Statistical Process Control

Up to now the class has covered basic business statistics and regression analysis. We

now move to an application of statistics known as Operations Management. It involves

(in part) the challenge of designing and operating processes that provide a service

package to the total satisfaction of customers. The failure to satisfy customers (be they

internal or external) is a process failure. Thus, evaluating process performance is an

important element of process analysis.

I. Total Quality Management

TQM is a philosophy that stresses three principles for achieving high levles of

process performance and quality: Customer Satisfaction, Employee Involvement, and

continuous improvement in performance.

1. Customer Satisfaction

Customers (internal or external) are satisfied when their expectations

regarding a service or product have been met or exceeded.

• Conformance to specifications

• Value

• Fitness for use – how well the service or product performs its intended

purpose.

• Support

• Psychological Impressions – atmosphere, image, or aesthetics

2. Employee Involvement

1
 • Cultural Change. Under TQM everyone is expected to

contribute to the overall improvement of quality. Thus one of

the challenges is to define customer for each employee. The

external customer(s) are often far removed from particular

employees. Thus the notion of internal customers is important

here.

• Teams.

3. Continuous Improvement

Continuous improvement involves identifying benchmarks of excellent

practice and instilling a sense of employee ownership in the process.

Generally firms will use a “plan-do-check-act” cycle in their problem-solving

process

II. Statistical Process Control

One practical type of continuous improvement is the use of statistical process control.

This is the application of statistical techniques to determine whether a process is

delivering what the customer wants. SPC primarily involves using control charts to

detect production of defective services or products or to indicate that the process has

changed and that services or products will deviate from their design specifications

unless something is done to correct the situation. Examples:

• A decrease in the average number of complaints per day at a hospital,

• A sudden increase in the proportion of bad lab tests,

• An increase in the time to process a lab test, chart, billing claim, etc.

2
 • An increase in the number of medication errors

• An increase in the absenteeism rate in a particular nursing unit.

• An increase in the number of claimants receiving late payment from an

insurance company.

Suppose that the manager of the accounts payable department of an insurance

company notices that the proportion of claimants receiving late payment has risen

from an average of .01 to .03. Is this a cause for alarm or just a random occurrence?

Note that if it is random, any resources devoted to “fixing” the problem would be

wasted, but if there is truly a problem, then it may be worthwhile to attempt to fix.

Variation of Outputs

Even if the processes are working as intended there will be variation in outcomes, but

it is important to minimize the variation because variation is what the customer sees

and feels. We can focus on the types of variation:

1. Common Causes -- these are purely random, unidentifiable sources of

variation that are unavoidable with the current process. Statistically, this

is referred to as “noise”

2. Assignable Cause – any variation-causing factors that can be identified

and eliminated. An employee that needs training, or a machine that needs

repair.

To detect abnormal variations in process output, employees must be able to measure

performance variables. One way is to measure variables – that is, service or product

3
characteristics, such as weight, length, volume, or time that can be measured. Another

way is to measure attributes – characteristics that can be quickly counted for acceptable

performance. Ex: the number of insurance forms containing errors that cause

underpayments or overpayments, the proportion of radios inoperative at the final test, the

proportion of airline flights arriving within 15 minutes of scheduled times, etc. The

advantage of attribute counts is that less effort and fewer resources are needed than for

measuring variables, but the disadvantage is that, even tough attribute counts can reveal

that process performance has changed, they may not be of much use in indicating by how

much.

Control Charts

In order to decide if the variation is out of whack, statistical process control methods use

control charts. These are time-ordered diagrams that are used to determine whether

observed variations are abnormal.

We can use the following control chart decision tree (taken from page 20 of Carey’s

Improving Healthcare with Control Charts), to help classify the different types of charts

that are possible.

4
 Type of Data

Measurement Data Count Data

Continuous variables: time, Discrete variables:
money, length, height, Number of errors, yes/no,
temperature pass/fail/ etc.

Each subgroup Each subgroup is Can only count Nonconforming

has more than one composed of a nonconformities units (defectives)
observation. single (defects), eg. Errors, are counted as
observation. complications, falls, percentages. An
needle sticks per entire unit either
subgroup, meets or fails to
Numerator can be meet criteria.
greater than Numerator can’t
denominator be greater than
the denominator,
e.g. mortalities, c-
sections, etc.

