
that the observed phenomenon will fall within the
given limits.”
20 CHAPTER 2. STATISTICAL PROCESS CONTROL

The predictability of a process that is statistically in control is the basis

for the control chart, a tool that can be utilized to distinguish between
situations where only common causes of variation affect the outcomes of a
process, and situations where also special causes are present. The control
chart will be discussed in Section 2.3 for the case of independent observa-
tions.
In the case of independent observations, the term ‘an in-control process’
is associated with a sequence of independently and identically distributed
observations. In most applications and in many textbooks, these require-
ments are adopted as necessary and sufficient conditions for an in-control
process. However, it must be noted that Shewhart’s original definition does
not require independence of successive observations from an in-control pro-
cess. It only demands that we can predict how the process may be expected
to vary in the future. In Subsection 2.5, we will extend the definition of an
in-control process to include cases where local deviations of the mean are
allowed, due to stationary serial correlation. Such a definition is entirely in
line with Shewhart’s definition of an in-control process.
In the next subsection, which is based on an article by Snee (1990), it is
shown that the philosophy behind SPC is not restricted to manufacturing
processes alone.

2.2.2 Statistical thinking

Application of the idea of improving quality by reduction of variability is
in most cases limited to manufacturing environments. However, the con-
cept can be applied at many other levels of an organization, as well. In
Snee (1990), the place of Statistical Process Control in the much broader
context of Total Quality Management is discussed. Successful implemen-
tation of such a strategy requires a new way of thinking, which he calls
statistical thinking. These ideas can provide a useful contribution to ef-
forts aimed at quality improvement at all levels in an organization, from
production to sales. Snee describes the essence of statistical thinking as
follows:

“... all work is a series of interconnected processes

and identifying, characterizing, quantifying, con-
trolling, and reducing variation provide opportuni-
ties for improvement.”

Even more insightful is the schematic presentation that Snee gave of sta-
tistical thinking in quality improvement, see Figure 2.1.
2.2. REDUCING VARIABILITY 21

reduce
common - change
cause process
variation
6 ? ?
develop
all work processes analyze satisfied
is a - are - process - process reduce - improved -
custo-
know- variation quality
process variable variation mers
ledge

? 6
remove
special - control
cause process
variation

Figure 2.1: Statistical Thinking in Quality Improvement

In the first two boxes of Figure 2.1, it is indicated why statistical think-
ing is a logical approach to follow in all activities that are aimed at improv-
ing quality. All work that is done can be viewed as a process. A process
can be defined as (see Nolan and Provost (1990)): “a set of causes and con-
ditions that repeatedly come together to transform inputs into outcomes”.
The inputs might include people, materials, or information. The outcomes
include products, services, behavior, or people.
In all such processes, variation is encountered. Careful analysis of this
variation, combined with knowledge of the process may lead to reduction
of variation. It is therefore important for a manager to realize that close
cooperation with those who work with the process (e.g. operators), is an
absolute necessity for successful quality improvement. The people that
work with the process posses much of the knowledge that is needed to
reduce variation.
Reduction of variation may be accomplished by one of two paths. Re-
duction of variation may be brought about by removing special causes of
variation, which is the responsibility of the people working with the process.
The other path is to reduce the effect of common causes of variation, which
requires the management to undertake action. Removal of common causes
requires a different approach than removal of special causes of variation.
22 CHAPTER 2. STATISTICAL PROCESS CONTROL

However,

“Deming and his colleagues point out that man-

agers typically treat all problem as due to special-
cause variation, when, in fact, more than 85% of
problems are due to defects in a system (common-
cause variation), which only management can change.
The result is that management spends too much
time ‘fire-fighting,’ solving the same problem again
and again because the system was not changed ”
Snee (1990) page 120.

