What method can be used to better remove defects during the production process?

Many people on this side of the ocean have greeted the hullabaloo over Japan’s miraculous achievements in manufacturing over the last decade with a yawn. “So what else is new?” they say. “Deming said this years ago.” W. Edwards Deming is the godfather, if not the real father, of plant productivity. He was mainly responsible for instituting in plants the statistical control procedures that the Japanese have so adroitly adopted. This article describes the Deming approach to productivity and quality: because only management has the authority to change the production system to eliminate them, product defects are a managerial responsibility.

To change the system, management first needs to distinguish abnormal from normal variation. It also needs to specify operationally what the system is supposed to produce. With these controls in hand, the organization can predict performance, costs, and quality levels, and managers can communicate effectively with customers and people on the shop floor. And this is most important, for when management sees the system and not the workers as the cause of problems, many of the morale-sapping results of poor decisions, targets no one believes in, and motivating slogans that implicitly blame workers disappear.

John Henry, president of Global Manufacturing Company, leaned back in his chair, sighed, and stared at the ceiling. On the desk in front of him was a report from two statisticians on the productivity and quality problems at Global’s Nightingale factory.

Henry and his vice presidents had known that things were bad. Customers were complaining, prices were too high, accounts receivable were escalating, repair calls were increasing, costs were up, worker morale was down, and the union was threatening to strike because of management’s incessant demands for better productivity. Also, most of the machines were not up to the job. But they hadn’t bargained, Henry thought ruefully, on what the statisticians would find. He picked up the report, sighed another time, and looked at it again.

“Your factory at Nightingale,” the report said, “is running along day after day sending out items, 15% of which (on the average) have one or more major defects… This proportion of major defects in your product may well explain some of your problems with sales and profits. The amount of rework your operators have to do along the production line is also stifling your profits.

“Your problems start this way. An operator on the line turns out an item. She looks it over. If she finds a major defect she may rework it herself because she knows that otherwise it may come back to her later to fix. But, she thinks, the inspector down the line might not spot that defect. If she does, she may rework it or send it back to the operator. But even if the inspector sees it, the supervisor may intercept the item on its way back to the operator and send it on through production to avoid getting caught short at a later stage down the line.

“From the operator’s point of view, why not take a chance with both minor and major defects? Send them down the line; chances are they won’t come back. From the inspector’s point of view, the supervisor may intervene, so fixing defects can be a waste of time. From the supervisor’s point of view, she can risk the defect. She can’t lose, and she might gain if she keeps her production record high.

“In other words, Mr. Henry, your operator’s job is to produce defects. She gets paid for them. This is the system, and the operator is not responsible for it. Management is.”

The point behind the statisticians’ memo to poor John Henry is that defects are not free: somebody makes defects and gets paid for making them. If a substantial proportion of the work force corrects defects, then the company is paying to correct defects as well as to make them. If the Nightingale factory is producing 15% defective products, then 15% of the total cost is spent making bad units. Obviously, low quality means high cost.

All the problems of Henry and his vice presidents stemmed from the mismanagement of quality. In other words, and this could be the statisticians’ second point, management achieves a high-quality product by improving the process. If managers can improve the manufacturing process, they can transfer resources from the production of defectives to the manufacture of additional good product.

Suppose management at Nightingale is able to improve the process by making some changes at no additional cost so that only 9% of the output is defective. What has management accomplished?

1. Productivity has risen. The factory now produces 6% more units at the same cost. (If the factory reworks defectives, then operators can use the time they would have taken to rework the 6% defectives to make more good product. This creates an additional—free—increase in productivity.)

2. Aggregate quality has improved. Now only 9% of the output is defective instead of 15%.

3. Capacity has increased. The factory produces 6% more good units with the same system—labor, machines, materials, and so forth.

4. The cost per unit is lower. The factory manufactures more units at the same cost.

5. The price can be cut.

One can see that process control (i.e., the proper management of quality) can alleviate John Henry’s problems. With improved quality, customers will stop complaining and returns will drop, sales people will be able to compete effectively due to a higher quality product and a lower price, service and repair calls will decline, accounts receivable will go down (because satisfied customers are likely to pay their bills), costs will decrease, productivity will go up, the union will stop threatening to strike, and management will have capital to maintain equipment properly.

Improving the process is the key to increasing productivity and quality and to reducing unit costs. Managers can achieve these goals through understanding the sources of variation in a process and using the appropriate operational definitions.

