Source:Lloyd. R.

Quality Health Care:

CHAPTER 9 A Guide to Developing and Using Indicators

2nd Edition, Jones & Bartlett Learning, 2019.

Understanding Variation
with Shewhart Charts
The details related to these differences are dis-
▸ Run Charts versus cussed in the remaining sections of this chapter.

Shewhart Charts
For many teams just beginning their quality mea- ▸ What Is a Shewhart
surement journey (QMJ) the run chart provides
an excellent starting point. It is easy to construct Chart?
with paper and pencil, it does not require a soft-
Like run charts, Shewhart charts are graphic
ware package in order to make one, and it can
displays of process variation as it lays itself out
be used with any type of data (i.e., time, money,
over time. FIGURE 91 shows the basic elements of
counts of errors, percentages, rates, scores, or days
a Shewhart chart and one of the tests to identify
between adverse events). Also, the four run chart
a special cause (i.e., a data point exceeded the
rules are easy to understand and apply. So, why
upper control limit [UCL], signaling too much
would I want to use a Shewhart chart instead of a
variation in the data, which, by the way, you
run chart?1 There are basically three reasons why
should recognize as an astronomical data point
Shewhart charts are preferable over run charts:
on the run chart). A run chart and a Shewhart
1. Shewhart charts are more sensitive chart look similar in that the indicator of interest
than run charts. and its values are plotted on the vertical or y
2. Shewhart charts have the added axis and the chronological order of the data are
feature of control limits and zones, organized by what are called subgroups (e.g.,
which run charts do not have. by individual patients, by day, week, or month)
3. Shewhart charts allow us to more along the horizontal or x axis. The data points
accurately predict process behavior, are then connected by a line and the mean of the
future performance, and process data points is then plotted as the centerline (CL)
capability than do run charts. on the Shewhart chart. The presence of control

© Michal Steflovic/Shutterstock

211

9781284023077_CH09_211_258.indd 211 31/07/17 5:59 PM

 212 Chapter 9 Understanding Variation with Shewhart Charts

Signal of a Upper Control

special Limit

60.0

50.0
Number of Patient Complaints

UCL=46.910
40.0 Data are plotted
in time order
30.0
CL=23.381
20.0 Centerline
(the mean)
10.0

0.0 LCL=0.148
Lower Control
-10.0 Limit
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Week

The unit of time is plotted along the horizontal axis

FIGURE 91 Elements of a control

limits on Shewhart charts are major points that data point actual values. By using the mean we
separate it from a run chart. are ensuring that the absolute value and the
Shewhart charts are more sensitive than distance of each data point from the CL will be
run charts because the run chart cannot detect considered in determining the variation in the
special causes that result from point-to-point indicator and if special cause variation exists.
variation. This is because the CL on the run Another reason why Shewhart charts are
chart is the median (i.e., the 50th percentile). more sensitive than a run chart is that Shewhart
The run chart basically allows you to classify charts have the added feature of control limits,
the data points as being only above or below which run charts do not have. The control
the median. The actual distance a data point limits are properly referred to as the UCL and
is from the CL is not an issue on a run chart. the lower control limit (LCL). They are also
Therefore, if one data point is 2 units above referred to as sigma limits. You will probably
the median and another point is 22 units hear someone refer to control limits, however,
above the median, they will both be treated as confidence intervals, confidence limits, or
the same because they are both on the same even standard deviation (SD) limits, which they
side of the median. The logic for this decision are not (Blalock, 1960; Carey, 2003; Daniel &
is related to the definition of the median and Terrell, 1989; Provost & Murray, 2011).
of a run (i.e., one or more data points on the The UCL and LCL basically define the
same side of the median). If these same two boundaries of process variation around the
data points (i.e., 2 and 22) were placed on a mean. The developer of the chart does not set or
Shewhart chart, however, you would notice a define the UCL and LCL. These are determined
discernable difference because the CL on the by mathematical formulae and the width of these
control chart is the mean or average of all the limits is dependent on the inherent variation

9781284023077_CH09_211_258.indd 212 31/07/17 5:59 PM

 What Is a Shewhart Chart? 213

that lives within the data. The only thing the UCL is 47 minutes, the lower control limit is
developer of the chart can place on the Shewhart 23 minutes and there are no special causes
chart is a target or goal and annotations as to detected. This means that the process is a stable
when improvements were introduced. and predictable. Therefore, if we do nothing
The control limits enable the Shewhart to change how this process works we can
charts to have increased precision over the run predict that patients will wait on the average
chart. A run chart will miss certain nonrandom 35 minutes with the possibility that the wait
patterns that would be detected on a Shewhart time could go up as high as 47 minutes or as
chart as special causes. According to Perla, low as 23 minutes. In light of the target of
Provost, and Murray (2011, p. 47), “The three having all patients seen by their doctor within
probability-based (run chart) rules are used to 20 minutes or less, however, you can see that
objectively analyze a run chart for evidence of we have our work cut out for us!”
non-random patterns in the data based on an This scenario provides a summary of how
α error of p < 0.05.” This means that run charts process capability for the wait time in a clinic
could miss a nonrandom pattern in the data can be based on the parameters calculated for a
approximately 5% of the time. Shewhart chart Shewhart chart (i.e., the UCL, LCL, and mean).
rules, on the other hand, will not miss detecting Classically, process capability is defined as, “The
a special cause. This is why it is recommended calculated inherent variability of a characteristic
that the terms special and common cause as well (indicator) of a product or service. It represents
as stable or unstable should be reserved for use the best performance of the process over a period
only with Shewhart charts and that the terms of stable operation” (ASQ, 2005, p. 78). Process
random and nonrandom patterns be applied capability is essentially aimed at determining
to run charts. whether under current operating conditions the
Shewhart charts also allow us to more process can meet the predetermined specifications
accurately predict process behavior and future or achieve the target or goal we have established
performance than do the run charts. On a (Blank, 1998; Carey, 2003; Kume, 1985; Provost &
run chart, if the variation is random the best Murray, 2011; Western Electric, 1985; Wheeler &
prediction of the future performance of an Chambers, 1992).
indicator is the median value. For example, Besides a verbal summary of the Shewhart
if a team is trying to improve the wait time chart parameters using the UCL, LCL, and mean
to see a doctor and have plotted the data on as described previously, process capability can
a run chart the median is the best estimate of also be defined statistically by “a single number
future performance. Let’s say that the median assessment of the ability of the process to meet
wait time is 27 minutes. If you were present- specification limits on the quality characteris-
ing this data to a team or a committee all you tic(s) of interest (ASQ, 2005, p. 78). When you
could say would be, “Ladies and gentlemen, the move to the statistical indices that capture pro-
median wait time is 27 minutes. The process cess capability it is necessary to have an upper
reflects only random variation. Therefore, if specification limit (USL) and a lower specifi-
we do nothing to change the current process cation limit (LSL), which are then compared
we can expect to have patients wait about 27 to the performance of the process as defined
minutes to see the doctor.” On a Shewhart by the UCL, LCL, and the mean.2 Although
chart, however, because we have the UCL, these indices have not been used extensively
LCL, and the mean as the CL, we have more in healthcare settings I believe that they have
precision. In this case, when you present the great utility. We have many physiological tests
data to the team or a committee you would be that have upper and lower preferred levels of
able to say, “Ladies and gentlemen, the average performance (i.e., specification limits). These
wait time to see a doctor is 35 minutes, the include such indicators as temperature, blood

9781284023077_CH09_211_258.indd 213 31/07/17 5:59 PM

 214 Chapter 9 Understanding Variation with Shewhart Charts

pressure, hematocrits, neutrophils, white and the interpretation of the chart and what can be
red blood cell counts, platelets, and clotting learned from it must come from the dialogue
factors.3 that emerges when people with subject matter
There are many useful books and articles on knowledge interpret the chart. This requires
the statistical theory behind Shewhart charts, how knowledge not a computer.
to construct them, and how to interpret the results.
I have provided only a brief introduction to the
key principles behind the Shewhart charts. This is
a very rich field of study that has been developed
▸ Key Questions about
over the past 100 years. Readers interested in Shewhart Charts
the detailed aspects of statistical process control
(SPC) and in particular Shewhart charts should There are three basic questions people typically
consult the rich variety of books and articles on ask as we start them on the road to using She-
this topic. Ones I have found particularly useful whart charts:
include Benneyan (2001); Benneyan, Lloyd, and 1. How many data points do I need to
Plsek (2003); Blank (1998); Carey (2003); Carey make a Shewhart chart?
and Lloyd (2001); Duncan (1986); Ishikawa 2. What is a sigma limit? And, why do
(1989); Mohammed, Worthington, and Woodall I need three of them?
(2008); Montgomery (1991); Provost and Murray 3. Do I apply the run chart rules to
(2011); Pyzdek (1990); Western Electric (1985); Shewhart charts?
Wheeler (1993, 1995); Wheeler and Chambers
(1992); and Woodall (2006). Each of these is discussed next.
If you wish to build a firm foundation in
Shewhart charts and SPC in general, I would How Many Data Points Do I Need
recommend that you read widely on this topic
and read what different authors have written. If to Make a Shewhart Chart?
one item you read seems too academic or math- As soon as the team begins to gather data they
ematical, read another author’s description and should start plotting the data points (dots) on a
use of SPC. As you read more of the literature chart. At first this will simply be a line graph. A
and different authors at some point there will be run chart requires less data because the median
a moment when you say, “Okay, I get it.” Fur- as the CL is not as sensitive to point-to-point
thermore, if you do not have a reasonably solid variation as is a Shewhart chart. Also the run
working knowledge of the theory and mechanics chart rules start to come into play with different
of Shewhart charts and how they are constructed, amounts of data. The trend rule can be detected
it will be rather difficult to successfully apply when you have five or six data points. As Provost
them to your improvement work. This becomes and Murray point out (2011, p. 87) “a trend
even more problematic when people say “No will remain a trend no matter the amount of
problem with the charts. We have software that additional data added to the graph.” The run
makes the charts for us.” This orientation creates chart rules related to a shift and too many or
several problems. Although it is easy to push too few runs, however, require more data to be
a few buttons on your computer and “make a detected. The general rule is that a minimum of
chart” this does not necessarily mean it is the 10 data points is necessary to properly determine
most appropriate chart for the indicator you are whether a shift has occurred on the run chart or
tracking. More important, the SPC software does whether too many or too few runs are present.
not help at all with interpreting what the chart When we move to using Shewhart charts more
is trying to tell you. The chart can come from data are usually required because (1) the mean
the machine in front of you with a keyboard but is now used as the CL and the absolute value of

9781284023077_CH09_211_258.indd 214 31/07/17 5:59 PM

 Key Questions about Shewhart Charts 215

any data point enters into the calculations of the pulled from the process when it is stable
UCL and LCL, and (2) the rules for detecting and predictable.
special causes of variation are more rigorous ■ If you have less than 20–30 subgroups of
and precise than the four rules for run charts. data you can still create a Shewhart chart but
But, the simple answer to the question of how the UCL and LCL should be referred to as
much data I need to make a Shewhart chart “trial” control limits (Carey, 2003; Carey &
is . . . it depends. I know some readers will Lloyd, 2001; Provost & Murray, 2011).
be thinking, “What kind of a lame answer is The trial limits can be used for learning
this? Just tell me how many data points I need but the use of the word “trial” is to remind
to make a Shewhart chart!” Because there are those using the chart that these limits may
many types of Shewhart charts, which are dis- change as more complete data are obtained
cussed in the next section, it must be realized and make the limits more reliable and stable.
that the different charts can be produced with The issue here is that when you have less
differing amounts of data. The subgroup, that than the recommended amount of data
is how you have organized your data along (i.e., 12–15 data points) the control limits
the x or horizontal axis of the chart, is key and CL (i.e., the mean) can change rather
to determining how much data you need to quickly and dramatically with the addition
make a particular type of Shewhart chart. For of each new data point. With a fewer num-
example, if you want to track the wait time of ber of subgroups you also run the risk of
each patient at a family practice clinic to see the committing a type II error (i.e., concluding
doctor then the subgroup is one patient and the that the chart indicates no special causes
one bit of data for this patient will be her wait when in fact one or more special causes do
time to see the doctor. If, on the other hand, exist). When you start to have more than
you decide that you want to track wait time by the recommended 20–30 subgroups of data,
day then the horizontal axis of your chart will say 40–50, you run the risk of committing
have Monday, Tuesday, Wednesday, etc. rather a type I error (i.e., finding special causes
than patient 1, patient 2, patient 3, etc. as the by chance alone). Additional detail on the
subgroup. Selecting day as the subgroup for theory and use of the type I and type II
a clinic could now provide upwards of 30–40 error concepts can be found in Carey and
patients’ wait times as possible observations (bits Lloyd (2001), Carey (2003), and Provost and
of data) within a single day. Having multiple Murray (2011). In summary, the underlying
data points in a subgroup or only one will play question here is how much data do you
a major role in deciding which Shewhart chart need to create a reasonably stable distri-
you can make. This is why it is very important bution? Different disciplines recommend
to make sure you have a well-thought-out data different amounts of data needed to form
collection plan. Again, more will be said about a distribution (e.g., from only a few data
these issues in the next section when I discuss points to over 500) but generally speaking
the types of Shewhart charts. a reasonably stable distribution of data for
All this being said, I do know that many improvement purposes occurs when you
people still want to have at least some general have 20–30 subgroups of data.
guidelines for organizing their data, so here are ■ As a general rule I also recommend not
a few that I offer to improvement teams as we using quarterly data for your improvement
begin to work on developing Shewhart charts:4 efforts. There is just too much variation
being aggregated in quarterly data to be
■ It is usually recommended that you have useful for improvement efforts. A quarter
20–30 subgroups of data before construct- consists of 3 months approximately 90 days
ing a Shewhart chart. These data should be and over 2100 hours. During this time a

9781284023077_CH09_211_258.indd 215 31/07/17 5:59 PM

 216 Chapter 9 Understanding Variation with Shewhart Charts

great deal of variation can occur. So when (i.e., the normal bell-shaped curve), skewed to
someone begins making conclusions about the left, or skewed to the right in which case the
the quarterly average or SD you should ask distribution will have a tail that is longer on one
them to provide the actual variation that side than the other. A distribution with a long
produced these summary statistics by day, tail skewed to the right will have a mean that is
week or at a minimum by month. greater than the median whereas a left-skewed
distribution will have the mean be less than the
median. Kurtosis, on the other hand, refers to
What Is a Sigma? And Why Do how spread out or peaked the distribution is.
For additional details on measures of central
I Need Three of Them? tendency, dispersion, and distribution shape
These two questions probably pose the most you can consult any basic stat book. Some of
challenging technical aspects of Shewhart chart the books I have on this topic come from my
construction. Some of you will really enjoy this undergraduate days and are just as relevant as a
issue and want to learn more whereas others statistics book published last year (Blalock, 1960;
of you will say, “I really don’t care about this Daniel & Terrell, 1989; Gonick & Smith, 1993;
statistical distinction just make the chart that is Levine & Stephan, 2005). These principles are
most appropriate for my indicator and tell me fundamental and have not changed over the years.
what it means.” Either position is fine. I do not FIGURE 92 provides examples of distributions
intend to go into great detail on this topic but with different characteristics of center, dispersion,
I do want to frame it properly so that you can and shape. Note that the normal (or Gaussian)
decide if you want to learn more or accept the distribution, which is popularly referred to as the
fact that these statistical principles have been bell curve, is typically not found in the real world
discussed, debated, and written about extensively of data collection and analysis. In a theoretical
for many decades. normal distribution, the mean, median, and
Let’s start with the basics. Whenever you mode are all at the same position and the data
have an array of data you need to consider three are distributed randomly and symmetrically
characteristics of the distribution these data about the mean. But it needs to be pointed out
create: the central tendency of the distribution, that not all symmetrical bell-shaped curves are
the dispersion or spread of the distribution, normal (Blalock, 1960, p. 80). You can have,
and the shape of the distribution. You were for example, three normal curves that have the
acquainted with these characteristics when same SDs but different means. Similarly, you
you took your first statistics class, which was could have several curves that have the identical
probably a number of years ago. So, this should means but very different SDs that in turn create
all sound rather familiar even if you have not different shapes for the distributions.
used the concepts in a while. Measures of central It is important that you have a comfort level
tendency include the mean (i.e., the arithmetic with the characteristics of distributions so that
average), the median (i.e., the midpoint of the you can more fully understand the character-
distribution or 50th percentile), and the mode istics of the data you have collected and their
(i.e., the most frequently occurring number). potential limitations. Although the Shewhart
Measures of dispersion include the minimum charts can accommodate both normally and
and maximum values, the range (i.e., the absolute nonnormally distributed data (Wheeler, 1995)
difference between the min and max values), having knowledge of the data you have gathered
the sum of the deviations, the mean deviation, is the first step toward creating and interpreting
the sample variance, and the SD. The shape of Shewhart charts.
a distribution is determined by skewness and With a few of the basics about distributions
kurtosis. A distribution can be symmetrical in hand it is now time to address in a little more