2.3 The Shewhart control chart

In the foregoing section, we argued that it is very important to detect the
presence of special causes of variation. A tool is needed for this, since the
effect of a possible special cause is hidden in the variation due to common
causes. Shewhart developed the control chart for this purpose. It will be
discussed in this section.
The control chart is based on the idea that if the process is in a state of
statistical control, the outcomes are predictable. Based on previous obser-
vations, it is possible for a given set of limits to determine the probability
that future observations fall within these limits.
Observations that are utilized for monitoring the process are usually
grouped according to time, amount of production, or some other descrip-
tive statistic. Sampling and subgrouping should be carried out with care.
In Subsection 2.3.1, it will be discussed how to take and group the observa-
tions. From each subgroup sample, descriptive statistics such as the mean,
the range or the sample standard deviation are computed.
In its original form, the control chart is a simple time plot of a sequence
of subgroup statistics. The points in the plot are compared to limits, which
indicate the bandwidth of the variation due to common causes. These limits
are called control limits. The width of these limits is such that, as long as
all points are within the control limits, it is reasonable to assume that the
underlying process is statistically in control. A point outside the control
limits is called an out-of-control signal, as it indicates that more variation is
present than can be attributed to the effect of common causes of variation.
However, due to the random nature of the observations, there is a small
probability that an out-of-control signal is encountered while the process is
2.3. THE SHEWHART CONTROL CHART 23

statistically in control. Such a signal is called a false out-of-control signal.

In Figure 2.2, an example of a control chart is depicted.

out-of-control
signal XX
z r
X
Upper Control Limit (UCL)

6 r
r r
r r
statistic
value of

1
r
2 3
r
4 5 6
r
7 8 9 10
-
subgroup
r number

Lower Control Limit (LCL)

Figure 2.2: Illustration of a control chart.

Control limits are not to be confused with specifications or other targets

for the process. They are simply a prediction of the variation that will
occur due to common causes. If the variation due to common causes is
relatively large, all points on the control chart may be within the control
limits, but the process outcomes might fail to meet the specification limits.
The presence of special causes does not necessarily mean that there is large
variation, or that the specifications are not met. It does mean that there is
some source of variation that causes the measurements to be more variable
than can be attributed to common causes.
Specification limits are agreements between manufacturer and its cus-
tomers concerning the tolerated deviation from target. They are in no way
related to the actual performance of the process. Control limits on the
other hand indicate the magnitude of the variability of the process when
only common causes are present. They reveal what Nelson (1988) calls ‘the
heartbeat of the process’.
It is not advisable to plot specification limits in a control chart. Not
only to avoid confusion: understanding the difference between the two con-
cepts is hard enough. A more important reason is that specification limits
relate to single products, whereas control limits are usually computed for
a grouped variable such as a sample mean. A sample mean that is both
within control limits and specification limits may give the impression that
the process is performing as required, also in situations where some of the
24 CHAPTER 2. STATISTICAL PROCESS CONTROL

individual observations exceed the specification limits.

2.3.1 Rational subgroups

Briefly, a control chart is a tool designed to judge whether a process is
statistically in control or not. The control chart utilizes samples of ob-
servations, mostly drawn in subgroups, to obtain information about the
behavior of the underlying process. Statistics that summarize the informa-
tion in the subsamples are compared to limits which represent the variation
due to common causes. If an out-of-control signal is observed, action is re-
quired to track down the special cause that is responsible. Sampling and
subgrouping of the observations must be carried out in such a way that
the search for a causal relation between the out-of-control signal and some
underlying disturbing phenomenon is facilitated.
The way the observations are sampled and grouped may have a large
effect on the behavior of a control chart. This phenomenon is illustrated in
Section 6.1, where serially correlated measurements result in control limits
that are too tight, leading to too many out-of-control alarms. Subgrouping
of the data should be carried out with care, especially in situations where
subsequent observations are serially correlated.
In addition to providing us with the control chart, Shewhart also gave
guidelines for sampling subgroups of observations from a process. To this
end, he introduced the concept of “rational subgroups”. A rational sub-
group is a sample in which all of the items are produced under conditions
in which only common causes are responsible for the observed variation.
Special causes do not occur within a rational subgroup, but only between
subgroups. If the observations can be grouped according to these require-
ments, then appropriate control limits can be determined that discriminate
between in-control situations and out-of-control situations.
Linking an out-of-control signal to a specific special cause of variation
is facilitated by preserving the time order in the data points on the control
chart. The time at which an out-of-control signal is given may give a hint
as to when the special cause has occurred.
A control chart must be sensitive enough to detect the effect of special
causes of variation, but must not generate too many false out-of-control
signals. In practice, a balance between these two must be struck by deter-
mining the width of the control limits. How this is done for the classical
Shewhart chart is the subject of the following two subsections. Following
Does and Schriever (1992), we distinguish two situations.
In Subsection 2.3.2, the situation is discussed where we have k rational
2.3. THE SHEWHART CONTROL CHART 25

subgroups of size n of past observations available. These observations are

used to set up the control chart. This is sometimes called “Phase I”, Does
and Schriever (1992) call it “analysis of past data”. In Phase I, the mag-
nitude of the variation due to common causes is determined, so that the
width of the control limits can be determined for what is sometimes called
“Phase II”. In Phase II, rational subsamples become available one by one,
as they are drawn online from the process. Does and Schriever (1992) call
this phase “performance of current control. Determining control limits for
the Shewhart chart in Phase II will be discussed in Subsection 2.3.3.