Subgroup Subgroup Equal area of Unequal are of Unequal or

Size >1 size =1 opportunity opportunity equal subgroup
size

X-Bar and I-Chart C-Chart U-Chart P-Chart

S-Chart

5
 The chart first splits the decision into two types – those using continuous variables

and those using count or discrete data. On the continuous side there are two further

classifications. When you have information about the subsamples (say you are

looking at the average LOS per week, and you have 100 patients per week to get that

average), then we use an X-bar and S-chart (Xbar for average, and S for standard

deviation). If, however, we only have information on the average (say we only have

the average LOS per week but not the individual observations that generated those

averages), then we use an I chart.

Control Charts for Variables

I chart. In the accompanying Excel spreadsheet (SPC Examples.xls) on the

worksheet labeled I-chart, is an example of using an Ichart. These data are the

Weekly average of delays between an abnormal mammogram and biopsy.

Presumably we’d want this number to be as low as possible. What is shown are

the average days delay per week over a 36 week period. After week 20 an

“intervention” was instituted that was intended to reduce the average delay. We

want to know if the intervention helped to reduce the delay. Note that all the

information we have is the average delay per week, we do not have the individual

data that went into making these averages. Thus we have to construct an I chart.

Basically what all of these charts do is construct a confidence interval that moves

through time, then by tracking how each new period’s data falls within that range

we can make judgments about how we are doing.

Week 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Average Delay 34 30 35 32 28 26 29 28 35 26 34 31 28 40 26 32 31 30 33 35

6
To construct the chart, we first calculate the average of the average weekly delay,

this is 31.15. The standard deviation is 3.70. Next we construct the upper and

lower control limits. Generally these are constructed to be 3 standard deviations

above and below the mean. So UCL = Xbar + 3S, and LCL = Xbar -3S. You can

also do this using “2 –sigma” limits as well. Basically it becomes a tradeoff

between type I and type II errors. Note that these give slightly different numbers

from what the book uses. The book uses a formula based on the range of the

data. I prefer using the standard deviation.

Doing this and putting it all in a graph gives us the following:

I Chart
Average Weekly Delays

45
40
Average Delay Per Week

35
30
25
20
15
10
5
0
0 5 10 15 20 25 30 35 40
Week

Average Delay Xbar UCL LCL

7
So the Xbar, UCL, and LCL were constructed on the first 20 weeks of data

(before the intervention). Plotted are the first 20 weeks, plus the 15 weeks that

followed. So what can we say?

Detecting Special Causes

We want to be able to distinguish “information” from “noise” or Special

(assignable) causes from common causes. That is when should we pay attention

and when should we ignore?

1. A special cause is indicated when a single point falls outside a control

limit. In weeks 26 and 30, note that we are below the LCL, that is we are

more than 3 standard deviations below the mean. It is pretty unlikely for this

to be a random event (less than a 1 percent chance), so we would say this is a

special cause – something different has happened here.

2. A special cause is indicated when two out of three successive values are:

a) on the same side of the centerline, and b) more than two standard

deviations from the centerline. The 2 sigma LCL is 23.7, so in weeks

21,22,23 we have two of three observations below this.

3. A special cause is indicated when eight or more successive values fall on

the same side of the centerline. So we get this in the above chart weeks 21

to 28 are all below the centerline.

4. A special cause is indicated by a trend of six or more values in a row

steadily increasing or decreasing. This is not shown in the above graph.

8
Using the above criteria we can know say something about the intervention. First

note that in the first 20 weeks of data there are no special causes – things are

pretty stable, but after week 20 we get a different picture, special causes are

detected from Tests 1, 2, and 3. So we could conclude that the “world has

changed” We could re-do the graph to show this:

I Chart showing Intervention

45
40
35
Average Delay

30
25
20
15
10
5
0
0 10 20 30 40
Week

Average Delay Xbar UCL LCL

These are the same data, but it now shows the new mean, UCL, and LCL after the

intervention. So now when we get future data, we compare them to the new

numbers, etc. This is the continuous improvement idea.

