STATISTICAL PROCESS CONTROL

Stoplight Charts
(With SPC Inside)
by Steven Prevette

become increasingly popular in recent years. Many

Q
uality Progress once published an article
titled “Exploiting the World's Most balanced scorecards employ the method, including
Recognized Standard,”1 which demon- ones used by ASQ.
strated the virtues of using the stop, caution and go The colors for the chart are established by com-
method. Stoplight charts, which use color coding paring current results to target thresholds. One
to judge performance as red, yellow or green, have threshold represents red, one represents green, and
anything in between is yellow. Software that auto-
mates the process is even becoming more available.
Load your target levels, load links to the appropri-
ate data sources (such as online analytical process-
In 50 Words ing cubes), and voilà, you have an automated sys-
Or Less tem that colors your performance red, yellow or
green, all untouched by human hands.
• Use stoplight charts with SPC to track defects. This operation is very quick, very efficient and
very harmful. Harmful? Let’s look at a classic
example used in many quality training sessions
• Stoplight charts are also a good way to report
over the years.
defect trends to management in a clear, under-
Tracking Defects
standable way.
I have a system in which I am counting defective
products in a day. Figure 1 charts the results. Fifty
• Eventually, stoplight charts can help prevent
items are produced each day, and the number of
defects and promote continual improvement. defective items is recorded and charted. As we
review the chart, we can see some days are better
than others. Days 16 and 22 are especially good
days, with only four defects each.

74 I OCTOBER 2004 I www.asq.org

 FIGURE 1 Number of Defects per Day

20
18
Number of defects observed

16
(out of a sample of 50)

14
12
10
8
6
4
2
0
Day 1

Day 2

Day 3

Day 4

Day 5

Day 6

Day 7

Day 8

Day 9

Day 10

Day 11

Day 12

Day 13

Day 14

Day 15

Day 16

Day 17

Day 18

Day 19

Day 20

Day 21

Day 22

Day 23

Day 24
Day of production

Let’s assume our company uses standard stop- ment. In the experiment, six workers each draw 50
light charts. After creating the chart, the next step beads at random from a bead supply. This is done
is determining our thresholds for red and yellow four times, for a total of 24 results. A paddle with
performance. Since the company has been having 50 holes in it is used to obtain the day’s produc-
problems with customer complaints of defects, we tion. White beads are the desired product—ideally
establish an achievable yet challenging goal of five 50 white beads per worker per day. Unfortunately,
or fewer defects as green performance. For the red there are a number of defective red beads in the
yellow boundary, we will double that and set the supply. Controls, incentives, quality control and
boundary at 10 to let us know days when we other management techniques are used to entice
should find out what went wrong. Six to 10 defects the workers into producing a quality product.
is yellow, and 11 or more is red. Never is the system changed—never are the red
Clearly, days 17, 19 and 20 were problems (see beads removed. Thus, the number of red beads
Figure 2, p. 76), and managers immediately imple- each worker produces varies in a random but over-
mented corrective actions. These actions worked, all predictable manner.
as day 22 was green. After throwing a pizza party So, if we had rewarded the workers and man-
for the workers, we again slip in performance and agers on days 16 and 22 for their green perfor-
have been at yellow for the past two days. The cor- mance, these workers would be left wondering
rective action management team is hard at work what they had done differently on those two days.
reviewing days 5, 9, 14, 17, 19 and 20 for root caus- The answer is nothing—it was a result of the ran-
es and adverse trends. Senior management has us dom variability in the system.
on its hot list of problem areas due to our erratic Next, consider the worker or manager who has
color performance. We sure hope we will be green had his or her pay withheld for red performance
again tomorrow. on day 19. What did he or she do differently? What
Anyone who has followed W. Edwards Deming’s can be done differently to avoid bad performance
work may have already guessed where these data again? If the process or system that created these
came from—performing Deming’s red bead experi- results is not changed, the answer again is nothing.

