Professional Documents
Culture Documents
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to Journal
of Business &Economic Statistics.
http://www.jstor.org
over, approximate normality of a histogram is often ulates well with current statistical practice and can be
erroneously assumed to imply a state of statistical con- implemented easily with standard computing packages.
trol for the process. For an example, see Deming (1982,
5. BOX-JENKINS MODELING FOR
p. 114).
PROCESS CONTROL
4. SUGGESTED EXTENSION OF TRADITIONAL A useful set of tools for implementing these decom-
PROCESS CONTROL positions is provided by the modern principles of time-
series modeling, illustrated, for example, by another
To call attention to common causes when they are
classic pioneering work, that of Box and Jenkins (1976).
present, we suggest an alternative to the dichotomy of
"a state of statistical control" versus "out of control." They stated their goals as follows:
The alternative is based on the familiar decomposition In this book we describe a statistical approachto forecasting
time series and to the design of feed forwardand feedbackcontrol
of regression analysis: Actual = Fitted + Residual. schemes. ... The controltechniquesdiscussedare closerto those of
(When the regression fit is used for forecasting, the the control engineer than the standardqualitycontrol procedures
parallel dichotomy is Actual = Predicted + Error.) developedby statisticians.This does not mean we believe that the
traditionalqualitycontrolchartis unimportantbut ratherthat it per-
Our experience suggests that in a wide range of ap- forms a differentfunctionfrom that with which we are here con-
plications in which processes are not in control in the cerned.An importantfunctionof standardcontrolchartsis to supply
sense of iid random variables, one can use relatively a continuousscreeningmechanismfor detectingassignablecausesof
variation.Appropriatedisplayof plantdataensuresthatchangesthat
elementary regression techniques to identify and fit ap- occur are quicklybroughtto the attentionof those responsiblefor
propriate time-series models. If we succeed in finding runningthe process.Knowingthe answerto the question,"whendid
such a model, we have reached a negative verdict about a changeof this particularkind occur?"we can then ask "whydid it
occur?"Hence, a continuousincentive for process improvement,
statistical control in the sense of iid and we obtain fitted often leadingto new thinkingabout the process,can be achieved.
values and residuals along with probabilistic assess- In contrast,the control schemeswe discussin this book are ap-
ments of uncertainty. We can regard the process as "in propriatefor the periodic,optimaladjustmentof a manipulatedvari-
control in a broader sense," a sense that is entirely able, whose effect on some output qualitycharacteristicis already
known, so as to minimizethe variationof that qualitycharacteristic
consistent with Shewhart's conception of a state of sta- about some targetvalue.
tistical control. The reason control is necessaryis that there are inherentdistur-
bances or noise in the process.... (p. 4)
Since successful time-series modeling decomposes the
actual series into fitted values and residuals, the tra- Our view is that Box-Jenkins time-series methods-
ditional purposes of process control can be served by and their many extensions-are applicable both to the
the two basic charts mentioned in Section 1: control schemes mentioned in this quotation and to the
traditional process-control objective of continuous
1. A Time-Series Plot of the Fitted Values (without process surveillance to detect special causes of vari-
computation of control limits). This plot can be re- ation. We believe that the basic philosophy of diagnostic
garded as a series of point estimates of the conditional checking espoused by Box and Jenkins leads almost
mean of a process-our best current guess based on inevitably to the approach to process surveillance out-
past data of the location of the underlying process. lined in this article.
2. Standard Control Charts (Shewhart, CUSUM, or To illustrate time-series modeling, we shall use a spe-
other) for the Residuals. Control limits are based on cial case of Box-Jenkins "ARIMA" models known in
the time-series model itself; for example, limits for pre- statistical quality control as the EWMA, explained by
diction errors would be based on the standard errors of Hunter (1986). In Hunter's use of EWMA, it was a
one-step-ahead forecasts. procedure for testing a state of statistical control, al-
though he did emphasize the predictive quality of the
This two-step approach appears to be an obvious union EWMA and the need for prediction for control and
of time-series modeling with traditional ideas of process went on to relate the EWMA to the control engineer's
control, but the possibility of the union appears not to modeling approach. In our approach, the EWMA
have been widely exploited. The closest approaches that emerges as a flexible time-series model that, for many
we have seen are those of Hoadley (1981) and Hunter but not all processes, may be a satisfactory approxi-
(1986). So far as we can tell, other applications of time- mation.
series concepts to process control have been oriented One appealing interpretation of EWMA is that the
toward more sophisticated testing of the traditional null process being studied can be decomposed into two com-
hypothesis of iid behavior. The power of proposed tests ponents:
may be assessed against specific time-series alternatives, 1. iid random disturbances, with mean 0
but these alternatives are not explicitly modeled.
