Some Perspectives and Challenges

for Control Chart Methods

ANDREW C. PALM James River Corporation, Neenah, WI 54957-0899

ROBERT N. RODRIGUEZ SAS Institute Inc., Cary, NC 27513

FRED A. SPIRING University of Manitoba, Winnipeg, ME R3T 2N2, Canada

DONALD J. WHEELER Statistical Process Control, Inc., Knoxville, TN 37919

Perspectives on Control 2 makes active use of computing technology, and

Chart Methods it is particularly well-represented in the semicon

IKE most successful statistical methods, control ductor and chemical process industries.

L charts were introduced to meet the practical

Group
methods.
3 has an academic interest in control chart
This group consists of researchers in
needs of a particular audience. Today, more than
university statistics, industrial engineering, and
70 years since Walter Shewhart constructed the first
business school departments who develop varia
control chart, there are three groups with distinct
tions of control charts that offer optimality or im
perspectives on control chart methods:
proved efficiency in a statistical sense. This group
Group 1 is primarily interested in the broad, effec often contributes to the Journal of Quality Tech
tive application of standard control chart methods nology and similar publications.
to many different fields, industries, and depart
ments within a company. This group is motivated Figure 1 summarizes these perspectives. Although
by competitive issues and changing industry stan there might be disagreement about the extent to
dards, and it is oriented by large-scale quality im which these groups intersect, we have used this clas
provement programs and by the "statistical think sification in our discussion to consider how these
ing" of the Deming approach. groups are challenged and how they can benefit from

Group 2 is interested in the development and ap each other.

plication of advanced control chart techniques to

industrial processes where on-line data are abun Statistical Challenges and
dant and where rapid response to out-of-control Enhancements
conditions is critical. This group is made up Group 2 has the most exposure to new,
of statisticians who are well-trained in statistical technically-challenging problems. Many individuals
process control and work closely with engineers in this group work within engineering and manufac
who are knowledgeable about the process. Group turing divisions of large, high-technology, capital
intensive companies whose management is idea
driven. For this group, the future of control charts
Dr. Palm is Director of Statistical Methods. He is a Mem
is closely connected to the future of on-line monitor
ber of ASQC.
ing and control of industrial processes. This work is
Dr. Rodriguez is Manager of the Statistical Quality Im beginning to merge with the already existing field of
provement R&D Group. He is a Member of ASQC. stochastic process control.

Dr. Spiring is an Associate Professor of Statistics and Di

Stochastic control has been successfully applied in
rector of the Quality Resource Center. He is a Senior Member
applications to aircraft and rockets, where the char
of ASQC.
acteristics of the "process" can be modeled quite ac
Dr. Wheeler is a consulting statistician. He is a Senior curately, and where the "disturbances" are Gaussian.
Member of ASQC. Stochastic control has also been applied to indus-

Journal of Quality Technology 122 Vol. 29, No. 2, April 1997

 A DISCUSSION ON STATISTICALLY-BASED PROCESS MONITORING AND CONTROL 123

the system rather than continually compensate for

Process Data and

them. Statisticians can help system designers un
Information Delivery
Statistical Computing
derstand issues of process variation as well as pro
vide technical consultation on data management and
analysis.

As these systems are being developed, another

challenge is to avoid the trap of believing that
newer, more complex (and hence more expensive) ap
proaches are always preferable to simpler methods.
For example, when studying various approaches to
process control, a premium should be placed on op
tions such as "Give the operators better instructions"
and "Convince management to pay more attention to
what is going on at the shop floor level" .
FIGURE 1. Three Perspectives on Control Chart Meth
There are also considerable challenges for Group
ods.
3. It has often been pointed out that statisticians
seeking to develop useful new process control meth
trial processes, but with somewhat less success. This
ods must understand the process generating the data,
should be expected because the process dynamics are
and that they must work closely with engineers. In
more difficult to model, the disturbances are more
varied in nature, and the economics are different.
order to meet the needs of Group 2, members of

However, just as the potential of using computers

Group 3 must be prepared to deal with the statis
tical problems and opportunities arising from the in
to handle business processes is finally being realized,
creasingly large volume and highly multivariate na
so too will computer control eventually have an ef
ture of process data. Unless statisticians are willing
fect on production processes. We will not consider
to consider new approaches and new measures of op
the issues of automated process control further in
timality and efficiency, the solutions will be devel
this discussion since they have been given consider
oped by problem domain specialists rather than by
able attention by numerous authors, including Box
statisticians, as has been the case in chemistry; see
and Kramer (1992); (1987, 1990); Mac
MacGregor
Gregor, Hunter, and Harris (1988); and Montgomery,
MacGregor (1995).
Keats, Runger, and Messina (1994).
New Frontiers for Control
For Group 2, the future of control charts is also Chart Applications
moving in the direction of techniques that are appro
priate for large multivariate data sets that possess One might question what, if anything, members

less than full statistical rank. Two multivariate pro of Group 1 have to say about the future of con

jection methods that are being applied extensively trol charts since they are characterized by their focus
by chemometricians and others in the chemical and on "classical" control chart methods that differ little
process industries are principal components analysis from Shewhart's original work. The answer is that

