ML20134A403
Text
'
O e'
[f /b w m.d'
~
/
/
(
>l -
g<j,y ON SOME STATISTICAL AIDS TOWARD ECONOMIC PRODUCTION BY W. EDWARDS Dd.'.tlNG I
Reprinted from INTERFACES Vol. 5.No.4. August.1975 e
pwA-tw-2Mt
. glz4 0508150297 850703 FOIA PDR PDR LEIGHTDe4-293 L
.I O,,
o e
IN TElt! 4 CES Vel, s.No 4 August 197s e
ON SOME STATISTICAL AIDS TOWARD ECONOMIC PRODUCTION' W. EDWARDS DEMING Consultant in Statistical Studies Washington DR. W. EDWARDS DDilNG is a consultant in statistaal studies, with a wide practice. He is known for his work in Japan. which creatM a revolution m quality and in methods of adminntration: Japanese manu-farturers created in his honour the annual Deming Pr re. Among other activities, he teaches statistical methods at New York
(,' n n e rs i t v. He is author of several bocks on statitucal methods, and 150 papers.
Anstu ct. This pape-covers mananment's responsibility for (1) design specification oc seruce ol!crM; th measurement bv simpic statis.
of product (2) f the amount of troubie with prr4uc. or with seruce that can be tical metho.3 o ascribed to tauses that only management en act on: W action on the causes so indicatert, tt shows bv principle and by esample how management may obserse week by week the effects of guided effort toward reducuen nf trcuble.
The paper upsets a number of commoniv accepiec prinop:es of actnmuttation.
For example, a job description, for best economy. snound require the prouucucn.
worker to achiese statistical control of his wora; to r:cet specifiestions without paying the high cost of inspection, rework, and replacement. Statisucal evidence of performance replaces opmion of foreman and superusor.
As a second pnnciple, it is decoratinne and costly to call the siuntion of a production. worker to a defective item when he is in a state of staustscal centrol.
t he fault for the defectne item is not cha*:: tab;e to the werker. but to the
.vstem. Fewer defectises can come only trec a change in the system. not frot:s efforu of the production. worker.
Third, it is better to shift to a totally dif ferent job a worker that has devel.
oped statistical control of bad habits in his present Jcb.
All varianon in quality characteristics f cimermon hardneu, color) causes loss, whether the sariation results in dcfecuse product or not. Economies in manu.
natural consequence of reduction m the sarianon of a cuautv-facture are a charactenstic. The author divides causes of sariat.on into two sources; ill the the responsibihtv of mana::ement; (2) special causes, system (common causes),
which are under the governance of the indindual e:noiovee. In the author s et.
perience, losses freus the suttm overshadow losses from special causes. The same pnnciples apply to sales and to service.
Purpose and Scope of this Faper One purpose of this paper is to pret.cnt a number of new. principles of training and administration that upset generally accepted conventions.
The new principles had their origin in the author's work in Japan, which 1950 (1), [2].
conimenced in Another purpose is to point out to management that most of the trouble with faulty produtt, recalls, high cost of production and service, is charge.
able to the system and hence to management. Effort to improve the per.
formance of workers will be a disappointment until the handicap of the system is reduced.
i The principles explained here will apply to any company, large or small, whether engaged in production of manufactured items or in service
'This part is based on principles taught in Japan since 1950. I am indebted to the editoe and to referees, and to students at New York University, for many helpful suggesuons in presentation.
Ccpynens 3 lsts. The lasuiute el biansteinent sesenas I
i i
7 i
c.
.1 j
I i
b
. (hotel, hospital, restaurant, retail store, wholesale, railway, motor freight, delivery. service, communication. including the postal service), at;ricultural or industrial, whether owned by private investment or by the government and in any country, whether it be developed, underdescloped or overdeveloped.
Many causes have contributed to devaluation of the dollar and to our precarious balance of payments, but one contributor, steadfastly avoided by the quality of many American products is no longer economists, is that competitive, here or abroad. Statisticians have failed in America to explain to people in management the impact that statistical methods could make on quality, prcduction, marketing, labor. relations, and competitive position.
Schools of buuriess teach words and goals, but not metheds.
The reader will note, I hope. that I write as a statistician, working with management on problems in industry and in researen in many disciplines.
1 am not a consultant in management. I am not an economist.
Road Blocks to Quality in Ameri:2 An obstacle that ensures dinppointment is the supposition all too preva-lent that quad:y control is som: thing that you install, like a new Dean, or a new carpet, or, sew furniture. Install it and iou hase it. This suppositien is unfortunately force. fed by the common language of cuality control engineers, some of whom offer to install a quality control syuem. Actually, quality contrcl, to be successful in any company, must be a learning. process. Sear by year, from the top downward and from the, bottom up, with accumulation of knowledge and experience, under competent tutelage.
A noths. - road. block is management's supposition that the production, workers are responsib!c for all trouble: that there would be no problems in production or in service if only the workers would do their Mbs in the way that they were taught. Pleasant dreams. The workers are h:ndicapped by the system.