9781284023077_CH09_211_258.indd 216 31/07/17 5:59 PM

 Key Questions about Shewhart Charts 217

FIGURE 92 Examples of distributions with different centers, spreads and shapes

detail the topic of sigma limits. As was mentioned Chambers, 1992, p. 60). Wheeler also
earlier in this chapter, the UCL and LCL are prop- points out that the common dispersion
erly referred to as sigma limits or alternatively, statistic for a distribution (i.e., the SD)
estimates of the SD (Provost & Murray, 2011, needs to be converted into sigma units by
p. 115; Wheeler, 1993, 1995). Although some the use of specific formulas. He concludes:
writers will call the UCL and LCL SD limits “By shifting from measurement units (i.e.,
and not even reference the term sigma, I think SD of a distribution) to sigma units, it is
it is important to use the term sigma to refer to possible to characterize how much of the
the estimates of variation in a Shewhart chart data will be within a given distance on
rather than SDs for several reasons: either side of the average. Thus, the sigma
■ This is how Dr. Shewhart (1980, 1986) units express the number of measurement
originally described the limits on the charts units which correspond to one standard
he developed. unit of dispersion” (p. 61)
■ The SD of a distribution is calculated dif- A final point to acknowledge is that if you
ferently than a sigma. The SD is a single calculate the SD of a distribution using the
number that represents the average distance traditional formula that you will find in
any individual data point in a distribution many software packages, multiply this
is from the mean. It cannot be a negative number by 3, and then add and subtract this
number and it will go, theoretically, from value (i.e., 3 SDs) to the mean, you will get
zero to a rather large positive value (but the incorrect UCL and LCL for a Shewhart
generally speaking the SD usually does chart. This becomes even more important
not go much beyond double digits). A when you realize that each type of Shewhart
sigma unit, on the other hand, is a “mea- chart has its own formula to compute sigma
sure of scale for the data” (Wheeler and values and that none of these formulae use

9781284023077_CH09_211_258.indd 217 31/07/17 5:59 PM

 218 Chapter 9 Understanding Variation with Shewhart Charts

the traditional SD formula for a sample based standard deviation is used to describe the
on the following formula: units of variation on a Shewhart chart or the
UCL and LCL in healthcare settings con-

∑ (x − x ) fusion occurs conceptually and statistically.

n 2
i
Sx = i =1 I would strongly encourage you, therefore,
n −1 to use the proper terms when constructing
and explain Shewhart charts. 5
n = The number of data points
x = The mean of the xi Now that we know a little more about what
xi = Each of the values of the Data a sigma is, the next question is, “Why do we need
three of them?” The answer to this question is
Note that if you were calculating the SD for found partially in statistical theory and partially
a defined population the formula would in practicality. According to Wheeler (1995, p. 14)
use N as the denominator (i.e., the total Shewhart’s use of 3 sigma limits (i.e., three above
number of observations in the population the mean and three below the mean for a total of
being observed) rather than n - 1, which is 6 sigma units) as opposed to any other multiple
typically used when calculating a SD for a of sigma did not stem from any specific math-
sample. ematical computation. Rather Shewhart said
If you do not use the appropriate formula that three “seems to be an acceptable economic
for computing a particular chart’s limits, you value,” and that the choice of 3.0 was justified by
will produce limits that are too wide or too “empirical evidence that it works” Provost and
narrow. This will then lead you to make the Murray (2011) provide a succinct summary of
wrong decision about the variation in your the rationale for using Shewhart’s 3 sigma limits:
data (i.e., you will see special causes when
they do not exist and miss them when they ■ The limits have a basis in statistical theory.
are actually present). ■ The limits have proven in practice to distin-
■ The third reason I prefer using the terms guish between special and common causes
sigma units and sigma limits with Shewhart of variation.
charts rather than SD units or SD limits is ■ In most cases, use of the limits will ap-
to avoid confusion. A majority of health- proximately minimize the total cost due to
care professionals have been exposed to overreaction and underreaction to variation
the concept of a SD but not to a sigma. We in the process.
have all been in meetings where someone ■ The limits protect the morale of the workers in
presenting data has proudly said, “We have the process by defining the magnitude of the
analyzed the data from last month for pro- variations that has been built into the process.
cedure X and discovered that the average Provost and Murray’s point about overreaction
length of stay is 4.3 days and the SD is 2.6 and underreaction to variation deserves a few more
days.” When this occurs most participants comments. There are basically two mistakes or
in the meeting either nod their heads and errors that you need to avoid when interpreting
say nothing or mumble a few words to the data. The first mistake (a type I error) is a risk
person sitting next to them about this is just of concluding that a data point requires special
what we heard last month. The SD is a very action when it is actually reflecting common cause
popular statistic presented in healthcare (random) variation. This leads to tampering with
management meetings. People hear the a process that is in fact stable and predictable.
number but most could not explain what Tampering (i.e., reacting one way a data point
it is, how it was calculated, or how to inter- then reacting another way to the next data point
pret it. But, it is a regular part of healthcare when they are part of a process that is stable and
management meetings. So, when the word predictable) leads to increases variation in a

9781284023077_CH09_211_258.indd 218 31/07/17 5:59 PM

 Deciding Whether a Special Cause Is Present 219

High

The combine total risk of a Type I and

a Type II Error is minimized when 3
sigma limits (SLs) are used.
Risk

Low

+/- 1SL +/- 2SL +/- 3SL +/- 4SL +/- 5SL +/- 6SL

FIGURE 93 Balancing the risk of a Type I and Type II error

nonlinear manner making things worse. (See the and Leavenworth (1988), Montgomery (1991),
reference to Rule 4 of the funnel demonstration Shewhart (1931), Wheeler (1995), and Wheeler
in Chapter 7 for more detail on tampering.) Type and Chambers (1992).
I errors happen most often when you decide to
use sigma limits on a Shewhart chart that are less
than three. On the other hand, the second mistake Do I Apply the Run Chart Rules
(a type II error) occurs when you basically do to Shewhart Charts?
the opposite of a type I error. In this case, you
The simple answer to this question is no. The
would conclude that a data point indicates no
four run chart rules (a shift in the data, a trend
need for action when it fact it reflects a special
in the data, too many or too few runs in the data,
cause. Type II errors lead to under controlling
an astronomical data point) should be applied
or what Provost and Murray call underreacting.
only to the run charts. Shewhart charts have
This happens most often when you decide to use
their own rules to determine whether special
sigma limits that are wider than plus and minus 3
causes are present. These rules ae explored next.
around the mean. As Carey and Lloyd (2001,
p. 67–68) point out, “The challenge, therefore, is
to balance the risk of tampering against the risk
of under controlling. In the first case, you will ▸ Deciding Whether
see special causes when they do not exist, and
in the second case, you will miss special causes
a Special Cause Is
when they are present. The combined total risk
of type I and type II errors is minimized when
Present
3 sigma limits are used.” FIGURE 93 provides a Much of the beauty of the Shewhart charts lies
visual of how the total combined risk of two types in their simplicity. They require just enough data
of error is minimized when the limits are set at (about 20 data points) to construct a reliable chart,
+/− 3 sigma. Those of you interested in exploring are easy to read, and allow you to determine
these issues further should refer to the works very quickly whether special cause variation is
of Blumenthal (1993), Deming (1994), Grant present in your data. Shewhart charts, according

9781284023077_CH09_211_258.indd 219 31/07/17 5:59 PM

 220 Chapter 9 Understanding Variation with Shewhart Charts

to Pyzdek (1990, p. 90), “are an operational defi- ■ Irving Burr recommended using no more
nition of a special cause,” which I think is a very than Detection Rules One and Four.
appropriate way to summarize the purpose of ■ Ellis Ott recommended the use of Detection
the charts. Shewhart (1931, p. 6) also captured Rules One, Two, and Four.
the purpose of the charts nicely when he wrote, ■ Lloyd Nelson recommends the routine use
“A phenomenon will be said to be controlled of Detection Rules One and Four, along
when, through the use of past experience, we can with Test 3 (trends) and Test 4 (sawtooth).
predict, at least within limits, how the phenome-
non may be expected to vary in the future.” This The selection of the most appropriate rules,
acknowledges the fact that no one can predict however, should be linked to the subject matter
the exact value of the next data point. But, if you being analyzed, the types of data being collected,
understand the differences between a process and the ability of those who own the processes
being in control (i.e., merely random variation) that produce the outcomes to actually move
and out of control (i.e., detecting special causes the relevant indicators in the desired direction.
in the data) then you will be well on your way to The application of the rules for special
understanding Shewhart’s notion of prediction causes to a Shewhart chart begins by dividing
within limits. He basically argued that in order the chart into zones. The area between the cen-
to understand the variation in a process you terline (CL) (the CL or otherwise known as the
needed to move away from static and aggregated mean or average) and the UCL is divided into
displays of data and look at the process from a three equal areas or zones. Because the control
more dynamic view by plotting the data over limits are referred to as sigma limits, each zone
time and understanding the inherent variability is the equivalent of 1 sigma. The area from the
in the process. Figure 7-3 (Chapter 7) depicts CL to the LCL is divided in a similar manner.
what Shewhart was recommending. These zones are labeled C, B, and A, respectively
For decades the Western Electric Statistical and emanate outward from the CL. FIGURE 94
Quality Control Handbook (1985) has served as provides an example of how a Shewhart chart
the standard reference for the special cause rules. is divided into six zones. The creation of zones
In fact, in many circles and even in several SPC is a very simple process that can be achieved
software packages the rules are frequently referred easily with any reputable SPC software program.
to as the “Western Electric tests for detecting A natural or random pattern of data will
special cause.” Although there are dozens of tests bounce around across the zones, between the
or rules to detect special causes, most experts in UCL and LCL, and include the following three
the field of SPC maintain that only a few of the characteristics:
tests are essential for a basic understanding of
what the charts are trying to tell you. Wheeler
Note: Each zone is equal to 1 sigma
(1995) and Wheeler and Chambers (1992) provide
excellent summaries and critique of the Western
UCL
Electric rules and the variations that have been Zone A +3 SL
proposed by leading SPC experts. Consider the Zone B +2 SL
Measure

following passage from Wheeler (1995, p. 139) Zone C +1 SL

X (CL)
on this issue: Zone C -1 SL
Zone B -2 SL
■ Shewhart used Detection Rule One. Zone A -3 SL
LCL
■ David Chambers remarked that “No data
set could stand up to the scrutiny of all of
Time
the detection rules in the Western Electric
Handbook.” FIGURE 94 Dividing the Shewhart chart into zones

9781284023077_CH09_211_258.indd 220 31/07/17 5:59 PM

 Deciding Whether a Special Cause Is Present 221

■ Most of the data points are near the CL A 3 sigma violation

■ A few of the data points spread out and
approach the UCL and LCL
■ None of the data points (or at least only a
UCL
very rare and occasional point) exceeds the A
control limits (Western Electric, 1985, p. 24) B
C
A natural pattern or random distribution of data CL
will exhibit these three characteristics simulta- C
neously. One of the first signals that a process B
has special causes, therefore, is the absence of A
LCL
any one of these characteristics.
Because these rules for detecting special
causes have grown primarily out of industrial A 3 sigma violation
and manufacturing applications, however, we
need to evaluate them in light of which rules FIGURE 95 Rule #1: A single data point that
are most appropriate in health, education, and exceeds the upper or lower control limit
social services settings. We have done this at
the Institute for Healthcare Improvement (IHI) statistical way to determine whether, in fact, it is
with our colleagues from Associates in Process astronomical. This is the only test that Shewhart
Improvement (API)6 and decided that five of the used to identify special causes and the reason
rules for detecting special causes on a Shewhart why Wheeler (1995) stated that “Shewhart used
chart are most appropriate for these disciplines. Detection Rule One.” Some texts refer to a single
The five IHI/API rules for detecting special causes data point that exceeds 3-sigma as a “freak” point
on a Shewhart chart are: (Pyzdek, 1990). Irrespective of the term being
used, a 3-sigma violation is a clear signal that
Rule 1: 1 point outside the +/− 3 sigma limits the variation of this single point is very different
Rule 2: 8 successive consecutive points above from the variation demonstrated by the rest of
(or below) the CL the data points on the chart.
When you detect a 3-sigma special cause do
Rule 3: 6 or more consecutive points steadily not overreact. The first thing you should do is
increasing or decreasing check the data to make sure that the data point
Rule 4: 2 out of 3 successive points in Zone is legitimate. For example, if someone used a
A or beyond different operational definition for this data point
it may in fact be a false positive. This data point
Rule 5: 15 consecutive points in Zone C on
might also be due to a data collection procedure
either side of the CL
that sampled the population differently than
Each of these rules is discussed next. the other data points. Finally, it could be due to
a stratification problem. In this case, data may
Rule 1: 1 Point Outside the +/− have been pulled from the afternoon shift when
the rest of the chart was based on data sampled
3 Sigma Limits (FIGURE 95) from the day shift. My point is that before you
This is usually referred to as a 3-sigma violation and see a 3-sigma violation as a true special cause,
is classified as a signal of instability in a process. investigate the methods used to gather that data
It is also one of the most easily recognized of all point. If the data point was based on the same
the tests because it is based on a single data point. operational definition as the rest of the data and
On the run chart this was a visual determination there were no sampling or stratification issues
of the “astronomical data point.” Now we have a then you do in fact have a special cause that