The Sources of Variation

Let’s look at a manufacturing process that produces steel rods. Although the average diameter of the rods is 2.00 inches, we can’t expect every rod’s diameter to be exactly that. We would expect some variation depending on how the measurement was rounded off.

Variation in a process is natural. In fact, we should all expect it and not be surprised when it occurs. But processes are subject to two sources of variation: normal and abnormal. Abnormal variation is due to a special or specific cause and may or may not be present in a process. In our example, let’s say that we produce a rod with a 1.96-inch diameter. Is the .04-inch discrepancy an abnormal variation in the process? Or is it a normal variation that we ought to expect? If it is an abnormal variation, we would want to intervene and, say, adjust a machine. If it is not, we shouldn’t intervene. In fact, by adjusting the machine without cause, we’d run the risk of throwing the process out of whack.

Some researchers estimate that abnormal variations cause 15% of the problems in a process, while normal variations cause the remaining 85%.1 Normal variations are common to all elements of a process—a whole group of workers, an entire department, and even a whole company—and create most of the high costs of production and service and low-output problems. Confusion between common and special causes of variation leads to frustration at all levels, more variation, and higher costs. Unable to distinguish between the two sources of variation, management may react by blaming the workers.

A worker is powerless to act on a normal cause of variation. Workers have no authority to sharpen the definitions and tests that determine acceptable quality. They cannot do much about machines or test equipment that is out of order. They can report such events, but management must do the follow-up and make the necessary changes. Workers cannot change the specifications and policy for procuring incoming materials either, and they are not responsible for the product’s design. These are all part of the system, and only managers can change the system.

It is hard to overestimate how high morale would go in most factories if management held workers responsible only for what they could control and not for the handicaps of the system.

What Kind of Variation Is It?

Because workers can’t be responsible for the system, managers need to be able to distinguish between abnormal and normal variation so they’ll know when and how to change the process. The only safe way to differentiate the two sources of variation in a process is through statistical signals that control charts generate.

Control charts

A system control chart has a center line that represents the process average, and two control limits, upper and lower. Suppose you want to examine the keypunching operation in a data processing department. First, according to statistical theory, you determine a sample size, let’s say 200 cards per day.2 Then you take random samples of 200 cards from each day’s output and inspect them for errors. Exhibit I shows how to construct a control chart for a keypunch operation.

Exhibit I Formulation of Control Chart for Keypunch Operation Note: Both of the points (day 8 and day 22) that lie above the UCL send a statistical signal to management to search for possible sources of abnormal variation on day 8 and day 22.

Exhibit Ia shows the percentage of keypunch cards that are defective. Exhibit Ib is a plot of “percentage defective” (column 4 in Ia) against “day” (column 1 in Ia). Exhibit Ic shows the computations you’ll need to construct the center line (in this example, the average percentage defective for the process) and the upper and lower control limits.

You construct the control chart (Exhibit Id) by connecting the points plotted in Ib and drawing the center line and upper and lower control limits across the points. Finally, you analyze the control chart. If a sample value falls within the upper and lower control limits, and if a trend or some other systematic pattern is absent, the variation is probably normal. If, however, a sample value falls outside the control limits, the variation is probably abnormal.

The chart shown in Exhibit I is just one of many kinds of control charts, each of which has a special purpose. (You can find examples of other charts in the sources listed at the end of the article.)

If the Variation Is Abnormal

By comparing Ib and Id the reader will see how difficult it is to differentiate between the two causes of variation with the naked eye. Exhibit Ib does not allow managers to distinguish between the two sources of variation, while Exhibit Id clearly shows that on days 8 and 22 something abnormal happened, not attributable to the system, to cause defective cards to be keypunched.

When a manager determines that the cause of the variation is abnormal, she should search for and eliminate the causes that are attributable to a specific worker or group of workers, a machine, a new batch of raw materials, and so on. Once management eliminates all assignable causes of variation it is left with a stable process that is in statistical control.

Let’s reexamine the keypunching operation shown in Exhibit I in more detail. Look at the control chart for the percentage of cards with errors (Id).

It is customary to base the control limits on a multiple of the standard error. Usually this multiple is 3 and the limits are called 3-sigma limits. This means that there are approximately 3 chances in 1,000 that the location of a point outside the limits is due to the natural random variation of the system. If we look at the charts in Exhibits I and II we can see that two points are outside the upper control limit, indicating that the process is not in statistical control.

