You are on page 1of 9

A Graphical Exploration of SPC

Editor’s note: This article is the first of a two-part illustrated discussion of the definitions and practices of process control, especially
those practices that are supported by the construction and interpretation of Shewhart control charts. The first article addresses issues
related to the structure of control charts. The second article (to be published in June) will focus on the interpretation of control
charts, with special emphasis on the sensitivity of seven rules that are commonly used to help quality professionals “hear the voice of
the process.”

I
T IS WELL KNOWN THAT DURING THE 1920S, not a statement about his intellectual ability
Walter Shewhart made significant contribu- (although his brilliance is unquestionable). Rather,
tions to a scientific basis for economically it is a statement about the intellectual poverty of
controlling the quality of manufactured the quality sciences during the past 65 years. Not
products. In 1931, he published Economic only has the technology of the quality sciences
Control of Quality of Manufactured Product, in changed very little at the level where it matters
which he described many of the statistical tools he most—namely, at the in-plant, practical (or
developed and used during the previous decade.1 applied) level—but many of those who are respon-
The content of that book is the foundation of what sible for using the methodology to improve quality
is now called statistical process control (SPC). often struggle to understand even the most basic
Part 1: SPC’s Imagine an automobile design engineer from statistical and probabilistic principles.
the 1930s walking into a design laboratory in For example, in training workshops devoted to
definitions 1996, with the intention of putting in a day’s work. elementary SPC—workshops that are frequently
and After recovering from the shock of encountering attended by engineers and quality professionals—
so much new equipment, new technology, new it is rare to find anyone who can provide a reason-
procedures materials, new manufacturing principles, much ably precise definition of an in-control process,
greater vehicle complexity, and substantially describe the fundamental decomposition of the
changed customer expectations, it would probably total variation of process output into its component
take the engineer several years before he or she parts, or understand the basic principles underlying
could make significant contributions to the process the use of control charts to signal the presence of
of designing a modern automobile. special-cause variation. When supervisors and
Now imagine an industrial physicist from the managers are quizzed about the fundamental prin-
by 1930s magically transported to a physics laborato- ciples of process control, their understanding of the
Robert W. Hoyer ry today. Again, even if the physicist successfully basic issues is practically nonexistent. (If this claim
and adjusted to the vast physical differences between is difficult to believe, engage any randomly select-
the two labs, it would take years of education and ed colleague from any of the previously mentioned
Wayne C. Ellis training to enable the person to make even the groups in a discussion of process control and
most basic contributions to his or her new col- observe just how quickly the notions of process
leagues’ activities. control and process capability become hopelessly
It is easy to envision similar scenarios in biolo- entangled.)
gy, chemistry, communications, telecommunica- The intention of this two-part series is certainly
tions, atomic energy, or aeronautics; the differ- not to critique the quality sciences. Rather, the
ences between state-of-the art expertise 65 years objective is to examine the structure, use, and
ago and today is so remarkable that it is difficult to abuse of the X chart as a tool for tracking the cen-
imagine anyone making the intellectual transition ter of a process over time and to define and discuss
in a reasonable amount of time. On the other hand, the probability structure of the standard rules used
if Shewhart were magically transported from his to identify out-of-control processes. This will be
job at Bell Labs in 1931 to any quality control done by:
department in the world today, he might find his • Defining and discussing what is meant by an in-
first few days on the job a bit disconcerting and it control process (covered in part 1)
might take him a week to learn how to use his • Defining the basic Shewhart control chart and
computer effectively, but he would certainly be in describing the constraints for its use (part 1)
his element in less than two weeks. • Discussing rules that signal the presence of spe-
This conjecture about Shewhart’s behavior is cial causes of process variation (part 1)