2.3.2 Phase I: setting up of the control chart

In order to be able to compute the control limits for Phase I, some formal-
ization of the foregoing is needed. Suppose that, from a certain manufac-
turing process, k rational subgroups of size n are sampled. Let us denote
the j th observation of the ith subsample by Xij , where i = 1, · · · , k, and
j = 1, · · · , n. For the time being, we will assume that for all i, the individ-
ual observations Xi1 , · · · , Xin are identically and independently distributed
within sample i, with distribution function Fi . Furthermore, observations
in different samples are also assumed to be independent. The independence
assumption will be loosened in subsequent chapters.
Note that the foregoing includes a formalization of the concept of ra-
tional subgroups. Distribution functions F1 , · · · , Fk are used to model the
joint effect of common and special causes of variation. Within each sample,
the distribution function does not change. However, due to the presence of
special causes, the distribution function may change over time.
If there are no special causes present, we assume that the distribution
functions do not change over time. Monitoring the process for special causes
of variation can then be formalized as testing the hypothesis

H0 : F1 = F2 =, · · · , = Fk ≡ F0

against the alternative

H1 : there are s, t ∈ {1, 2, · · · , k} such that Fs 6= Ft .

In many cases, it is assumed that the distribution functions F1 , · · · , Fk are

known except for a few parameters (for example the expectation and the
variance of a normal distribution). Control charts are set up for each of
26 CHAPTER 2. STATISTICAL PROCESS CONTROL

these parameters. An out-of-control signal on one of these charts is an

indication of the presence of a special cause of variation.
The control chart for a one-dimensional parameter ηi of Fi is set up in
the following way. Let Ti = T (Xi1 , · · · , Xin ) be an estimate of ηi based
on the sample at time i. Furthermore, let Mk and Vk be statistics based
on the k samples such that Mk = M (X11 , · · · , X1n , . . . , Xk1 , · · · , Xkn ) is
a consistent estimator of the location of the distribution of Ti under H0 ,
and Vk = V (X11 , · · · , X1n , . . . , Xk1 , · · · , Xkn ) is a consistent estimator of
the spread of the distribution of Ti under H0 .
In the control chart, realizations of Ti are plotted against i. These
values are compared to control limits of the form

L = Mk + c(n, k, p)Vk , (2.1)

where c(n, k, p) is the pth percentile of the null distribution of (Ti − Mk )/Vk .
For the LCL and the UCL of a control chart, different constants c(n, k, pLCL )
and c(n, k, pUCL ) must be determined.
In situations where a location parameter or a spread parameter of the
distribution function of Ti is known, these values are used in (2.1) instead
of their estimates.
Note that the elements of the sequence {T1 , · · · , Tk } are mutually inde-
pendent, but this does not, in general, hold for Ti , Mk , and Vk , since they
are (partly) based on the same set of observations.
In most literature on SPC it is assumed that Fi is a normal distribution
function with expectation µi and variance σi2 . The mean and/or variance
of the observations may change over time due to the presence of special
causes of variation. With these assumptions, the process is in control if
and only if µi = µ for some µ and if σi = σ for some σ for all i = 1, · · · , k.
It is for this reason that in a lot of cases, a production process is monitored
using two control charts, one for the standard deviation, and one for the
mean of the process.
Shewhart did not consider statistical arguments to determine the con-
stants c(n, k, pLCL ) and c(n, k, pUCL ). He decided to choose, “based on eco-
nomic considerations”