(PCA) and partial least squares (PLS). An excellent for this group, the key issues are the breadth and

overview of this direction was recently provided by effectiveness of application of control charts, cou
MacGregor (1995), also see Nomikos and MacGregor pled with competitive issues which motivate their

(1995) and Kourti and MacGregor (1996). use. Nowhere are these issues more evident than in
the American health care industry, where there is a
One challenge for members of Group 2 will be to rapidly emerging interest in control chart methods,
make sure that process improvement is balanced with and consequently it is worth discussing this case in
product quality. For example, data collected for pur some depth.
poses of process control may not be directly relevant
to the quality of the product leaving the process. The health care industry is driven by a desire
Even if this is the case, the data may be stored or to improve patient outcomes in the face of capi
managed in a way that makes it difficult to assess tation, cost reduction, competition, and changing
quality improvement. Also, it may be more econom health care standards. Measures of quality are be
ical to permanently remove some disturbances from ing required to demonstrate that patient care can be

Vol. 29, No. 2, April 1997 Journal of Quality Technology

124 A DISCUSSION ON STATISTICALLY-BASED PROCESS MONITORING AND CONTROL

maintained and improved as health care expenses are Proportion 01 C - Sections In 1995

brought under control. Hospitals joining managed 3.52<7 Urnit8:

1.00
care networks can succeed in winning contracts if
UDL
they can demonstrate a high level of patient sat
isfaction. The Joint Commission on Accreditation
of Healthcare Organizations (JCAHO) now requires
hospitals to improve organizational performance, and
patient satisfaction is one of nine measures of per
formance. The Health Plan Employer Data and In
formation Set (HEDIS), developed by the National

0
.25

P=.167
Committee for Quality Assurance, includes patient ..

satisfaction, and it is designed to help consumers

and employers compare performance of managed
lkbrb3 blbt:L1
care plans.
Medical Group ldantlftcation Number

As a result, health care managers and profes FIGURE 3. ANOM for Proportion of Caesarean Sections.
sionals are adopting continuous quality improvement
(CQI) programs, they are studying the Deming ap
Health care providers must also make simultane
proach, and they are applying SPC methods. Con
ous comparisons of rates across medical groups and
trol charts for attributes are proving especially useful
hospitals, and they have begun to apply the analy
due to the prevalence of count data. Figure 2 illus
sis of means (ANOM), a generalization of the Shew
trates a chart constructed by a health care provider
hart chart, to attribute data. (It is interesting to
to report the rate of office visits performed each
note that, by comparison, ANOM has received rel
month by a clinic. The rate charted for each month
atively little attention in manufacturing since the
is computed by dividing the number of visits by the
early 1980s, and there are very few published exam
membership for that month (expressed in thousand
ples of ANOM for attribute data; one exception is
member years).
the article by Ramig (1983).) F igure 3 displays the
results of an ANOM in which caesarean section rates
The number of visits is analogous to the number
are compared for 29 medical groups during 1995.
of defects in u charts used in manufacturing, and the
For each group, the value plotted is the number of
monthly membership is analogous to the number of
caesarean sections performed divided by the number
inspection units (both are measures of opportunity
of deliveries. The decision limits correspond to the
for an event to occur). Two sets of control limits are
ex = 0.05 level of significance. See Rodriguez (1996)
displayed in Figure 2 because a change in the system,
for further details.
resulting in a shift in the average rate of office visits,
was known to have occurred in September 1994. Although the charts in Figure 2 and Figure 3
are based on universal principles of statistical pro
U Chart lor OfIice Visits per 1,000 Members: Clinic E cess control, there are a number of special statisti
cal issues in health care applications. In the first
3000
place, deciding what to measure or count is often
a challenging problem. It is also difficult to aggre

2800
I! UCL gate count data and determine appropriate "mea
I sures of opportunity" for rates and proportions in

12800
::0
0 such a way that meaningful comparisons can be made
across groups. It is a substantial misuse of statistics
2400
LCL
to compare rates, collected at different locations, by
different people, and using different reporting sys
tems. Fortunately the validity of comparative per
a: 2200
formance statistics, such as survival rates for coro
nary artery bypass procedures across hospitals, is be
2000