In my experience, it is something new and incomprehensible to a man in an executive position that management could be at !ault in the production.
end. Production and quality, in the view of mans.;ement, are the responsibili, ties of the production. worker. Research into faults of the system, to be the corrected by management, is not what a manager is trained for. Result:
faults of the system stay put, along with rejections and high costs of pro.
duction.
Management usually discharges its responsibilities (sweeps them under the rug) by turning the job over to a department of quality control. This would be a happy solution and good administration if it solved anything, but it seldom does:.the job lands on people that try hard but have not the neerssary competence, and the management never knows the difference.
As a result, one finds in most companies not quality control, but guerrilla sniping-no provision nor appreciation for the statistical control of quality in the broad sense of this paper.
People in management need to know encueh about quality control to be able to judge whether their quality control departments are doing the job.
Statements by management of aims desired in quality and production are not quality control, nor are they action on improvement of the system.
Neither are periodic reviews and evaluations of quality and production. They are necessary but not sufficient.
Exhortations, pleas. and platitudes' addressed to the rank and file in an organitation are not verv effective instruments for the improvement of 2
INTEM. ICES August 1975
r
(.
l
[
quality. Something more is required.
1 should mention here also the costly fallacy held by many people in l
management that.i technical man (a statistician, for example) must know '
~
all about a process and all about the business in order to work in the company.
l l
All evidence points to the contrary. Competent men in every position, from top management to the humblest worker. if they are doir'g their best, know all there is to know about their work except how to improtic it. Help toward im.
provement can come only from some other kind of knowledge. Help may come from outside the company. or from better use of knowledge and skills already j
L within the company, or both.
l Loss from Variation. Two Sources of Variation It is good management to reduce the variation of anv quality. characteristic l
(thic -
or measure of performance), whether this characteristic be in a state of control or t, and even when no or few defectives are produced. Reduction in variation me s greater uniformity and dependability of product, greater output per hour, greater output per unit of raw material, and better competi.
tive position (5). [7].
Causes of variation and of high cost. with loss of competitive position, may be usefully subsumed under two categories:
Faults of the syster,
$pecial cataes
-(commors ce environmental casats) 25 %
15 ",
These faults stav in the system until re.
These causes are specific to a certain duced by manactment. Their combined ef.
worker or to a roachine. A statitoral sgnal fett is usually cast to measure. Sorne indivi.
detects the esistence of a special ca use.
dual causes must he isolated by judgment. which the worker can usuallv identify and Others may be identified by expettract't:
Cotte(f.
some b" records on operations and materials suspect d of being offenden (see reference to juran).
Both types of cause require attention of mana;ement. Common causes get their name from the fact that they are common to a whole group of workers: they belong to the system (2].
No improvement of the system, nor any reduction of special causes of variation and trouble, will take place unless management attacks common causes with as much science and vigor as the pecduction. workers and engineers attack special causes [3].
The percentages shown, are intended only to indicate that, in my exper-ience. problems of the system overshadow special causes. The percentages will fluctuate as special causes are eliminated one by one, and as faulu of the system are reduced or eliminated.
Confusion between the two types of cause leads to frustration at alllevels, and leads to greater variability and to higher costs--exactly contrary to what is needed.
Fortunately, this confusion can be eliminated with almost unerring accu.
l j.
racy. Simple statistical techniques, distributions, run. charts. Shewhart control.
charts, all explained in many books, provide signals that tell the operator l
when to take action to improve the uniformity of his work. They also tell j
him when to leave it alone. Results of inspection, without signals lead to l
frustration and dissatisf action of any conscientious worker.
What is not in the books, nor known generally amongst quality. control engineers. h that the same charts that send statistical signals to the production.
worker also Indicate the totality of f.iutt that belon;;s to the system itself.
INTEMACES August $975 l- - -
I l
i correction of which is management's responsibilitv [2} The production worker can obserse from his charts whether attempts by mana.:er::ent to improve the system hr.ve had an effect.. f anagement can give themselves the same test.
i hamples appear further on.
Removal of a special cause of variation, important though it be. is not it merely reduces the variation to a level that improsement of the system:
identifies the system. but leaves it unimproved.
Mechanical feed. backs enat hohl dimensions and other quality. character.
istics within bounds are sometimes helpful but may be wasteful of material and of machitic. time. They do not improse the setem. Better understanding of the func: ion of feed.back systems. so as to use them effectively, and to supplement them, will be an important step for manacement.
"We rely on our experience." is the answer that came from the manager of quality in a large company recently, when I inquired how he distinguishes This answer is between the two kinds of trouble, and on what principle 3.
. 3 centinue to pile up self incriminating: it is a guarantee that this company S
about the same amount of trouble as in the past. Whv should i, change.
" Bill." I asked of the manager cf a large compar.y eneaged in motor.
freight. "how much of this trouble (shortage and dam *.ge. 7911 examp ts in one terminal alone in 1974) is the fault of the drisers? ' His repiv. "All of it.'
is again a guarantee that this level of loss will cominue un:il statistical methods detect some of the sources of trouble with the system for Bill to wotk on.