9781284023077_CH09_211_258.indd 221 31/07/17 5:59 PM

 222 Chapter 9 Understanding Variation with Shewhart Charts

requires investigation. Why is this data point are observed as a gradual movement of the data
statistically different from the rest of the data? over time, which is demonstrated as a shift in the
The presence of a true special cause provides the process. Ideally this shift would be in the desired
opportunity for learning. direction but the shift could also be in the oppo-
site direction. The data are neutral. They do not
know if they are in the direction of goodness or
Rule 2: 8 Successive Consecutive away from it. This is why it is important that you
Points Above (or Below) the apply the statistical decision rules that allow you
to know when there is a true signal in the data of a
Centerline (FIGURE 96) special cause and when it is just random variation.
People generally find it easy to detect a 3-sigma Although the rule of 8 is a classic Western
violation (Rule 1). But, as the Western Electric Electric rule you will see other alternatives offered
Statistical Quality Control Handbook points out (e.g., 7 in a row, 9 in a row or even one approach
(1985, p. 26) the data can reflect instability even that favors a spread of 8 to 10 in a row). Wheeler
when all the data points fall between the UCL and (1995) lists all the various options defining a
LCL. A shift in the process is one such indication of shift that have appeared over the years and offers
instability. Most writers refer to this rule as “eight commentary on which ones he has seen used most
consecutive data points on the same side of the often. My point in even mentioning these alterna-
centerline.” When such a pattern is observed, it tives is that you will hear a variety of opinions on
signals that there has been a shift in the process. the number of data points used to define a shift
Another way to think of this rule is that it reflects and also on what constitutes a trend (rule 3). The
a run of data that has lingered too long above or challenge is if you define a shift with say seven data
below the mean, which indicates a nonrandom points you may see special causes when they do
pattern. This test is a variation on the run chart not exist (i.e., a type I error). If, on the other hand,
shift rule but you will notice that it requires eight you choose to use 9 or 10 data points as a shift you
data points in a run whereas the run chart rule may fall prone to a type II error, which is missing
required six data points to determine a shift. a special cause when it is present. The rule of 8 has
This test is one of the original Western Electric been regarded as a solid practical rule and it is the
four primary tests and it is a frequent signal on one I and my colleagues at IHI have decided to
healthcare charts. As teams work continuously use. It is neither too lenient nor too conservative
on improvement strategies, their work typically for health and social service application. Unless
produces results that are not immediate and dra- you like to get into rather heavy statistical theory
matic in nature (i.e., the 3-sigma rule) but rather debates about which rule is “the best” I’d suggest
that you accept a set of rules that are practical and
UCL appropriate for your work.
A
B
C
CL
Rule 3: Six or More Consecutive
C
B
Points Steadily Increasing or
A Decreasing (FIGURE 97)
LCL
This rule detects a trend in the data that Provost and
Murray (2011, p. 117) define as “a small, consistent
Too many data points in a row below the
centerline signals a downward shift in the process. drift in a process.” When deciding if a trend exists,
duplicate points (i.e., repeating values) should be
FIGURE 96 Rule #2: 8 successive consecutive points ignored. This rule engenders considerable debate.
above (or below) the centerline First, there is the popular definition of a trend. We

9781284023077_CH09_211_258.indd 222 31/07/17 5:59 PM

 Deciding Whether a Special Cause Is Present 223

UCL
A

C
CL
C

A
LCL

Downward trend Upward trend

FIGURE 97 Rule #3: 6 or more consecutive points steadily increasing or decreasing

see a trend in fashion, a trend in food, a trend in is detected when there is “a series of consecutive
the stock market, which is usually referring to data points without a change in direction.” At the
the fact that the stock market closed higher than IHI we have decided to use the rule of 6 as initially
it started the day. I regularly hear the weather defined by Nelson (1985) and then by Pyzdek
reporter on the Chicago TV stations referring to (1990) as a common practical basis for detecting
a “trend in the temperature.” In this case, the trend an upward or downward trend in the data. The
is usually a comparison of today’s temperature final point I will make about this particular rule
to the average temperature for the past week or is that like Rule 2 this rule engenders considerable
month or the comparison of today’s temperature debate. Wheeler (1995, p. 137), for example, states
to the temperature on the same day a year ago. that “all of these tests (for a trend) are problem-
The point is that there are very popular usages atic.” He offers a number of reasonable statistical
of the word trend and then there are statistical principles as to why he maintains this perspective.
definitions. As we analyze Shewhart charts we Others will argue with you about a trend because
definitely should be using a statistical definition. they are (1) wanting to see a trend, (2) are using
But, I have been in many meetings where people a popular definition of a trend, or (3) have some
interpreting either static or dynamic displays of other statistical reference that says their trend is
data have devised their own definitions of a trend. preferred over the one you propose. So, once again,
Over the next week make a mental note of how unless you are ready for these debates I would
often you hear your coworkers or people in the suggest that you accept the rule of 6 as a trend and
media look at data and declare a trend is present. see how well it fits with your analysis of the data.
People will conclude that there is a “trend” in the
data when in fact they are merely comparing two
data points. If the second data point is higher than Rule 4: Two out of Three
the first and in the direction of goodness then this Successive Points in Zone A
gets labeled as an upward trend. Deming (1992)
had a very good bit of guidance: when you have or Beyond (FIGURE 98)
two data points, “it is very likely that one will be Another of the classic Western Electric rules for
different from the other.” instability is when two out of three consecutive
The Western Electric handbook does not data points are more than 2 sigmas away from the
specify how many data points are needed in order CL. In this particular case, the single data point
to identify a trend. They merely indicate that a trend not in Zone A or beyond can be anywhere on the

9781284023077_CH09_211_258.indd 223 31/07/17 5:59 PM

 224 Chapter 9 Understanding Variation with Shewhart Charts

Rule #2: Two out of three

consecutive data points that fall
in Zone A or beyond

UCL
A

C
CL
C

A
LCL

Rule #2: Two out of three

consecutive data points that fall
in Zone A or beyond

FIGURE 98 Rule #4: 2 out of 3 successive points in Zone A or beyond

chart. The deciding criterion is whether two out of tail(s) of the distribution when you should in
the three successive data points are in Zone A or fact be observing less and less the further you
beyond on the same side of the CL. This is one of go out. There certainly are more complex sta-
the rules that is more difficult to explain in words tistical explanations of why this rule detects a
than pictures. Observing Figure 9-8 will help in special cause. But as Wheeler points out (1995,
understanding this rule. The primary question I p. 135) “this rule provides a reasonable increase
get with this rule, however, is “so what?” “Why in sensitivity without an undue increase in the
is it that two out of three data points in Zone A false alarm rate.”
constitutes a special cause?” First, envision the
static normal curve. Slightly over two thirds of
the data (68.26%) will fall within ± 1 SD of the Rule 5: 15 Consecutive Points
mean. When you go out to ± 2 SDs of the mean in Zone C on Either Side of the
you will find 95.46% of the data. This means
that by the time you are out to ±3 SDs from the Centerline FIGURE 99
mean you should be observing 99.73% of all the This test is generally described as reflecting an
data in the distribution. But, because the normal issue with stratification. Stratification usually
curve theoretically extends infinitely in either indicates that two or more different causal sys-
direction you do not account for 100% of the tems are present in every subgroup. This pattern
data. Now let’s get back to the two out of three of stratification is also known as “hugging the
data points in Zone A of a Shewhart chart. As centerline” because there is a run of 15 or more
you go out the tails of the normal distribution data points within 1 sigma of the CL (i.e., in Zone
you should expect to see less data. The two out C above or below the CL) and the variation is
of three rule, therefore, is signaling that you relatively small for these data points compared
are observing too much data bunching in the to the width of the UCL and LCL. Stratification

9781284023077_CH09_211_258.indd 224 31/07/17 5:59 PM

 Deciding Which Shewhart Chart Is Most Appropriate 225

UCL
A

C
CL
C

A
LCL
FIGURE 99 Rule #5: 15 consecutive points in Zone C on either side of the centerline

occurs most often because the data collection

plan was flawed. For example, you will find a ▸ Deciding Which
stratification pattern when two separate distri-
butions of data have been collected (e.g., day shift
Shewhart Chart
turnaround time was combined with night shift
turnaround time) or the sample of data points was
Is Most Appropriate
drawn from two different distributions of data. Although there is only one way to make a run
Even though we have these rather specific chart, there are numerous ways to make a She-
statistical rules for determining special cause whart chart. The basic design and look of any
variation that are grounded in decades of testing in Shewhart chart is essentially the same as shown
manufacturing settings, I think it is important that in Figure 9-1 (i.e., data plotted over time, the
when we apply these rules to healthcare situations mean of the indicator as the CL and the calcu-
we apply them with a serious dose of common lation of the UCL and LCL). Furthermore, the
sense. For example, if we are trying to improve charts are all grounded in established statistical
food tray delivery time we may be willing to fully theory and are all interpreted in terms of the
accept six data points constantly increasing as a fundamental ideas related to common and
trend. But, on the other hand, if we are dealing special causes of variation. But, there are many
with wrong site surgeries we may not want to different types of Shewhart charts and the user
wait until we have six occurrences of wrong site needs to know which one is most appropriate
surgery to declare a trend and then take action. As for the indicator being studied. The variety of
my colleague Dr. Ray Carey (2003, p. 19) wrote: Shewhart charts is summarized by Benneyan
“When the well-being of patients is at risk, a case et al. (2003, p. 16):
can be made for using 2-sigma limits as ‘early
warning limits’ or for using 6 rather than 8 points There are at least a dozen different
to detect a shift.” In these situations, clinicians types of control charts in common use
would still be looking for signals of special cause in manufacturing and other industry,
so they do not over react to a single data point. with three or four new types being
But, they would use the data not necessarily to developed each year. The various types
justify changing the system but rather as a basis differ by the statistic plotted (e.g., av-
to investigate potential instability in their process erages, percentages, counts, moving
that could cause harm to patients. Wheeler and averages, cumulative sums, etc.) and
Chambers (1992) refers to this as having a process the distribution assumed (e.g., normal,
on the “brink of chaos.” Statistical decisions must binomial, Poisson, geometric, etc.). All
be moderated with and filtered through a healthy have different formulae for calculating
dose of common sense and rational thinking. centerlines and control limits.

9781284023077_CH09_211_258.indd 225 31/07/17 5:59 PM

 226 Chapter 9 Understanding Variation with Shewhart Charts

If there are all these different types of Shewhart ■ Blood glucose readings
charts how do you decide which one is the most ■ The number of procedures or tests performed
appropriate for your data? The decision involves ■ The number of surgeries done each day
two rather simple steps: (1) deciding on the ■ Financial measures such as revenue, oper-
type of data you have collected and (2) deciding ating margin, or expenses
which of the various Shewhart charts is the most ■ Duration of a surgical procedure in minutes
appropriate for this type of data. or hours

Attribute data are essentially counts of events

that can be placed into discrete categories. Unlike
▸ Types of Data measuring a patient’s weight on a continuous scale,
attribute data are looking at characteristics that
The first step in selecting a Shewhart chart is to
can be classified and placed into categories or
determine the type of data you are collecting.
“buckets.” For example, any time you are measuring
There are basically two types of data: (1) variables
mortality you are using attribute date (the patient
data (also known as continuous, interval, ratio,
is either alive or dead). Similarly, pregnancy is an
or measurement data depending on your back-
attribute indicator. There are only two outcomes:
ground and training) and (2) attributes data
the woman is either pregnant or not pregnant.
(also known as classification or count data). The
A woman does not proudly announce that she
term used to identify the type of data is a matter
is 53.9% pregnant. This is essentially a binomial
of taste and preference. Most SPC books will
outcome. Attribute data can be further divided
use the terms variables or continuous data and
into two subdivisions, defectives and defects.
attributes, classification, or count data. What is
Defectives (also known as nonconforming
more important than the terms you choose to
units) require that you have a count of the total
use are the concept the terms are capturing and
number of items or events being observed or
how you apply them to your data. In this text,
produced and the number of items from this total
I use the terms variables and attributes data as
that were not acceptable. The unacceptable items or
the primary categories. FIGURE 910 provides
events become the numerator and the denominator
examples of these two types of data.
is the total number of items or events observed.
Variables data can be measured along a
When you know how many items out of the total
continuous scale. In Figure 9-10 this type of data is
are unacceptable you can either plot the number
depicted as money, time, weight, length, and tem-
of defective items on your Shewhart chart or you
perature. Consider the ruler as a form of variables
can compute the percent of defectives. When we
data. It has equal appearing intervals that can be
compute a percentage, therefore, we are basically
divided into as many subdivisions as your calibration
determining what proportion or percentage the
instruments will permit. With variable data you
numerator is of the denominator. The standard
can perform all the mathematical function. Data
terminology used in most SPC books to define
measured this way can be either counts of whole
defectives is that you know both the occurrence
numbers or they can have decimals or fractional
of the defective product or service (the numer-
parts. Examples of variables data include:
ators) and the nonoccurrences (defectives plus
■ Wait times in the emergency department (ED) nondefectives which when added together form
■ Turnaround time for a lab test the denominator). Knowing these two pieces
■ Blood pressure readings allows you to calculate a percentage or proportion
■ Newborn weight (measured in grams or of defectives. Keep in mind that when you use
pounds and ounces) percentages you are comparing the same types of
■ international normalized ratio (INR) and items, products, or services. If you are looking at
prothrombin times (PTs) the percentage of food trays delivered late to the

9781284023077_CH09_211_258.indd 226 31/07/17 5:59 PM

 Types of Data 227

Variables Data

© Pedjami/Shutterstock

© Ultrashock/Shutterstock

© Butterfly Hunter/Shutterstock

© Paul Velgos/Shutterstock © Lipskiy/Shutterstock

Attributes Data

Defectives Defects
(occurrences only)
(occurrences plus
Nonconformities
non-occurrences)
Nonconforming Units

© HeinzTeh/Shutterstock

FIGURE 910 Examples of variables and attributes data

9781284023077_CH09_211_258.indd 227 31/07/17 5:59 PM

 228 Chapter 9 Understanding Variation with Shewhart Charts

patient, for example, you will have the number of report that she just stuck herself. After you try to
late food trays as the numerator and total number calm her down and explain the next steps you will
of food trays produced as the denominator. In take, you do not say, “Oh by the way, how many
this case, you have trays divided by trays—like times didn’t you stick yourself today?” Similarly, if
divided by like. The only attribute that is different a nurse asked a patient, “How many times didn’t
for classification purposes is whether the food you fall today?” she would probably get a rather
tray was delivered late to the patient. This is an confused look from the patient.
important distinction to keep in mind because When you are dealing with defects you need
as we move next to define defects, this condition to remember that a count of the number of falls,
will not hold. In summary, data classified as de- needle sticks, or medication errors gives you a
fective can be divided into one of two categories numerator but you do not have a denominator. So
(i.e., a binomial situation) when you know both you cannot compute a percentage. So, you either
the occurrences and the non-occurrences of an just count the number of defects as whole numbers
event. Examples of this form of classification in- (e.g., the total number of falls today was eight)
clude conforming/not conforming to standards, or you create a falls rate. A rate is a ratio (i.e., it
harm/no harm, go/no-go, pass/fail, OK/not OK, has a numerator and a denominator just like a
complete/incomplete or present/absent. proportion or percentage) but the two numbers
Defects pose an interesting challenge. Defects you are using to form this ratio are not alike. For
occur and can be counted. But, how do you count example, when we compute an inpatient falls rate
all the nondefects? Stated differently, you know by month we have the number of inpatient falls
when a defect occurs (the occurrence of an event) (including multiple falls) for the month as the
but you do not know when the nondefects or numerator and the denominator is usually the
nonoccurrences happen. I know, at this point you total number of inpatient days for the month. Now
are thinking, “This makes no sense.” When I first we have falls divided by days, two unlike things.
heard this statement it did not make a lot of sense The resulting number is reported as so many
to me either. Examples should help to clarify this falls per 1,000 patient days. Any time you report
concept. Look down at the rug in your office or that there are so many defects per 1,000, 10,000
in your family room. How many spots, stains, or or 100,000 units (e.g., inpatient days, medication
snags do you find in the carpet? For argument’s orders, lab tests, or surgeries) you have just created
sake I’ll imagine that you found three dirt spots, a rate. Note that when you see the little word per
two coffee stains, and four snags on the carpet. included with the name of a measure you know
Now, count the number of nonstains on the carpet? that it is a rate and not a percentage. Most of the
How did you do? You cannot count the nonstains patient safety indicators as well as epidemiology
or blemishes on the carpet. This is an unknown. indicators are constructed as rates (e.g., patient
Similarly, when the highway department records fall rate, restraint rate, surgical site infection
traffic accidents they do the same thing. They can rate, ventilator-associated pneumonia (VAP)
count how many accidents occurred today on a rate, needle stick rate, or medication error rate).
particular segment of the highway but they have The other characteristic of a rate is that the
no idea how many nonaccidents there were today. numerator of a rate can be larger than the denom-
There are times, therefore, when we know the inator. For example, if you had a 20-bed unit and
occurrence of an event when the nonoccurrences each patient fell two times you would have 40 falls
are unknown and unknowable. for 20 patients. What do you call this? Is it 200%
In health care, we experience this situation with falls? No. If you wanted to use a percentage you
patient falls, needle sticks, nosocomial infections, would have to make a different indicator, which
medication errors, and liability cases. We know would be the percent of patients who fell once or
only when the event happens. Think of needle more while they were with us. In this case, we do
sticks. A staff member comes into your office to not care about the total number of falls, which

9781284023077_CH09_211_258.indd 228 31/07/17 5:59 PM

 Types of Shewhart Charts 229

OK?
If Yes, then the car is
fit to be shipped out!