QUALITY PROGRESS I OCTOBER 2004 I 75

STATISTICAL PROCESS CONTROL

Workers will begin posting Dilbert cartoons, man- as the first bar chart and the first stoplight chart.
agers will be perplexed, and the business will fail. However, you gain a completely different level of
“But we were green yesterday,” will be the dying knowledge about the process from this chart. The
gasp of the chart maker as the final door of the data are stable about a center line of 9.1 and do not
building is closed and padlocked. go outside the control limits. As we question the
worker who created the outcome of 16 defective
Taking Control of the Process products on day 19, we find he or she did nothing
So what can we do differently to understand different than the worker who created the outcome
what the data are telling us and bring an end to the of four defective products on day 22. Both workers
cycle of randomly colored results? A first step in say they produced their work using the equipment
understanding the results of the red bead experi- and procedures provided by the company.
ment (or any stable and predictable process) is to Further, no amount of root cause analysis, failure
plot a control chart. Figure 3 (p.78) is a count chart mode and effects analysis (FMEA), corrective ac-
(c-chart) of the results. The upper control limit tion management orPareto analysis will be able to
(UCL) is plotted as a red line, and the lower control explain why there was a result of 16 on day 19,
limit (LCL) is plotted as a green line. which then dropped to four on day 22. All I can do
Note that Deming generally used a percent chart is tabulate that 16 red beads (and perhaps which
(p-chart) or np-chart during the red bead experi- red beads) fell into the holes in the paddle on day
ment. The c-chart is used here because the calcula- 19, and four fell in on day 22. No analysis of the
tions are easy enough to be done in real time while differences between these two results will tell you
conducting a red bead experiment, and the c-chart how the process can be improved. The control
control limits are close enough to the p-chart and np- chart shows the only way to improve performance
chart results that they yield the same interpretation. is to do so across all 24 outcomes, by making the
The control chart contains exactly the same data systems change of removing the red beads in the