2. A random walk, which is the sum of a fraction a
In the regression decomposition here proposed, each
of all past iid random disturbances.
fitted value is conditioned only on past data rather than
on all of the data, as in signal-extraction theory. We To bring out our main point, we shall consider an
propose the regression decomposition because it artic- application in which individual observations rather than
1 only for illustration. All points are within these limits 60 70 80 90 100 110 120
(the maximum is 8.2 and the minimum is 6.1). (If the 8.25+
data are viewed without regard to time sequence, they Y
-.-------
conform very closely to the normal model, providing
another illustration of nonrandom but approximately
normal data.) 6.75+
- LCL
8.25+ IJ,=;
Y .+ + + . + + +S QUEE
120 130 140 150 160 170 180
7. 50+
8.2+ 2
y - -A
6.75+ ----
7.50+
- - -- ---CL
6.00+ rrr
+ + + + + + +
60SECUE 6.75+
0 10 20 30 40 50 60
8.25+ .L ;L
6.00+
+ + +SEQUENCE
180 190 200
7. 50S
7. 50+
Even if the process were deemed stationary, however,
6.75.
three-sigma limits should logically be based on the mar-
ginal rather than the conditional standard deviation.
6.00+
Any concern with control limits, however, diverts at-
+++^ tention from the basic observationthat the conditional
120 130 140
mean is changing constantly. These changes of the con-
8.25+ -UCL
ditional mean are essential to understanding and control
of the process.
r7.500+~0-
A common approach for dealing with nonrandom
6.75+
data like these is to base control limits on the mean of
moving ranges of successive observations (Wadsworth
6.00+ et al. 1986). We see in Figure 2 that these limits are
LCL
++
180 190
+ +GSEGU CE
200
much tighter and show many points out of control, oc-
curring mainly at the peaks and troughs of the waves
Figure 1. Sequence Plot of SeriesA, Box and Jenkins (1976): of the data.
197 ConcentrationReadings orna Chemical Process,SingleRead- These control limits based on moving ranges alert the
ings TakenEvery Two Hours.Data are coded by subtractionof a
constant 10 fromeach observaition.CLis the eer line
e center e at a height user that the process is not close to being random, but
equal to the sample mean, ancI UCLand LCLshow the locationof they provideno real guidancein understandingwhat is
conventionalthree-sigmacontr.ol limits. happening.
For many users, data like these suggest a series of Y,is 0. As estimatedby the MinitabARIMA procedure,
loosely connectedepisodes, each invitingad hoc expla- we have
nations. (This parallels the after-the-fact "explana-
Fitted Y, = Yt_1 - .705A,_1.
tions" offered by stock-marketanalysts to "explain"
changesof the Dow-Jones IndustrialIndex!) The pro- The quantity.705 is an estimate of what is called 0 by
cess of achievinga state of statisticalcontrol is some- Box and Jenkins; 1 - .705 = .295 is an estimate of
times picturedas findingspecial explanationsfor each whatis often calleda in discussionsof the EWMA. The
episode, making a correction, observing the process standarderrorof the estimated0 is .05, andthe standard
further,findingfurtherspecialexplanationsfor further deviationof residualsis .318. (We have suppressedthe
episodes, and so forth, until controlis finallyattained. constant term, for two reasons: We have judged such
Here, however, as in the exampleof the Dow-Jones a term unlikelya priori, and the sample evidence sug-
IndustrialIndex, simple time-series modeling unifies gests an estimatedconstantnear 0.)
the pictureimmediately.
Box and Jenkins fitted two models, each of which 5.1 Common-Cause Chart: Fitted Values
describesthe data about equallywell. We considerone
of these models(the modelunderlyingEWMA),namely It is useful to display the fitted values or estimated
a first-orderintegratedmoving average, called "AR- conditionalmeans of the process, called "FITlED" in
IMA(0,1,1)," which, as pointed out previously,can be Figure 3, on a time-sequenceplot. This plot gives a
interpretedas a random-walktrend plus a randomde- view of the level of the process (estimatedconditional
viation from trend. This model specifies that the ob- mean) and of the evolution of that level throughtime.
served Yt is the sum of an unobservedrandomshock We see the systematicbehaviorof the processthat per-
A, plus a (proper) fraction a of the sum of all past vades the entire period of observation.This behavior
random shocks A,_ , At-2, .... In the current appli- may aid in real-timecontrolor in better understanding
cation we have assumed that the mean of changes of how the process is working.