Not OK?
If No, then the car is classified as being “defective” but
we do not know why it is defective (not fit to be shipped)
until we inspect it and count the number of specific
“defects” that make the car “not OK” or defective.

FIGURE 911 Defectives versus Defects

Ed Aldridge/Shutterstock

includes duplicates. All we are concerned with is defects: one headlight does not come on (defect
at the time of discharge did this patient fall once 1), the driver’s side door does not close flush with
or more, yes or no? Basically if the indicator is the the body (defect 2), and the driver’s seat moves
percentage of patients who fell we do not care if backwards but not forward (defect 3). The entire
they fell more than once. But, because patients car is classified as defective but three defects have
can fall more than once, and we are concerned been discovered. The next car is also classified as
about this problem, we generally do not use the defective but it has only one defect (the oil pressure
percentage of patients who fell as a binomial in- warning light on the dashboard does not go out
dicator (i.e., fell/did not fall). If we are concerned after the specified period of time). In summary,
about the magnitude of the falls and severity we defects or nonconformities are the specific things
typically track all falls, which means that we have that make a product or service defective. Once you
the possibility of having a numerator that is larger understand the distinctions between defectives and
than the denominator. When this can occur, we defects you will be well on your way to selecting
normalize the total number of falls by creating a the most appropriate Shewhart chart. To help you
rate (e.g., 3.2 falls per 1,000 inpatient days). A final in building your skills in differentiating between
point about defects is that they usually occur less defectives and defects refer to EXERCISE 91. For each
often than indicators measured by a percentage. indicator listed decide if it is describing a defective
FIGURE 911 provides an easy way to remember or defect. The answers to this exercise can be found
the differences between defectives and defects. When at the end of this chapter.
cars come off the assembly line they get inspected.
If the car is determined to be acceptable by the
inspectors it is fit to be shipped to a dealer. But, if
the inspectors find one or more things wrong with ▸ Types of Shewhart
the car it is not fit to be shipped. In this case, the car
would be classified as being defective. This fit to ship
Charts
determination is a binomial decision: the car is okay After determining whether your data are variables
or the car is not okay to be shipped. At the end of or attributes, the next step is to decide which She-
the shift the inspectors take all the defective cars and whart chart is most appropriate for the type of data
provide a summary of why each car was classified you have collected. Seven basic control charts are
as defective (i.e., as a nonconforming unit). This is regularly described in the literature and taught in
where the defects come into the picture. The first most classes and seminars on SPC. After working
defective car has a summary report pasted to the with the charts for over 15 years, however, I have
windshield. It reads that this defective car has three found that five of the seven charts are the most

9781284023077_CH09_211_258.indd 229 31/07/17 6:00 PM

 230 Chapter 9 Understanding Variation with Shewhart Charts

EXERCISE 91 Defective or defect? You make the call!

Defective Defect
Indicator (Classification) (Count)

1. Number of accidents per 1,000 employee days

2. Number of errors per 25 food trays

3. Percentage of acute myocardial infarction (AMI)

patients receiving aspirin within 24 hours of arrival
in ED

4. Percentage of inpatient deaths each month

5. Number of surgical complications per 1,000

surgeries performed

6. Proportion of hand hygiene observations done

incorrectly

7. Number of falls per 1,000 patient days

8. Number of medication errors per 10,000 doses

dispensed

relevant and frequently used with healthcare, that you read beyond what I summarize in this
social services, and educational indicators. I focus chapter. As I mentioned earlier in this chapter,
on these five Shewhart charts but encourage you by reading the explanations of different authors
to explore the full range of charts as discussed in describing Shewhart charts and their uses you
the Western Electric Statistical Quality Control will build knowledge on how to use them with
Handbook (1985), Wheeler (1993, 1995), Wheeler your own improvement efforts.
and Chambers (1992), Carey and Lloyd (2001),
Carey (2003), Duncan (1986), Pyzdek (1990),
Kume (1996), and Provost and Murray (2011).7
FIGURE 912 presents the Shewhart chart
▸ Defining the Key Terms
decision tree with the five control charts that Before addressing the details related to each of
have the most relevance to health care, social the reviewing the five basic Shewhart charts
service and educational indicators. Two of the shown in Figure 9-12, however, is it necessary
five charts are used with variables data (i.e., X-bar to review three key terms that play a critical
and S chart and the XmR chart) and three of the role in helping you work your way successfully
charts are used with attributes data (i.e., p-chart, through the Shewhart chart decision tree shown
u-chart, and c-chart). Each of the five charts is in Figure 9-12. These key terms are subgroup,
described next and examples of how to apply observation, and area of opportunity and are
the charts are offered. I would suggest, however, summarized in FIGURE 913.

9781284023077_CH09_211_258.indd 230 31/07/17 6:00 PM

 Defining the Key Terms 231

Variables data Decide on the type of Attributes data

data

More than one

Occurrences and
observation per
Yes No nonoccurrences?
subgroup? No Yes

Is there an equal
Yes area of No
opportunity?

X-bar & S XmR c-chart u-chart p-chart

Average and Individual The number of The defect The proportion

Standard measurement defects rate or percentage of
deviation defectives

FIGURE 912 The Shewhart chart decision tree

Subgroup Observation Area of Opportunity

How you organize your data (e.g., The actual value (data) you Applies to all attributes or
by day, week, or month) collect counts charts
The label of your horizontal axis The label of your vertical Defines the area of frame in
Can be patients in chronological axis which a defective or defect
order May be single or multiple can occur
Can be of equal or unequal sizes points Can be of equal or unequal
Applies to all the charts Applies to all the charts sizes

FIGURE 913 Defining Subgroup, Observation, and Area of Opportunity

Subgroup axis of the chart. The subgroups will be arranged in

chronological order of occurrence. When deciding
The subgroup defines how you have organized your on a subgroup you should strive to select them so
data and usually captures some dimension of time that if special causes exist the chances for differences
such as when patients show up for an appointment, between subgroups will be maximized, whereas the
day of the week, week, or month. The subgroup chances for differences due to special causes within
will be the label you place on the horizontal or x

9781284023077_CH09_211_258.indd 231 31/07/17 6:00 PM

 232 Chapter 9 Understanding Variation with Shewhart Charts

a subgroup will be minimized (Duncan, 1986; month we have 30 or 31 days, upwards of

Montgomery, 1991). The traditional subgroups 90 shifts in a hospital, and approximately
for Shewhart charts have been: 720 hours in which to deliver care. Why
would we want to aggregate all this activity
■ An individual patient as the subgroup in into a monthly average or monthly total?
which case you would order the patients Monthly data frequently lead people down
along the x axis of the chart in the order the path of judgment or accountability not
that they presented themselves in the office. quality improvement (QI). In my view, a
Patient 1 arrived at 9:00 a.m., patient 2 at primary reason we have so many health-
9:25 a.m., patient 3 at 9:50 a.m. and so on. care indicators structured around monthly
■ A day as your subgroup in which case you subgroups is that this is how financial and
would have Monday, Tuesday, Wednesday, etc. resource allocation systems are organized.
across the x axis. Then each day you would In health care, work is being produced
select either all of the patients or a sample of every minute of the day not in monthly
them and record their wait times. blocks. Patients are waiting to be seen,
■ A week the subgroup and you would label the have tests performed, or surgery started.
x axis as Week 1, Week 2, Week 3, etc. You Their focus is on minutes or possibly hours
would then have to decide if you were going not months. Administrators and managers
to track the wait time for all the patients in a think in terms of months but patients think
week or just a sample. Usually when a week about the here and now not in monthly
or even a 2-week period is selected as the aggregates.8 The other challenge with using
subgroup and patient wait time is the indicator month as a subgroup is that the variation
of interest you would probably want to draw in the indicator of interest is usually not
a sample of the patients. A total enumeration visible because the data is aggregated into
would probably provide more data than you an average. Although administrators, man-
need and create data collection challenges. agers, and policy makers frequently rely on
■ A month is a frequently used period of aggregated data and summary statistics to
time for a subgroup. But it is not always make decisions, no customer, patient, or
the best block of time in which to think service user cares about the average. They
about improvement or understanding are concerned about why they or their loved
variation. Remember that the Shewhart one are not getting service or treatment
charts are designed to help you understand now. A patient takes no comfort in being
the variation in a key process indicator as told that the average wait time to see the
close to the production of the indicator as doctor last month was only 49 minutes. Or
possible. In manufacturing, they evaluate a physician waiting for her stat lab result to
products and services on an hourly, shift, come back will rightfully be irritated if she
or daily basis. Although they may aggregate is told “We don’t have your result quite yet
the key indicators for management reports but don’t worry, the average time to get a
by month or quarter the ability to improve result last month was only 63.5 minutes.”
quality and insure reliability does not come My point is that although we have a ten-
by looking at monthly or quarterly averages. dency to fall back on making month the
It comes by looking at production almost subgroup for many healthcare indicators
as it happens. In health care and many there is no reason to do so. I have made
social services, there is a strong tendency many charts for teams that are collecting
to track indicators by month or even quar- monthly data. But each time this happens
ter. This is what I refer to as the tyranny I make sure we have a discussion about
of monthly data. Think of it this way: in a what is the smallest unit of time that we

9781284023077_CH09_211_258.indd 232 31/07/17 6:00 PM

 Defining the Key Terms 233

could gather data on. Using month as a occurrence and nonoccurrences of the events
subgroup should be a fallback option not being studied. If you answer “yes” to this question
the first choice. you will be able to calculate either the proportion
■ Finally, it is probably quite evident at this or percentage of defectives and proceed to make
point that it is not advisable to use quarters a p-chart. If you respond that you do not have
as a subgroup choice because quarterly data the occurrences and nonoccurrences you are
represent a very long time period and the left with having only the occurrence of an event
variation you are interested in understanding when the nonoccurrences are not know. As was
has been aggregated and therefore severely mentioned previously, this gives you a count of
dampened. Quarterly data can lead only to the defects and you will make either a c-chart or
judgment not improvement. a u-chart. The decision to make a c- or u-chart is
based on your answer to the following question:
“Is there a relatively equal area of opportunity
Observation for the defect to occur?” If you respond “yes” to
As the term implies this is the actual piece of data this question you will make a c-chart, which is
or the measurement that you record or observe a plot of the number of defects occurring within
(e.g., turnaround time for a lab test or medication each subgroup (e.g., a count of the number of
order, blood glucose readings for a patient, or falls occurring each day). If you respond “no” to
time to administer beta blockers to heart attack this question (i.e., there is not an equal area of
patients in the ED). The vertical axis label on the opportunity for a fall to occur) then you would
chart defines the observation and the units of make a u-chart, which would be a plot of the
measurement along the y axis show the potential defect rate by subgroup (e.g., 3.2 falls per 1,000
distribution of these values. An observation patient days). So, it really does not matter if you
can be classified as either variables or attribute respond “yes” or “no” to the area of opportunity
data (e.g., time, money, weight, a percentage of question. Consider it essentially a filtering ques-
defectives, a count of defects, or a defect rate). tion that will help you select the correct chart for
For example, if your indicator is wait time to see your indicator. As each chart type is explained
the doctor in a clinic your observation will be the use of these terms is demonstrated.
the actual wait time in minutes that occurs from The terms used in the Shewhart decision tree
when the patient checks in at the registration (Figure 9-12) and summarized in Figure 9-13 are
desk until she is seen by the doctor. This amount not only central to understanding SPC theory but
of time will be what gets plotted on the chart. also from a practical perspective, understanding
Therefore, the dot on the chart, or the “doink” the terms subgroup, observation, and area of
as I like to refer to it, represents the quantitative opportunity are essential in the operation of SPC
aspect of the indicator you are observing during software packages. Many of the SPC software
the defined period of time (i.e., the subgroup). packages I have used explicitly ask you to identify
the subgroup and the observation or some variant
of these terms. Although most software appli-
Area of Opportunity cation do not ask you the “area of opportunity”
All Shewhart charts must have a subgroup and an question, understanding this concept is critical
observation clearly defined or the chart cannot to selecting the most appropriate chart for your
be constructed. When we move to the right side indicator. With these basic terms in mind, we
of the decision tree (Figure 9-12) and consider can start using the Shewhart chart decision tree
the attributes charts, discussed in detail later, a to understand the conditions that will lead us to
third term comes into play. Notice that the first select each chart. We will start on the left side of
decision point when dealing with attributes the decision tree and address the variables data
charts is determining whether you have both the charts then move over to the attributes charts.

9781284023077_CH09_211_258.indd 233 31/07/17 6:00 PM

 234 Chapter 9 Understanding Variation with Shewhart Charts

X-Bar and S Chart If the subgroups are of unequal size, however

(e.g., on Monday, we sample 10 patient wait
The left side of the Shewhart decision tree times, on Tuesday, we had 15, and Wednesday,
(Figure 9-12) follows a pathway to two charts. we collected 20 wait times) the UCL and LCL
The first one is referred to as the average (X-bar) will not be straight lines. Instead, they will be
and SD (S) chart. It is the most powerful of the what are called “stair-step” control limits as
five Shewhart charts because it has multiple shown in FIGURE 915. With an unequal sub-
observations of continuous data that have been group size the amount of data varies within
organized into subgroups. In this case, the “doink” each subgroup and so the dots on the chart
on the chart (i.e., the plotted dot) has multiple (i.e., the observations) each have their own
“doinkettes” (i.e., observations) going into it. For individual UCL and LCL calculated. With more
an X-bar and S chart the subgroups can be of data the limits are tighter and with less data in
equal size or unequal size. If the subgroups are each subgroup the limits are wider as shown in
of equal size (e.g., a stratified random sample of Figure 9-15. Day 4 in Figure 9-15, for example,
15 patients is selected each day and their wait has tighter limits indicating that there is more
times to see the doctor are recorded) then the day being collected on this day. Day 9, on the
UCL and LCL on the chart will be straight lines other hand, has wider limits due to less data
as shown in FIGURE 914. being collected on this day.