JAN94 MAR94 MA'194 .lJU14 SEP94 NOV94 JAN95 MAA95

ing questioned as health care organizations increas
ingly use this information in advertising. Finally,
FIGURE 2. u Charts for Rate of Office Visits. as pointed out by Benneyan and Kaminsky (1995),

Journal of Quality Technology Vol. 29, No. 2, April 1997

 A DISCUSSION ON STATISTIC ALLY-BASED PROCESS MONITORING AND CONTROL 125

there is a natural tendency to use control charts as tical professionals, including Balestracci and Barlow
management or industry "report cards" rather than ( 1994); Benneyan ( 1995); Benneyan and Kaminsky
for genuine quality improvement (not unlike the mis (1995); Berwick, Godfrey, and Roessner (1991); and
informed use of control charts that occurs in manu Spoeri (1991).
facturing applications).
Although Group 3 is insufficiently represented in
Formal continuous improvement and quality ini new applications of control chart methods, this group
tiatives within educational institutions also repre is potentially in the strongest position to provide
sent new frontiers for the application of quality tools. Group 1 with training and consulting. Group 3 can
Possibly five years behind the health care industry, also contribute articles and updated textbooks that
the evolution of quality within education faces many motivate and illustrate the use of control charts in
hurdles similar to those encountered by health care new application areas.
institutions. Traditional control charts and their use
currently represent a significant change in the way There are also research opportunities for Group
information is represented and used within most ed 3 within this intersection. It is not clear, for in
ucational institutions. Here innovation is in the use stance, whether the usual binomial and Poisson mod
of traditional control charts. As quality initiatives els are appropriate for the count data encountered
evolve within the educational field more sophisti in health care applications (in some cases, individ
cated and innovative control charts will result. ual measurement charts for rates appear to be at
least as effective). As the analysis of means contin
It is extremely difficult for educational institutions
ues to be applied in this area, the need for additional
to make the transition from gathering vast amounts
methods of simultaneous inference will emerge. An
of data to effectively monitoring processes and iden
opportunity for time-series methods is illustrated in
tifying changes in the process and their impact. Ed
Figure 4, which displays a control chart maintained
ucational institutions routinely gather information
by a hospital system for the monthly variation in
such as the teaching loads of faculty members, the
emergency room visits; see Rodriguez (1996). Here,
number of returns to the bookstore, the number of
distinct control limits have been used to adjust for
travel claim forms handled per month, and the num
known higher rates of emergency room visits in warm
ber of new enrollments. Few of these variables would
weather. However, improved methods for seasonality
be plotted, and none would include control limits.
adjustment and prediction can also be useful. Very
"What to measure" and "how to measure it" , in ad
little has been written about this problem in the tra
dition to the types of inferences that are possible,
ditional control chart literature since manufacturing
are issues that need to be learned as quality initia
data are typically collected over much shorter time
tives mature. In many situations these issues will be
intervals. On the other hand, health care applica
unique to the educational field.
tions are by no means the only area in which control
charts are maintained over long periods of time. In
Resources for New Applications
One of the major obstacles for Group 3 is a lack Emergency Room Visits per 1000 Member Years

of statistical training and domain-specific examples

120
and case studies. For historical reasons, application
areas such as health care are not represented in tradi
tional textbooks on statistical quality improvement,
and few resources are available for professionals in
these areas who are considering the use of control
chart methods. One of us recently conducted an ex UCL
tensive on-line search for articles on "statistical qual
ity improvement" and "statistical process control"
lCL
that apply to health care. The search yielded a to
tal of less than 75 publications, the great majority
of which were authored by members of the health
40
care field and were published in health care jour JAN90 AUG90 MAR91 0CT91 MAY92 DEC92 JU193 FEB94 SEP94 APR95
nals or conference proceedings. However, some valu
able early contributions have been made by statis- FIGURE 4. Emergency Room Visits.