c(n, k, pLCL ) = −3

and

c(n, k, pUCL ) = 3.
2.3. THE SHEWHART CONTROL CHART 27

These values turn out to work well in a lot of practical cases. The underlying
statistical arguments for a control chart for the mean of normal observations
are the following. Suppose that we are testing the hypothesis H0 : µ1 =
· · · = µk ≡ µ where µ is known, assuming a constant known variance σ 2 .
Furthermore, assume that the sample means Ti = 1/n(X1i + · · · + Xni )
√
are plotted in a control chart with limits LCL = µ − 3σ/ n, and UCL =
√
µ + 3σ/ n. Then we have pLCL = (1 − pUCL ) = 0.00135, so that under these
assumptions a false out-of-control signal is quite unlikely.
If an out-of-control signal is generated in Phase I, a search is initiated
for a responsible special cause of variation. If this can be found, action
should be taken to prevent it from re-occurring. In cases where a special
cause is found and removed, the corresponding sample does not provide
information about the in-control state of the process. Therefore, in such
cases, it should be removed from the data set, and the remaining k − 1
subsamples should be compared to re-estimated control limits.
This procedure should be repeated until no out-of-control signals are
generated, or when underlying special causes either cannot be found or
cannot be removed. At the end of Phase I, we have a data set at our
disposal of, say, m ≤ k subsamples that provides information concerning
the variability that can be attributed to common causes of variation. This
information is needed for Phase II, when samples are drawn online.

2.3.3 Phase II: current control

In Phase II, we have an estimate of the in-control distribution available
based on m samples from Phase I. This distribution function will be denoted
by F0 . Each time a sample Xf 1 , · · · , Xf n becomes available at time f > k,
we want to test the hypothesis

H0 : Ff = F0

against the alternative

H1 : Ff 6= F0 .

The derivation of the control limits for Phase II is analogous to Phase I,

with this difference: Mm and Vm are independent of Xf 1 , · · · , Xf n . This
will result in different constants c(n, k, pLCL ) and c(n, k, pUCL ) for the LCL
and UCL, respectively.
28 CHAPTER 2. STATISTICAL PROCESS CONTROL

The control charts discussed in this section are Shewhart-type control

charts. Their performance under various levels of first-order autocorrelation
(including the special case of independence) will be studied in Chapter 3.
More efficient control charts are also developed, such as the EWMA control
chart and the CUSUM control chart. Their performance for various levels
of first-order serial correlation will be discussed in Chapter 4 and Chapter 5,
respectively.

2.4 Serial correlation

In the previous sections, we made the assumption that the measurements
that are used to monitor the process are independently distributed. This is
standard practice in most literature on SPC. In this thesis, it is investigated
how the performance of control charts is affected by serial correlation in
the observations. This is motivated by many practical situations, where the
assumption that the observations are (approximately) uncorrelated cannot
be justified, see also Chapter 1. In this section, we will shortly review some
of the types of serial correlation that will be considered in subsequent chap-
ters. We will restrict ourselves to serial correlation that can be successfully
modelled using ARIMA(p,d,q) models. This class of models is discussed in
a large number of texts on time series, such as Box and Jenkins (1976),
Pandit and Wu (1983), or Montgomery, Johnson and Gardiner (1990). We
refer to such texts for a deeper discussion of time series analysis.
Throughout this thesis, a serially correlated sequence of variables will
be denoted by {Y1 , · · · , YT }, whereas a sequence of independent random
variables will be denoted by {X1 , · · · , XT }.
Before we turn to the discussion of ARIMA models, we introduce some
of the terminology that is used in time series analysis. This will facilitate
the discussion in subsequent chapters. Much of it is taken from Ander-
son (1976).
A time series model that has proved to be useful in practice is the AR(1)
model. This model deserves and will receive most of the attention in subse-
quent chapters. It is one of the simplest special cases of ARIMA(p,d,q) mod-
els. It will be discussed in Subsection 2.4.2. More general ARIMA(p,d,q)
models will be discussed in Subsection 2.4.3