X-bar chart: patient wait time

60.0
UCL
50.0
Wait time in minutes

40.0

30.0 LCL
20.0

10.0

0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Week

S chart: patient wait time

18.0
16.0 UCL
14.0
Standard deviation

12.0
10.0
8.0
6.0
4.0
2.0
0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Week

FIGURE 914 X-bar and S chart with straight control limits due to equal subgroup sizes

9781284023077_CH09_211_258.indd 234 31/07/17 6:00 PM

 Defining the Key Terms 235

140

137
UCL = 135.22
134
Average (mmHg)

131

128 CL = 127.333

124

121
LCL = 119.44
118

115

20
16
UCL = 12.071
Sigma

12
8
CL = 6.129
4
0 LCL = 0.186
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

Day

FIGURE 915 X-bar and S chart with stair-step control limits due to unequal subgroup sizes

When you make an X-bar and S chart most In Figure 9-15 the indicator of interest is a
software programs will give you the option of patient’s systolic blood pressure. The patient re-
producing two charts as shown in Figures 9-14 corded several blood pressure readings each day
and 9-15. The top chart is the X-bar chart or (a minimum of three and a maximum of five each
average chart and the bottom chart is the S chart day). As a result the subgroups are of unequal sizes,
or SD chart. The X-bar chart is considered to which produces the stair-step control limits. If the
be the primary chart. Both charts have three patient had recorded exactly the same number
main components: (1) the CL or average, (2) of blood pressure readings (observations) each
the UCL, and (3) the LCL. The X-bar chart day (e.g., four) then the UCL and LCL would be
will show the average of the data within each straight lines. As you will see in subsequent ex-
subgroup and the lower chart (the S chart) amples, stair-step control limits will also be found
shows you the SD for each subgroup (i.e., each on p- and u-charts. As was mentioned above, on
dot) plotted on the X-bar chart. In Figure 9-14, those days when more data were collected (e.g.,
for example, Week 1 on the X-bar chart has an Day 4 in Figure 9-15) the control limits will be
average wait time of 38 minutes and an SD of 5 tighter. On days when fewer data were collected
minutes (seen on the bottom chart). On Week (e.g., Day 9) the limits will be wider.
2 the average wait time is 39.7 minutes and the The upper chart in Figure 9-15 reveals
SD is 7.2 minutes. So for each week we can see the average systolic blood pressures by day
what the average wait time is and the amount and the overall average. The lower chart shows
of spread around that average for this week as the SD for each day as well as the average SD
measured by the SD. across all 25 days.9 The average systolic blood

9781284023077_CH09_211_258.indd 235 31/07/17 6:00 PM

 236 Chapter 9 Understanding Variation with Shewhart Charts

pressure shown in the top chart for the patient (1992), Western Electric (1985), and Provost and
in Figure 9-15 is 127. Note that the decimal Murray (2011).
places on the chart can be ignored in this case
because this is entirely too finite a reading for
blood pressure results. The degree to which you XmR chart or the I-chart
can control the decimal places on a chart will The XmR chart is also known properly as the
depend on the software being used. The average Individuals and Moving Range chart. But it
UCL is 135 whereas the average LCL is 119. can also be referenced as the Individuals chart
Because this chart reveals only common cause or simply the I-chart. The key characteristic of
variation, the way to describe the performance of this chart is that each subgroup contains one
this patient’s systolic blood pressure this chart is and only one individual observation or bit of data
as follows: “On the average this patient’s systolic (i.e., the “doink” on the chart has only 1 bit of
blood pressure is 127. It could go up as high as data and no “doinkettes” as we discovered in the
135 on any given day or down to 119 and that X-bar and S chart). In the Shewhart decision tree
is the natural rhythm of this patient’s systolic (Figure 9-12), this decision point is identified
blood pressure process.” by the question “More than one observation
The lower chart is the S-chart. This chart for each subgroup?” When the answer is “no”
has two primary purposes. First, it helps you then the chart of choice is the XmR chart. Like
to understand the variation that exists within the X-bar and S chart, you will typically get
each subgroup (i.e., day). For example, the SD two charts when you request this type of chart
for Day 4 is around 4. On Day 9 the standard from your SPC software. The X chart shows the
deviation is about 3 mmHg. As you look at values for the individual data points as well as the
each day, therefore, you will see that there is a average for all the individual data points. The mR
different average and standard deviation which chart documents the “moving range.” The XmR
reflects the variation in this patient's blood chart is typically used when you are interested
pressure over time. The second purpose of the in answering questions such as:
S-chart (i.e., the bottom chart) is that the av- ■ How many surgeries do we do each week?
erage SD (the CL) is used in the calculation of ■ What is the cost of each knee replacement
the control limits for the average (upper) chart.
surgery?
In this case, the average systolic blood pressure ■ How long does each patient wait before
(i.e., CL) is 127. What is important to realize is
being seen by the doctor?
that if the SD chart reveals wide variation, then ■ How many home care visits do we conduct
the average SD will likewise be large. Because
each week?
the average SD is used to compute the UCL and ■ How many calories do I eat each day?
LCL, a large SD will also contribute to making ■ What is the length of stay for each coronary
the control limits of the top chart (the average
artery bypass graft (CABG) patient?
chart) wider. The relationship of the two charts
must be understood together. This chart is used frequently to address questions
I do not intend to elaborate on the statistical related to volume, frequency of events, or financial
formulae for the calculation of the control limits. issues. Note that you are not interested in finding
It is important to realize, however, that each out what percentage of surgeries started late (this
Shewhart chart has a different set of formulae to would be considering a late surgery start classified
calculate the chart’s UCL and LCL. For additional as a defective, which would require a p-chart),
details on calculating the statistical parameters but rather you merely want to know how many
for the various Shewhart charts readers should surgeries are done in the course of a day or a
refer to Wheeler (1995), Wheeler and Chambers week. In this case, the day or week becomes the

9781284023077_CH09_211_258.indd 236 31/07/17 6:00 PM

 Defining the Key Terms 237

subgroup (x axis label) and the total number of dashed lines that divide the chart into three areas
surgeries completed each day or week becomes above and below the CL. As was discussed earlier
the individual observation for that week (i.e., the in this chapter, the zones are used to assist in
dot on the chart). In short, the XmR chart can identification of special causes. Typically, and you
be used in many situations. Remember, however, will see exceptions to this point, the zones are used
that the indicator being placed on the XmR chart when you have a chart with equal subgroup sizes.
is not being classified as a defective or defect. The bottom chart is referred to as the moving
When you use the XmR chart you will usually range chart. The moving range is derived by calculating
be asking a more neutral question such as “How the simple difference between two successive data
many of procedure X do we do?” points on the Individuals chart and then plotting
FIGURE 916 provides an example of an XmR this difference on the mR chart. This is also referred
chart. In this particular example, the indicator of to as creating an artificial subgroup of 2 since each
interest is the total number of U.S. dollars saved subgroup on the chart initially contains only 1 bit
each month as a result of implementing a new of data. These steps are highlighted in Figure 9-16
transcription system for radiology. Note that like by the circles drawn around each neighboring data
the X-bar and S chart there are two charts. The point on the top chart and the corresponding arrows
top portion of the chart provides a plot of the that point to the mR value between the coupled
individual data points along with the average of data points on the lower chart. Notice that the first
all the data points and the UCL and LCL. This three data points on the mR chart (Months 2–4)
chart also has the zones identified. These are the are relatively close together. This is due to the fact

5750.0
5500.0 UCL = 5470.10
5250.0 A

5000.0
B
4750.0
Dollars saved

4500.0 C

4250.0 CL = 4360.90
C
4000.0
3750.0 B

3500.0 A
3250.0 LCL = 3251.70
3000.0

1500.0
UCL = 1362.79
Moving range

1000.0

500.0 CL = 17.20
0.0 LCL = 0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Month

FIGURE 916 XmR chart for the total amount of dollars saved each month in radiology transcription

9781284023077_CH09_211_258.indd 237 31/07/17 6:00 PM

 238 Chapter 9 Understanding Variation with Shewhart Charts

that there are small differences between the first does not have a neighboring data point to be com-
four data points that have been coupled together pared with until the data for Month 2 is posted.
on the Individuals chart. If you look at data points Therefore, there is no moving range for Month 1
for Months 15 and 16, however, you see a very on the mR chart, which is identified by the triangle
different picture. The difference (i.e., the range) surrounding Month 1. This will always be the case
between Months 15 and 16 is much larger ($938 when you use an XmR chart.
to be exact) than any of the ranges found when the EXERCISE 92 (You make the call: Is it an X-bar
first four data points were compared. In short, the and S chart or XmR chart?) will test your ability to
individual values when coupled together produce determine whether a particular indicator should
an artificial subgroup of two, which you must have be placed on an X-bar and S chart or an XmR
in order to calculate a range and subsequently the chart. Answers to this exercise may be found at
moving ranges. One final thing to note about the the end of this chapter.
XmR chart. The mR chart will always have one less When we move to the Attributes side of the
data point on it than the Individuals chart. This Shewhart chart decision tree (Figure 9-12), we
is due to the fact that the first data point (Month 1) need to address two questions:

EXERCISE 92 Is it an XmR (I) or X-bar and S? You make the call!

X-Bar and XmR

Indicator S Chart (I Chart)

Time to clean an inpatient room (in minutes)

Patient satisfaction scores for subgroups of 15 patients in the outpatient clinic

Average turnaround time for all STAT labs done each day

Cost for each normal delivery

A diabetic patient’s 3x a day blood sugar readings

Average length of stay for a subgroup of 20 intensive care unit (ICU)

patients

The distance (in feet) that a sample of 10 knee replacement patients can
walk in 15 seconds

■ Do we have the occurrence and nonoccur- equal opportunity for a defect to occur?”
rences of an event? If “yes” then we make a If we have an equal opportunity for a
p-chart (i.e., a percentage chart) defect to occur we make a c-chart. If not,
■ If the answer is “no” meaning that we then we make a u-chart. The details are
have only the occurrence of an event (i.e., explained next.
a defect when we do not know the non- We start with the first question and address the
defects) then we need to ask, “Is there an use of the p-chart.

9781284023077_CH09_211_258.indd 238 31/07/17 6:00 PM

 Defining the Key Terms 239

P-Chart denominators for each dot on the chart are not

equal and spread from a minimum of 326 cases
The p-chart derives its name from the fact that (Month 3) to a maximum of 1,041 (Month 15).
either a percentage or a proportion is what you Notice that the smallest denominator (326 for
actually are plotting on the chart. Most of the Month 3) has the widest control limits whereas
time the percentage will be the statistic of interest the largest denominator (1,041 for Month 15) has
rather than the proportion. When you make a the tightest set of control limits. The numerators
p-chart, or any other attributes chart, you will get go from a low of 75 readmissions (Month 3) to a
only one chart (unlike the variables charts which high of 249 in Month 16. If the distance or spread
gives you two charts). The p-chart is used to between the stair-step limits is relatively small
monitor the proportion or percentage of defectives this means that the denominators are relatively
when you know the occurrence of the defective close in size.
product, unit, event, or service (the numerator) Finally, it should be noted that if the sub-
and the nonoccurrences (the denominator, groups (the denominators) were of equal size, the
which is the total being observed). This chart is control limits on the p-chart would be straight.
used frequently in healthcare settings because But because most healthcare indicators that
we track many indicators that look at accuracy, are defined as percentages differ from one sub-
completeness, errors, or the percentage of some- group (i.e., time period) to another (e.g., we do
thing done or not done (e.g., cesarean sections, not have the same number of deliveries each
completed history and physical reports, proper month, produce the same number of food trays
hand washing, or compliance with a standard each day, or have the same number of patients
protocol). FIGURE 917 provides an example of visit at clinic each day) we usually do not have
a p-chart with stair-step control limits. In this equal subgroups when calculating percentages
case, the indicator is the percentage of hospital or proportions. Therefore, most p-charts will
readmissions for home healthcare patients. The generally have stair-step control limits.

35.0

30.0
Percentage of readmissions

25.0
UCL = 24.59

20.0 CL = 20.64

LCL = 16.68
15.0

10.0 Denominator = 1041

(tightest limits)
Denominator = 326
5.0 (widest limits)

0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
Month

FIGURE 917 p-chart on the percent of hospital readmissions for home healthcare patients

9781284023077_CH09_211_258.indd 239 31/07/17 6:00 PM

 240 Chapter 9 Understanding Variation with Shewhart Charts

C-chart ■ Medication errors

■ Central line infections
The c-chart and the u-chart are the Shewhart
charts of preference when you are tracking defects. Figure 9-1 was used to show the elements of a
The c-chart is the appropriate chart when you Shewhart chart. It also provides an example of
have an equal area of opportunity for a defect a c-chart. In this example, the customer service
to occur. As shown in Figure 9-13, an area of manager of a large medical group is interested in
opportunity has the following characteristics: charting the total number of patient complaints
received each week at 17 sites of care. Because
■ It applies to all attributes or count charts each patient could file more than one complaint,
(c, u, and p) complaints are viewed as defects. The alternative,
■ It defines the frame or area in which a considering the registration of a complaint as
defective (i.e., a nonconforming unit) or a defective, is not selected because this would
a defect (i.e., a nonconformity) can occur preclude counting more than one complaint from
■ It can be of equal or unequal sizes. an individual patient. Remember that defectives
A manufacturing example may help clarify this are based on the binomial distribution (i.e., the
concept. Imagine that you work on a paint line at patient complained or did not complain). If you
an automobile manufacturing plant. If you were approached this indicator as a defective you
assigned to paint the hoods of a single model would not be concerned with the magnitude of
of a car (e.g., a Ford Taurus) there would be an the complaint problem (i.e., the total number
equal area of opportunity for a paint blemish of complaints) but rather with the fact that a
because all Ford Taurus hoods are the same size. patient complained or did not complain and
In this case, we would make a c-chart and plot you do not care if a patient complained more
the total number of paint blemishes (defects) than once. Measuring complaints as a defective
on each hood you paint. Because each hood has would produce a percentage of patients who
the same number of square inches of surface complained (a p-chart). As a defect, however,
area there is a constant area of opportunity for we are concerned with the magnitude of the
a paint blemish. problem so we count the total number of com-
The challenge now becomes determining plaints, including multiple complaints from the
when this equal area of opportunity condition same patient. The c-chart is selected because the
exists in healthcare settings. One of the more volume of patients seen at the 17 clinics Monday
frequently used examples of how this might occur through Friday remains fairly constant each week
is with monitoring patient falls. If you conclude and the number of sites included in the study
that there is basically an equal opportunity for does not change. These two conditions allow the
a patient to fall each day of the week at your manager to assume an equal area of opportunity
hospital, rehabilitation facility, or long-term care for a complaint to occur. She merely counts the
facility, then you would merely count the num- total number of complaints received each week
ber of falls occurring each day, week, or month and plots this number on a c-chart. The chart
and plot the number of falls on a c-chart. Other produces a CL (average number of complaints)
indicators that could be placed on a c-chart if and a UCL and LCL.
the equal area of opportunity assumption was A frequent challenge with using the c-chart
met include the number of: for healthcare applications is that the condition
of equal area of opportunity may not be met.
■ Patient restraints
Frequently in healthcare settings there are few
■ Lawsuits
indicators that have equal areas of opportunity.
■ Patient complaints
The severity of a patient’s condition can change
■ Needle sticks

