Featured Product
This Week in Quality Digest Live
Health Care Features
Etienne Nichols
How to give yourself a little more space when things happen
Chris Bush
Penalties for noncompliance can be steep, so it’s essential to understand what’s required
Jennifer Chu
Findings point to faster way to find bacteria in food, water, and clinical samples
Smaller, less expensive, and portable MRI systems promise to expand healthcare delivery
Lindsey Walker
A CMMS provides better asset management, streamlined risk assessments, and improved emergency preparedness

More Features

Health Care News
Showcasing the latest in digital transformation for validation professionals in life sciences
An expansion of its medical-device cybersecurity solution as independent services to all health systems
Purchase combines goals and complementary capabilities
Better compliance, outbreak forecasting, and prediction of pathogens such as listeria or salmonella
Links ZEISS research and capabilities in automated, high-resolution 3D imaging and analysis
Creates one of the most comprehensive regulatory SaaS platforms for the industry
Resistant to high-pressure environments, and their 3/8-in. diameter size fits tight spaces
Easy, reliable leak testing with methylene blue
New medical product from Canon’s Video Sensing Division

More News


Health Care

Graphical Principles for Rapid Quality Improvement

Forget attributes data. Variables data is where improvement lies.

Published: Wednesday, December 2, 2009 - 06:00

I’ve spent most of the past two years living in China where I have learned much on how enterprise is managed over there. Many people have said that this century belongs to Asia. That may be, but they have a lot to learn and change before that happens. They cannot depend on cheap rote labor to claim the century.

All of the graphical examples that I present here are based on real data from Chinese companies. The data were stored in a computer or in paper files or were gathered based on my questions. These graphs did not exist before I made them. I have also “sanitized” them to protect the identity of the various companies involved.

Tables of spreadsheet data that look at yesterday’s failures or successes are the norm in China. Defects are recorded and reported using the latest in expensive, automatic equipment for shop floor control systems. One hundred percent final test with reinspection and rework of failed items supposedly guarantee shipped quality. They do not.

What’s the problem? Pass/fail and go/no-go procedures provide little information for rapid improvement. Using these types of attribute data guarantees higher cost and lower quality and longer time to fix than using the variables data that remains ignorantly locked in expensive automatic equipment.

The inefficiencies are not confined to China. Management throughout the world still thinks that quality is no defects, no complaints, and no problems. I’ll show you why they are wrong.

Remember: The only reason to collect data is to take action.

Figure 1

Data in business exists because someone somewhere asked a question. “Did we ship our production quota yesterday?” “Is OPEX on target today?” “Were there any customer returns last shipment?” “Will you make this quarter’s budget?”

Some questions are better than others because they are based on theories that pertain and are more useful to the situation at hand. Whether you realize it or not, every question is preceded by a theory or point of view: “If I make my quota, then I will not need help from corporate.” “If there are few customer returns, then I will not need help from the customer.” “If this, then that.”

Theory, questions, data, action—every learning or improvement or problem-solving cycle uses the same four steps (see figure 1). They use different words for these steps to appeal to different constituencies or egos—PDCA, DMAIC, TQDA, PDSA, etc.—but they are all the same.

Improvement works counter-clockwise in these cycles: If you want to improve your actions, you must get better data. If you want better data, you must ask better questions. If you want better questions, you must embrace better theory.

Better theory includes asking for variables data over time.

Figure 2


Data are classified into two types (see figure 2):

Attributes data, which is called discrete data by some

Variables data, which is called continuous data by some

Attributes data are counts of things that have operationally defined attributes. For instance: operationally define what a defect is, and then count the number of defective parts produced today; operationally define what an on-time delivery is, and then count the number of on-time deliveries; do the same for the number forecasts within 10 percent of budget, the number of change orders, or the number milestones missed.

Attributes data are inefficient and contain less information compared to variables data. For example, for a given decision risk, 96 attributes counts are the equivalent of 32 variables measures. Three times more samples are needed for the attributes data to have the same risk as the variables data and even then do not provide the precision needed to quickly take action. If you use attributes data, you do not know enough about the process that you are managing.

Variables data are measures of things that can be considered continuous. For instance: volts, amps, centimeters, milliseconds, Renminbi, degrees Celsius, etc.

Variables data are preferable to attributes data. In virtually all modern equipment, variables data are gathered but attributes data are reported out. (4.945 volts are measured, but only “Pass” or a green light is reported out.) You’ve already spent the money on the equipment, but you don’t take advantage of the information buried inside it.