The QC. Circle movement in Japan (3 million members: 4 to 8 workers to study and resiw the to a circle) gises to production wori,ers the chanc, system of p. eduction at the local level, for greater cutput and better quality.
Japanese workers are not handicapped by the rigidits ci the Amer:can produc.
tion.line. The QC Circles represent partial decentrr.hcatica of manage:nent's responsibility to find local faults in the system, and to take action on them.
The QC. Circles in Japan bear no relationship to sug;estion bous, common The boost in morale of the production. worker, if he were to perceive everywhere.
'a genuine attempt on the part of management to improve the system and to hold the production. worker responsib!c only for what the production. worker is responsible for and can govern, and not for handicaps placed on him by the sutem, would be hard to overestimate. It has nct been tried. I believe, outside Japan.
It is now clear that the term rero defects can cnly be a theatrical catch-word, a nostrum. The management of many concerns have adopted it outright or in equisalent form and have posted it all oser the plant for everyone to see, especially visitors, expecting magic. Empty words they are till the manage-ment acknowledges responsibility toward reduction of common causes. One eliminating 8 inspectors company that I know of reduced their defects bvout of 10. This is a with claims attached.
Thur. Mall Sketch of the State of Stat lstical Control 9ome understandime of the concept of statistical control, invented by Shewhart (6) is necessary as background. A state of statistical control is a state of randomness. Simple tests of randomneu are the Shewhart charis-run charts.
Icharts. R charts. The up am! down inovements on a chart are to be die regarded hv the production.worter unten there is indication of a special cause.
A point that falls outside the controllimits is a statistical signal that indicates INTEMA ff!
Aueust (M$
4 L
the existence of a special cause of variation. This the production. worker can almost alway identify readily and correct.
Control limits are not specification limits. Control limits are set by simple statistical calculations from the output itself. What the control limits do is to send out signals that if heeded will minimize the net loss from the two kinds of mistakes that the production worker can make:
1.
Adjust the machine or his work when it would be better to leave it alone.
2.
Fail to adjust the machine or his work when it needs adjustment.
The only rational and econcmic guide to minimum loss is statistical signals.
The production. worker himself may in most cases plot the s:atistical charts that will tell him whether and wher to take action on his work. He requires only a knowledge of simple arithmetic. But the prcduction worker cannot by himself start his own chart, and still less a mosement for use of charts..\\fana;:ement must start the movement, and stay on the job.
Some processes in nature exhibit statnt' cal control. Radio.actise disinte.
gration is an example. The distribution of t; ne to failure of vacuum tubes and of many other pieces of complex apparatus furnish further examples.
But a state of statistical control is r.ot a natural state for a manufacturing proc.
ess. It is insteld an achievement. arris ed at by elimination one bv one, by deter, mined effort, of special causes of exce:nve sariation.
Figure I shows the results cf inspection from a precess that is not in stastical control. The upper pinel (O r.erm;c of 5 mccesdve items indicates the existence of special causes. There are points below the. lower control limit and too many points on the border cf the upper control lima. The lowcr panel (range (R) shows a downward trer.d wh;ch, although it rnas indeed be a trend toward greater uniformity, indicate;5 ne.erthele:s also the exhtence of one er more (possibly additional) special causes, which again the worker must discover and correct. The charts thus ir.dicate the existence of special causes.
climination or reduction of which is the responsibility of the operator. The reader may turn to Figures 2 and 4 to see a state ut statistical contro!.
33.0 uct.
- 33 C3 p
X 32.S CL e 32.49 32.0 0
LCLe3I96 y
av 31.S 2.0 UCLel96 R
A^ -
e"-
CLe0 92S e
10
9'
N LCL e o 0 *I : 3 4 3 0 7 8 9 :G iS 20 Ficunt 1. A control ch1rt, showing the exister.ce of special causes of vari.
ation. Taken from W. Elsards Dernmg, Elemenray Pnncartes of the Statistical Ontm: of Oulity (Union oi lapnese Scten, tists and Eninneers T:
1950),p.31.
INTEMA CES August 131%
i.
i t
i 6
A process has an identity only if it is in a state of statistical control. The control limits, and the site of the sample, then enable the manufacturer to predict rationally the level and rance of variation of product that will come off the line tomorrow. The same principles and rules are applicable to service-orpnizations. Statistical control thus provides a basis for doing bu:iness in a rational way. The manufacturer knows what quality he can produce, and at what cost.
He will not walk into heavy loss by taking a contract fer uniformity that he cannot meet, or can meet only by inspection and rework. alwass costly and unsatisfactory. He can make no rational prediction about his product and costs when his processes are not in stanstical control.
Results of inspection are too often unreliable-wor:e. are sources of strife
-because of mistrust of the measurements whether made by use of imtruments or by visualinspection. Measurement must be viewed as a process the product of which is figures. A method of measurement cannat be d!anified as a methcd.
unless, with some operators at least, it shows a state of statistical control. A control chart is a powerful scientific tool.