/
Quality Progress May 1996 65
Using the graphical equivalent of a strobe light, the behavior of
Figure 1. Coil Spring Compression Output the process is captured in eight different situations, as shown in
Figures 2 through 9. These figures show the compression dis-
tances of the springs, but the springs are being processed by
eight different machines. The output of the machines is sup-
posed to be identical, but a close examination reveals that it is
not. Hence, there are actually eight different processes.
The process depicted in Figure 2 is in control and on target.
The fact that the process is on target is interesting but not actu-
ally relevant to its control status. Notice that the process dis-
played in Figure 3 is also in control, even though it is substan-
tially off target.
The process pictured in Figure 4 is out of control. Over time,
it changed in a way that caused the compression distance to
Figure 2. In-Control Process decrease. The variation of process output remained constant,
but the process center gradually migrated in the direction of
smaller values. Perhaps the temperature of the oven in which
the springs were tempered gradually increased, causing an asso-
ciated decrease in the average compression distance of the steel
coils. Typically, this is described as tool wear, an effect that
causes process output to gradually change as the tool wears out.
The process displayed in Figure 5 is also out of control but
for an entirely different reason. The sudden shift in the center of
the process from one value (in this case, the target value) to a
larger value was the result of a sudden shock to the process.
Perhaps process input was changed; the new steel bars from
which the springs were manufactured might have had a greater
lead content than did the steel bars that were purchased from a
different supplier.
The out-of-control behavior of the process depicted in
Figure 6 suggests a cyclical pattern that might be the result of
changes in humidity in the plant. For example, it might be the
case that, as the humidity increases during the day, the compres-
sion distance of the springs produced at that time decreases; all
other things being equal, as the humidity decreases at night, the
compression distance of springs made at that time increases. In
• Presenting the sensitivities (probabilities) of the rules used to any event, there is a distinct change in the center of the process
detect special causes (part 2) over time.
• Examining the robustness of the X chart (part 2)
• Suggesting how to bend the rules when certain assumptions Figure 3. In-Control, Off-Target Process
required for using an X chart are not satisfied (part 2)

Process control
To illustrate the principles of process control, a hypothetical
process whose output is a small coil compression spring will be
examined. The process characteristic of interest is the compres-
sion distance, X, when a 50-lb. load compresses the spring. The
lower specification limit (LSL) and upper specification limit
(USL) for X are LSLX = 5 cm and USLX = 7 cm. The target
value, T, is TX = 6 cm. The distribution of process output for
this characteristic at any particular moment in time is the nor-
mal distribution displayed in Figure 1.
Notice that the process is somewhat off target. Its process
capability (Cpk) is fair (Cpk = 1.33), but it could be improved to
Cpk = 1.67 by centering it at X = 6 cm and without having to
use procedures designed to reduce the variation of the compres-
sion distances. While the capabilities of both the process and
process output are interesting, it is the control status of the
process that is the focus here.
To determine whether the coil compression spring process is
in control, its behavior must be assessed in a time continuum.
66 /
Quality Progress May 1996
acteristic has a normal distribution. This information indicates
Figure 4. Out-of-Control Process Trend something about the process characteristic but, because there
are infinitely many normal distributions, the particular normal
distribution of interest is not known. But suppose that, in addi-
tion to knowing the distribution is normal, one knows that the
process’s mean () is X = 6.2 cm. While this sheds even more
light on the distribution of the process characteristic, there are,
nevertheless, infinitely many normal distributions that are cen-
tered at X = 6.2 cm. If one also knows that the standard devia-
tion () of the process characteristic is X = 0.2 cm, there is
only one normal distribution that fits that description (see
Figure 1). In other words, a normal distribution is completely
characterized by its mean and variance or by its mean and stan-
dard deviation.

Figure 5. Out-of-Control Process Shift

The process shown in Figure 7 is a quality nightmare.


Concurrent with an unwanted decrease in the center of the
process over time is an unwanted increase in process variation.
In this case, the process owner should search for root causes
that induce trends rather than shifts. This type of process behav-
ior, while not uncommon, is very difficult to detect and remedy.
Being out of control isn’t always bad. For example, the
processes displayed in Figures 8 and 9 demonstrate patterns of
process improvement even though both are out of control. In
other words, there is often a positive explanation for being out
of control. W. Edwards Deming often stated that any company
engaged in continuous improvement strategies will be able to
substantiate its improvements by observing a large number of
associated out-of-control processes.2
At this point, a number of processes have been examined and
classified as either in control or out of control. Based on this Figure 6. Out-of-Control Cyclical Process
examination, the following definition of “in control” can be
made:
A process is in control if and only if the distributions of all
of its measurable characteristics are constant over time.
Because many important probability distributions—including
normal distributions—are completely characterized (i.e., deter-
mined) by their mean and variance, it is possible to use the fol-
lowing less precise but more practical definition:
A process is in control if and only if the centers of all of its
measurable characteristics are constant over time and the
variances of all of its measurable characteristics are con-
stant over time.
Do not, however, confuse the issues of process control and
process predictability. If a process is in control, it follows that,
subject to the constant variation of the distribution of its output
at each moment in time, its behavior will be predictable. On the
other hand, many processes that exhibit predictable behavior are
simply not in control (Figure 6, for example). Although this
issue is much more complex than is suggested here, the second,
less precise definition is adequate for most quality profession-
als’ purposes.
Suppose it is known that a certain measurable process char-