Recall that this type of chart only had data on the mean per week and so we had to

tread each sample as a data point and use the standard deviation (as opposed to

9
 the standard error) to construct the limits. This makes them larger than the would

be if we had sample information. This is what the X-bar chart does:

X-Bar and S-Chart When each subgroup has more than one observation then we

can use this information to our advantage by accounting for the sample variations.

In the worksheet titled “X-Bar and S-Chart” is an example of this. Here we have

lab turnaround time from lab to ED using a sample of three tests each day for 23

consecutive weekdays (you don’t have to have the same number of observations

per period, but as we’ll see it is easier if you do).

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

test1 86 90 101 76 102 81 75 92 93 109 70 80 85 69 106 89 85 95 72 95 75 60 77

test2 73 82 74 71 76 82 50 65 71 92 84 79 63 71 93 95 101 89 60 84 97 110 55

test3 75 95 89 105 115 55 95 93 82 76 67 58 110 112 82 73 68 88 97 61 115 56 99

xbar 78.0 89.0 88.0 84.0 97.7 72.7 73.3 83.3 82.0 92.3 73.7 72.3 86.0 84.0 93.7 85.7 84.7 90.7 76.3 80.0 95.7 75.3 77.0

st dev 7.0 6.6 13.5 18.4 19.9 15.3 22.5 15.9 11.0 16.5 9.1 12.4 23.5 24.3 12.0 11.4 16.5 3.8 18.9 17.3 20.0 30.1 22.0

So xbar is the average for each day, sdev is the standard deviation for each day. If

we take the average of the average (Xbarbar) we get 83.28, and if we take the

standard deviation of all the sample means (Sbar) we get 15.99. Now we can

construct the UCL and LCL as:

UCL = Xbarbar + 3*Sbar/sqrt(n)

LCL = Xbarbar – 3*Sbar/sqrt(n)

Where n is the size of the sample from each day – so if the sample sizes are the

same for each period the UCL and LCL will be the same across the chart, but if

the sample sizes vary, then the UCL and LCL will also vary. Doing this and

graphing gives us the Xbar chart:

10
 X-Bar Chart
CBC Turnaround Time

110.0

100.0

90.0
Minutes

80.0

70.0

60.0

50.0
0 5 10 15 20 25
Day

xbar Xbarbar UCL LCL

Note that things here look pretty stable: there are no observations outside the 3-

sigma limits. The two sigma limits are 101.7 and 64.8 and no observations are

outside of them either. There are not eight successive values above or below the

centerline, and there is not a trend of six or more.

Note that we also can (and should) look at what is happening to the variance over

time. This is the sbar chart. Basically we do the same thing with the standard

deviation as we did with the mean. We know the standard deviation for each

period, and we can construct the average of the standard deviations and look at

how day to day observations bounce around the standard deviation. First we

construct the average of the standard deviations and then the standard deviation of

the standard deviations. Then use 3 times this standard deviation to construct the

11
 UCL and LCL. Note that if the LCL is calculated to be negative, we set it equal

to zero since negative values do not make sense.

Sigma Chart CBC turnaround time

40.0

30.0
minutes

20.0

10.0

0.0
0 5 10 15 20 25
Day

st dev Sbar UCL LCL

Again things look pretty stable. In practice one would first want to look at the s-chart to

make sure the process was stable, and then go to the xbar chart, but both can help identify

abnormalities in the process. A good way to think about it is that the xbar chart looks at

variations overtime (or across subgroups) while the s chart looks at variation within

groups.

Control Charts for Count or Attribute Data

P-chart The p-chart is probably the easiest to deal with. In this case we have a

percentage or proportion of something that we are tracking over time. On the worksheet

titled “p-chart” data for Readmission Rates after Congestive Heart Failure 1998-2000. In

January 2000 an intervention occurred (a case management protocol), and so we want to

know if things have improved.

12
So we know how many patients were admitted for heart failure and how many of them

were later readmitted, and thus we know the proportion of readmit for each month. To

construct the control chart, we first calculate the total proportion of readmission for the

period prior to the intervention (1998 and 1999) this is Pbar= .125, then to construct the

UCL and LCL we calculate Sigma: = Sqrt [(p)*(1-p)/(n)]. This should look somewhat

familiar (think back to the standard error when doing hypothesis tests on a proportion).