The model just fitted can be interpretedas follows:
Each observationcan be thought of as a randomdis-
-
turbanceplus a random-walktrendor driftthat reflects
FITTED
7.50+ 9 a certainfractionof the sum of all past randomdistur-
bances. Thus a part of each disturbancecontinues to
7.00+ affect the process into the indefinitefuture. The fitted
values in Figure3--P'lED-are estimatesof the un-
6.50+
derlying random-walk trend;they follow a randomwalk
+
0
+ +
+
+
+
+ SEC.CE withoutdrift. The chartof FI'IT'EDrepresentsthe first
10 20 30 40 50 60
of the two charts proposed by us. Each point is an
FITTED -
7.50+
estimate of the local level of the process itself, as dis-
tinguishedfrom the observedreadings.
7.O0 1 e
CL To illustratethe type of control decisionsthat could
be based on this plot, suppose that the most desirable
6.50S level of the processis 7.0, andthat increasingdeviations
+ 4 + + + +4 *SEQUENCE
fromthatlevel entailincreasingeconomicloss fromless-
60 70 80 90 100 110 120
-
than-optimalproduct.Supposefurtherthat at a certain
FITED
7.50+
known cost it is possible at any time to recenter the
process 7.0. Then one can make an economic calcula-
7.00+
tion to balance the expected loss of bad productover
some specified period of time against the cost of re-
6.50+ centering.This calculationwill define actionlimitsboth
+ + + + + + SEQUEiCZ
below and above 7.0 at which the process should be
120 130 140 50 160 170 180
recentered.
FITTED- Note that these action limits are conceptuallydiffer-
7.50+
ent from traditionalcontrollimits. They are not signals
7.00+
- CL that it is time to look for special causes; rather, they
are signalsthat a specific correctiveaction is needed.
6.50+
i80
+ +
190
+SEQUECE
200
5.2 Special-Cause Chart: Residuals
The second graph is essentially a standardcontrol
Figure3. Common-CauseChart:Sequence Plot of FittedValues
for the Data of Figures 1 and 2. Estimatesare based on the AR- chartfor residuals.To facilitatevisualizationof control
IMA(O, 1,1) model with0 constant. limits, we plot in Figure 4 the standardizedresiduals
ST.S -
For many purposesof control, whether real-timecon-
2.50+
trol or process surveillance,either would be suitable.
Froma theoreticalpointof view, the two modelshave
very differentimplicationsfor the long-runbehaviorof
the process, assumingcontinuationof the basic condi-
tions now being observed. ARIMA(0,1,1) is nonsta-
tionary, but ARIMA(1,0,1) is stationary. Given the
0 10 80 30 40 50 60
STRES
0.00+
a. so*
-
- ARIMA(0,1,1) fitted previously,there is no tendency
of the process to revert to its mean level as it wanders
away. Given the ARIMA(1,0,1) for the same process,
there is a relativelyweak tendencyto mean reversion.
2.50+
Moreover,prospectivecontrollimitsas we look farther
into the futureare constantfor ARIMA(1,0,1) but ever
0.00+ wideningfor ARIMA(0,1,1).
At one time we contemplatedthe concept of a sta-
tionary process as a possible extension of the concept
O ._~~~~-
O O4CLU+LCL
of iid in defininga state of statisticalcontrol. The pres-
ent example illustrates, however, that data may not
-a.50.
0.00+ -
permita sharpdistinctionbetween stationaryand non-
++
- ~
LC.~~~~~LCL
~~~~- stationary.Hence we propose only that some time-se-
SEEUECCE + + +
810 t30 140 160 170 180 ries model be used to define statisticalcontrol in an
STRES) with approximatecenter line and upper and extended sense. Even a nonstationarymodel permits
2.5UL0+
probabilisticprediction.
Of course, if the weight of statisticalevidence and
0- t t t tt backgroundknowledgesuggeststhatthe processshould
-2.50+
be regardedas nonstationary,the problemconfronting
+ +?
LCL
+ SENCE
managementis inherentlymore difficult, since a non-
180 190 200
stationaryprocess has no tendency toward mean re-
version.Forexample,a stationaryARIMAmodelabout,
Figure4. Special-CauseChart:SequencePlotof Residualsfor
the Dataof Figures1 and 2. Estimatesare based on the AR- say, a linear trend suggests disasterin the absence of
1,1)modelwith0 constant.
IMA(O, intervention.At the same time, the chart of residuals
could suggest that special causes are occurringalong
the road to ruin of the process itself.