Exhibit II Control Charts for Keypunchers

What should management’s next step be? To bring the process under control, management should investigate the points that were out of control to remove assignable causes of variation from the process. Let’s say that management found that on day 8, a new keypunch operator had been added to the work force, and that the one day it took the worker to acclimate to the new environment probably caused the unusually high number of keypunch errors. To ensure that this assignable cause would not be repeated, the company instituted a one-day training program.

Investigation of day 22 showed that the night before, the department had run out of cards from the regular vendor and did not expect a new shipment until the morning of day 23. Consequently, the department purchased one day’s supply of cards from a new vendor. Management found that these cards were of inferior quality, which caused the large number of keypunch errors. To correct this assignable variation, management instituted a revised inventory policy and operationally defined acceptable quality for keypunch cards.

After eliminating the days for which assignable causes of variation were found, managers recomputed the control chart statistics:

Exhibit IIb shows the revised control chart (IIa shows the original chart). The process is now stable, in statistical control.

A stable process that exhibits only variation due to inherent system limitations allows a manager to determine its capability, that is, what is normal. Here are some of the advantages of achieving a stable process:

1. Management knows the process’s capability and can predict its performance, costs, and quality levels.

2. Under the present system, productivity is at a maximum and costs are at a minimum.

3. Management can measure the effects of changes in the system with greater speed and reliability.

4. If management wants to alter specification limits it has data to back up its argument.

The capability of the process becomes a given. A stable process that produces an unacceptable number of defects will continue to do so as long as the system, as currently defined, remains the same. And only management is responsible for changing the system.

Normal variation

Once a process reaches stability, which is not a natural state but an achievement, management is ready to act on the system to improve productivity and quality. Managers can improve the system by:

1. Shifting the process average. For example, management may want to decrease the percentage of defects or increase the average output.

2. Changing the amount of variation. Given the economic demands of the marketplace, management may want to decrease the amount of variation to obtain a more consistently uniform product or increase it to obtain a less uniform product.

Certain inputs and procedures, such as labor, training, supervision, raw materials, machines, and operational definitions, define the system. To improve the system, management must alter these factors. Again we stress that only management has the responsibility and authority to make these changes. Workers on their own cannot affect the system.

How can management set about changing the keypunch process to improve productivity and quality? By instituting training procedures that reduce the average percentage of defective cards and the amount of common variation (resulting in narrower control limits), management can help employees produce more error-free cards consistently.

Exhibit IIc shows the new control chart after management instituted training and procedural changes. The average percentage of keypunch cards with errors has decreased from .017 to .008 and the process variation has decreased as well.

It is important to stress that the concepts we’ve been discussing encompass more than just control charts. Companies may use control charts without any understanding of the approach we’re concerned with, namely, management’s responsibility for improving the system, no habitual dependence on final inspection, elimination of slogans, elimination of arbitrary work standards, and so on.

We have come full circle. We know that improving the process increases productivity and quality. By distinguishing between abnormal and normal variation, and by eliminating the abnormal variation, managers can obtain statistical control. But this on its own isn’t sufficient to upgrade productivity and quality.

If management fully understood sources of variation as well as saw that its responsibility is to improve the process, but did not understand operational definitions, its efforts would still be in vain.

What’s Being Produced?

If management can’t precisely define its products, how can it sell them, describe what it wants to people on the shop floor, or improve the production process? It can’t. Without an operational definition, people can’t do business. Here’s an example of the confusion that the absence of a precise idea of what’s being produced can cause:

“The label on a blanket reads ‘50% wool.’ What does this mean? Half wool, on the average, over this blanket, or half wool over a month’s production? What is half wool? Half by weight? If so, at what humidity? By what method of chemical analysis? How many analyses? Is the bottom half of the blanket wool and the top half something else? Is it 50% wool? Does 50 per cent wool mean that there must be some wool in any random cross-section the size of a half dollar? If so, how many cuts should be tested? How do you select them? What criterion must the average satisfy? And how much variation between cuts is permissible? Obviously, the meaning of 50% wool can only be stated in statistical terms.”3

What is an exact or true definition of a term? For example, what is “exactly round”? No one definition exists that will help us tell if something actually is round. The dictionary is no help either. Webster’s says that a figure is round if it has “every part of the surface or circumference equidistant from the center.” This definition is very useful for formal logic, but if we try to use it to determine if our disk is round, we will have insurmountable difficulty. The dictionary provides a concept, not a definition for use in industry.