Note that the n is the sample size for each period which varies, thus the UCL and LCL

will vary across the chart. So the UCL is Pbar + 3 times the sigma for each month and

the LCL is the maximum of Pbar – 3sigma and zero.

P Chart
Readmission Rate

0.35
0.30
0.25
Percent

0.20
0.15

0.10
0.05

0.00
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 25 26 27 28 29 30 31 32 33 34 35 36

Month

Percent Readmission Pbar UCL LCL

So note that prior to the intervention there is only common cause variation (all the

variation is noise), but after month 24 we get 9 consecutive months below the centerline

(test 3) and so conclude that the plan seems to have been successful – at least for a while.

13
Note that the last 3 months of 2000 show a percentage back above the centerline. So

further tracking would be needed before concluding things were better.

Subgroup sizes for P-charts. P charts are likely to be especially sensitive to small

sample sizes. One simple rule is that the subgroup size should be big enough such that

there is at least on event or occurrence in every subgroup – so there are no zero percent

occurrences. Alternatively, some argue it should be large enough to get a positive LCL.

The American Society for Testing and Materials (ASTM) has set guidelines for p-charts:

The lower limit on subgroup size for p-charts may not yield reliable information when:

1. Subgroups have less than 25 in the denominator, or

2. the subgroup size n multiplied by Pbar is less than one.

U-Chart. When we have count data and different sample sizes for each period – where

there is an unequal area of opportunity. For example on the worksheet labeled U-chart

are data that show the number of code blues as well as the number of patient days per

month, from April 1992 to June 1995. Note that you have a count -- the number of code

blues and you also have a varying number of patient days. In months with a higher

census you’d expect more codes even if things were still “normal” so you want to

account for this to the extent you can. Also an x-bar chart probably would not be

appropriate since the count is not really normally distributed. Likewise one could

calculate the proportion of code blues and do a pchart, but since codes are such rare

events most of the proportions would be close to zero and so it would be difficult to pick

up any action. A U-chart is generally more powerful than the pchart since it will take all

this information into account.

14
First we calculate Ubar – the average proportion of codes per patient day: Ubar =

number of defects for all subgroups/total number of observations in all subgroups. In this

example we get Ubar = .0042 or about 4 codes per 1,000 patient days. Then the sigma =

sqrt(Ubar/n) where n is the number of observations in each period. So sigma will vary

across subgroups. Then the control limits = Ubar ±3*Sigma. Then we get:

U Chart
Ratio of Code Blues per Patient Day

0.012

0.01

0.008
Ratio

0.006

0.004

0.002

0
0 1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839

Month

codes per day Ubar UCL LCL

15
Looking at the graph does not reveal any special causes – but it is close to test #4 where

we have a trend of 5 in a row of increasing, but officially you need 6 in a row.

C-Chart The final case to discuss is the c-chart. This is an alternative to the U-chart, but

when there is equal opportunity for defects (or when the opportunity is unknown). So

suppose on the code blue data we only knew the total number of codes per month, but not

the patient days. Now we have to assume that codes are equally likely across months and

we look at how the actual counts vary across months. This is done on the C-chart

worksheet. Now we first calculate Cbar = average number of defects over the period,

cbar = 5.72. Then the standard deviation = sqrt(cbar) This is assumes the count data

follows the hypergeometric distribution. So now we get the following C-chart. Note that

the UCL no longer varies across the sample but is constant.

C-Chart
Count of Code Blues

14
12

10
Count

8
6
4

2
0
0 1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536373839

Month

codes Cbar UCL LCL

16
We generally get the same picture here, but the U-chart generally is more powerful than

the C-chart since it has more information in it. Similarly the Xbar chart is more powerful

than the I chart. But sometimes you just don’t have the information needed to do the U

chart or Xbar chart.