(STRES) with approximatecenter line and upper and
lower three-sigmalimits.
It is seen that two individualpoints-observations 43 5.4 Applications in Business and Economics
and 64-breach the three-sigmalimits. By contrast,the
We have used a "public"data set to illustratethe
firstset of controllimitscalculatedpreviouslysuggested
mechanicsof our basic proposalsin an applicationthat
no points out of control, but the second set, based on
will be familiarto many readers. Even here, however,
mean ranges, suggested many individualpoints out of
it results in identificationof apparent out-of-control
control. These out-of-control points did include obser-
vation 64, which was at the top of a wave, but not points that are easy to miss given the generallysatis-
observation 43. factoryfit by ARIMA models.
We have had substantialexperience in applications
In addition, the preceding plot reveals perhaps two
to business and economics, and these have confirmed
or three short intervals for which run counts would
the fruitfulnessof time-seriesmodeling.One interesting
suggest some suspicion of lack of control. Note that run illustrationis based on a reportby BaRon (1978), who
counts are applied only after the residuals from the
studied the numberof air tourist arrivalsin Israel by
time-seriesmodel have been isolated. As appliedto the
months for 20 years startingin January1956. Except
originalseries, run counts would suggestthat the proc- for an upwardtrendand seasonality,BaRon'sdatacon-
ess was nearlycontinuallyout of control, sometimeson
the high side and sometimeson the low side. veyed the same visual general impressionas the series
presentedpreviously.Based on observationof the data
and knowledge of the historicalbackground,BaRon
5.3 Other Simple Models
classifiedthe 240 months into 26 segments of varying
The ARIMA(0,1,1) modelfittedpreviouslyis not the lengths; for example, segment 9 extended from Feb-
only reasonablemodel for the data. Box and Jenkins ruary1961to July 1962.Withinthese segments,he used
also fit ARIMA(1,0,1) with a nonzero constant. The three subclassifications:"regular,""shortmonotonic,"
two models give nearly identical fitted values and re- and "unusual."February-December1961, for exam-
sidualsand hence about the same degree of overallfit. ple, was identifiedas "unusual,"and labeled "Eichman
trial: disturbed growth," whereas January-July 1962 Although there can be no completely satisfactory sub-
was "regular." BaRon also picked out nearly 40 "ex- stitute for statistical skills in time-series analysis, we
treme" months. believe that much can be accomplished by people with
In a discussion of BaRon's paper, Roberts (1978) limited skills. For example, modern computational tools
reported that the series could be well fitted by a mul- make possible relatively automated implementation of
tiplicative seasonal model (0,1,1) by (0,1,1)12 on the time-series modeling; automatic fitting of Box-Jenkins
logs, with a constant term. Only one residual turned up models has for some time been available in software of
as an unmistakable outlier, and that was for the month general purpose software packages like Minitab, Stat-
of October 1973-the Yom Kippur War. The model graphics, SCA, or Systat now available on the IBM PC
itself is the so-called "airline" model used by Box and and compatibles. [There are even programs that at-
Jenkins (1976) to fit a monthly time series of interna- tempt to automate model identification, as well as fit-
tional airline passenger travel from 1949 to 1960. This ting! See, e.g., Shumway (1986).]
suggests that many common causes underlying BaRon's It is reassuring to know that precise model identifi-
data were similar to those underlying foreign travel cation may not be essential to effective process control;
worldwide and that special causes were rare. several alternative models may fit the past data, at least,
Our approach has also been used by several hundred about equally well. In particular, the two models men-
students in the Executive Program and other programs tioned in our example in Section 5-ARIMA(1,0,1) and
of the Graduate School of Business at the University ARIMA(0,1,1)-offer reasonably good fit for a wide
of Chicago who have applied it to data arising in various range of applications. (As one referee has suggested,
areas of organizations for which they worked, ranging the EWMA model may be a good all-purpose starting
from finance and marketing to production and research point.)
and development. These applications typically find sys- If a process can be modeled, then the traditional
tematic time-series variation, including autocorrelation, objectives of quality control-or surveillance-can be
seasonality, and trend. Many of them also find at least better served. Moreover, a control mechanism can be
one outlying residual, and most of these students are designed, and if a principal disturbance to the process
able to identify plausible special causes for these resid- can also be modeled, we have the adaptive control sit-
uals. uation described by Box and Jenkins (1976).