Figure 3


These data come from an expensive automatic optical inspection (AOI) machine that was positioned after wave solder (see figure 3). As the name implies, this machine optically scans each printed circuit board (PCB) for defects.

Some defects/errors are:

  • No solder (error type 20)
  • Missing component (error type 2)
  • Solder splash (error type 12)
  • Bridging (error type 21)
  • Wrong component (error type 7)


They only reported the number of each defect in a spreadsheet table. You can see that error type 12 (solder splash) occurred on 488 PCBs. Since the only reason to collect data is to take action, what should they do about the solder splash?

Attributes data are inefficient in that they give little of the information that was originally collected by the expensive AOI. Engineers must now spend time on failure analysis to answer the where, why, how, how much, and when questions.

Attributes data is minimally actionable.

Figure 4


In figure 4, I am using a scatterplot for X and Y values for the location of each operationally defined defect. As in all of these examples, I assume that you know how to use the tool. I am only showing perhaps a unique use of it.

For the AOI to identify possible defects, it “looks” at specific X and Y coordinates and compares what is sees with what the engineering drawing says should be and not be there. So the where question is answered with no additional expenditure of time or money. I could have plotted the time the defects occurred because that is collected as well.

As you can see, 200 PCBs had error type 2 (missing component). This error was confined to one spot on the board. This is very useful information. We know that the missing component was localized to one spot on the board. We can immediately ask questions concerning auto insertion problems, bad leads, supplier issues, the pattern over time (increasing), whether the errors were consecutive (no), and so forth.

Plot the variables data and you more quickly improve performance.

Figure 5


Incoming inspection made 960 measurements of support heights on 120 printer platens coming from one supplier using two-cavity injection molding machines. This column chart in figure 5 shows that there were zero failures and 960 passes. The platens were put into production where they caused rework in the line.

Attributes data give a false sense of security. Just because there are zero defects, doesn’t mean that your operating costs are minimized.

Figure 6a

Figure 6b

This box and whiskers plot in figure 6b shows the variables data for the eight bottom row supports of the printer platen shown in figure 6a. There is a target height of 4.22 mm with an upper spec of 4.32 mm and a lower spec of 4.12 mm.

This plot also shows the differences in the cavities. It shows that the further right we go in the die, the more variation there is in the row support heights. A good process engineer should adjust the process parameters accordingly to have minimum variation on the target height of 4.22 mm.

Plot the variables data and more quickly improve performance.

Figure 7


This matrix plot in figure 7 shows the cavity differences in the top, middle, and bottom rows on the far left supports. If the two cavities were the same, the red and black cavity data points would overlap. A good process engineer should be able to look at these patterns and make the appropriate corrections to minimize the differences and reduce assembly and reliability costs.

Plot the variables data and more quickly improve performance.

Figure 8


This LED manufacturer tested and binned product based on customer performance specifications. The final test/bin equipment measured the forward voltage and light intensity, and put the LEDs in the appropriate bin for disposition. A spreadsheet report summarized the number that failed and the number that passed (see figure 8).

How do you improve? Only with a lot of extra failure analysis, time, and money.

Figure 9


In figure 9, Vf1-A is forward voltage applied to the LED; Iv-A is the intensity of the light. You can see in the scatterplot of figure 9 that the bin designated with the blue cross is out of spec. The other colors and shapes are bins that have different markets and selling prices. An engineer should be able to adjust the cloud to optimize selling price.

Plot the variables data and more quickly improve performance.

Figure 10


In figure 10 we have fewer defects, which results in congratulations all around. But something is happening to create a separate cloud of performance. Was it different dice, a double peak based on design, phosphor mixing, what? Whatever the case, these variables data shown in this scatterplot shed more light on what should be done to improve the process. In the case of binning, some bins have a higher selling price than others. Is the manufacturer managing the process to get the highest price for its products?

Figure 11


Cash conversion cycle (CCC) days are equal to days inventory plus days sales (receivables) outstanding minus days payables outstanding. It is typically reported as an average or in “aged” bins as shown in figure 11. The sooner you can get your cash from your customers before you have to pay your suppliers, the better for your business cash flow.

No matter how low the CCC days are, you should be trying to make it lower. A negative CCC indicates that you get your money from your customers before you have to pay it to everyone else. This is Dell’s business model and many internet-based businesses try to emulate it.

The problem with using averages or attribute-based bins is that they don’t give you enough information to quickly focus on what to do to make them lower.

