Chapter 5

Methods and Philosophy of Statistical Process Control

IMPORTANT TERMS AND CONCEPTS Action Limits Out-of-control process
Assignable causes of variation Out-of-control-action plan (OCAP)
Average run length (ARL) Pareto chart
Average time to signal Patterns on control charts
Cause-and-effect diagram Phase I and phase II applications
Chance causes of variation Rational subgroups
Check sheet Sample size for control charts
Control chart Sampling frequency for control
charts
Control limits Scatter diagram
Defect concentration diagram Sensitizing rules for control charts
Designed experiments Shewhart control charts
Factorial experiment Statistical control of a process
Flow charts and operations process charts, and value stream Statistical process control (SPC)
mapping
In-control process Three-sigma control limits
Magnificent seven Warning limits

EXERCISES

5.1.
What are chance and assignable causes of variability? What part do they play in the operation and interpretation of a Shewhart
control chart?

“Chance” or “common” causes of variability represent the inherent, natural variability of a process - its background noise. Variation
resulting from “assignable” or “special” causes represents generally large, unsatisfactory disturbances to the usual process
performance. Assignable cause variation can usually be traced, perhaps to a change in material, equipment, or operator method.
A Shewhart control chart can be used to monitor a process and to identify occurrences of assignable causes. There is a high
probability that an assignable cause has occurred when a plot point is outside the chart's control limits. By promptly identifying
these occurrences and acting to permanently remove their causes from the process, we can reduce process variability in the long
run.

5.2.
Discuss the relationship between a control chart and statistical hypothesis testing.

The control chart is mathematically equivalent to a series of statistical hypothesis tests. If a plot point is within control limits, say
for the average , the null hypothesis that the mean is some value is not rejected. However, if the plot point is outside the control
limits, then the hypothesis that the process mean is at some level is rejected. A control chart shows, graphically, the results of
many sequential hypothesis tests. x

NOTE TO INSTRUCTOR FROM THE AUTHOR (D.C. Montgomery):

There has been some debate as to whether a control chart is really equivalent to hypothesis testing. Deming (see Out of the Crisis,
MIT Center for Advanced Engineering Study, Cambridge, MA, pp. 369) writes that:
“Some books teach that use of a control chart is test of hypothesis: the process is in control, or it is not. Such errors may derail self-
study”.
Deming also warns against using statistical theory to study control chart behavior (false-alarm probability, OC-curves, average run
lengths, and normal curve probabilities. Wheeler (see “Shewhart’s Charts: Myths, Facts, and Competitors”, ASQC Quality Congress
Transactions (1992), Milwaukee, WI, pp. 533–538) also shares some of these concerns:
“While one may mathematically model the control chart, and while such a model may be useful in comparing different statistical
procedures on a theoretical basis, these models do not justify any procedure in practice, and their exact probabilities, risks, and
power curves do not actually apply in practice.”

On the other hand, Shewhart, the inventor of the control chart, did not share these views in total. From Shewhart (Statistical
Method from the Viewpoint of Quality Control (1939), U.S. Department of Agriculture Graduate School, Washington DC, p. 40, 46):
“As a background for the development of the operation of statistical control, the formal mathematical theory of testing a statistical
hypothesis is of outstanding importance, but it would seem that we must continually keep in mind the fundamental difference
between the formal theory of testing a statistical hypothesis and the empirical theory of testing a hypothesis employed in the
operation of statistical control. In the latter, one must also test the hypothesis that the sample of data was obtained under
conditions that may be considered random. …
The mathematical theory of distribution characterizing the formal and mathematical concept of a state of statistical control
constitutes an unlimited storehouse of helpful suggestions from which practical criteria of control must be chosen, and the general
theory of testing statistical hypotheses must serve as a background to guide the choice of methods of making a running quality
report that will give the maximum service as time goes on.”

Thus Shewhart does not discount the role of hypothesis testing and other aspects of statistical theory. However, as we have noted
in the text, the purposes of the control chart are more general than those of hypothesis tests. The real value of a control chart is
monitoring stability over time. Also, from Shewhart’s 1939 book, (p. 36):

“The control limits as most often used in my own work have been set so that after a state of statistical control has been reached,
one will look for assignable causes when they are not present not more than approximately three times in 1000 samples, when the
distribution of the statistic used in the criterion is normal.”