2.4.1 Some time series analysis terminology

We define a time series as a set of observations that are ordered in time.
We will only consider discrete series, with observations drawn at equal
2.4. SERIAL CORRELATION 29

time intervals. A sequence of observations {y1 , y2 , · · · , yT } can be viewed

as a single realization of some underlying stochastic process. Each yi is the
realization of some random variable Yi , which has an associated probability
density function fYi (·). We assume that for any set of Yi ’s, say Yj1 , · · · , Yjr , a
joint probability density function fYj1 ,···,Yjr (·) exists. If a statistical process
is such that fYi+n1 ,···,Yi+nm (·) is independent of i for any positive integer m
and for any choice of n1 , · · · , nm then the probabilistic structure does not
change over time and the process is said to be strictly stationary. Otherwise,
it is nonstationary. If the definition of strict stationarity only holds for
m ≤ p for some positive integer p, then the process is said to be stationary
of order p.
A Gaussian process is defined by the property that the probability den-
sity function associated with any finite subset of {· · · Y−2 , Y−1 , Y0 , Y1 , Y2 , · · ·}
is multivariate normal. For a Gaussian process, a sufficient condition for
strict stationarity is stationarity of order 2, since all moments of higher
order are then precisely fixed. Stationarity of order 2 is sometimes also
called weak stationarity.
It is important to make the distinction between (weakly) stationary
models and nonstationary models, since observations from the former wan-
der around a fixed mean. Observations from a nonstationary model may
wander away from a specified target value if no action is taken. That is,
nonstationary processes require some form of Automated Process Control
(APC) if the quality characteristic should fall between certain specification
limits. Stationary processes are mean reverting and need not be controlled
by APC techniques. However, the performance of such processes can be
improved by applying APC, see Box and Luceño (1997b).
Weak stationarity implies that for all i

E(Yi ) = µ

and

Cov(Yi , Yi−k ) = γk ,

where µ is the constant mean of the process, and γk , the autocovariance

at lag k (integer), is also constant. In particular, Yi has constant variance
σY2 = γ0 .
Furthermore, we have for all integers k,

γ−k = γk
30 CHAPTER 2. STATISTICAL PROCESS CONTROL

since
Cov(Yi , Yi+k ) = Cov(Yi+k , Yi ) = Cov(Yi , Yi−k ).

Hence it is only necessary to consider γk for k > 0. The set {γ0 , γ1 , · · ·} is

sometimes called the autocovariance function. We define ρk , the autocorre-
lation coefficient at lag k by
γk
ρk = .
γ0

The set {ρ0 , ρ1 , · · ·} is called the AutoCorrelation Function (ACF) of the

process. An estimate of the ACF can be used to identify which models
within the class of ARIMA(p,d,q) can be used to model the random behav-
ior in a given set of observations.

2.4.2 The AR(1) process

A process {Yt } is said to be a first-order autoregressive process (AR(1)
process) if it is generated by
Yt − µ = φ(Yt−1 − µ) + εt for t ∈ ZZ, (2.2)

where φ is some constant satisfying φ ∈ (−1, 1), and {εt } is a sequence of

i.i.d. disturbances, εt ∼ N (0, σε2 ) for t ∈ ZZ . Subsequent observations of
model (2.2) are serially correlated, since
Cov(Yt , Yt−k ) = φk σY2 ,

where
σε2
σY2 = Var(Yt ) = .
1 − φ2

Model (2.2) is strictly stationary since |φ| < 1.

Obviously, E(Yt ) = µ for all t in model (2.2). A model that includes
the possibility of a shift in E(Yt ) ≡ µt is
Yt − µt = φ(Yt−1 − µt−1 ) + εt for t ∈ ZZ. (2.3)

First-order autoregressive models are useful when disturbances affect not

only the current outcome of the process, but also have an (exponentially de-
clining effect) on future outcomes. Such situations occur for example when
a tank containing raw material is refilled from time to time with raw mate-
rial of varying quality, see also the example considered in Subsection 1.1.3.
2.4. SERIAL CORRELATION 31

2.4.3 ARIMA processes

More general AR processes are of the form

Yt − µ = φ1 (Yt−1 − µ) + · · · + φp (Yt−p − µ) + εt for t ∈ ZZ, (2.4)

If we introduce the backward shift operator B, where BYt = Yt−1 , then (2.4)
can be rewritten as

φ(B)(Yt − µ) = εt ,

where φ(B) is a polynomial in B of degree p. Autoregressive models of

order p are stationary if all roots of the polynomial φ(·) are outside the
unit circle (see Box and Jenkins (1976), Section 3.2).
A Moving Average (MA) process of order q is generated by

Yt = µ + εt − θ1 εt−1 − · · · − θq εt−q .

Moving average models of any order are always stationary.