FIGURE 2 Number of Defects per Day

20

18
Number of defects observed (out of a sample of 50)

16
Red

14

12

10
Yellow

4
Green

0
Day 1

Day 2

Day 3

Day 4

Day 5

Day 6

Day 7

Day 8

Day 9

Day 10

Day 11

Day 12

Day 13

Day 14

Day 15

Day 16

Day 17

Day 18

Day 19

Day 20

Day 21

Day 22

Day 23

Day 24

Day of production

76 I OCTOBER 2004 I www.asq.org

first place, or at least getting a can of white spray ering the leading reasons and causes for poor sys-
paint, one of my favorite suggestions from an audi- tem performance. There are many quality tools,
ence member during a red bead experiment I con- such as FMEA, barrier analysis, theory of con-
ducted. straints and root cause analysis, that also may help
At the end of every red bead experiment, unless here. Whatever the method, it is important to real-
performance improves, the company inevitably goes ize the system is long-term stable and to analyze
out of business as the customer finds the defect rate is the long-term performance of the system, not just
unacceptably high and stops buying beads. Certainly, the most recent results.
this is a bad result, and we ask if something could
have been done prior to losing customers. In an at-
mosphere of executive summaries, sound bites and
stoplights, how can we let management know action
is needed, and what sort of action? Rather than comparing
Get to the Bottom of It
Rather than comparing each day’s result to a
each day’s result to a
threshold, compare the stable baseline to a thresh-
old. We know from customer response the current
threshold, compare
level of 9.1 is not acceptable, and management con-
curs. Since the overall performance is not accept-
the stable baseline to
able, I will color the chart a color other than green.
The color red generally implies I need to stop, but
a threshold.
in the stable case it does no good to stop what I am
doing, as I would produce no beads at all. So I will
use the color yellow to indicate “stable at an unac- This creates a separation between the colors yel-
ceptable level.” The system, or process, is in need low and red, which implies what action is needed.
of improvement. If the performance level were If a chart is yellow, the chart has been showing sta-
acceptable, including potential results out at the ble performance, and only a change to the system
control limits, then the chart would be “stable at an that creates that performance will move it from yel-
acceptable level” and green. low to green. Deming called this a common cause
Assume on day 25 there are 26 red beads. That is system.2 If the chart is red, something has changed,
above the upper control limit and is not an expect- there is a statistically significant trend, and special
ed result of this stable process. I need to alert man- causes are at work. We declare all yellow results as
agement to stop and correct whatever changed to belonging to the system, and the system is in need
produce this result. Stopping is a natural reaction of improvement. We declare red results as special
to the color red. Therefore, I will use red to indicate cause, and in most cases the workers within the
an adverse trend. There is an immediate problem; system can make the appropriate corrections.
we should stop and take corrective actions. This application of yellow and red immediately
If there is an improving trend on the control integrates with systems thinking and a Deming
chart, I will set the color to green. This provides the based management approach. Deming estimated
feedback to management that movement has been most troubles and most possibilities for improve-
made in the right direction. Once the trend ends, I ment add up to these proportions: 94% belong to
review the new control chart baseline average. If the system (responsibility of management), and 6%
the chart is stable at an acceptable level, it receives are special. If Deming’s proportions hold, then 94%
a green color. If it is not acceptable, it is stable at an of the nongreen charts will be yellow, and 6% red.
unacceptable level and receives a yellow. Table 1 (p.79) is an overview of the four chart
To improve from a yellow rating, we must under- outcomes of acceptable or unacceptable level,
stand the long-term performance of the system. adverse or improving trend, and the relation to
The Pareto chart becomes a useful tool for discov- chart colors and action.

QUALITY PROGRESS I OCTOBER 2004 I 77

STATISTICAL PROCESS CONTROL

It’s Up to Management a statistically significant trend occurs. The chart will

Management owns the decision of whether a ignore single lucky results, such as days 16 and 22.
stable chart is acceptable or unacceptable, and it
must determine whether improvement is needed. Continual Improvement
Analysts can help management make this decision Another characteristic of this methodology is it
by gathering benchmark data, performing cost- forms a continual improvement cycle. If I start out
benefit and risk analyses and conducting customer with a problematic system and institute these new
interviews and surveys. However, the final deci- charts, there will likely be sporadic reds on the
sion is management’s. It may choose to execute a charts. I can then remove the special causes of the
policy of continual improvement, pick a small reds through corrective action and gaining control
number of stable systems for improvement, do the over the process.
improvement and move on to others. It is not nec- The chart will next settle into a phase of being
essary to make a new decision on each update as consistently yellow as it is stable and predictable
to whether a stable system is yellow or green. This but performing in an unacceptable manner. Changes
is a one-time decision that remains in effect until a will be made to the system and the impact of those
trend occurs or other priorities change and force a changes monitored on the control chart. Once a statis-
re-evaluation of this system. tically significant trend in the improving direction
I could also incorporate numerical targets. In the occurs, the chart will be set to green to reflect this suc-
red bead example, I could establish the value of five cess. When the chart stabilizes at a new baseline, new
as the target. The major difference in this methodol- average and control limits will be added, and those
ogy compared to other stoplight methods is I’m will be reviewed by management to see if an accept-
comparing the baseline average to the target, not able level has been reached. If further improvement is
the current result. Since the current baseline aver- needed, the chart is reset to yellow, and the next sys-
age is above five, the chart will remain yellow until tem improvement is implemented.