Clearly, Shewhart understood the value of statistical theory in assessing control chart performance.

My view is that the proper application of statistical theory to control charts can provide useful information about how the charts
will perform. This, in turn, will guide decisions about what methods to use in practice. If you are going to apply a control chart
procedure to a process with unknown characteristics, it is prudent to know how it will work in a more idealized setting. In general,
before recommending a procedure for use in practice, it should be demonstrated that there is some underlying model for which it
performs well. The study by Champ and Woodall (1987), cited in the text, that shows the ARL performance of various sensitizing
rules for control charts is a good example. This is the basis of the recommendation against the routine use of these rules to
enhance the ability of the Shewhart chart to detect small process shifts.

5.3.
Discuss type I and type II errors relative to the control chart. What practical implication in terms of process operation do these two
types of errors have?

Relative to the control chart, the type I error represents the probability of concluding the process is out of control when it isn't,
meaning a plot point is outside the control limits when in fact the process is still in control. In process operation, high frequencies
of false alarms could lead could to excessive investigation costs, unnecessary process adjustment (and increased variability), and
lack of credibility for SPC methods.

The type II error represents the probability of concluding the process is in control, when actually it is not; this results from a plot
point within the control limits even though the process mean has shifted out of control. The effect on process operations of failing
to detect an out-of-control shift would be an increase in non-conforming product and associated costs.

5.4.
What is meant by a process that is in a state of statistical control?

The statement that a process is in a state of statistical control means that assignable or special causes of variation have been
removed; characteristic parameters like the mean, standard deviation, and probability distribution are constant; and process
behavior is predictable. One implication is that any improvement in process capability (i.e., in terms of non-conforming product)
will require a change in material, equipment, method, etc.

5.5.
If a process is in a state of statistical control, does it necessarily follow that all or nearly all of the units of product produced will be
within the specification limits?

No. The fact that a process operates in a state of statistical control does not mean that nearly all product meets specifications. It
simply means that process behavior (mean and variation) is statistically predictable. We may very well predict that, say, 50% of the
product will not meet specification limits! Capability is the term, which refers to the ability to meet product specifications, and a
process must be in control in order to calculate capability.

5.6.
Discuss the logic underlying the use of three-sigma limits on Shewhart control charts. How will the chart respond if narrower limits
are chosen? How will it respond if wider limits are chosen?

The logic behind the use of 3-sigma limits on Shewhart control charts is that they give good results in practice. Narrower limits will
result in more investigations for assignable causes, and perhaps more false alarms. Wider limits will result in fewer investigations,
but perhaps fewer process shifts will be promptly identified. Sometimes probability limits are used - particularly when the
underlying distribution of the plotted statistic is known. If the underlying distribution is unknown, care should be exercised in
selecting the width of the control limits. Historically, however, 3-sigma limits have been very successful in practice.

5.7.

What are warning limits on a control chart? How can they be used?
Warning limits on control charts are limits that are inside the control limits. When warning limits are used, control limits are
referred to as action limits. Warning limits, say at 2-sigma, can be used to increase chart sensitivity and to signal process changes
more quickly than the 3-sigma action limits. The Western Electric rule, which addresses this type of shift is to consider a process to
be out of control if 2 of 3 plot points are between 2 sigma and 3 sigma of the chart centerline.

5.8.
Discuss the rational subgroup concept. What part does it play in control chart analysis?

The concept of a rational subgroup is used to maximize the chance for detecting variation between subgroups. Subgroup samples
can be structured to identify process shifts. If it is expected that a process will shift and stay at the new level until a corrective
action, then sampling consecutive (or nearly) units maximizes the variability between subgroups and minimizes the variability
within a subgroup. This maximizes the probability of detecting a shift.

5.9.
When taking samples or subgroups from a process, do you want assignable causes occurring within the subgroups or between
them? Fully explain your answer.

I would want assignable causes to occur between subgroups and would prefer to select samples as close to consecutive as
possible. In most SPC applications, process changes will not be self-correcting, but will require action to return the process to its
usual performance level. The probability of detecting a change (and therefore initiating a corrective action) will be maximized by
taking observations in a sample as close together as possible.