A useful class of models for time series is formed from a combination
of MA and AR processes. A mixed autoregressive moving average process
containing p AR terms and q MA terms is abbreviated to an ARMA(p,q)
process, and is given by

Yt − µ = φ1 (Yt−1 − µ) + · · · + φp (Yt−p − µ)
+ εt − θ1 εt−1 − · · · − θq εt−q for t ∈ ZZ,

Or, rewritten using the backward shift operator B:

φ(B)(Yt − µ) = θ(B)εt , (2.5)

where φ(B) and θ(B) are polynomials of degrees p and q, respectively. Such
processes are stationary if the roots of the polynomial φ(·) lie outside the
unit circle. The class of ARMA(p,q) models can be used to model a wide
range of stationary time series, with only a few parameters to estimate. In
practice, values of p and q larger than 2 are rarely encountered.
However, many time series encountered in practice are nonstationary.
In order to fit a stationary model it is necessary to remove the nonstationar-
ity first. In many cases, this can be obtained by taking successive differences
of the observations one or more times. That is, in case of first differences,
32 CHAPTER 2. STATISTICAL PROCESS CONTROL

we consider {y2 − y1 , y3 − y2 , · · · , yT − yT −1 } and try to fit a stationary

ARMA(p,q) model to these new observations. If taking first differences
is not enough to obtain stationarity, the observations are differenced once
more, and so on, until stationarity is obtained. Taking differences can be
expressed in terms of the backward shift operator as (1 − B)yt . The nota-
tion (1 − B)d yt is used to indicate that successive differences were taken d
times.
If we replace Yt − µ in Equation (2.5) by random variables that are
differenced d times, we have

φ(B)(1 − B)d (Yt − µ) = θ(B)εt .

Such models are called AutoRegressive Integrated Moving Average (ARIMA)

models of order (p,d,q). The term ‘integrated’ is used because the station-
ary model which is fitted to the differenced data has to be summed or
‘integrated’ to provide a model for the nonstationary data. ARIMA(p,d,q)
models are capable of describing certain types of nonstationary time series.
An important special case is the IMA(1,1) model, which is often encoun-
tered in practice. In Chapter 8, an example of an IMA(1,1) process is
discussed.
The ARIMA(p,d,q) models were popularized by Box and Jenkins (1976).
Box and Jenkins (1963) themselves developed one of the first control charts
to account for serial correlation. By assuming that the quality characteris-
tic was drifting away from its target value according to an ARIMA(0,1,1)
model, they derived a chart with action limits that are determined such
that the total costs of running the process are minimized. The cost min-
imization procedure required trading off the cost of being off-target with
the cost of resetting the machine.

2.5 Control charts for serially correlated data

In Section 2.3, it was discussed that the control chart is the tool to detect
special causes of variation. The control limits that are drawn on the control
chart bound the variation due to common causes of variation. For the
proper placement of control limits in the case of serial correlated data, it
is of crucial importance to decide which process behaviors are part of the
process, and which are attributed to special causes. Gilbert, Kirby and
Hild (1997) state the following.
2.5. CONTROL CHARTS FOR SERIALLY CORRELATED DATA 33

“For example, if autoregressive behavior is a nor-

mal, unchangeable part of the process, then a chart
that gives out-of-control signals because of the pres-
ence of autocorrelation is not very useful. On the
other hand, if autocorrelation in a process is a symp-
tom of a problem that should be addressed, then the
control chart should detect the presence of the au-
tocorrelation.”

Crowder, Hawkins, Reynolds and Yashchin (1997) share this view. They
argue that the cause of autocorrelation should be assessed before the data
is analyzed and interpreted. As argued in Chapter 1, we will consider
cases where autocorrelation in process data is unremovable and part of
the process. Control charts should not signal because of autocorrelation,
but give out-of-control signals because of the presence of special causes of
variation.
Consequently, the definition of an in-control process that is most com-
monly used in practice and Shewhart’s original definition do not necessary
agree. In practice, the term ‘an in-control process’ is more often than
not associated with a sequence of independently and identically distributed
observations. As was discussed in Subsection 2.2.1, Shewhart’s original def-
inition of an in-control process only requires that we can predict (within
statistically determined limits) how the process may be expected to vary
in the future. A process that exhibits serial correlation is predictable. For
this reason we extend the definition of an in-control process to include ob-
servations which may be serially correlated. Alwan (1988) refers to such
processes as being ‘in control in a broader sense’.
34 CHAPTER 2. STATISTICAL PROCESS CONTROL