The first step in many plants is to achiese reasonably good statistical control of some of the main operations, including inspection. The next step is for management to work on the common causes of variation and of defective products. Textbooks on quality control (except for Juran (3]) teach only detection of special causes (Shewhart's assignable causes) and acceptance sampling (for statistical disposition of product alreadv on hand). These are importantmeth and remmal of special causes stops short of the main part of the there:
problem, namelv, faults of the syste....The explanation is simple. The usual term the remaining sources of sariation. lumped together,
[6] himself, is thatonce control is established, constitute " chance causes." variation to chance. This is the correct view for the production worker in a state cf controh he should indeed leave the remaining vanation to chance. Likewise for a group of workers, or a line, or a process: ups and downs in a state of control are not a basis for action on the process.
The contribution that I am trying to make here is that management management must not leave the remaining varia.
must take a different view:
tion to chance.
Famittar Conscquences of Faults of the system Recalls of automobiles, electrical apparatus, and of other items, familiar enough to people in America, for possible hazards from failu to carry out adequate tests of components and assemblies oser the ranges of jolt, stress, dust, speeds, voltages, corrosion. Hiely to be met in practice, or failure to heed the results of such tests, is chargeable to the system; hence to mana;;ement. Or, as sometimes happens, mana;ement sometimes goes ahead l
with production, test or no test, to beat a competitor to the market. place. No amount of care and skill on the part of the production. worker can oscrcome a fault in de>ign. Where is the statistician's report on the performance of parts 1
and assemblics that give rise ib trouble >and to recalls?
If one enquires Ahether enore experimentation in adsance would have oserpaid its cost, or whether it is better husiness to tinh into the market place INTEltt.f CES Augun IMS
(,
l
and take a chance. I would offer the following remarks: (a) no dollar.value can be placed on the unhappiness of a customer and the loss of future business oser a detective item of an unsatisfactory sersice calh (b) the costliest experimentation on the performance of a product is the tests the customer caries out for himself. (c) cost. benefit analysis has important uses, but also serious limitations. If the Japanese manufacturers had depended on cost.
benefit analysis in 1949 to decide whether to learn and use the statistical control of qu.ility. they would. I surmise. hase given it a negative vote, or would still be studying the matter.
Partial List of Usual Faults of the System (Common or Environmental Causes of Variation)
The reader may make addiuons and subtractions to suit his own situation.
1.
Hasty design of component parts and assemblies. Inadequate tests of protot> pes. Hastv production.
- 2. Inadequate testing of incemine mat: rials. Specifications that are too stringent, or too loose, or meaningless. Waising specifica-tions.
3.
Failure to know the capabilities of processes that are in a state of statistical control, an.1 to use this information as a basis for contracts, both for quantity and quality.
4.
Failure to provide production wcriers with statistical signals that will tell them how they are doing and when to make some change.
5.
Failure to use these charts as a meuure of the faults of the system, and of the effect of actit o taken by management ta reduce them.
6.
Failure to write job. descriptions that take account of the capability of the process.
7.
Inadequate training of workers with the help of statistical controls.
8.
Settings of machines chronically inaccurate (fault of the crew responsible for settings).
9.
Instruments and tests not reliable. Consequent demoralization and loss from false reports and fahe signals. Loss from needless retesting.
- 10. Smoke, noise, unnecessary dirt. poor light, humidity confusion.
The production. worker is helpless to reduce any of these causes of trouble.
Economic considerations mint of course gosern the decision of management to reduce or eliminate a fault of the s) stem. An easy way out is to say that it would cost too much.
A Word on Due Care Statistical control and its consequences, if explained by statisticians to the legal profession in industry and in government, would clear up many problems about safety and reliability. The most that a manufacturer can put into the uniformity and dependability of a device is (a) to achiese and maintain statistical contrni, at the nght lesel and spread, of the most important quality. characteristic of the main outgoing components and assem-blics, and incoming ingredients, and (b) to be able to demonstrate by adequate statistical records and charts, along with action taken on special causes and on cominon causes that he It.as dune 50.
IN TElif A CES August 1975 7
E.
In spite of scrupulous care and intellicent use of statistical controls. it is inevitable that a defective item will get out now and then. An unfortunate fre'ak of this kin ; cannot be viewed as an r.ct of carcicssness on thh part of
+
the manufacturer. He can do no more than to e.sercise due care.
Somo New Principles in Administration This paper upsets some well. accepted principles of administration, which when examined under the logic of statisucal inference turn out to be bad practice-that is, demoralizing to the rank and file of production workers, and costly. For example, it turns out to be bad practice to draw the attention of a producuan worker to a defect in his work when he is in a state of statistical control. Why? The production worker is helplcss: he cannot do better. It is as if he were drawing blindfohled handfuis from a mixture of black and white beans. The number of black beans in a handful may be 0, or it may be 1. or 2, or more. The laws of chance apply. He cannot siter these laws, once he achieves statistical control. He will on!v make things worse (increase the proportion of black beansi if he tries to adjust his work except on statistical sienal. To draw his attention to an error or to hold him on the job until he corrects is to charge him with a fault of the svstem.