Subgroup sizes for C-charts and U-charts. The ASTM suggests that, to provide

reliable information, the subgroup size for U-chart should at least be equal to one divided

by the average of nonconformities (Ubar), but will be “most useful” when the subgroup is

at least equal to four divided by Ubar. For example, if the average number of medication

errors at a hospital is four per 1000 (.004), The U-chart would be most useful when the

subgroup size (the number of medication orders) was at least 4/.004 or 1000.

For C-charts the subgroup size should be large enough that the average

count of nonconforming items (cbar) is at least greater than one, but preferably greater

than 4.

III. Process Capability

Statistical process control helps mangers achieve and maintain a process distribution

that does not change in terms of its mean and variance. The control limits on the

control charts signal when the mean or variability of the process changes. A process

that is in statistical control, however, may not be producing services or products

according to their design specifications because the control limits are based on the

mean and variability of the sampling distribution, not the design specifications.

17
Process Capability refers to the ability of the process to meet the design specifications

for a service or product. Design specifications often are expressed as a target and a

tolerance. For example, the administrator of an ICU lab might have a target value for

the turnaround time of results to the attending physicians of 25 minutes and a

tolerance of ± 5 minutes because of the need for speed under life-threatening

conditions. The tolerance gives an upper specification of 30 minutes and a lower

specification of 20 minutes.

The administrator is also interested in detecting occurrences of turnaround times

of less than 20 minutes because something might be learned that can be built into the

lab process in the future.

Lower Upper
Specification Specification

20 25 30

The above process is capable

Upper
Lower Specification
Specificatio

20 25 30

Process is not capable

18
 The idea here is kind of in reverse from the control charts. There we let the data

decide what the limits were and looked to see if there were any outliers. But now we

are saying, “lets define what our limits are and then see if our data fit into them”. If

they do not, then we change the process until they do. The above diagrams show one

process that is “capable”, that is, it is working within the specifications. The bottom

diagram is not.

There are two measures commonly used in practice to assess the capability of a

process: Process capability ratio and process capability index.

Process Capability Ratio. A process is capable if it has a process distribution

whose extreme values fall within the upper and lower specifications for a service or

product. As a general rule, most values of any process distribution fall with ± 3

standard deviations. [Specifically 68.26% are within one SD, 95.44 are within two,

and 99.73 are within 3] In other words, the range of values of the quality measure

generated by a process is approximately 6 standard deviations of the process

distribution. Hence if a process is capable, the difference between the upper and

lower specification, called the tolerance width, must be greater than 6 standard

deviations. The process capability ratio, Cp is defined as:

Upper Specification − Lower Specification

Cp =
6σ

where σ is the standard deviation of the process distribution.

19
 A Cp value of 1.0 implies that the firm is producing three-sigma quality (.26

percent defects) and that the process is consistently producing outputs within

specifications even though some defects are generated. Values greater than 1 imply

higher levels of quality achievement. Firms striving to achieve greater than three-

sigma quality use a critical value for the ratio greater than 1. A firm targeting six-

sigma quality will use 2.0, a firm targeting 5 sigma quality will use 1.67, etc.

Process Capability Index. The process is capable only when the capability ratio

is greater than the critical value and the process distribution is centered on the

nominal value of the design specification. For example, the lab process may have a

process capability ratio greater than 1.33 for turnaround time. However, if the mean

of the distribution of process output, x , is closer to the upper specification, lengthy

turnaround times may still be generated. Likewise if x is closer to the lower

specification, very quick results may be generated. Thus, the capability index

measures the potential for the output of the process to fall outside of either the upper

or lower specifications.

The process capability index, Cpk, is defined as:

⎡ x − lower specification upper specification − x ⎤

Cpk = Minimum : ⎢ , ⎥
⎢ 3σ 3σ ⎥
⎣ ⎦

We take the minimum of the two ratios because it gives the worst-case situation.

If Cpk is greater than the critical value (say 2 for six sigma quality) and the process

capability ratio is also greater than the critical value, we can say the process is

20
capable. If Cpk is less than the CV, either the process average is close to one of the

tolerance limits and is generating defective output, or the process variability is too

large.

Example:

The intensive care unit lab process has an average turnaround time of 26.2

minutes and a standard deviation of 1.35 minutes. The target value for this service is

25 minutes with an upper specification limit of 30 minutes and a lower specification

limit of 20 minutes. The administrator of the lab wants to have four-sigma

performance for her lab. Is the lab process capable of this level of performance?