Vol. 29, No. 2, April 1997 Journal of Quality Technology

126 A DISCUSSION ON STATISTICALLY-BASED PROCESS MONITORING AND CONTROL

the automotive industry, for instance, control charts run. From a practical standpoint, the inability to
are being used to track defect and customer satisfac manage large or even moderate amounts of data is
tion data on a monthly and yearly basis. the main obstacle to scaling up the use of control
charts, and resources tend to be allocated accord
Information Technology ingly. This is reflected in the fact that specifications
for process control systems often devote hundreds of
In many process control environments, the chief
pages to data communication and management re
impact of computer technology has been to deliver
quirements, and only tens of pages to statistical anal
on-line measurements at increasingly higher rates,
ysis requirements.
and as indicated earlier, this is a motivating factor
for Group 2. One consequence of automated data A third problem is the delivery of process and
collection is that it is now possible to recognize pro quality information that is appropriate for individ
cess autocorrelation that was previously undetected. uals throughout the organization at different levels
We will not consider this as a statistical issue since of responsibility, ranging from plant operators to the
it is addressed by other discussion teams and has CEO. Graphical displays (including control charts)
also been considered by other authors, including Box that aid operators and engineers in the interpreta
and Kramer (1992); MacGregor (1987); Schneider tion of process variability have long been regarded
and Pruett (1994); Wheeler (1991); and Woodall and as essential for this purpose; for a recent discussion,
Faltin (1993). see Roes and Does (1995). This problem can also
be addressed with graphical user interfaces to data
Standard control chart methods are used effec access, analysis, and presentation, and the develop
tively by Group 1 in a great many real-time appli ment of software to provide effective interfaces is still
cations. Often "real-time" is rather vaguely defined in an early stage.
and can mean that measurements arrive as often as
every five seconds or simply once every few hours. Figure 5 illustrates an interface development facil
Timely notification of out-of-control conditions and ity (referred to as PFD for "process flow diagram" )

rapid response are critical, and consequently infor that was introduced specifically for process control

mation technology is used to link data bases and applications in which there are many steps; see SAS

automated data collection systems to software that Institute (1995). At build time, the system developer

provides the results of statistical process control. So can represent the process as a series of nested flow

phisticated examples of these systems can be found diagrams in which the steps and their sequences are

in the semiconductor, food and beverage, automo represented as nodes and arrows. At run time, the

tive, and chemical industries. user can drill down through the diagram to the ap
propriate level or step of the process and then click on
One important problem in this arena is traceabil a box to request an analysis, such as a control chart,
ity (i.e., rapid identification of the sources of an out for that particular step. This interface is data driven
of-control condition by linking the subgroup identi in the sense that for each step, both the analysis and
fier to information about raw material and earlier the node attributes (such as color) are programmat
steps in the process that is available on-line). This ically linked to the most recent set of measurements.
can be achieved through database access and query The interface exploits the fact that a flow diagram is
techniques, and it is a major step forward in the a well-established visual metaphor for a complex pro
broad-based application of control charts. cess, and it provides the appropriate level of detail
for a variety of users, ranging from an operator who
Another problem is the management of data and
is responsible for a particular combination of steps
control limits for large numbers of processes that
to a plant manager who needs to identify the critical
are continually evolving over time. Although there
problem areas.
are significant benefits to constructing control charts
manually in training exercises and in occasional ap In organizations where Group 1 is well supported
plications, many companies have recognized that by information systems, control charts are regarded
large-scale applications require software systems that as components of information delivery technology
compute, save, update, and display the control lim rather than statistical technology, and they are of
its. This requirement is further compounded in short ten implemented by systems developers who have not
run process control, where multiple sets of control been trained in statistical process control. The risk
limits must be maintained and accessed from run to is that such systems can greatly amplify the negative

Journal of Quality Technology Vol. 29, No. 2, April 1997

 A DISCUSSION ON STATISTICALLY-BASED PROCESS MONITORING AND CONTROL 127

Soluble Dry
material material

Weigh, screen
&mix

Liquid
material

Absorption
&mixing

Percent Liquid (Individual Analyses)

No 3.;; Limit ..
r=-----r- Rrr n= 1Q
58
UCL= 55.5

54
"D
' :;
cr
::i 52
No "if. LCL= 50.8
50

o 2 4 6 8 10 12 14 a 18 20 22 24 26 28 XJ
Packaging O"ily ""mpleo (July 1, 1886)

FIGURE 5. Graphical User Interface based on a Process Flow Diagram Metaphor.

consequences of poor statistical thinking, For ex provement with removal of out-of-control points from
ample, it has been our experience that automated a chart; and attempts to automate the detection of
process control systems rarely enforce the notions process shifts by means of complex rules that defy
of rational sampling and rational subgrouping; if subsequent diagnosis.
measurements are subgroup ed, the samples are typ
ically chosen arbitrarily or by convenience. The Again, Group 3 has the potential to deal with
consequences of incorrect subgrouping are discussed these conceptual problems. Through collaboration
by W heeler and Chambers (1992) and by Wheeler with the developers of information systems as well
(1995). Other common problems include the failure as process engineers, statisticians can contribute the
to distinguish between control limits and specifica skills needed to interpret, evaluate, and synthesize
tion limits; the failure to recognize multiple compo the flood of process data in which industry is drown
nents of process variation; confusion of process im- ing.

Vol. 29, No. 2, April 1997 Journal of Quality Technology