Yet it is common practice in industry. whether it be preduction or sersice, to bring to the attention of a man any output th::t is discesered to be defectise. In an example that I could cite, a production. worker, whether in a state of statistical control or not, reworks on his own time, all the defectives that impection discovered in the product that he turned out durint his shift. This is what some people call quality con:rol. The reason gisen to me upon enquirv is that this procedure is a continual reminder to the production. worker that defectives will not be tolerated: : hat he is respon.
sible for tM work that he turns out. flow can he imprme if he doesn't know about his mistakes)
Like so man obviom solutiom to problems, this one is aho wrong.
The fallacy in this principle is'demomtrated hv dependable day to. day figures on rejections.
A job 4!cscription, for best economy, should require achievement of natistical control of a dimension (4]. Under this requirement, the production.
worker is in chart;e of his own process, and can achieve in his work maximum economic uniformity and output. This is very different from askine a product lon worker to force a dimension of individual pietes to stay within specified limits.
An economic level and spread of the control limits would produce a distribution for indisidual pieces that rarely if ever estends beyond the specification and produces a defective item. It is the responsibility of the foreman or higher supervision to remove obstacles to an economic level and spread. This mistht mean better setting of the machine, or better maintenance, or incoming materials better suited to the rirht spread, b*one of this refine.
ment in job. descriptions will take place without understanding and action on the part of manacement (see Example 1).
A state of statistical control can exist in a climate of mild but uniform carelessness. This degree of carelessness is part of the sistem, the clitnate.
I In my experience, workers seldom know the cost of carelessness nor the cost of a mistake (see Example 2). Only management can teach them.
To call to the attention of a worker to a careless act, in a clirnate of general carelessness, is a waste of tiene and can only generate hard feelings, because the condition of general carelessness belongs to everybody and is the f ault of management, not of any one worker, nor of all workers.
8 INTiltf.1CES Aupu 1975 l
l t
l,:,
l l
l l
A general reminder, pmted in a factory 4o that everyone can see it, to explain to the workers the cou of a defectise item, or the cost of a mistake.
l may be helpful in improving the svstem. Meetines illustrated with moving pictures to show how defects nie made, cau.:ht in the act. are also helpful.
l Continuing education on the job to rehearse principles of the job and the cost of defective work belong in the systern. This is management's job:
workers cannot institute it.
A worker who is in a state of control but whose work is unsatisfactorv presents a difficult problem. It is usually uneconomical to try to retrain him l
on this same job. It is more economical to put him into a new job in which the training may be more expert than it was in his present job.
Figure 2 provides an illustration. An experienced man in golf hoped to improve his score by taking lessons. The chart shows that the lessons i
accomplished nothing. His techniques were engrained: his teacher was unsuccessful in dislodging them and replacin:t them with better ones.
l Curiously, so long as a man has not reached a state of statistical control.
l there is hope. Figure 3 shows asera:;e scores (r) in golf for a beginner. His scores, before the lewons, were obviously not in a state of controh there are points outside the control limits. Then came lessons. His scores thereupon showed a state of statistical control with the desired results. viz.. an aver:ge score considerably below what his asera;;c was before the lessons. Here.
l.
lessons chan;;ed the system.
l BEFORE LESSONS LESSONS AFTE R LESSCNS UCCet Controf life.f' l
? A f\\ w !).- - -... ! a h P
.A y by q 4 6 dy8 y
a:.-
,0 FacurtE 2. Averge scores in golf for an expenenced golfer, before and after lessons. Taken from W. Edwards !>: ming, Elementary Principles of the Statistical QnrroI of Quality (Utiicn of 1apa.
nese Scienti:ts and Eni;in:ers Tokyo,1950), p. 22.
BEFORE LESSONS LES$cNS AFTER LESSCNS fn uct uct j
6 kw LCL b
O L et.
i Ficunc 3. Average weekly scores in golf for a beginner who took lesso'is before he reached a state of stattstical control. Taken from the book cited in Figure 2.
l INTElifACIS AuCunt 1975 9
l L
e Ten production. workers may all be in statistical control as individuals.
all at different levels. Their combined output will also be in control.
Imprct:ement co nes about by studying the individual t.:orkers, transferring to anoth:r jnb seith a fresh start anyone that is out of line on the side of poor performance.
It is my observation that training in industry is deplorable. A new employee simply goes to work. Written instructions for the job, if they exist, are in many cases incomprehensible. What happens is that the new worker gives up on the instructions for fear of being furth:r confused. His co-workers come to the rescue, instructions or no instructions, and in a few days he is running along with the herd. The service industries (restaurants, hotels, laundries, etc.) provide horrible examples. The argument is that instruction and training are too costly, and hat it is all lost if an employee quits the job.