Typical examples are a sudden decrease in mean of
length of stay in a hospital, which was traced to the 7. ENGINEERING AND ECONOMIC ISSUES
imposition of the diagnostically related groups program;
an extreme negative bank stock return related to a pub- More difficult than model identification and fitting is
lic announcement of a write-off of nonperforming loans; the interpretation and exploitation of fitted values for
a sharp decline in company sales of a capital-equipment real-time process control, especially when process re-
manufacturer in the third quarter of 1980, subsequently centering is costly, and for better understanding of un-
traced to the lagged effects of the economic recession derlying common causes that make the process behave
that had occurred earlier in the year; a time-of-day as it does. Real-time process control has received con-
effect in the occurrence of computer crashes; and an siderable attention, for example, by Box and Jenkins
extremely high reading of hardness of steel coils, traced (1970) and Box, Jenkins, and MacGregor (1974). Sys-
to a scheduling error by which a billet of mixed steel tematic study of common causes has attracted little at-
was inadvertently processed into a coil. tention. We hope, in future work, to examine questions
like these in some detail.
One interesting economic illustration encountered by
several students from a variety of perspectives was the
ACKNOWLEDGMENTS
great increase in volatility of interest rates starting in
the fall of 1979, for which the special cause appears to We are indebted to two referees and an associate
be a change of policy by the Federal Reserve System, editor for many excellent suggestions.
a switch in emphasis from targeting interest rates to
[Received November 1986. Revised March 1987.
targeting monetary aggregates.
6. TIME-SERIES MODELING IN PRACTICE
REFERENCES
A practical limitation on the use of the approach
BaRon, RaphaelRaymondV. (1978), "The Analysisof Single and
advocated in this article is that implementation requires Related Time Series Into Components:Proposalsfor Improving
some skill in analysis of time series, whereas the im- X-11," in Seasonal Analysis of Economic Time Series (Proceedings
plementation of the standard Shewhart procedures en- of the Conferenceon the Seasonal Analysis of EconomicTime
tails only the most elementary statistical knowledge. We Series,Washington,DC, Sept. 9-10,1976), Washington,DC: U.S.
believe that in many applications, the ability better to Departmentof Commerce,Bureau of the Census, pp. 107-158,
171-172.
sort out special causes from common causes justifies Berthouex,P. M., Hunter, W. G., and Pallesen, L. (1978), "Moni-
the use of the more elaborate machinery required by toringSewageTreatmentPlants:Some QualityControlAspects,"
our approach. Journal of Quality Technology, 10, 139-149.
Box, G. E. P., and Jenkins, G. M. (1976), Time SeriesAnalysis, Page, E. S. (1954), "ContinuousInspectionSchemes,"Biometrika,
Forecastingand Control(2nd ed.), San Francisco:Holden-Day. 41, 100-115.
Box, G. E. P., Jenkins,G. M., and MacGregor,J. F. (1974), "Some Roberts,HarryV. (1978), Commenton "TheAnalysisof Singleand
Recent Advances in Forecastingand Control:Part II," Applied Related Time Series Into Components:Proposalsfor Improving
Statistics,23, 158-179. X-11,"by R. R. V. BaRon, in SeasonalAnalysisof EconomicTime
Deming, W. E. (1982), Quality,Productivityand CompetitivePosi- Series(Proceedingsof the Conferenceon the SeasonalAnalysisof
tion,Cambridge,MA:MITCenterforAdvancedEngineeringStudy. Economic Time Series), Washington,DC: U.S. Departmentof
Eisenhart,Churchill(1963), "RealisticEvaluationof the Precision Commerce,Bureauof the Census,pp. 161-170.
and Accuracyof InstrumentCalibrationSystems,"Journalof Re- Shewhart,W. A. (1931), EconomicControlof Qualityof Manufac-
searchof the NationalBureauof Standards,C-Engineering and turedProduct,New York: Van Nostrand.(Republishedin 1981,
Instrumentation, 67, 2. witha dedicationby W.EdwardsDeming,by the AmericanSociety
Hoadley, B. (1981), "The QualityMeasurementPlan," Bell System for QualityControl,Milwaukee,WI.)
TechnicalJournal,60, 215-271. Shumway,R. H. (1986), "AUTOBOX(Version1.02)," TheAmer-
Hunter,J. S. (1986), "TheEWMA,"Journalof QualityTechnology, ican Statistician,40, 299-300.
18, 4. Wadsworth,HarrisonM., Stephens,KennethS., andGodfrey,Blan-
Montgomery,Douglas C. (1985), Introductionto StatisticalQuality ton A. (1986), ModernMethodsfor QualityControland Improve-
Control,New York:John Wiley. ment, New York:John Wiley.