How then can we define a term that is understandable at the shop level? Operational definitions are of two types: one for attributes, e.g., success versus failure, and one for variables, e.g., sales volume. An operational definition for an attribute consists of:

1. A criterion to be applied to an object or group.

2. A procedure to select the object under study.

3. An operation, such as measuring or observing the object.

4. A record of the result.

5. A test of the object to decide whether it conforms to the criterion.

6. A yes or no decision about whether the object meets the criterion.

To derive an operational definition for a variable, managers would take the same first four steps they took to derive a definition of an attribute. (Steps 5 and 6 for attributes do not apply to variables.)

Now, the question is, What is the significance of operational definitions to the productivity of a company? We know how important it is that producers and users understand each other. Without operational definitions, a specification is meaningless. Conflict and confusion between companies and between departments in a company arise from managers’ failure to state in advance, in meaningful terms, the specifications for an item or its performance. Think of the productivity and quality problems that can arise when an inspector who is responsible for finding defects is inconsistent over time in her judgments, or when inspectors are inconsistent with each other. The workers don’t know what is acceptable or what is defective. They need an operational definition of a defective product.

Let’s suppose we manufacture round disks. Are the disks round? Why do we care? If a disk is too far from round, it will jam the customer’s machine, cause equipment damage, and cause downtime. If we want to remain in business, we had better care.

Let’s write down an operational definition of round for the disk. Since we are measuring an attribute (round versus not round), we will work on the first type of operational definition.

Step 1: First we want to derive a criterion for the object.

a. “Use calipers that are in reasonably good order.” (You perceive at once the need to question every word.)

“What is ‘reasonably good order’?” (We settle the question by letting you use your calipers.)

“But how should I use them?”

“We’ll be satisfied if you just use them in the regular way.”

“At what temperature?”

“The temperature of this room.”

b. “Take 6 measures of the diameter about 30 degrees apart. Record the results.”

“But what is ‘about 30 degrees apart’? Don’t you mean exactly 30 degrees?”

“No, there is no such thing as exactly 30 degrees in the physical world. So try for 30 degrees; we’ll be satisfied.”

c. If the range between the 6 diameters does not exceed .007 centimeters, we will declare the disk to be round. We have determined the criterion.

Step 2: Let’s select a particular disk. (We could at this point specify some sampling scheme.)

Steps 3 and 4: Take the measurements and record the results in centimeters—3.365, 3.363, 3.368, 3.366, 3.366, and 3.369.

Step 5: The range is 3.369 to 3.363, or a 0.006 difference. We test for conformance by comparing the range of 0.006 with the criterion range of less than or equal to 0.007 (from Step 1).

Step 6: Because the disk passed the prescribed test for roundness, we declare it to be round.

If a company has workers who understand what round means, and a customer who agrees, the problems the company may have had satisfying the customer will disappear.

Let’s look at another example where operational definitions improve understanding within the company. In this example we measure a variable (sales), so we use the second type of operational definition.

A salesperson is told that her performance will be judged in respect to the percentage of change in this year’s sales over last year’s sales. What does this mean? Average percentage change each month? Each week? Each day? For each product? Percentage change between December 31, 1980 and December 31, 1981 sales?

How are we measuring sales: gross, net, gross profit, net profit, and so forth? Is the percentage change in constant or inflated dollars? If it’s in constant dollars, what is the base year? If it’s in inflated dollars, is it at last year’s prices or this year’s prices? Under what economic conditions?

A loose definition of percentage change can only lead to confusion, frustration, and ill will between management and the sales force—hardly the way to improve productivity. How should management operationally define a percentage change in sales?

Step 1: A percentage change in sales is the difference between 1981 (January 1, 1981 to December 31, 1981) sales and 1980 (January 1, 1980 to December 31, 1980) sales divided by 1980 sales:

S80 is measured in constant dollars, with 1979 as the base year, using June 15, 1979 and June 15, 1980 prices to derive the constant dollar prices, and total unit sales less returns (due to any cause) as of December 31, 1980 for each product.

S81 is measured in constant dollars, with 1979 as the base year, using June 15, 1979 and June 15, 1981 prices to derive the constant dollar prices, and total unit sales less returns (for any reason) as of December 31, 1981 for each product. (Pi79 remains the same for all products.)

This procedure for computing the percentage change in sales between 1980 and 1981 will be in effect regardless of the economic conditions. Further, management may revise the definition of a percentage change in sales after the 1985 sales evaluation, but not before unless the sales force and sales management agree.