In contrast, a girl that runs a lift in J.tpan, or is conductor on a bus, spends two months in training on how to handle people, this in spite of her genteel background of culture.
Training or the lack of it is part of the system. Training can be improved only by management, certainly not by the workers.
8 f.
Defective
~
tone.ficat on for indiv.dua# ite.as f
.......k w-s~,a e g
2
.s
'v' yV
~
Q
\\
- 10 0
-4...A i
Lower contros 1.mit 0.pnt,t.e cf a j
g C.str.At.cn ef a for,nem.ouw.we 0
10 15 20 2S CRCIN At. NUMBER OF SAVPt.C (in erw of prou.t.on)
FicURE 4. Chart for I for test Cf uniformity of whee's turned out by a production worker.
1 Example 1.
This example illustrates how a small chans;e in the system could virtually eliminate the possibility of defective items. The ordinates in Figure 4 are the meam (M of samples of n = 3 for tests of uniformity of finished wheels.
l The test is the running balance of the wheel. Observations:
l 1,
The production. worker is in a state of control with re:pect to his l
own work (which _is the only work that he is responsible for). No j
point is11s outside the control limits.
2, lie is under the handicap of the system. Ife cannot best the system j
and the capability of his process: he will once in a while produce a defecthe wheel, even though he i,s a good werker and in a state of l
control.
3.
He is meeting the requirements of his job. lie can do no more.
lie has nothing further to offer, 10 INrritr.< cts, August 1975
4.
The main trouble lies in the system. The central line in Figure 4, which fall at about 125 gram <ms. represents the contribution of the system to the total trouble. This handicap is built in. Il the faults e
of the sy.t'.m were reduced to 75"" of thei present !qvel the upper tail of the distiihutton of individual pieces would drop well below the specification limit, and the entire production would be accepted; economies in production would be realized.
The reaction of manacement on the abose paragranhs was the usual one, namely, that they did not hase in mind this kind of quality control when they went iruo it. They were looking for esenthing to clear up, once the production worten put their best effcr:s into the job. Esentually, however, patience paid off.
Charts like Figure 4 are to be seen almost anywhere, but interpretation of them in terms of a quantitatise measure of the faults of the system are rare.
Example 2.
The second example de m with a service industry. motor frei;ht.
terminal for Drivers of trucks pick up sh;pniuts and bririe them into a reload and onward mosement. Other driven deiner. A large wnu,ans in momr freight mav have anyw!'ere from ten to forn terminals in ar nest lar:;c cities. There is a long chain of operations between the request of a shipper to the carrier (usually by telephor.3) to come and pict up a 'stupment, and placement of the shipment on the platform of the carrict, ready for re! cad and line. haul to the terminal that serves tlie destination of the shipment.
Esery operation offen a chance for the dt" o mLe -
Nake. The :ah!c shows ti types of mistakes, plus all ethert * haugh the frer;uenn n! mistakes is low, the total loss is substantial.
In mistake No. I, the driver signs the shipping. order for (e.g.) 10 cartons, but someone che finds, later on in the chain of operations. that there are only 9 cartons: one carton missing. Where k it? There may have been only 9 cartons in the first place: the shipping:ctder was written incorrectly; or, more usual, the driser lef t one carton on the shipper's premises. Let us list some of the sources of loss from mistake No.1:
1.
It costs about $25 to search the platform for the missing carton, or to find the truck (by now out on the read) and to search it.
2.
It costs SIS on the average to s:nd a driver back to the shipper to pick up the missing carton.
3.
It costs $10 to segregate and hold the 9 cartons for the duration of the search.
4.
If the carrier does not find the carton, then the shipper may legitimate-ly put in a claim for it. The i.orier is responsible for the 10th carton.
Its value may be anywhcre from $10 to $1,000, with the pcssibility of an amount esen greater.
It is obvious that Mistake No. I may be costly. Any one of the 7 mistakes will on the average lead to a loo of $50, There were a total of 617 mistakes on the record, and they caused a loss of $31,000 for claims alone. Multiplied by 20, for 20 ternunals, the total low f rom the 7 mistakes was $620.000. This amount is a minimum. h does not inc!ade the expemes of sc..che, nor administration. Moreover, some mistake, are not included in the total of 617, but they nesertheless cause loss.)
INTEMA CES Augtut \\MS Il
. = _
e
.. ~ ~.
-- ~.
p_
e The 7 types of mistake a
Type of Description mistake 1
Short on pick up 2
Over on pick up jl 3
Failure to callin (by
[
telephone) on over, short, arid damaged cartens on dehvery 4
Incomplete bd! ofIsding 5
Improperly marked cartons 6
Incomplete s!;nsture on delivery receipt 7
Other There were 150 drivers that worked all year long. Figure 5 shows the distribution of the 150 drivers by number of mistakes, al! 7 mistakes combined.