/
Quality Progress May 1996 67
Figure 7. Nightmarish Out-of-Control Process Trend Figure 8. Out-of-Control Process Improvement

Finally, if one knows that the process characteristic is nor-


mally distributed, that its mean is X = 6.2 cm, that its standard
Figure 9. Another Out-of-Control Process Improvement
deviation is X = 0.2 cm, and that it is in control, then one
knows precisely how that characteristic is behaving at the pre-
sent time and for as long as the process remains in control.

Variance decomposition
To illustrate the concept of variance decomposition, the out-
of-control process depicted in Figure 6 will be revisited. To
evaluate 100% of the process output during this time period, the
output has been projected to the upper right-hand corner of the
illustration in Figure 10. Because the projected distribution
resides in the margin of the display, it is called the marginal dis-
tribution of process output. This projection depicts the total
variation of process output during the time period of the display.
By comparing this total variation with the smaller variation of
process output at a specific point in time (see Figure 11), it is
easy to see that:
[Total variation] = [Variation at a specific point in time]
+ [Some nonnegative number]
Furthermore, it is apparent that the more out-of-control the
process is, the larger the nonnegative number must be in order
to maintain this equality. If a process were in control (such as patterns of the statistic with the occurrence of one or more spe-
that in Figure 2), the projection of all process outputs to the cial causes of variation.
upper right-hand margin would yield a combined output whose
total variation is identical to the variation at a specific point in The control chart
time. In other words, the nonnegative number would be zero. Any quality professional who is responsible for evaluating
This equation, called the fundamental equation of SPC, can processes could have easily determined the control status of
be expressed as follows: those depicted in Figures 2 through 9. Unfortunately, most qual-
[Total variation] = [common-cause variation] + [special- ity professionals don’t have such a complete picture of the
cause variation] process at their disposal. Usually, the process is producing
Since the total variation of process output can be decom- numerous units of output, and quality professionals attempt to
posed into the component parts of common- and special-cause track the distributions of the measurements of important charac-
variation, control charts based on that decomposition can be teristics by sampling a small number of units at specifically des-
used to help bring processes into control. A control chart is a ignated time or production intervals (e.g., every hour or every
picture of the common-cause variation of a process. Having 5,000 units). For example, the recommended sample size, n, for
such a picture enables the process owner to associate unusual variable data is usually between n = 1 and n = 10. The fact that
68 /
Quality Progress May 1996
Figure 10. Total Variation of Process Output
Two Common Misconceptions About SPC
Most quality professionals recognize the following equa-
tion as the fundamental equation of statistical process con-
trol (SPC):
[Total variation] = [common-cause variation] +
[special-cause variation]
In Out of the Crisis, W. Edwards Deming said 94% of the
total variation of processes that he encountered could be
attributed to common causes, with the remaining 6% being
due to special causes.1 Based on this conjecture, he con-
cluded that 94% of all process variation is the responsibility
of management, with the balance being the responsibility
of process owners (production workers).
Deming’s observation that most reductions in common-
cause variation necessitate a significant investment of
resources and, hence, depend on higher-level decision
making is appreciated. Nevertheless, it is easy to imagine
that he purposely exaggerated the focus of responsibility
in this instance to make a very important point: that
process owners can be held responsible for only a small
portion of the total variation in a typical production environ-
ment. On the other hand, achieving quality superiority
using continuous improvement is invariably an all-one-
SPC provides a great deal of information about processes with team endeavor, so to suggest that variance reduction ini-
such a meager investment in data is the primary justification for tiatives can be partitioned and assigned to different strata
its status as one of the most powerful and efficient process in the organizational hierarchy distorts an equally important
improvement tools. point: that variance reduction is the responsibility of the
In any event, if the distribution of measurements of a product entire team.
or service characteristic can be completely determined by its Another implication of Deming’s numbers—assuming
mean and standard deviation (or by its mean and range), then at that they describe the typical process today—concerns
the designated time or production interval, relatively small sam- the effectiveness of process control as a continuous
ples of data can often be used to determine how the process is improvement strategy. Because the whole point of process
behaving. Figure 12, for example, is the distribution of the control (of which SPC is a significant part) is to remove all
weight (in kilograms) required to compress the spring a fixed root causes of special-cause variation, SPC actually
distance (in this case, 5 cm). The natural variation in the attacks only 6% of the total variation in a typical process. In
weights is the common-cause variation. Assuming that the other words, if SPC were 100% effective, the typical
process would retain more than 90% of its original varia-
process output is fairly uniform over time, samples of n = 5
tion.
springs are taken every 30 minutes and the mean, X , and stan-
Some companies that have implemented SPC have
dard deviation, sX, are computed for each sample. Figure 13 complained that they made significant investments in
process control but seemed to have gotten very little in
Figure 11. Comparison of the Total Variation With the return. They had unrealistic expectations. Other compa-
Variation at a Specific Point in Time nies that have implemented SPC, however, believe that it is
the critical difference between marginally acceptable qual-
ity and outstanding quality.
Reference
1. W. Edwards Deming, Out of the Crisis (Cambridge, MA:
Massachusetts Institute of Technology, Center for Advanced
Engineering Study, 1986), pp. 314-316.