Step 2: The salesperson and her sales records are the object under study.

Steps 3 and 4: The sales manager will use all 1980 and 1981 invoices and sales return slips to compute the net number of units sold for each product in 1980 and 1981. The sales manager will record the computations and results.

The prior definition of sales might not suit another manager and sales force; however, if the sales manager adopts it and the sales force understands it, it is an operational definition.

Operational definitions are not trivial. If management doesn’t operationally define many critical variables and attributes so that workers as well as customers agree, serious problems will follow. The control chart becomes a useless managerial tool due to an entirely new source of variation: measurement variation. It is management’s responsibility to operationally define the characteristics being charted. If inspectors don’t agree with each other, or with themselves from day to day, chaos will develop. Workers do not know what is expected of them. Their output is OK for Inspector 1 and not for Inspector 2; an employee’s work may have been passed by Inspector 1 yesterday but may not be today.

Top Management’s Job

Numbers of people have recently written guidelines that tell management what it should do to improve productivity:

Create an institution that has a constant purpose and long-term, top management commitment.

Break down barriers between departments.

Create an environment in which people are not afraid to report problems.

Defuse built-in levels of defects, mistakes, poor materials, and so forth.

Do not blame productivity and quality problems on the workers.

The reader no doubt is familiar with these. Additional managerial guidelines that may not be so obvious follow, however, from the approach we’ve outlined here.

1. Don’t expect inspection to solve the quality problem. By the time the inspection is made, the product is already acceptable or defective. You cannot inspect quality into a product.

Mass inspection does not cleanly separate good items from bad. A better way is to monitor small samples of product for control charts to achieve or maintain statistical control. In this way managers might eliminate the need for inspection and put inspectors’ talents to other uses. Sellers and customers could also compare their instruments and tests; sellers and customers could begin to speak the same language. Inspection under pressure is often a farce: whether it is coming in or going out, anything passes. And because divided responsibility means that nobody is responsible, 200% inspection is less reliable than 100% inspection.

2. As a matter of policy, stop awarding business to the lowest bidder. Without a measure of the quality being purchased, price has no meaning.

To judge quality, purchasing managers require education and experience in evaluating statistical evidence of quality. If purchasers become experts in assessing quality, most of them will drastically reduce the number of vendors they deal with. A vendor that does not know its costs, nor whether it can repeat today’s distribution of quality tomorrow, is not a good business partner.

3. Eliminate targets, numerical goals, slogans (“zero defects”), pictures, and posters that supervisors so often plaster in plants urging people to increase productivity. Unfortunately, such “productivity improvement” programs leave the defects right where they are. They do not uncover or correct faults of the system, nor do they provide the statistical signal managers need to take corrective action. They do not answer the critical question, “How can we improve productivity?”

4. Eliminate work quotas. Work quotas do not take into account normal variations in the system. They do not include a way to detect the need for corrective action or a way to assign responsibility to management or to management’s delegate on the line. For example, a bank manager may determine the number of customers he thinks a teller ought to handle in an hour, the number of computations of interest and penalty someone ought to compute in an hour, and a similar figure for every other activity. However, the standards don’t say anything about the quality of work or give the manager any way to understand the variation in the process. Standards do not indicate what action managers should take or how to improve the process.

5. Institute training programs in statistics so that managers and supervisors can understand how to manage quality. Supervision is part of the system and is, of course, the responsibility of management. Statistical methods are vital aids to foremen and production managers to indicate causes of waste, low productivity, and poor quality. Managers can also use them to determine when employees are fully trained and when further training would help.

These guidelines indicate what top management must do to improve productivity and quality. Though following each of the guidelines will not produce tangible results, at the same time, a company that starts today fully committed will shortly realize impressive gains.

A close second for quick results would be to drive out fear, to help people feel secure, and to help people get over the fear of reporting trouble with equipment or with incoming materials. Managers can achieve this goal within two or three years and reap powerful economic results.

1. See, for example, Joseph M. Juran, Quality Control Handbook, 3d ed. (New York: McGraw-Hill, 1974).

2. A discussion of how to compute a sample size can be found in numerous texts, some of which we’ve listed at the end of the article.

3. W. Edwards Deming, Quality, Productivity, and Economic Position (Cambridge: M.I.T. Center for Advanced Engineering, 1982).

How can we reduce defects in a process?

How can we prevent defective products in production?

What are the three ways we deal with defects?

Which practices help for eliminating defects in the process and always deliver products and services that meets customer specifications?