We postu' ate the fo!!owing mechani:m which dill distribute errors at randcm to druers. We imagine a huge bowl of bla:k and white beads ther.
oughly mixed. Each driver scoops up a sample of 1.000 cr more (the number of trips that an average driver makes in a year), and returns the beads to ill be the bowl for more mixing. The number of black beads in a scoop w a random variable, following the Poisson distribution. The total number of mistakes in Figure 5 is 617, and there were 150 drivers. An estimate of the mean number of mistakes per driver would be i
~
(1)
Y = 617/150 = 4.1 30 -
- 0 65 si 20 b.n e
$M e
3$
E to-e, r4 ee 3
I S
to IS 20 25 i
e X. INMBER OF 4tSTMS 1
I 4
Ficunt 5. ne distnbution*of drivers by number of restar.es, all 7 errors combined.
INTElf fA CES Aurust 1975 I
l2
- - - ~,. _, ---.
= _ _ - - -.
L 1.
i l
The upper and lower 3-sigma limits for these samples would be easy to calculate by use of the square. root. transformation, by which (d1 + 1.5)a = 12
'[ upper limit]
(2) and (81 - 1.5)s = 0
[ lower limit)
(3)
One may perform the same calculations instantly by use of the hiosteller.
Tukey ocuble square root paper (S].
We interpret the upper limit to mean that a driser that made 12 or l
more mistakes in the year is not part of the system. He contributes more than his share. He is a special cause of loss. I may add here that other statistical models that I have tried lead to about the same conclusions.
Drivers that made 0, I, 2. 3, or 4 mistakes are far more numerous than the Poisson distribution would allow. I accordingly consolidate the drivers l
I that made 0,1,2. 3, or 4 mistakes, and postulate that they too form a sep-arate group. There are then three ;;roups of drivers:
A.
Drivers that made 12 or more mistakes.
B.
Drivers that made between 5 and 11 mistakes.
C. The extra careful group, drivers that made 0,1, 2, 3. or 4 mistakes.
What have we learned from this simple statistical model?
1.
The 7 drivers with 12 or more mistakes accounted for 112 617 or 132. of the mistakes. They could reduce ' heir rates of mistakes to avua:;e if l
they knew that they were outliers.
2.
Drivers that made 5 to 11 mistakes measure the losses that arise from the system itself. They make the system what it is. They account for (425 112)/617 or about 51% of the mistakes. Clearly, about half the losses from the 7 types of mistake arise from the system as it is.
5 The 102 drivers of Group C accounted for only 192/617 :: 31% of the mistakes. This Group C is worth studying: how do they do it? Did they have easy routes or easy conditions (e.g., day. time pick ups, inside pick ups), or do they have a system of their own? These are questions to pursue. If these men have a system of their own, then they should teach the others. (Enquiry turned up no evidence of easy routes.)
No problem with people is simple. It would be. wise for the management to defer criticism of Group A, to determine first whether these drivers worked unusually difficult routes, or whether they had achieved excessive mileage (high productivity). As it turned out, they had.
l Ilere we encounter an important lesson in administration. This company had been sending a letter to a driver at every mistake. It made no difference whether this was the one mistake of the year for this driver, or the 15th: the letter was exactly the same. A letter sent to a driver in Group B or C is demoralizing: the driver's interpretation thereof-and he is absolutely correct
-is that he is blamed for faults of the system.
The management had failed to see that they face three distinct types of prob!cm. What was needed was a separation of responsibilities for improve.
ment-special causes, to be corrected by the drivers of Group A: the system itself, to be improved by the rnanagement: study of Group C: and examination l
of the accuracy of their records of mistakes.
IN n RfACIS August 1975 13
a
,g One might pause here in passing to ask two questions: (1) what does the manager of the terminal think of the driver to <hom he has sent in one year 15 warnings of dis <.iplinary action?.\\ fore important, (2) what does the driver think of the manager?
Example 3.
A small manufacturer of shoes was having trouble with his sewing machines, rent of which was very co tly. The operators were spending a lot of their time rethrea' ding the machines, a serious loss.
The key observation was that the trouble 'was common to a!! machines and to all operators. The obvious conclusion was that the trouble, whateser it was, was common, environmental, affecting all rnachines and all operators.
A few tests showed that it was the thread that caused the trouble. The owner of the shop had been purchasing poor thread at bargain prices. The loss of machine. time had cost him hundreds ci times the difference between good thread and what he had been buying.11argain prices for thread turned out to be a costly snare.
Better thread eliminated the problem. Cnly the management could make the change. The operators could not go s it and buy better thread.
even if they had known where the trouble lay. They work in the system.
The thread was part of the system.
Prior to the simple in estigation that found the cause. pedestrian but effective, the owner had supposed that his trouhics all came from inexperience and carelessness of the operators.
Example 4.
,The work of every one of 50 production workers on a certain production.
line is in statistical control.The mana::er of per:ennel came forth with a plan, immediately hailed by the management. to award momhh a prize and half a day off to the operator on this line whose production the month before showed the smallest proportion of defectise product.