9781284023077_CH09_211_258.indd 240 31/07/17 6:00 PM

 Defining the Key Terms 241

quickly, the census can show fluctuations, the actually means. An example should help clarify
clinic is not a 7 days a week operation, the volume how these are opposite sides of the same coin. If
of orders may change rapidly, and the ED may we have 21 inpatient falls, this number becomes
have to go on bypass because there are no more the numerator of the rate-based statistic. When we
inpatient beds available on a Friday or Saturday place this count of 21 falls over the total number
night. So if the assumption of an equal area of of patient days for the month (e.g., 4,775) we have
opportunity is violated, what do we do next? The a ratio of two different numbers that produces
answer is simple—you make a u-chart. a result of 0.00439 (i.e., 21/4775 = 0.00439).
Because the number of inpatient days is in the
thousands we multiple the resultant value of
U-Chart 0.00439 by 1000 to produce the inpatient falls rate
This chart is used frequently in health care, of 4.4 falls per 1,000 inpatient days. The number
especially now that there has been a more con- of spots you slip the decimal point on the resultant
centrated effort to track patient safety indicators. ratio depends on how large your denominator is.
The u-chart, like the c-chart, is used to track In this case, we had 4,775 inpatient days so we
defects. The difference is that the u-chart is slip the decimal point three places to the right
selected when you conclude that there is not an by multiplying the value of 0.00439 by 1000. If
equal area of opportunity for the defect to occur. you had patient days in the tens of thousands you
Let us return to the paint line at the Ford plant would slip the decimal point four places to the
for a moment. Although you have in the past right and have 43.9 falls per 10,000 inpatient days.
painted one model of car at a time, today you Or you could go out to 100k inpatient days and
have been told that the line will have a mixture say “I’m sorry but we had 439 inpatient falls per
of cars and a mixture of hood sizes. So, how do 100,000 inpatient days.” Or if you really wanted
you count the paint blemishes on the hoods to depress the senior management team or board
of a Ford Escort, a Taurus, a Mustang, and an you could report 4,397 inpatient falls per 1 million
Expedition? Each hood has a different number inpatient days. You can adjust the result of the ratio
of square inches, takes a different volume of of 21/4,775 very easily for any value you place in
paint to cover the surface of the hood, and has the denominator position. The general rule for
a varying probability of experiencing a paint rates, however, is that the denominator you use
blemish. The u-chart takes care of this problem should be based on the volume you are observing
very quickly by computing a defect rate. The on a regular basis. In the case of inpatient falls,
number of paint blemishes is used as the nu- most hospitals are dealing with inpatient days
merator and the number of square inches of the that are in the thousands so this is what should
hood’s surface is used as the denominator. The be used to calculate the final rate-based statistic
resultant ratio provides the number of blemishes of 4.4 falls per 1,000 inpatient days. If, on the
per so many square inches of hood area. The other hand, you were tracking medication errors
rate essentially normalizes the differences in you would most likely be justified in making the
denominator size (i.e., the area of opportunity number of errors per 10,000 doses dispensed or
for a blemish to occur). scripts written because an average-size hospital
One technical point about rates. Explaining will general dispense 10,000 or more doses each
a rate-based statistic can be a little challenging. It month. Finally, if your measure was the neonatal
is much easier to say, “This past month we had death rate for a state, province, or region then the
21 inpatient falls” than to say, “This past month proper denominator size might be per 100,000
we experienced an inpatient falls rate of 4.4 falls live births.
per 1,000 inpatient days.” Some in your audience Because it is an extremely rare to have the
may struggle with what this rate-based statistic same number of medication orders each week,

9781284023077_CH09_211_258.indd 241 31/07/17 6:00 PM

 242 Chapter 9 Understanding Variation with Shewhart Charts

the same patient census, or the same number on SPC. The ASQ, for example, offers public
of central line days in the ICU, the u-chart is seminars on SPC. You may want to check with
used more often in healthcare settings than the your local ASQ chapter to see when such courses
c-chart. Furthermore, because epidemiologists will be offered. The IHI also offers workshops
frequently produce rate-based statistics (e.g., on building effective measurement systems and
the neonatal death rate or the VAP rate) the SPC. The various program offerings that I and
use of terms associated with the u-chart should my colleagues teach throughout the year can be
sound familiar to many healthcare professionals. reviewed on the IHI home page (www.ihi.org).
Examples of u-chart applications are provided Finally, if you have the opportunity to attend a
in the case study chapter (Chapter 10). local or national quality conference (e.g., the
TABLE 91 provides an overview of the five IHI National Forum on Quality Improvement
charts just described and offers examples of in Healthcare or the IHI-BMJ International
indicators that could be placed on each type Forum on Quality and Safety in Healthcare),
of chart. Other useful tables that summarize make sure that you sit in on sessions that are
how charts should be set up and their various discussing Shewhart charts and SPC. Hearing
uses can be found in Statistical Quality Control about control charts from multiple sources will
Handbook (Western Electric, 1985) and Benneyan be very beneficial.
(2001).10 Readers wishing to gain additional You can also test your knowledge of the var-
insights about the selection of control charts ious charts by completing the You Make the Call
should consult Wheeler (1995), Montgomery exercise found in EXERCISE 93. When I teach my
(1991), Pyzdek (1990), Ishikawa (1989), Duncan classes on Shewhart chart applications, I give the
(1986), Carey and Lloyd (2001), Carey (2003), participants this exercise at the end of the class
and Provost and Murray (2011). to provide a final test of their understanding of
the selection of appropriate Shewhart charts. It
gives them a chance “make the call!” and tests
▸ You Make the Call their control chart knowledge. The indicators
listed in this exercise are taken from actual teams
Now that you are familiar with the basic ideas I have had the opportunity to facilitate or coach.
behind the Shewhart charts, the next step is to Start the exercise by determining the subgroup.
apply this knowledge to your own indicators. Remember that the subgroup is the label for
The study questions in BOX 91 will serve as a the horizontal axis and reflects how you have
quick overview of some of the central issues organized your data (e.g., by day or week). Next
related to Shewhart chart development and as decide if you have variables or attributes data.
a test of your current knowledge. If you struggle Finally, list the chart you think is most appropriate
with some of the questions you can review the for this situation. You may want to refer to the
material presented in this chapter and then Shewhart Decision Tree shown in Figure 9-12 to
explore some of the listed references for addi- assist you in thinking through the chart options.
tional explanations. Another way to enhance The answers to the You Make the Call exercise
your knowledge base is to attend workshops can be found at the end of this chapter.

9781284023077_CH09_211_258.indd 242 31/07/17 6:00 PM

 You Make the Call 243

TABLE 91 Shewhart chart summary

Type of Data and Data Examples of Indicators

Type of Control Chart Collection Issues Used on This Type of Chart

X-bar and S chart Variables data ■ Actual turnaround time

This is known as the Average The X-bar and S chart usually for five lab tests or three
(X-bar) and Standard Deviation involves drawing a sample of pharmacy orders each day
(S) chart. Most SPC software observations (e.g., 3–10 per ■ Blood pressure readings
programs will give you two subgroup). Rational subgrouping (e.g., three to five per day)
charts when you select this is frequently used with this ■ Diabetes monitoring (e.g.,
chart: one for the X-bar portion chart. The statistical principles three fasting blood sugar
and one for the S portion. This behind this chart are based readings each day)
is considered to be the most on the assumptions of the ■ Anesthesia time for a
statistically powerful of all the normal (Gaussian) bell-shaped sample of cases each day
charts. The X-bar and S chart distribution. ■ Patient satisfaction scores
can have straight or stair-step
control limits.

XmR chart Variables data ■ Patient wait time to see

This chart is known as the The XmR chart is used when you the physician or to be
Individual values (X) and have a single observation for each seen in the ED
moving range (mR) chart. subgroup (i.e., n = 1). Sampling ■ The number of days to
Sometimes it will be referred typically is not done but might be mail a patient bill after
to as the Individuals or I-chart. if the process being monitored discharge
It does not have the statistical has an extremely large volume. ■ The number of calls
rigor or power of the X-bar Because this chart frequently uses coming into a clinic each
and S chart because each dot aggregates as the plotted number day
on the chart is representing (e.g., days in accounts receivable ■ Average length of stay
only one observation. This this month), it is important to by week for a particular
chart is used frequently to make sure that the data are diagnosis-related group
answer questions related to consistently collected from one (DRG)
volume, for example, “How time period to the next. This chart ■ The number of surgeries
many surgeries did we do this is used to evaluate questions done each week
week?” The XmR chart does related to process outcomes ■ Operating margin by
not address the question as to (volumes), with no concern as month
whether these surgeries were to whether the outcomes of the ■ Pounds of laundry
started on time (this would process are acceptable or not each day
require a p-chart). Instead, acceptable. ■ Average turnaround time
the XmR chart is answering a by day
neutral question, “How many?” ■ The number of food trays
or “How much?” The XmR produced
chart will always have straight ■ Patient satisfaction score
control limits.

(continues)

9781284023077_CH09_211_258.indd 243 31/07/17 6:00 PM

 244 Chapter 9 Understanding Variation with Shewhart Charts

TABLE 91 Shewhart chart summary (continued)

Type of Data and Data Examples of Indicators

Type of Control Chart Collection Issues Used on This Type of Chart

p-chart Attributes data ■ Percentage of cesarean

The p-chart is used frequently These data are classified as sections
in health care to compute the defectives or nonconforming ■ Percentage of late food trays
percentage (or proportion) units because they reflect the ■ Percentage of incomplete
of defective products or percentage (or proportion) of charts
services. The p-chart requires things or events that do not meet ■ Percentage of late surgery
being able to count both specifications or criteria (the starts
the numerator and the numerators). The denominators ■ Percentage of bills that are
denominator. The p-chart is usually (but not always) are of inaccurate
the weakest of the attributes varying sizes, which produce ■ Percentage of mortality
charts because it is based on stair-step control limits. Data of ■ Percentage of staff
the binomial distribution (i.e., this type reflect the binomial turnover
there are only two outcomes distribution. The denominators ■ Percentage of patients
such as yes/no, acceptable/ need to be sufficiently large (e.g., responding “Very Good” to
not acceptable, or complete/ usually greater than 12) to enable a survey question
not complete. The p-chart a reasonable percentage to be ■ Percentage of x-rays that
can have straight or stair-step calculated yet not too large (e.g., had to be redone
control limits. over 5,000). ■ Percentage of did not
attends (DNAs) at an
outpatient clinic

c-chart Attributes data ■ The number of falls

The c-chart is used to count The key to using a c-chart is that ■ The number of restraints
the number of defects that there should be an equal area of ■ The number of needle
occur within an equal area opportunity for a defect to occur. sticks
of opportunity when the This condition frequently makes it ■ The number of lawsuits
nondefects are unknown. difficult to use this chart in health filed
In this case, each observed care because the conditions under ■ The number of ventilator-
unit (e.g., a patient) can have which we provide care do not associated pneumonias
multiple defects (e.g., falls). always remain constant. One way ■ The number of
Generally speaking, defects to address this inequality in the nosocomial infections
are the specific reasons why a area of opportunity is to apply ■ The number of
product or service is classified stratification. For example, if the medication errors
as defective (i.e., a defective conclusion is that there is not an ■ The number of returns to
product or service will suffer equal area of opportunity for an surgery
from one or more defects). inpatient fall because the hospital ■ The number of surgical
Generally speaking, indicators functions differently on weekends site infections
appropriate for a c-chart than weekdays then separating the ■ The number of violent
should be considered “rare data by weekdays versus weekends events in a mental health
events.” The c-chart will always may be sufficient to conclude that ward
have straight control limits. there is a relatively equal area of ■ The number of central line
opportunity for a fall during each of infections
these periods. The c-chart is based
on the Poisson distribution.

9781284023077_CH09_211_258.indd 244 31/07/17 6:00 PM

 You Make the Call 245

TABLE 91 Shewhart chart summary (continued)

Type of Data and Data Examples of Indicators

Type of Control Chart Collection Issues Used on This Type of Chart

u-chart Attributes data ■ Medication errors per

The u-chart is used to track The Poisson distribution is also 10,000 doses dispensed
defects when the area of used as the frame of reference for ■ VAP per 1,000 vent days
opportunity is not equal. For this chart. The u-chart presents ■ Total falls per 1,000 patient
this reason, the u-chart is rates (e.g., so many falls per 1,000 days
typically used more often in patient days). Knowledge of how ■ Total readmissions per
health care than the c-chart. to collect data to form rates is 1,000 discharges
This chart is based on rates essential. ■ Bloodstream infections
rather than simple counts. The per 1,000 line days
u-chart can have straight or
stair-step control limits.

BOX 91 Shewhart charts study questions

■ When is it appropriate to use Shewhart charts? Should I use them in place of descriptive statistics?
■ What is the relationship between Shewhart charts and tests of significance?
■ How many data points do I need to make a Shewhart chart? What do I do if I do not have
enough data?
■ Which is better, attributes or variables data?
■ What is a subgroup? Do I have to have one to make a Shewhart chart?
■ Can I make a Shewhart chart with only single data points?
■ Do my subgroups have to be of equal size when I make Shewhart charts?
■ Much of the data I get does not have the date on it. So, does it really matter if the data points are
not in chronological order?
■ I still don’t get this distinction between a SD and a sigma limit. Why aren’t they the same? Does it
really matter? My spreadsheet software will give me a SD. Why can’t I just multiple this number by 3
and then add and subtract this product from the mean to get the control limits?
■ Why do I have to use 3 sigma control limits? Why can’t I use two or maybe 1.5 sigma limits?
■ Do defects add up to make defectives or is the other way around?
■ When I make a p-chart, does the size of the denominator make a difference? Can I have, for
example, 4 or 5 in my denominator?
■ What is the difference between a proportion, a percentage, and a rate?
■ Should I view common cause variation as “good” variation and special cause variation as “bad”
variation?
■ Do I really have to investigate a special cause? Can’t I just remove the data point from the chart and
get on with making changes?

9781284023077_CH09_211_258.indd 245 31/07/17 6:00 PM

 246 Chapter 9 Understanding Variation with Shewhart Charts

EXERCISE 93 You make the call! Selecting the right chart

Type of Type of
Situation Subgroup? Data? Chart?

1. Each day you record the number of films processed in V or A

the radiology department.

2. Each day you record the number of films requested V or A

and the number that cannot be found in the radiology
library.

3. The number of inpatient restraints each month is V or A

placed over the total inpatient days each month.

4. Each day you pull a stratified random sample of V or A

15 complete blood counts (CBCs) and record the
turnaround time (in minutes) for each CBC.

5. The number of minutes it takes to get a stat med V or A

order administered to the patient (order time to
administration time).

6. Every 2 weeks you pull a sample of 30 medication V or A

orders and count the total number of orders that have
one or more errors.

7. The wait time in the ED (door to discharge) is tracked V or A

for each patient.

8. The clinic receptionist notes the time of check-in for V or A

each patient. The physician notes the time when he/
she first sees the patient in the exam room. An analyst
compiles the data daily and reports the percentage of
patients who had to wait more than 30 minutes.

9. The director of surgery keeps track of the total number V or A

of surgical procedures performed each week.

10. The dietary department records the number of food V or A

trays that come back uneaten each day and the total
number of trays they produced for that day.

11. You are interested in the average time patients spend V or A

in your waiting area, so every day a student randomly
picks eight patients and measures their actual waiting
time in whole minutes.

(continues)

9781284023077_CH09_211_258.indd 246 31/07/17 6:00 PM

 Additional Shewhart Charts 247

EXERCISE 93 You make the call! selecting the right chart (continued)

12. The ICU nurses want to evaluate the ventilator- V or A

associated pneumonia (VAP) rate. So every 2 weeks
they record the total number of pneumonia episodes
and the total number of vent days.

13. Each week patient satisfaction scores for three units V or A

are compiled and an average is calculated for the three
units.

14. The finance department tracks the total number of V or A

business days it takes to process a vendor’s request
for payment. Process time starts when the request
for payment is received in the finance department
and ends when the payment is sent (electronically or
posted in the mail) to the vendor.

15. Every week each medication order is checked against V or A

five potential types of errors. The total number of errors
for the week is divided by the total number of orders
submitted that week.

16. You know the number of people who come to the ED V or A

complaining of chest pain and the number who are
actually diagnosed with an AMI or unstable angina.