contains the five observations per sample (the white dots) and
the means of the samples (the black dots) over 10 time periods
from a process that is in control.
While it is difficult to depict the standard deviations in this
illustration, one could fairly easily envision each sample’s
range, RX, as the largest distance between a pair of white dots in
the sample. Notice, for example, that the range of the third sam-
ple is fairly small and the range of the fourth sample is relative-
ly large.

/
Quality Progress May 1996 69
Don’t think that constructing these marginal distributions is
Figure 12. Distribution of Compression Weights an esoteric statistical exercise that has no practical value.
Marginal distributions are valuable for three reasons:
1. Customers want a picture of the marginal distribution of X
because that is precisely what will be shipped from the suppli-
er’s plant (or will be delivered to the next worker on the assem-
bly line). While the supplier cares about the behavior of its
processes, customers do not purchase processes—they purchase
process output.
2. Knowledge of the marginal distribution of X —including
knowledge of the mean and standard deviation of X—is instru-
Figure 13. Samples (n  5) and Means From an mental in the construction of the X chart. One should be careful
In-Control Process to avoid casual use of terms like “three sigma.” There are so
many important and different standard deviations (sigmas) in a
basic discussion of SPC that misunderstandings will result if
someone fails to ask, “To which three sigma are you referring?”
In this discussion, the standard deviation of X is of interest. In
other discussions, the standard deviation of RX, the standard
deviation of p, or the common-cause standard deviation might
be of interest. For that reason, it is always important to complete
the phrase, “the standard deviation of ________.”
3. There is no possibility that anyone could understand the
usual rules for reading and interpreting a control chart without
understanding the marginal distribution of the statistic that is
plotted in the chart. (This topic will be discussed in much
greater detail in the second article next month.)
To track the center and dispersion of the in-control process
displayed in Figure 14, simple time-series plots of X and sX are
used (see Figure 15). Standard practice dictates that certain
guidelines be used to interpret the information in the chart.
Naturally, the mean, X , of the values of X that have been plot-
ted in the chart serves as a standard against which to compare
those values. In addition, for no better reason than it was
Shewhart’s choice in the 1920s, a distance of three standard
In a real-world situation, it is usually impossible to observe deviations of the statistic on either side of the mean defines the
the distribution of X (i.e., the measurements of process output outer boundaries of acceptable values of the statistic when the
that have been depicted in Figure 13). The only available infor-
mation about the process is the 50 white dots, so the quality pro- Figure 14. Data, Means, and Marginal Distributions
fessional must use that relatively small amount of information to From an In-Control Process
determine how the entire process is behaving. In Figure 14, only
the sample data and the means have been displayed for each of
the 10 time periods. (Although the actual marginal distributions
of X and X would be unknown to the quality professional, they
are known in this situation because this is a constructed exam-
ple.)
To appreciate the nature of the two marginal distributions,
imagine that this process has been in control for a long time. At
each time period, a sample of n = 5 measurements has been
obtained and plotted. The sample mean, X , has also been calcu-
lated and plotted. Now, if all of the values of X (the white dots)
are pushed up the target axis to the upper right-hand margin and
stacked and if a smooth curve is drawn through them, that curve
would be the marginal distribution of X. In this instance, the
marginal distribution of X is identical to the distribution of
process output at each time period (see Figure 13) because the
process is in control (i.e., there is no special-cause variation).
The marginal distribution of X is obtained in a similar man-
ner. All of the X values (the black dots) are pushed up the target
axis to the upper right-hand margin and stacked. Then, a
smooth curve is drawn through the stack. This curve is the mar-
ginal distribution of X .
70 /
Quality Progress May 1996
pling distribution of B has a mean, B, and standard deviation,
Figure 15. X and s Chart for Compression Weight B. Since there are usually relatively few sample observations
available, B and sB are calculated from the sample data, and
these values serve as estimates of B and B, respectively.
Now, according to Shewhart, “We usually choose a symmet-
rical range characterized by limits, B  tB, symmetrically
spaced in reference to B....We are still faced with a choice for
the value of t. Experience indicates that t = 3 seems to be an
acceptable economic value. Hence the method for establishing
allowable limits of variation in a single statistic, B, depends
upon theory to furnish the expected value, B, and the standard
deviation, B, of the statistic. It depends upon empirical evi-
dence to justify the choice of limits, B  3B.”3
In other words, the limits in all control charts have been set at
B  3B because, in 1931, Shewhart’s experience indicated
that it was “an acceptable economic value.” While it is probably
not worth making an issue of it, 3B is a rather peculiar choice,
especially since it would have been more intuitively plausible
and less misleading to use a probability confidence interval
about the center of the statistic (say, a 99% confidence interval
or a 99.8% confidence interval about B). Instead of making
control limits symmetric in distance from the center as mea-
sured in standard deviations, it would have been somewhat
more useful to have made them symmetric in distance from the
center as measured in probability. Then, for every control
chart—no matter what statistic is plotted in it—the probabilities
associated with each of the rules for detecting special causes of
variation would be comparable.