Was this a good idea? What was wron:;? Why should the statistician advise the management to drop the idea? The answer is that it would not improve the performance of the workers, nor improve quality.
- Why not? Every operator has already put into the job all that he has to offer: the work of each one is in a state of control.
This award would not be an award of merit. What harm would come from it? It would produce frustration and dissatisfaction amongst conscien.
'tious workers. Their efforts to find out what they are doing that is wrong, and why their work is not as good as that of the man that won the prize, would be a fruitless chase. They would try out changes in their operations.
the only effect of which would be greater variability, not less.
The award would be a lottery. There would he no harm that I know of
)
in introducing a lottery for excitement, provided management calls it a lottery, not an award of merit.
Tlds is an example of administration without statistical judgment.
The plan seemed to be a great idea until examined by the theory of proba.
bility, with reference to special causes and common causes.
What the personnel msn could do, if he wishes to offer a prire and be effective, is to reward a man that contrises ways to imprese the system, to decrease the per cent defective for the group by some stated amount of economie importance.
14 INTERTACES August 1975
I c
.N 9-Stanagement could make good use of the figures on defectives for the 50 workers. but not to award a prize. The 50 proportions of defectives furnish a basis, by use of the simp!c statistical technique called char: for fraction defectise, to cliscover which worker is any cucht to be transiened and trained in other jobs. The same chart, even if the 50 workers were not in statistical control with each other, woubt indicate how much of the overall fraction defective arises from the system itself. beyond reach of the workers, and correctible only by manacement.
Concluding Remarks The principles expounded here. :.nd the examples of application, are all simple, set the economic.;ains frem conectise acticn bs manacement are considerable. The examples all belone to the stati tical control of quahn.
Did the solutions require a statistician) Couldn't other people have done as well? One answer is that other people nad their chance.
Some people would call this work operations re<ench. Some would call it systems analysis: others. industrial en:f r: erme. To me. it is just a statistician trying to be positise and helpftai in the me of statistical methcds.
When will schools of busines and ot!.er acasmic departments cet into the business of teachim: modern prinap'.es of adnunistration and mangement?
Without statistical lo.:ic. manauement lerns worcs and coals,5 u nc,t.nethcds by which to reach these :oals. nor meaningfu' lan:uage hv which to describe a goal or to measure advancement tow.ird i: or away from it.
Rz m t.sc s 1.
W. Edwards Demint. E!cr:entarv P :-c le: i s the Sta:n:ical Cc n:rol et C,u.shrv.
Unico of J:cance Soentius and E.neineen, ro. s
'. N 2.
W. Edwardt Demin t "On the c*e of :: e-n ? / :1:.:!-it! o t al Mr d ml. uii.
e.e n t' of quah t v.'
.4 tl India No.1. Julv 19'6. "On sc,me statisocal iecie in ::ie mar.4:
Congreu c,n Qtality Control. New Delhi.17 \\f aren l*.71.
3.
J. \\l. Juran. Qualary Centrol Ha tuoce. srcGraw llut. New Yerk. 1951 l#40; pas;es !! 4 5: 26. 10 11. and' other pares in thri % :.t:ed i:nc.er P:. en.
m.e = 2 : a*
cited be!cne to the 3rd ediuona I suonely remmend the w hme baos for recNe of all respontibihties and disciplines. see aLo Hy htt. 'Pareto r e uuted.' QWm P cress, sci.
vd. No. 2. pp. 29 30 4.
This procedure was fint de ct:hed. *o far as I know. bv J. ?.f. Juran in a meeting of the American Sooety for Quahty Centrol in New York at ' cut 29 scars n as it was formahted bv Irvinit nurt. " Spec;funt the dr itec d;uribut;cn rather thn maucurn and cainimum 1. min. Industrial DuMits Control. vol 2 4. No.
IM7 pp. 94101.
5.
Kenichi Kos anart. "3tatistical Q.4! sty Cons ol in /enete Indust *v," Proceedings of the Americm. Soocty for Quality Control. P.c.choter 1952.
G.
Walter A. Shewhart. The Econo nse Co trrel of Stantstactured Product, Van Natt-rand New York. 1931. Stanstacal afethod frem the l'ieu7at t of qs.shty Co strof. The Graduate School. Department of Agriculture. Washington. 1939.
7.
An easy reference for the nonnantucan is my chpier entitled. "Makirg thinr1 richt." in the bcek, judith Tanner, et 11.. Statturcs, Guide to the Uni nown, Holden. Day.
t 1972.
8.
See alnwst any t e s t.bm k in stauthcal methods. The oririnal reference it Fred.
erick Mosteller and John W. Tukev. "The u*c1 a rwi usefuln mi of binnmial ornbabilit, paper. fournal of the.f rnervan Statisteral.f m cistron. s ol. 44 194): pp. 17 6412. The double square root paper is manufactured by t!.e Codes Dora Ccmpany of.Netwol, Maas.
o
/N TERT.4 Cf f August 1975 15 i
i