▸ Additional Shewhart ■
■
Multivariate Shewhart-type Charts
P primed chart (p′-chart)
Charts ■ U primed chart (u′-chart)
Provost and Murray (2011) do a very good job of
In addition to the five basic Shewhart charts not only describing these alternative charts and
described previously, there are many other charts provide examples of their use. I do not intend to
that have their roots in manufacturing but have go into depth about these various charts but I do
proven to be very useful in certain healthcare want to make a few comments about the t- and
situations. Some of these alternative Shewhart g-charts that are being used more and more in
charts include: healthcare improvement work.
■ Median chart The t- and g-charts are designed to address
■ t-chart the occurrence of rare events. I know, you are
■ g-chart wondering, “What is the operational definition
■ Moving average chart of a rare event?” When I was first learning
■ Cumulative Sum chart (CUSUM) about these charts the instructor used a simple
■ Exponentially weighted moving average example. He would ask, “What is the probability
chart (EWMA) of looking out the window and seeing a car go
■ Standardized Shewhart Chart by?” Everyone would respond, “High.” Then he

9781284023077_CH09_211_258.indd 247 31/07/17 6:00 PM

 248 Chapter 9 Understanding Variation with Shewhart Charts

would ask, “What is the probability of looking happens (e.g., a fall, a pressure ulcer, a surgical
out the window and seeing an accident occur?” site infection) you basically reset the counter and
Everyone would respond, “Low.” He then would begin counting the number of days again until
proudly announce, “You now understand a rare the next fall occurred. This is the same approach
event.” Now this is a pretty casual explanation of that factories use to track the number of days that
a rare event but I think it helps to set the context have gone by without an accident in the factory.
for thinking about rare events. If you wish to get If you never had an event you would never have
very statistical about rare events you can study a dot on the chart. Because you are counting the
what is called the “rare event rule for inferential number of days that have gone by since the last
statistics.” Within this body of statistical theory event (i.e., a defective or a defect) the horizontal
you will be reacquainted with probability theory axis will not have Monday, Tuesday, Wednesday
that you were exposed to relatively early in your or January, February, March, etc. marked. When
statistical training. Most of the time this is ex- an event occurs this is when you place the date
plained by using the probabilities associated with of the event on the horizontal axis, which will
rolling various combinations on dice or getting not be occurring in equal periods of time. The
a particular combinations of cards while playing indication of improvement on a t-chart is when
blackjack or poker or betting on a roulette wheel. you observe an ever-increasing run of days
At the IHI we use a practical approach to without the occurrence of an event.
defining rare events that is grounded in statistical As healthcare providers have become more
theory but does not require detailed compu- focused on safety indicators and reducing harm
tations. Simply stated, if you have more than the t-chart has become increasingly popular.
25% of the data on a p-, c-, or u-chart at zero But a word of caution is in order. The statistical
(or conversely at 100%) you need to consider basis for properly calculating the limits on a
moving to a t- or g-chart. With 25% or more of t-chart are a little involved. First, you need to
the data points at zero the use of the traditional realize that a distribution of rare events does not
rules for detecting special causes on a Shewhart follow a normal Gaussian bell curve. A Poisson
chart become questionable (Provost & Murray, distribution is a better referent for rare events.
2011). It also is a practical issue. If you do not The Poisson distribution is appropriate as a
have sufficient nonzero data for an attribute referent for the c- and u-charts as well as the
chart the LCL may not exist, which makes the time between chart. In the case of the t-chart,
interpretation of the chart difficult. In these however, the form of the Poisson distribution
situations, you should consider moving to the is actually an exponential distribution, which
time between chart (t-chart) or the cases between is in turn highly skewed. Second, the skewness
chart (g-chart). The t-chart (the t part of the of the exponential distribution is not a major
name refers to “time”) or time between chart problem and is addressed by transforming the
shows you how much time has gone by since the time between events (i.e., days gone by) into
last adverse event. Nelson (1994) provided the a quasi-normal or symmetric distribution by
details on how this chart is constructed. When performing what is called a Weibull transfor-
you use this chart you have to reorient the way mation. Third, once the data are more or less
in which you explain the chart. For example, the approximating a normal distribution the UCL
horizontal axis on the t-chart is a discontinuous and LCL can be calculated by using the formulae
time sequence. If you start next Monday to begin for the XmR chart. Finally, after the limits and
tracking patient falls but a fall does not occur CL are calculated they are transformed back
until Wednesday then you would place a dot to their original state for plotting on the chart.
on the chart’s vertical axis at 2 (i.e., 2 days have I know, this all sounds rather complicated.
gone by before a fall occurred). When an event The detailed steps for constructing the t-chart

9781284023077_CH09_211_258.indd 248 31/07/17 6:00 PM

 Using Shewhart Charts Effectively 249

are clearly discussed by Provost and Murray sudden proclaim, “I think I get it!” Once you
(2011) and Nelson (1994). Also take heart reach this point it is now time to start applying
in knowing that any reasonably good SPC this knowledge to actual improvement oppor-
software package will do all the calculations tunities. But be careful. I have seen some people
for a t-chart quickly and easily once you cre- become so enthusiastic about the various charts
ate the time between data file. Your biggest that they start making graphs on any process that
challenge will be to explain how to interpret produces data. It is at this stage that I remember
the t-chart. But because its use is growing an old adage—if you give a child a hammer, the
in popularity in healthcare setting, it is well whole world looks like a nail! The charts play a
worth your time to gain more knowledge of valuable and central role in all QI efforts. It is
the time between chart. important to realize, however, what they can do
The g-chart (or geometric chart) is similar in and what they cannot do.
principle to the t-chart. It too is a chart for rare First, appreciate the fact that the charts do
tracking events except that instead of plotting not answer the following questions:
the amount of time (e.g., days) between a rare ■ What is the reason for a special cause?
event, the g-chart plots the number of cases ■ Should a common cause process be improved?
that are regarded as being successful against ■ What should I do to improve the process?
cases considered to be failures. A failure in
this situation might be a surgical site infection, The answers to these questions do not come
patients experiencing a medication error, or from the charts or statistics. They come from
a return to surgery within 24 hours. Like the the will, ideas, and ability of the team to execute
t-chart success is determined by having a long tests of change. I have seen too many teams feel
run of successful cases with no failures or ad- that once they have created a chart their work
verse events. Although the t-chart is modeled is finished. I think that this occurs because the
after an exponential distribution the g-chart chart is a tangible thing that can be pointed to
referent is a geometric distribution. Again, and shown to others. Improvement strategies,
the steps for computing the limits and the CL on the other hand, are not as finite or discrete.
on a g-chart are nicely laid out in Provost and Developing improvement strategies is actually
Murray (2011). There is also a considerable much more difficult than mastering control
body of literature on both the exponential and chart theory and construction because you are
geometric distributions that can be found in the dealing with people, behaviors, and culture
ASQ's Journal of Quality Technology. The g-chart not numbers.
is also becoming a standard offering in most Second, after you make a chart and decide
SPC software packages. Once again, however, whether the process exhibits common or special
the challenge is making sure you have at least cause variation, you then need to decide how
a moderate foundation in being able to explain you are going to approach the variation you
the chart and how to interpret it. have identified. Do you need to merely reduce
variation in the process or fundamentally redesign
the process and change the way in which work
▸ Using Shewhart Charts is envisioned and delivered? All improvement
strategies emanate from an understanding of
Effectively variation. If the process exhibits special cause
variation the appropriate decision is to investigate
At some point after reading various books on the special cause(s) and determine why they have
run and Shewhart charts and listening to others made the process unstable and unpredictable.
explain control chart theory, you will all of a Just as we would investigate a patient safety event

9781284023077_CH09_211_258.indd 249 31/07/17 6:00 PM

 250 Chapter 9 Understanding Variation with Shewhart Charts

(i.e., a sentinel event) by conducting a root Notes

cause analysis, we also need to do the same
thing when a special cause is detected on a run 1. Historically these charts have been known
or Shewhart chart. Ignoring a special cause will as control charts. Shewhart himself even
guarantee that it will rear its ugly head at some referred to them as control charts as have
point in the future. We cannot predict exactly many writers since Shewhart’s time. But as
when a special cause will occur but you can Blank (1998, p. 1) points out, “It is important
be sure that it will pop up again if you choose to understand that SPC does not control
to ignore it.11 processes. People control processes. SPC
The other aspect of a special cause is that not is merely a tool that provides you with
every special cause is negative and undesirable. information you need to reduce variation
Remember that special causes are not bad and and tell you whether or not your processes
common causes are not good. The key point can meet the customer’s expectations.” In
is the special causes make a process unstable more recent times, the charts have been
and unpredictable. It is very likely that you will referred to more and more as Shewhart
observe a special cause that you want to emulate charts (Provost & Murray, 2011, p. 113)
(e.g., when lab turnaround time is much faster to emphasize their use primarily in un-
than it has been or the past 15 days). In this derstanding variation and to facilitate
case, you want to investigate why the process learning about process capability rather
worked so well on those days and see whether conveying images of “control.” The term
these conditions can be replicated. Common Shewhart chart is also used to recognize the
causes on the other hand are not inherently significant contributions of Dr. Shewhart to
good. Common cause variation merely means the field of SPC. A final note on the use of
that the process is stable and predictable (i.e., the word “control.” The ASQ was originally
predictable within the boundaries of the UCL called the American Society for Quality
and LCL). Just as you can have a special cause Control (ASQC). In 1997, the membership
that you might want to emulate, you can also voted to drop the word “control” from the
have common cause variation that is unac- organization’s name. This was to recognize
ceptable (e.g., when a patient’s blood pressure that quality was becoming a broader con-
is running at a very high level and staying cept and used in many other fields besides
there or when the wait time to see your family manufacturing where initially in the early
physician is consistent and predictable but it 1900s control was used as a key operative
is at such a high level that it is predictably bad word. Shewhart’s book, Economic Control
and unacceptable). of Quality of Manufactured Product (1931)
QI starts with making the correct decision provides a classic reference to the initial
about the variation that lives in your data. Walter use of the term “control.” So for a variety
Shewhart introduced the control chart and the of reasons I use the term Shewhart chart (s)
notions of common and special causes of variation in this text rather than control charts.
in 1924 (Hare, 2003). Since then SPC has become 2. The USL and LSL are frequently referred
the foundation for all successful QI initiatives. to in manufacturing as “tolerance limits”
It is a key component of the Baldrige criteria, Six and are also frequently referred to as the
Sigma, Lean, and International Organization for voice of the customer (VOC, i.e., what the
Standardization (ISO). Without a clear under- customer wants, needs, or expects from
standing of variation and its causes, however, the product or service). There are many
individuals and organizations will continue to different types of indices that have been
suffer from numerical illiteracy. developed to capture statistically process

9781284023077_CH09_211_258.indd 250 31/07/17 6:00 PM

 Notes 251

capability. The three basic process capability precision for many of its indicators as the
indices are the process capability index manufacturing industry. So, when I use the
(Cp), the minimum process capability term process capability I am using it in a
index (Cpk), and the process capability general sense to describe the variation in
index to the mean (Cpm). The traditional the process as defined by the mean (CL)
statistical use of process capability (Cp) and the UCL and LCL. These numbers
is to indicate whether or not the process define how well the process is performing
can meet the predetermined specifica- relative to the target or goal.
tions (Blank, 1998). There are numerous 4. Some of my colleagues may disagree with
variations on the Cp statistic, all of which these guidelines. I have found over the years
are designed to help the quality control that there are two general issues that need
(QC) researcher investigate special causes to be balanced against each other: statisti-
and get the process to perform as closely cal purity and practicality. The science of
as possible to the expectations of the improvement (SOI) is as Shewhart referred
customer (i.e., the specifications). to it an “applied science.” Therefore, in my
3. I have calculated a Cp and Cpk statistics work I have always tried to balance the
only once for a healthcare indicator. It precision of statistical requirements with
was when I was helping to set up an out- a heavy dose of practicality. For example,
patient clinic designed to manage patients I have worked with wonderful people in
on anticlotting medication (i.e., warfarin the National Health Service (NHS) of
sodium). Several key indicators are used Scotland for over 12 years. During this
in assessing clotting issues. The PT, along time we have developed a variety of health
with its derived measures of prothrombin and social service measurement systems.
ratio (PR) and INR, are assays evaluating Most of the data are collected monthly and
the extrinsic pathway of coagulation. many of the indicators were not collected
This test is also called “ProTime INR” historically. So, we were starting out with
and “PT/INR” (MedlinePlus medical no data on selected indicators and had
encyclopedia, https://medlineplus.gov/ to build charts as we went along. In this
ency/article/003652.htm). Because there case, trail control limits were essential.
are defined therapeutic limits associated We also made a very practical decision to
with these measures they can be regarded use the first 6 months of data as baseline
as USL and LSL. These values would be set for indicators that had no history. Again
on the Shewhart charts as reference lines. some would argue that this is not enough
Then the patient’s actual results on the PT data to establish a baseline but it was
and INR would be plotted on the chart, sufficient to get us started on the road to
and the UCL and LCL of the patient would improvement.
then be compared to the USL and LSL. 5. In one of my measurement workshops,
Because we had both an USL and a LSL and a few years ago this confusion was high-
control limits the capability statistics could lighted very clearly. A young woman
be calculated to determine how well the near the front of the room raised her
patient was conforming to the therapeutic hand after I was done explaining that a
limits (USL and LSL) of the drug. But in sigma was not equivalent to a SD. She
most instances in healthcare settings, there had a bit of a wrinkled brow and looked
is only a single target or goal rather than concerned. She said, “I was told that the
the USL and LSL. Healthcare simply does UCL and LCL were calculated as SD. Is
not currently function at the same level of this not correct?” I drew the formula for

9781284023077_CH09_211_258.indd 251 31/07/17 6:00 PM

 252 Chapter 9 Understanding Variation with Shewhart Charts

the SD on a flipchart and asked her if this for his 4-day seminars. They have written
is what she used to calculate the three a number of key books on QI including
sigma control limits. She said “Yes, that is The Improvement Guide (Langley, Moen,
what I was told to use.” I then proceed to Nolan, Nolan, Norman, & Provost, 2009),
politely tell her that the control limits on Quality Improvement Through Planned
her charts were wrong. The UCL and LCL Experimentation (Moen, Nolan, & Provost,
would either be too wide or too narrow 2012), and The Health Care Data Guide
if she used the SD of the data. She got a (Provost & Murray, 2011).
very strange look on her face, was quiet 7. When Dr. Ray Carey and I first started
for a moment, then burst out, “But this teaching control chart applications to
means I have been giving the charts to healthcare professionals in 1992, we
the senior management team and to the taught the traditional list of seven control
board! What am I going to do?” I asked charts. In 1995, we wrote a book that
her if anyone had ever noticed or com- described these seven charts and their
mented on the fact that the limits were not use (Carey & Lloyd, 2001). In December,
properly calculated. She responded, “No.” 2002, Dr. Carey and I taught a minicourse
I suggested that she learn how to make and two workshops on control charts
the charts correctly with SPC software at the 14th National Forum on Quality
(she was merely using Excel with no SPC Improvement in Health Care sponsored
add-on software that properly computes by the IHI. This was the first time in the
sigma limits) and then submit the correct 12 years that we had been teaching for
charts to the senior management team the IHI that we reduced the number
and board the next time around. She still of charts we taught from seven to five.
looked a little perplexed, however. She was The sessions were well received, and the
concerned that she would lose credibility participants found the more simplified
with the management team when they approach to be appealing. The two charts
found out the charts were wrong. I told her we dropped were the X-bar and R chart
that unless she tells them that her original and the np-chart. Our reasoning for doing
charts had the wrong limits it does not this was that the X-bar and S chart can
sound like anyone on the board or the be used in any situation that calls for the
senior management team had sufficient X-bar and R chart (when the subgroup
grounding in Shewhart charts to actually is greater than 2). The np-chart, which
discern that the charts were different. I is a count of the number of defectives,
told her to let me know how it went when requires equal subgroup sizes (i.e., the
she showed them the correct charts. She denominators), which do not happen very
wrote back and said that no one asked often in healthcare settings. The p-chart
any questions. can be used effectively, however, in any
6. API develops methods, works with leaders situation where an np-chart could be used.
and teams, and provides education and If there are equal subgroup sizes then the
training to help organizations improve their p-chart will have straight control limits.
products and services and to build their If, on the other hand, the measure has
capability for ongoing improvement. The unequal subgroups then the p-chart will
principals of API have worked in industrial, have what is known as “stair-step” control
educational, health, and social service limits. In this case, the control limits are
settings. They have worked extensively different for each data point. The closer
with Dr. Deming and provided support in size the denominators the smaller the