Figure 12

Figure 12 is a short-term look at the volatility of the same CCC data. This gives management an efficient visual display of patterns by business unit and patterns over time. Two business units have consistently negative CCCs. Get your money from the customer before you have to pay it out to the suppliers.

Figure 13


The data in figure 13 are from a business unit that reported CCC days every month in the management review meetings. They only showed the past month’s performance, not the four-years worth of monthly data presented in this spreadsheet.

Figure 14


Figure 14 shows the same data plotted over time with an individual and moving range control chart. We can quickly see that CCC days were creeping upward for about 24 months before putting into action a plan to reduce CCC. Half of the information in any set of data is in the time series: patterns over time that are increasing, decreasing, jumping, and so forth.

Plot the variables data over time and more quickly improve performance.

Figure 15


Many companies consider forecasts that are within plus or minus 10 percent to be okay, such as the one shown in figure 15. If you are within 10 percent, you don’t owe any explanation. If you are outside the 10 percent, you must explain why you didn’t make the forecast. This is attributes thinking applied to management. The forecast is good, or it is bad.

It is very difficult to take action using attributes data.

Figure 16


This company was predictably 21-percent below its “profit before fixed expenses” forecast, as seen in figure 16. But the variation in the forecasting system was too large to be acceptable.

Figure 17


A customer requires capability data for new sample approval for critical characteristics. If the Cpk was below 1.33, the sample failed. The reported summary in figure 17 indicates that only one critical characteristic passed the threshold.

The question now is what do we have to do to increase the Cpk of the failing characteristics?

Figure 18


In figure 18, I have plotted Cp vs. Cpk to graphically make visible the performance of many measures so that management can take action. If Cp and Cpk are both above 1.33, management knows that the process is minimally capable. If they are below 1.33, incoming inspection should reject the shipment.

Figure 19


Statistician Donald Wheeler’s “effective cost of production” plotted Cp vs. Cpk. I have plotted the Cp and Cpk from the previous chart on a contour plot generated from Table 1 of Wheeler’s The Process Evaluation Handbook (SPC Press, 2000). This should really drive home the need for management to take action. They are spending more than twice what they should be to produce these parts.

Figure 20


Automated shop floor control systems collect lots of data. They are all the rage in China. But if you don’t use the data, even to understand what is really being collected, you have spent a lot of money with little or even negative benefit. In a one-month period, about 100,000 parts had a defect code 23 (see figure 20). You would like to prevent this defect from being made. What is the reason for the defect? Well, 96,496 had the reason “Not Applicable.” How do you take action on this? You wasted your money and cycle time.

Figure 21


If you see a pattern like this, you must improve your measurement process so that any stage that inspects the same characteristic gets the same answer. "No trouble found" (NTF) or "no fault found" (NFF) are immediate signs of bad management.

Remember: Plot the variables data and more quickly improve performance.


About The Author



William W. Scherkenbach's executive career in operations and quality management spans world-class enterprises, multinational high-technology organizations, and entrepreneurial ventures, with consistent success in improving quality while reducing cost. Dell, Lexmark, Tokai Rika, Samsung, Liteon Technologies, Ford, General Motors, U.S. departments of defense and energy are a few of the organizations that have benefited from his technical and transformational leadership. He has extensive European and Pacific Rim supply chain experience.
Deming Medal,
Engineering Society Gold Medal,
ASA Fellow,
ASQ Board of Directors (past),
American Insurance Congress Board of Directors (past).


LED Binning

Regarding LED binning, I have learned the following. Correct me if I'm wrong.
(1) The LED binning is based on the spec of various applications from different customers. With this nature, every bin is possible to be sold if some customer demands. On the other hand, every bin is also possible to be stocked if it meets no spec of any customer. And the price of different bins may not purely be based on the Cpk, or characteristics, of the LED. For example, for white LED, the people in western countries might prefer "coldish" white, while the people in eastern countries might prefer "warmish" white. Therefore, different customers may have different spec for even the same application. For an OEM manufacturer, the customer's requirements and orders are changing rapidly. The bin once "sellable" in plan might become "unsellable" the other day.
(2) The "two groups" phenomenon in fact inherits the characteristics from the incoming material (dice). Unless the LED manufacturers can resolve the supplier quality issue, they just suffer this problem. That said, the characteristics of the dice determines the characteristics mentioned here of the LED.

Excellent Article!

It's great to see Bill Scherkenbach writing again! As always, great insight. In a manufacturing environment, many of the things we can track are continuous; wherever possible it will always be better to track the actual measurements.