Figure 16. Normal Distribution Probabilities

process is in control. Virtually all control charts used in industri-


al and service settings share the same characteristics.
To avoid confusion when important characteristics of a con-
trol chart are discussed later, it is useful to describe exactly what
is meant by the term “control chart.” The description here will
be framed in terms of a generic “Blob,” or B, chart. As is the
case with all charts, the B chart gets its name from the fact that:
• Samples are obtained from a process at specific points in
time.
• From each sample, a value of B (whatever that is) is
computed.
• The values of B are plotted in the chart.
Specifically, the generic B chart consists of five lines:
• A vertical scale of B values
• A horizontal scale that designates sequential points in time or
in a progression of the number of units produced
• A centerline (CL), where CLB = B
• A lower control limit, where LCLB = B  3sB
• An upper control limit, where UCLB = B + 3sB
Calculations to determine the values of the chart’s centerline
and control limits must use data that have been obtained when
the process was not subject to special-cause variation. With that
in mind, it should be apparent that a control chart for a process
characteristic is a picture (a model) of the common-cause varia-
tion of the process characteristic.
Using terminology similar to that used by Shewhart, the sam-

/
Quality Progress May 1996 71
Figure 17. Distributions of X for Normal, Uniform, and Exponential Process Distributions

Normal distributions other words, knowing only two parameters—the mean and
In many (but hardly all) circumstances, the marginal distribu- standard deviation—provides complete knowledge of the
tion of the statistic plotted in a control chart is very close to a distribution. Two normal distributions with the same mean
normal distribution. For that reason, it is useful to review some and standard deviation are identical in every respect.
of the characteristics of normal distributions: • Normal distributions are unimodal and symmetric.
• There is an infinite number of different normal distributions. Therefore, the mean and median of a particular normal distri-
• The normal distribution is a two-parameter distribution. In bution are identical.
72 /
Quality Progress May 1996
• Normal distributions are bell-shaped. (The importance of this distributions of measurements of plate glass strength, nickel
fact, however, should not be exaggerated.) There are many plating thickness, tank leakage, or length of time required to
other distributions besides normal distributions that are bell- resolve a customer’s complaint are all likely to be skewed.
shaped (e.g., all of the t distributions). More than a few data • There is no substitute for knowing the data and then being
analysts have been misled and have misled others by erro- able to assess the process and the process output accordingly.
neously assuming that all bell-shaped distributions are nor- Because quality professionals are essentially in the continu-
mal distributions. ous improvement business, it is imperative that they develop
The distribution displayed in the three parts of Figure 16 is a creative ways of displaying and assessing processes.
normal distribution for which a number of useful probabilities Nevertheless, when customers back their trucks up to their
have been displayed. The tick marks on the X axes are at the suppliers’ shipping docks, they expect to load process output,
mean, X, and at points that are one, two, and three standard not the process itself, into their trucks. Quality—defined as
deviations, respectively, on either side of the mean (i.e., at X satisfying or exceeding customer needs and expectations—
 1X, X  2X, and X  3X. dictates that every quality professional who had or will have
The quality professional’s interest in normal distributions a vested interest in that process output should have complete
stems from two important applications: knowledge of what will be loaded into those trucks.
1. Normal theory. This theory states that if the measure-
ments of an in-control process characteristic are normally dis- Until next time...
tributed, then, no matter what sample size is chosen, the distrib- Part 1 of this SPC series started with the unsubstantiated
ution of X based on samples of that size will also be normally claims that the technology of the quality sciences has been