The first step is to check to see if the process is capable by applying the process

capability index:

Lower Specification = (26.2-20)/3(1.35) = 1.53

Upper Specification = (30-26.2)/3*(1.35) = .94

So the minimum is .94

Since the target value for four-sigma is 1.33 (4σ/3σ), the process capability index

tells us the process is not capable. But note this doesn’t tell us if the problem was the

variability of the process , the centering, or both.

Next we look at the process variability with the process capability ratio:

Cp = (30-20)/6(1.35) = 1.23

So this does not meet the four-sigma target of 1.33. Thus, there is too much

variability. Suppose the administrator initiated a study and found that two activities:

report preparation and specimen slide preparation were identified as having

21
 inconsistent procedures. When these procedures were modified to provide more

consistent performance, new data were then collected and the average turnaround was

now 26.1 minutes with a sd of 1.2.

Now: Cp = (30-20)/6(1.20)= 1.39

So we have process capability.

But note the capability index still has problems:

Lower: (26.1-20)/3(1.2) = 1.69

Upper: (30-26.1)/3(1.2) = 1.08

Thus we have 3 sigma capability, but not 4 sigma. The variability is OK, but we are off

center – 26.1 is still too high.

IV. Six Sigma

Six Sigma is a comprehensive and flexible system for achieving, sustaining, and

maximizing business success. It uses the statistical analysis described above along with a

focus on understanding customer needs, and attention to managing, improving, and

reinventing business processes.

Motorola is credited with developing Six Sigma in the 1980s to improve its

manufacturing capability in a market that was becoming very competitive. Management

noticed they were getting many complaints and competitor products were outperforming

its products. They responded to this by soliciting new ideas and benchmarking their

competitors and followed with extensive changes to employee compensation and reward

programs, training programs, and critical processes. At one plant, after 10 months, the

defect rate improved 70 percent and the yield improved 55 percent.

22
 The procedures for achieving those results were documented and refined and

became known as Six Sigma. The notion being that your variation in output be so low

that your range covers six standard deviations of the target range – or that a process

generating Six-Sigma quality would have only .002 defects per million opportunities.

General Electric is credited with popularizing the application of the approach to

nonmaufacturing processes such as sales, human resources, customer service, and

financial services. The concept of eliminating defects is the same, although the definition

of “defect” depends on the process involved. Some of the challenges in converting to

nonmanufacturing processes involve:

1. The “work product” is much more difficult to see because it often consists of

information, requests, orders, proposals, presentations, meetings, invoices,

designs, and ideas. This makes it difficult for people working in diverse

functional areas such as sales, marketing, and software development to

understand that they are actually part of a process that needs analysis.

2. services processes can be changed quickly. Service processes in many

companies evolve, adapt, and grow almost continuously. This makes the

analysis part difficult.

3. hard facts on process performance are often hard to come by. The data that do

exists are often anecdotal or subjective.

The Six Sigma Improvement model is a five-step procedure that leads to improvements

in process performance.

Define – determine the characteristics of the process’s output that are critical to

customer satisfaction and identify gaps.

23
 Measure – quantify the work the process does that affects the gap.

Analyze – use the data on measures to perform process analysis, which may be

focused on incremental process improvement or major process redesign.

Improve – Modify or redesign existing methods to meet the new performance

objectives

Control – Monitor the process to make sure that high performance levels are

maintained.

Implementation.

Implementing Six Sigma requires much time and commitment.

Top-Down Commitment

Measurement System to Track Progress

Tough Goal Setting

Communication

Customer Priorities

Education – employees must be trained in the whys and how-tos of

quality and what it means to customers, internal and external. This is accomplished by
“train-the-trainer” programs. Firms using Six Sigma develop a group of internal teachers
who ten are responsible for teaching and assigning teams involved in a process
improvement project. Green Belts – devote part of their time to teaching and helping
teams with their projects and the rest of their time to their normally assigned duties.
Black Belts – are full-time teachers and leaders of teams. Finally Master Black Belts are
full time teachers who review and mentor Black Belts.

24