9781284023077_CH09_211_258.indd 252 31/07/17 6:00 PM

 Answers to the Chapter 9 Exercises 253

“steps” between each of the control limits. 11. There are many good examples of how
If there are large differences between the people have ignored special causes when
denominators the “steps” will be greater they first occurred and then decided to
between the individual data points. deal with them when they popped up
8. I wrote a commentary in JAMA a few again. The terrorist attacks on our nation
years ago titled “A Matter of Time” (Lloyd on September 11, 2001 provide a classic
& Goldmann, 2009) highlighting how example. Several years prior to 9/11
clinicians, researchers, patients, and the World Trade Center was bombed
improvement specialists all have very by terrorists. Although this seemed to
different views of time. To these four draw the nation’s attention for a while,
categories I could add management time, interest in this special cause soon faded
which focuses on monthly aggregates into the “old news” category and steps
of data. were not taken to extricate the factors
9. The control chart examples presented that led to the special cause. The condi-
in this chapter have been developed to tions for 9/11 were still existing within
demonstrate the five different charts. The our system. The September 11 special
substantive importance of the various cause, however, generated a completely
charts is not the focus of this chapter. The different reaction. Our nation mobilized
charts have been developed for heuristic not only to investigate the special cause
purposes, and the clinical or operational but also take steps to literally try to
impacts of the indicators presented on eliminate the origin of the special cause.
the charts are not the primary objective Every day there are stories in the news
in this chapter. Analysis and interpreta- that should prompt a discussion as to
tion of control charts are addressed in whether the event is a special cause or
Chapter 10. part of a common cause system. All too
10. The idea for creating this table came from often, however, we overreact to a special
Dr. James Benneyan of Northeastern cause and want to change the system
University in Boston. In a paper titled without fully investigating the reasons
“Design, Use, and Preferences of Statistical why it occurred. Other times, however,
Control Charts for Clinical Process Im- we ignore a special cause and “hope” that
provement” (September 16, 2001), he used it will not happen again. Hope is not a
a table to summarize the various charts. plan. Knowing how to appropriately react
After reading this paper, I realized that to common and special causes is a much
the table was something I had not used better approach than hoping a special
to summarize the control charts. I believe cause will not pop up again.
a table format works nicely to augment
the utility of the decision tree shown Answers to the Chapter 9
in Figure 9-12 and the textual details.
Dr. Benneyan has written extensively Exercises
on the topic of control charts in health This section provides the answers to the exer-
care and I would encourage readers to cises presented earlier in this chapter. The first
review his work. He can be reached at the EXERCISE 91 deals with differentiating defectives
following address: MIME Department, 334 from defects. EXERCISE 92 provides indicators
Snell Engineering Center, Northeastern that could be placed on either an X-bar and
University, Boston, MA 02115; phone 617- S chart or an XmR chart. The answers to these
373-2975; email benneyan@coe.neu.edu. two exercises are shown here.

9781284023077_CH09_211_258.indd 253 31/07/17 6:00 PM

 254 Chapter 9 Understanding Variation with Shewhart Charts

EXERCISE 91 Defective or defect? You make the call! (Answers)

Defective Defect
Indicator (Classification) (Count)

1. Number of accidents per 1,000 employee days *

2. Number of errors per 25 food trays *

3. Percentage of AMI patients receiving aspirin within *

24 hours of arrival in the ED

4. Percentage of inpatient deaths each month *

5. Number of surgical complications per 1,000 surgeries *

performed

6. Proportion of hand hygiene observations done incorrectly *

7. Number of falls per 1,000 patient days *

8. Number of medication errors per 10,000 doses dispensed *

EXERCISE 92 You make the call: Is it an X-bar and S chart or XmR chart? (Answers)

X Bar and S XmR

Indicator Chart (I Chart)

Time to clean an inpatient room (in minutes) *

Patient satisfaction scores for subgroups of 15 patients in the *

outpatient clinic

Average turnaround time for all STAT labs done each day *

Cost for each normal delivery *

A diabetic patient’s 3x a day blood sugar readings *

Average length of stay for a subgroup of 20 ICU patients *

The distance (in feet) that a sample of 10 knee replacement *

patients can walk in 15 seconds

9781284023077_CH09_211_258.indd 254 31/07/17 6:00 PM

 Answers to the Chapter 9 Exercises 255

The final EXERCISE 93 brings together the of this situation is that if the target is to have
key issues related to selecting the most appro- all patients be seen in 30 minutes or less, the
priate Shewhart chart for different measurement 30-minute target actually needs to be the UCL
situations. In this exercise, the subgroup, type of of the X-bar and S chart not the average. If 30
data, and type of chart all need to be specified. minutes is the average on the chart you will
Depending on how you interpret the word- naturally have some patients waiting more
ing describing the situations in Exercise 9-3, than 30 minutes and some waiting less. A
you might think that a type of chart other target is useful on a chart but it needs to be
than that I have listed could be selected. A understood in light of the actual variation in
key leaning point for this exercise is that slight the process and the capability of the current
changes to the wording of the situation could process to achieve the target. The Shewhart
lead you to selecting a different chart. For chart can help you determine the magnitude
example, take a close look at situations 8 and of improvement needed to achieve the target
11 in Exercise 9-3. The wording for situation and but in the case of improving wait time, this
8 points you to select a p-chart because they is best accomplished by not turning variables
decided to focus on patients who had to wait data into attributes.
more than 30 minutes. Even though they had The most appropriate chart for each situation
variables data (i.e., time) they basically turned described in Exercise 9-3 is shown here. Note
it into attributes data because of the 30-minute that situation 16 is a trick question. Did you
target. They have taken the more powerful determine that a chart cannot be identified?
form of data (variables data) and relegated it Why? Because there is no subgroup identified in
to a binomial condition, over 30 minutes and the situation description. Remember, a Shewhart
under 30 minutes. They will never understand chart must have a subgroup and an observation
the true variation in wait time. What is the as minimum requirements. In this situation,
longest wait? You have no idea. All we know is there is no subgroup. But if the situation had
that a certain percentage of patients had to wait been worded as follows then we would have
more than 30 minutes. The longest wait could a subgroup: “You know the number of people
be 31 minutes or 13,184 minutes. The more who come to the ED complaining of chest
appropriate approach is found in situation 11. pain EACH MONTH and the number who are
Here they are taking a sample of eight patients actually diagnosed with an AMI or unstable
each day and recording their actual wait times. angina.” Now you would be able to determine
The chart of preference in this situation is which chart is most appropriate. In this situ-
the X-bar and S chart. We will now have the ation, the Shewhart chart of choice would be
average wait time for a given day and the SD the p-chart because we know the denominator
from this average. We can lay a separate line (i.e., the number of people coming to the ED
of the chart showing the target of 30 minutes. complaining of chest pain) and the numerator
This gives us much more information about (i.e., the number who were actually diagnosed
the process variation and how capable it is of with an AMI or unstable angina). Without a
achieving the target, which cannot be deter- subgroup, however, we cannot make a decision
mined by using the p-chart. The final aspect about which chart is most appropriate.

9781284023077_CH09_211_258.indd 255 31/07/17 6:00 PM

 256 Chapter 9 Understanding Variation with Shewhart Charts

EXERCISE 93 You make the call!: Selecting the right chart (Answers)

Type of Type of
Situation Subgroup? Data? Chart?

1. Each day you record the number of films processed in Day V XmR
the radiology department.

2. Each day you record the number of films requested Day A p-chart
and the number that cannot be found in the
radiology library.

3. The number of inpatient restraints each month is Month A u-chart

placed over the total inpatient days each month.

4. Each day you pull a stratified random sample of Day V X-bar & S
15 CBCs and record the turnaround time (in minutes)
for each CBC.

5. The number of minutes it takes to get a stat med Stat med V XmR
order administered to the patient (order time to order
administration time).

6. Every 2 weeks you pull a stratified sample of 30 Two weeks A p-chart

medication orders and count the total number of
orders that have one or more errors.

7. The wait time in the ED (door to discharge) is tracked Patient V XmR

for each patient.

8. The clinic receptionist notes the time of check-in for Day A p-chart
each patient. The physician notes the time when he/
she first sees the patient in the exam room. An analyst
compiles the data daily and reports the percentage
of patients who had to wait more than 30 minutes.

9. The director of surgery keeps track of the total Week V XmR

number of surgical procedures performed each week.

10. The dietary department records the number of food Day A p-chart
trays that come back uneaten each day and the total
number of trays they produced for that day.

11. You are interested in the average time patients spend Day V X-bar & S
in your waiting area, so every day a student randomly
picks eight patients and measures their actual waiting
time in whole minutes.

9781284023077_CH09_211_258.indd 256 31/07/17 6:00 PM

 References 257

EXERCISE 93 You make the call!: Selecting the right chart (Answers) (continued)

Type of Type of
Situation Subgroup? Data? Chart?

12. The ICU nurses want to evaluate the ventilator- Two weeks A u-chart
associated pneumonia (VAP) rate. So every 2 weeks
they record the total number of pneumonia episodes
and the total number of vent days.

13. Each week patient satisfaction scores for three Week V XmR
units are compiled and an average is calculated for
the three units.

14. The finance department tracks the total number of A request V XmR
business days it takes to process a vendor’s request for payment
for payment. Process time starts when the request
for payment is received in the finance department
and ends when the payment is sent (electronically or
posted in the mail) to the vendor.

15. Every week each medication order is checked against Week A u-chart
five potential types of errors. The total number of
errors for the week is divided by the total number of
orders submitted that week.

15. You know the number of people who come to the Unknown* A Unknown*
ED complaining of chest pain and the number who
are actually diagnosed with an AMI or unstable
angina.

*NOTE: Item 16 is a trick question. A subgroup is not specified. Without a subgroup you cannot make a decision about the most appropriate chart. If
this description indicated that “You know the number of people who come to the emergency department EACH MONTH . . .” you would have a subgroup.
The chart of choice would then be a p-chart.

Benneyan, J., R. Lloyd, and P. Plsek. “Statistical Process Control

References as a Tool for Research and Health Care Improvement.”
ASQ Statistics Division. Glossary and Tables for Statistical Journal of Quality and Safety in Healthcare 12, no. 6
Quality Control, 4th ed. Milwaukee: ASQ Press, 2005 (December, 2003): 458–464.
Benneyan, J. “Design, Use and Performance of Statistical Blalock, H. Social Statistics. New York: McGraw-Hill, 1960.
Control Charts for Clinical Process Improvement.” Blank, R. The SPC Troubleshooting Guide. New York: Quality
Unpublished paper, Northeastern University, September Resources, 1998.
2001. For access to this paper contact Professor James Blumenthal, D. “Total Quality Management and Physicians’
C. Benneyan, Ph.D., MIME Department, 334 Snell Clinical Decisions.” Journal of the American Medical
Engineering Center, Northeastern University, Boston Association 269 (1993): 2775–2778.
MA 02115; tel: 617-373-2975; fax: 617-373-2921; Carey, R. Improving Healthcare with Control Charts. Milwaukee:
e-mail: benneyan@coe.neu.edu. Quality Press, 2003.

9781284023077_CH09_211_258.indd 257 31/07/17 6:00 PM

 258 Chapter 9 Understanding Variation with Shewhart Charts

Carey, R., and R. Lloyd. Measuring Quality Improvement Practitioners.” Quality and Safety in Health Care 17
in Healthcare: A Guide to Statistical Process Control (2008): 137–145.
Applications. Milwaukee: Quality Press, 2001. Montgomery, D. C. Introduction to Statistical Quality Control,
Deming, W. E. The New Economics, 2nd ed. Cambridge, 2nd ed. New York: John Wiley & Sons, 1991.
MA: Massachusetts Institute of Technology, Center for Nelson, L. “Interpreting Shewhart Average Control Charts.”
Advanced Studies, 1994. Journal of Quality Technology 17 (1985): 114–116.
Deming. “Quality, Productivity and Competitive Position.” Nelson, L. “A Control Chart for Parts-Per-Million Noncon-
Indianapolis, Indiana, August 11–14, 1992. forming Items.” Journal of Quality Technology 26,
Duncan, A. J. Quality Control and Industrial Statistics, 5th ed. no. 3 (1994): 239–240.
Homewood, IL: Irwin Press, 1986. Perla, R., L. Provost, and S. Murray. “The run chart: a simple
Gonick, L., and W. Smith. The Cartoon Guide to Statistics. analytical tool for learning from variation in healthcare
New York: Harper Perennial, 1993. processes” British Medical Journal Quality and Safety in
Grant, E., and R. Leavenworth. Statistical Quality Control. Healthcare 20, no. 1 (2011): 46–51.
New York: McGraw-Hill, 1988. Provost, L P., and S. Murray. The Health Care Data Guide.
Hare, L. “SPC: From Chaos to Wiping the Floor.” Quality San Francisco: Jossey-Bass, 2011.
Progress (July 2003): 58–63. Pyzdek, T. Pyzdek’s Guide to SPC. Vol. 1: Fundamentals.
Ishikawa, K. Guide to Quality Control. White Plains, NY: Milwaukee: Quality Press, 1990.
Quality Resources, 1989. Shewhart, W. Economic Control of Quality of Manufactured
Kume, H. Statistical Methods for Quality Improvement. Tokyo: Product. New York: D. Van Nostrand, 1931. Reprint,
Association for Overseas Technical Scholarship, 1996. Milwaukee: Quality Press, 1980.
Langley, G. L., R. Moen, K. M. Nolan, T. W. Nolan, C. L. Shewhart, W. Statistical Method from the Viewpoint of Quality
Norman, and L. P. Provost. The Improvement Guide: A Control. New York: Dover Publications, 1986.
Practical Approach to Enhancing Organizational Perfor- Western Electric Company. Statistical Quality Control
mance, 2nd ed. San Francisco, CA: Jossey-Bass, 2009. Handbook. Indianapolis: AT&T Technologies,
Levine, D., and D. Stephan. Even You Can Learn Statistics: Inc., 1985.
A Guide for Everyone Who Has Been Afraid of Statistics. Wheeler, D. Understanding Variation: The Key to Managing
Upper Saddle River, NJ: Pearson-Prentice Hall, 2005. Chaos. Knoxville, TN: SPC Press, 1993.
Lloyd, R., and D. Goldmann. “A Matter of Time.” Journal of the Wheeler, D. Advanced Topics in Statistical Process Control.
American Medical Association 302, no. 8 (2009): 894–895. Knoxville, TN: SPC Press, 1995.
Moen, R., T. Nolan, and L. Provost. Quality Improvement Wheeler, D., and D. Chambers. Understanding Statistical
Through Planned Experimentation, 3rd ed. New York: Process Control. Knoxville, TN: SPC Press, 1992.
McGraw-Hill, 2012. Woodall, W. “The Use of Control Charts in Health-Care
Mohammed, M., P. Worthington, and W. Woodall. “Plotting and Public-Health Surveillance.” Journal of Quality
Basic Control Charts: Tutorial Notes for Healthcare Technology 38, no. 2 (2006): 89–104.

9781284023077_CH09_211_258.indd 258 31/07/17 6:00 PM