distributed. intellectually dormant for the past 65 years and that the techni-
Even when n = 1, the marginal distribution of the values of cal content of quality improvement is largely misunderstood by
X in an X chart will be normally distributed, provided that the many quality professionals. In the ensuing discussion, an in-
process is normally distributed and in control. Thus, it is not control process was defined, the decomposition of the total vari-
necessary to invoke the central limit theorem in this situation. ation of process output into common-cause variation and spe-
2. Central limit theorem. This theorem states that if the cial-cause variation was described, the basic structure of a con-
measurements of an in-control process are not normally distrib- trol chart was outlined, normally distributed marginal distribu-
uted, then, for a “large” sample size, the distribution of X based tions of process output and control chart statistics were
on samples of that size will be approximately normally distrib- explained, and the normal theory and central limit theorem were
uted. stated and illustrated.
Needless to say, the important, practical question to ask when This discussion sets the stage for an extended examination of
using the central limit theorem is, “How large is ‘large?’” The those principles and procedures that enable quality profession-
answer is certainly not as clear cut as is implied in elementary als to use control charts to hear the voice of the process. That is
statistics textbooks. For example, if the distribution of process the topic of the second article in this two-part series, which will
output is symmetric and unimodal, relatively small samples, appear next month.
say, n = 5, are adequate. On the other hand, if the process output
is seriously skewed (which often occurs in plating processes, for References
example), even samples as large as n = 25 might not be suffi- 1. Walter A. Shewhart, Economic Control of Quality of
cient. Manufactured Product (New York, NY: D. Van Nostrand Company,
The normal theory and central limit theorem are illustrated in Inc., 1931).
2. W. Edwards Deming, Out of the Crisis (Cambridge, MA:
Figure 17. In the first column, the process output is normally Massachusetts Institute of Technology, Center for Advanced
distributed; in the second column, it has a uniform distribution; Engineering Study, 1986), pp. 322-323 and pp. 337-338.
and in the third column, the measured characteristic is exponen- 3. Shewhart, Economic Control of Quality of Manufactured
tially distributed. The three distributions of X displayed in each Product, pp. 275-277.
column are based on samples of size n = 5, n = 10, and n = 25.
The distributions in the first column illustrate the normal theory. Robert W. Hoyer is the president of Decision Dynamics Inc. in Ann
The distributions in the second and third columns demonstrate Arbor, MI. He received a doctorate in statistics from the Virginia
the central limit theorem. Polytechnic Institute in Blacksburg.
It is noteworthy that, had the process output been uniformly Wayne C. Ellis is a professor of statistics at Eastern Michigan
distributed (i.e., not even close to bell-shaped), samples as small University in Ypsilanti. He received a doctorate in mathematics from
as n = 5 would have ensured that X would be normally distrib- the University of Michigan in Ann Arbor.
uted. If the process output had been exponentially distributed,
the distribution of X would still be skewed to the right, even
with samples as large as n = 25.
This discussion of normal distributions can be concluded by What did you think about this article?
noting that: Quality Progress needs your Excellent Circle #321
• Nothing about normal distributions makes them “optimal” as feedback. On the postage-paid
reader service card inserted toward Good Circle #322
compared to other distributions of process output. While the
the back of this magazine, please
natural distributions of many processes are normally distrib- circle the number that corresponds Fair Circle #323
uted, there are thousands of processes whose measurable with your opinion of the preceding Poor Circle #324
characteristics are not normally distributed—and some of article.
those distributions are seriously skewed. For example, the

/
Quality Progress May 1996 73

You might also like