FIGURE 3 Control Chart of Number of Defects per Day

20

18
Number of defects observed (out of a sample of 50)

16
Average for days
14 1 - 24 = 9.1

12

10

0
Day 1

Day 2

Day 3

Day 4

Day 5

Day 6

Day 7

Day 8

Day 9

Day 10

Day 12

Day 13

Day 14

Day 15

Day 16

Day 17

Day 18

Day 19

Day 20

Day 21

Day 22

Day 23

Day 24
Day 11

Day of production

78 I OCTOBER 2004 I www.asq.org

 Figure 4 is a flowchart that represents the contin- TABLE 1 Overview of Four Possible
ual improvement cycle established by this method- Chart Outcomes
ology.
Control Management
The Payoff chart result Decision Color action needed
In practice, this system has worked well where Level is Green None
use of statistical process control (SPC) has been Stable acceptable
accepted, although some managers express frustra- (common cause) Level is not Yellow Improve the
tion when they are still yellow even though last acceptable system
month was good. The system is reasonably insensi- Trend is in Red Correct the
tive to random noise in the data, avoiding both adverse direc- problem
false declarations of improvement and false decla- Trend tion
rations of adverse trends. The method requires (special cause) Trend is in Green Keep the
some patience in waiting for assurance that the improving trend going
trend is significant, but the payoff is avoiding false direction
alarms and false payoffs.
One practical consideration is that charts go from
green (stable at an acceptable level) immediately to
red (adverse trend) when a trend
develops. One workaround to ex-
periment with is to set a chart to red FIGURE 4 A Method To Color-Code Performance
only if there is both an adverse trend
Indicator Results for Executive Summary
and the current level has exceeded
• Benchmarks. Collect initial data, perform
the acceptable threshold. This al- • Customer expectations. initial baseline control chart.
lows detection of a trend away from • Management philosophy.
• Risk analysis.
the current stable level without • Cost benefit analysis. Control chart update.
immediately leaping to red if the
level is still good. The disadvan- Establish target.
tage is it removes the linkage be- (Optional)
tween red as a special cause and Does a
No statistically Yes
yellow as a common cause varia- Evaluate significant trend
baseline average and exist currently?
tion. control limits.
Finally, we have run across No Is trend
some instances in which the cus- in improving
direction?
tomer establishes a specification
for performance. In this case, we
Yes
can reserve the red color for violat- Is improvement No

ing specifications and merge the needed?

Red Green
SPC red and yellow criteria into
just yellow. In addition, we can set Yes
the acceptable baseline by examin-
Yellow Green Identify special causes,
ing the likelihood of violating the take corrective actions.
specification. This concept is simi-
lar to the Six Sigma concept of Determine common causes,
implement system change.
achieving stable performance at If trend has ended, evaluate need
six standard deviations from the for new average and control limits.
specification limits.
This system of merging stop-
light presentations with SPC offers Management review, publish charts and continue to next update.

QUALITY PROGRESS I OCTOBER 2004 I 79

STATISTICAL PROCESS CONTROL

advantages of both methodologies. The use of red, minimizes the random flickering of colors as data
yellow and green colors offers the quick executive randomly cross thresholds and also links the con-
summary of performance and quickly highlights cepts of special cause and random cause variation
areas in which action is needed. The use of SPC to the red and yellow.

REFERENCES

1. Paul Palady, “Exploiting the World’s

Most Recognized Standard,” Quality
Progress, February 2001, pp. 54-61.
2. W. Edwards Deming, Out of the Crisis,
Massachusetts Institute of Technology,
1982.

RESOURCES

For more information on statistical

process control:
Duncan, Acheson J., Quality Control and
Industrial Statistics, McGraw-Hill, 1986.
Ott, Ellis R., Process Quality Control: Trouble-
shooting and Interpretation of Data,
McGraw-Hill, 1975.
Prevette, Steven, “Cleaning Up With SPC,”
Quality Progress, September 2001,
pp. 104-107.

STEVEN PREVETTE is a quality engineer

for Fluor Hanford and a statistics lectur-
er at City University, both in Richland,
WA. He has a master’s degree in manage-
ment science operations systems from the
Naval Postgraduate School in Monterey,
CA. Prevette is an ASQ member, a certi-
fied quality engineer and chair of Colum-
bia Basin Section 614.

Please
comment
If you would like to comment
on this article, please post your
remarks on the Quality Progress
Discussion Board at www.asq.org,
or e-mail them to [email protected].

80 I OCTOBER 2004 I www.asq.org